1.1 Why microphone arrays?
UNIVERSITÀ DEGLI STUDI DI PARMA
CORSO DI DOTTORATO DI RICERCA IN INGEGNERIA
INDUSTRIALE
XXVII CICLO
DEVELOPMENT OF A SOFTWARE FOR
PANORAMIC VIDEO AND AUDIO RECORDING
WITH MICROPHONE ARRAYS
Coordinatore:
Chiar.mo Prof.
AGOSTINO GAMBAROTTA
Tutor:
Chiar.mo Prof.
ANGELO FARINA
Dottorando:
Dott. Ing.
SIMONE CAMPANINI
a Giovanna Emanuela
e
ad Ivo ed Elsa
Contents
Contents
v
List of Figures
ix
1 Introduction
1
1.1
Why microphone arrays? . . . . . . . . . . . . . . . . . . . . . . .
2 Microphone arrays: the state of the art
2.1
2.2
2.3
7
The beam forming technique . . . . . . . . . . . . . . . . . . . . .
7
2.1.1
Limits of the beam forming theory . . . . . . . . . . . . .
9
2.1.2
Advanced beam forming techniques . . . . . . . . . . . . .
10
Array for noise maps elaborations . . . . . . . . . . . . . . . . . .
13
2.2.1
Brüel & Kjær . . . . . . . . . . . . . . . . . . . . . . . . .
13
2.2.2
gfai tech . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
2.2.3
Nittobo . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
Directive microphoning . . . . . . . . . . . . . . . . . . . . . . . .
20
2.3.1
Acoustic Magic . . . . . . . . . . . . . . . . . . . . . . . .
20
2.3.2
Yamaha . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
2.3.3
Squarehead Technology . . . . . . . . . . . . . . . . . . . .
23
2.3.4
mh acoustics
. . . . . . . . . . . . . . . . . . . . . . . . .
24
2.3.5
Customs . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
2.3.5.1
Planar Array . . . . . . . . . . . . . . . . . . . .
26
2.3.5.2
Cylindrical Array . . . . . . . . . . . . . . . . . .
26
3 Virtual microphones
3.1
2
A theoryless approach . . . . . . . . . . . . . . . . . . . . . . . .
29
30
vi
CONTENTS
3.1.1
Computation of the filters . . . . . . . . . . . . . . . . . .
31
3.1.2
Experimental characterization of the array . . . . . . . . .
36
3.1.3
Synthesis and test of virtual microphones . . . . . . . . . .
39
4 The optical part
41
4.1
The camera . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
4.2
The panoramic frame . . . . . . . . . . . . . . . . . . . . . . . . .
42
4.2.1
Mirror and lens . . . . . . . . . . . . . . . . . . . . . . . .
42
4.2.2
The unwrapping technique . . . . . . . . . . . . . . . . . .
47
5 The software
5.1
5.2
51
3D Virtual Microphones System . . . . . . . . . . . . . . . . . . .
®
5.1.1
From Pyhton to OS X App . . . . . . . . . . . . . . . . .
53
5.1.2
The filtering engine . . . . . . . . . . . . . . . . . . . . . .
55
5.1.2.1
The logic module . . . . . . . . . . . . . . . . . .
56
5.1.3
The recording/playback engine . . . . . . . . . . . . . . .
57
5.1.4
The multimedia engine . . . . . . . . . . . . . . . . . . . .
57
5.1.5
The timecode manager . . . . . . . . . . . . . . . . . . . .
63
5.1.6
The control interface . . . . . . . . . . . . . . . . . . . . .
64
5.1.7
The actual structure . . . . . . . . . . . . . . . . . . . . .
65
How to use the 3D VMS . . . . . . . . . . . . . . . . . . . . . . .
67
5.2.1
System requirements . . . . . . . . . . . . . . . . . . . . .
67
5.2.2
Installation . . . . . . . . . . . . . . . . . . . . . . . . . .
67
5.2.3
Things to do before launching the application . . . . . . .
69
5.2.4
The main dialog . . . . . . . . . . . . . . . . . . . . . . . .
71
5.2.5
The Playback activity . . . . . . . . . . . . . . . . . . . .
72
5.2.5.1
JACK routing . . . . . . . . . . . . . . . . . . . .
75
5.2.6
The Realtime activity . . . . . . . . . . . . . . . . . . . .
76
5.2.7
The Recorder activity . . . . . . . . . . . . . . . . . . . .
80
5.2.7.1
81
Output file formats . . . . . . . . . . . . . . . . .
6 Endings
6.1
52
Future developments . . . . . . . . . . . . . . . . . . . . . . . . .
83
85
CONTENTS
A File format descriptions
vii
87
A.1 Matrix files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
87
A.1.1 The file metadata.xml . . . . . . . . . . . . . . . . . . . . .
88
A.2 Microphones presets . . . . . . . . . . . . . . . . . . . . . . . . .
89
B LTC/SMPTE
91
Bibliography
93
viii
CONTENTS
List of Figures
2.1
Schematic representation of a plane wave incident on a linear microphone array. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
2.2
Steering of the synthesized virtual microphone. . . . . . . . . . . .
9
2.3
Normalized beam patterns by varying the microphone capsules
number. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
2.4
An example of two noise maps . . . . . . . . . . . . . . . . . . . .
13
2.5
The Brüel & Kjær Wheel Array. . . . . . . . . . . . . . . . . . . .
14
2.6
Example of Brüel & Kjær’s PULSE™ acoustic map. . . . . . . . .
15
2.7
Brüel & Kjær spherical beam-former. . . . . . . . . . . . . . . . .
15
2.8
gfai tech Ring 48 . . . . . . . . . . . . . . . . . . . . . . . . . . .
17
2.9
gfai tech Star 48
. . . . . . . . . . . . . . . . . . . . . . . . . . .
17
2.10 gfai tech Spiral 120 . . . . . . . . . . . . . . . . . . . . . . . . . .
18
2.11 gfai tech Sphere 48 . . . . . . . . . . . . . . . . . . . . . . . . . .
19
2.12 gfai tech’s free configurable 24 channel microphone bundle . . . .
19
2.13 Examples of acoustic maps produced by gfai tech Acoustic Camera. 20
2.14 Nittobo’s Noise Vision™ system. . . . . . . . . . . . . . . . . . . .
20
™
I Array Microphone. . . . . . . .
21
2.16 Acoustic Magic Voice Tracker™ II Array Microphone. . . . . . . .
21
2.17 The Yamaha PJP-50R Conference Microphone System. . . . . . .
23
2.15 Acoustic Magic Voice Tracker
2.18 The discontinued Yamaha PJP-100UH Conference Microphone System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
2.19 The Yamaha YVC-1000. . . . . . . . . . . . . . . . . . . . . . . .
24
2.20 The Squarehead Technology’s Audio Scope system: on the left the
array, on the right the control station.
. . . . . . . . . . . . . . .
24
2.21 mh acoustic Eigenmike® EM32. . . . . . . . . . . . . . . . . . . .
25
x
LIST OF FIGURES
2.22 The IED Planar array with the control laptop. . . . . . . . . . . .
27
2.23 The IED Cilindrical array. . . . . . . . . . . . . . . . . . . . . . .
28
3.1
Scheme of signal processing. . . . . . . . . . . . . . . . . . . . . .
31
3.2
Polar plots of virtual cardioid microphones of various orders (target
patterns Qn ). . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3
33
Impulse response measurements from D source positions to the M
microphones. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
3.4
Scheme of the Least Squared method . . . . . . . . . . . . . . . .
35
3.5
Regularization parameter in dependence of the frequency. . . . . .
36
3.6
Frequency response and polar patterns for capsule n.1. . . . . . .
37
3.7
Variation of the distance during rotation. . . . . . . . . . . . . . .
38
3.8
Improvement of polar patterns of a 3rd order virtual cardioid due
to delay compensation. . . . . . . . . . . . . . . . . . . . . . . . .
38
1st order figure of eight. . . . . . . . . . . . . . . . . . . . . . . .
39
3.10 3rd order cardioid. . . . . . . . . . . . . . . . . . . . . . . . . . .
40
3.11 6th order cardioid. . . . . . . . . . . . . . . . . . . . . . . . . . .
40
4.1
The Sony® IPELA® SNC-CH210 IP-camera with black case. . . .
44
4.2
The Wision WS-M8P31-38B IP-camera in bulk version without lens. 44
4.3
The 0-360° Panoramic Optic™ hyperbolic mirror. . . . . . . . . .
4.4
The optical path in an omnidirectional vision system with hyper-
3.9
45
bolic mirror. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
4.5
Scheme of the field depth. . . . . . . . . . . . . . . . . . . . . . .
46
4.6
The mirror and the camera in test mount . . . . . . . . . . . . . .
46
4.7
The image of the mirror seen by the camera. . . . . . . . . . . . .
48
4.8
The polar coordinates system. . . . . . . . . . . . . . . . . . . . .
48
4.9
The optical calibration procedure. . . . . . . . . . . . . . . . . . .
49
4.10 The output (cartesian) coordinates system. . . . . . . . . . . . . .
49
4.11 The unwrapped frame. . . . . . . . . . . . . . . . . . . . . . . . .
50
5.1
The structure of 3D VMS. . . . . . . . . . . . . . . . . . . . . . .
52
5.2
BruteFIR filtering scheme. . . . . . . . . . . . . . . . . . . . . . .
56
5.3
The 3D VMS main dialog window. . . . . . . . . . . . . . . . . .
58
LIST OF FIGURES
5.4
GStreamer architecture. . . . . . . . . . . . . . . . . . . . . . . . .
5.5
The GStreamer pipeline implemented in the Recording activity of
xi
59
3D VMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
62
5.6
The 3D VMS application icon. . . . . . . . . . . . . . . . . . . . .
68
5.7
Hardware connections layout: Eigenmike® + external IP-camera. .
70
5.8
Hardware connections layout: IED Cylindrical Array (with embedded IP-camera). . . . . . . . . . . . . . . . . . . . . . . . . . .
71
The 3D VMS default JACK routing dialog. . . . . . . . . . . . . .
73
5.10 The 3D VMS Playback activity. . . . . . . . . . . . . . . . . . . .
76
5.11 The JACK connections dialog . . . . . . . . . . . . . . . . . . . .
77
5.12 The zoom function: no zoom . . . . . . . . . . . . . . . . . . . . .
78
5.13 The zoom function: increasing zoom level . . . . . . . . . . . . . .
79
5.14 The zoom function: zooming further . . . . . . . . . . . . . . . .
79
5.15 The 3D VMS meter monitoring array capsules output levels. . . .
80
5.16 A 3D VMS Realtime session with VU-meter and timecode display.
80
5.17 The timecode display. . . . . . . . . . . . . . . . . . . . . . . . . .
81
5.9
5.18 A 3D VMS Recorder session with camera monitor, VU-meter and
timecode display. . . . . . . . . . . . . . . . . . . . . . . . . . . .
81
5.19 The 3D VMS Recorder activity. . . . . . . . . . . . . . . . . . . .
82
6.1
Example of post-processing with 7 fixed microphones (La Bohème
- Theater Regio, Turin, 20 May 2010 ). . . . . . . . . . . . . . . .
84
B.1 An example of SMPTE frame. . . . . . . . . . . . . . . . . . . . .
91
Chapter 1
Introduction
Multichannel recordings are usually performed by means of microphone arrays.
In many cases sparse microphone arrays are used, where each discrete microphone
is employed for capturing one of the channels.
However, also the usage of dense microphone arrays has a long history, dating
back to the first MS-matrixed microphones setups and passing through the whole
Ambisonics saga.
A dense microphone array is employed differently from a sparse array, generally
speaking each channel is obtained by a combination of the signals coming from
all the capsules, by means of different matrixing and filtering approaches.
During last years at the Industrial Engineering Department of Università degli
Studi di Parma (IED) with the collaboration of the spinoff AIDA srl., a new
meta-theory was developed that provides a framework capable of representing all
the known microphones array, based on the model of Virtual Microphones.
The strict collaboration with the Centro Ricerca ed Innovazione Tecnologica
(CRIT, Research Center and Technology Innovation) of RAI, the national Italian television broadcasting company, quickly individuates some key applications
- mainly in live or off-line broadcast contest - of this technology that conduct
to the realization of some ad-hoc dense arrays and of a software to manage all
the acquisition and post-processing system, that is the subject of this PhD thesis
work.
2
Introduction
1.1
Why microphone arrays?
The use of microphone arrays is the basis for making audio recordings (or acoustical measurements) capturing information about the spatial distribution of the
wavefronts impinging onto a listener.
In a sparse microphone array the microphones are placed at large distances, with
the goal to sample the sound field at points where the sound is significantly different.
On the other hand, in a dense array, the microphones are usually small and close
each other, so that the minimum distance between two microphones is significantly smaller than the wavelength for any allowed direction of arrival of the
wavefront The intended usage of the signals captured by the two types of microphone arrays is completely different: in a sparse array, the signal coming from
each capsule has to be kept well separated from all the others, and is tailored to
become the feed for a single loudspeaker in the playback system. The only adjustments possible for the sound engineer is to apply to each signal some amount of
gain, some spectral equalization, and perhaps some delay. But each signal is always processed independently from the others. The only exception to this rule is
the so-called mixdown, which occurs whenever the playback system is equipped
with a number of loudspeakers which is smaller than the number of recorded
channels: in this case the channel count is reduced by simple mixing rules.
In a dense array, instead, the whole set of signals are treated as an unique entity, and the processing required for computing the output signals (speaker feeds)
involves generally complex matrixing operations, so that each output channel
includes some amount of information captured by all the microphones. Furthermore, heavy filtering is usually required, instead of the simple gain and delay
adjustments employed for sparse arrays.
This means that, inevitably, the dense array technology requires more effort: all
the microphones must be reliable and high quality,1 the processing algorithm requires much more computational power,2 and the risk to get artifacts is vastly
larger.
1
2
A noisy capsule affects the behaviour of the whole system
As each output signal requires to apply specific filtering to all the input channels, and each
filter is generally computationally very heavy.
1.1 Why microphone arrays?
3
Whilst the sparse array approach is perfectly compatible with a minimalistic,
audiophile philosophy, avoiding almost completely dangerous digital processing
effects, the dense array approach cannot work without massive and powerful filtering and matrixing.
These factors explain the comparatively larger success encountered till now by
the sparse array technology, and the limited success of dense microphone arrays.
However, regarding stereo (2-channels) recordings, the prominent dense array
approach has been the M-S technique,3 which conquered favoritism for some applications in the film-video-tv industry; mostly thanks to its excellent mono compatibility, obtained simply by discarding the S signal, when not needed, and by
the versatility of adjusting the amount of ambience, obtained changing the relative gain of the M and S signals. Generally speaking, however, most high-quality
stereo recordings make use of sparse arrays, employing typically two spaced microphones, such as in the ORTF method4 or in the A-B method.5 Going to
surround recordings for horizontal sound systems (quad, 5.1, 7.1, etc.), again we
have seen a modest success of one particular type of dense microphone arrays,
the 1st order Ambisonics method [2]. This makes use of just 3 channels (W , X,
Y ) for capturing the spatial information in the horizontal plane, with the possibly to scale up to full periphony adding a 4th channel (Z). Different types of
dense microphone arrays have been developed for surround recording, which can
be subdivided in three categories:
• Tetrahedrical probes (Soundfield™ microphone and similar)
• Multiple M-S probes (Schoeps™)
• native B-format arrays (Nimbus™)
The latter, in theory, already provides the required W , X and Y signals as the
outputs of the three capsules (one omni capsule for W , and two figure-of-eight
capsules for X and Y ), but in practice some amount of post- processing is always
required, for ensuring perfect gain and phase matching among the three channels.
3
See [6].
Modestly spaced capsules, approximately 170 mm.
5
Here the capsules can be several meters apart
4
4
Introduction
So all of these microphone arrays require some processing for delivering the canonical B-format signals. For years the processing has been analog, but nowadays
it is better performed digitally, and hence these microphone arrays are already
embraced by the meta-theory described in chapter 3.
Of course, also for surround recordings, the vast majority are performed with
sparse microphone arrays. Also these can subdivided in two main categories:
Distributed arrays several microphones spread in the orchestra and picking up
the voices of the vocalists, followed by amplitude-panning over the surround
channels
Compact arrays for example Williams’, INA- 5, OCT, etc.
In this case the most successful approach was the first, giving full control to
the sound engineer to place each soundtrack at will on the surround panorama
by panning.
The second approach, indeed, is the preferred one for minimalist/audiophile surround recording, as the signal of each capsule is directly the feed for the corresponding loudspeaker, without the need of any further processing. But this
comes at the price of requiring an ideal placement of the microphone array, inside a venue with excellent acoustics, and making it impossible to fix anything
after the recording is done. And finally in the last years novel massive recording/playback techniques emerged, capable of driving advanced sound systems
equipped with dozens or hundredths of loudspeakers, and employing advanced
digital sound processing methods.6 Here the channel count increases dramatically, ranging from 16 (3rd order HOA) up to several hundredths (large WFS
systems). In this field, dense arrays are usually preferred, making it possible to
pack a large number of capsules in a single, compact and light object, which can
be conveniently manipulated and placed.
These dense arrays were built of various shapes, with or without a solid object
over which the capsules are mounted, and in many different shapes: line, planar,
circle, sphere, pseudo-random cloud, and recently cylinders. The system usually
contains also the digital interface, ensuring that all the signals are sampled with
6
WFS - Wave Field Synthesis [3], HOA - High Order Ambisonics [5], SPS - Spatial PCM
Sampling [8], etc.
1.1 Why microphone arrays?
5
perfectly matched gain and phase, and the same digital sampling clock. The massive processing required is performed digitally on a standard PC, which nowadays
provides enough computational power even for very large numbers of capsules,
thanks to the raw computing capabilities of modern CPUs and the refinement of
advanced processing algorithms, such as the FFT7 and the partitioned convolution.8
However, some sparse microphone arrays did survive also for these massive multichannel applications, such as the bulky microphone system developed by the
Japanese television broadcast company NHK in cooperation with the McGill
University for the 22.2 three-dimensional surround system.9
At IED a single, unified meta-theory was developed for describing the processing performed in any kind of dense microphone array. The approach is perfectly
accurate whenever the processing being performed is perfectly linear and time
invariant, as it is the case for many classical approaches.
Whenever a not linear approach is employed, such as the steering methods based
on analysis of the acoustical three-dimensional scene and consequent continuous
change of the filters,10 the approach adopted is still useful for representing each
temporary state of the variable processing system. This remains a rigorous approach whenever the processing is performed over chunks of data, provided that
an independent evaluation of the system’s filtering is performed for each chunk.
The approach becomes just approximate, indeed, whenever the processing employs continuously-changing filters: this would require to recompute a complete
digital filtering network for each data point of the sampled sound.
7
A fast, open source implementation is described in [10].
An implementation example in [21].
9
See [23].
10
As in the Dirac and Harpex-B algorithms [20], [1].
8
6
Introduction
Chapter 2
Microphone arrays: the state of
the art
Having already introduced what is a microphone array and why this kind of
probes have captured the attention for many years, in the following sections a
panoramic, divided by field of application, of the systems currently available on
the market by various manufacturers will be given, preceded by a description of
the so far universally adopted technique of beam forming.
As conclusion, it is reported a short description of the most interesting arrays
developed at IED during the years 2011-2014.
2.1
The beam forming technique
With set of M real microphone capsules, it is possible to synthesize one or more
virtual microphones, with a certain orientation and order, by the application of
the beam forming technique, widely employed in the electromagnetic field, but it
can be applied with some modifications in the acoustic field, too.
With reference of figure 2.1, representing a plane wave that invests a linear
microphone array, if λ is the wavelength of the incident sound, it is possible to
state that a certain wave front, taken as reference, will reach the left capsule
with a delay ∆t respect the time instant in which the same front has invested the
center capsule.
8
Microphone arrays: the state of the art
Signal
Wavefront
λ
θ
θ
l
Figure 2.1: Schematic representation of a plane wave incident on a linear microphone
array.
Let l the distance between nearest capsules and θ the incident angle respect
the normal to the plane on with the capsules lie, this delay can be expressed in
the following way:
∆t(θ) = T
l
sin θ.
λ
(2.1)
It should be now clear that summing the signals coming from the M capsules
with a progressive advance from right to left multiple of ∆t(θ), it will be possible
to obtain a sort of virtual orientation of the synthesized microphone,1 the so
called steering. In equations:
yθ (t) =
−1
1 MX
xk (t + k∆t(θ)) .
M k=0
(2.2)
The desired effect is to obtain the maximum phase cancellation for waves
coming from a direction different from θ, and, conversely, to leave the maximum
amplitude to waves coming from θ direction.
This theory is obviously extensible to planar and three-dimensional arrays,
but in any case the only modifications introduced to captured signals are on
amplitude and phase.
1
Given by the sum of M signals.
2.1 The beam forming technique
9
MICROPHONE
OUTPUT SIGNALS
t2
OUTPUT
SIGNAL
t1
t2
P
t1
0
INCIDENT
PLANE WAVE
SIGNALS AFTER
TIME DELAY BLOCKS
Figure 2.2: Steering of the synthesized virtual microphone.
2.1.1
Limits of the beam forming theory
In the previous paragraph, two important hypothesis were made to simplify the
understanding:
1. incident waves don’t suffer any kind of attenuation;
2. incident front wave is planar, that is the array is placed in far field respect
the sound source.2
Also the spatial aliasing phenomenon was ignored, because the sound field
is sampled in the space: the spatial counterpart of Shannon sampling theorem
states that the minimum wavelength the array can sample correctly is given by:
λ sin θ > 2l,
(2.3)
that becomes λ > 2l in the worst case.
Finally, the equation 2.2 was written ignoring the uniqueness or not of the
solution found for every θ: let the focus direction θ = 0, in this case the equation
2.2 become
yθ (t) =
−1
1 MX
xk (t);
M k=0
(2.4)
it can be noted that incoming signals xk (t) can be written, referencing to
picture 2.1, as a function of the wave arriving to the capsule on the right
2
This implies that the distance between sound source and array is almost equal to the major
dimension of the array.
10
Microphone arrays: the state of the art
!
klT
xk (t) = x t −
sin θ
λ
!
kl
= x t − sin θ ,
c
(2.5)
and, combining the results of equations 2.4 and 2.5,
−1
1 MX
kl
yθ (t) =
t − sin θ .
M k=0
c
!
(2.6)
Passing to frequency domain and by the application of the sum and time shift
properties, it can be easily obtained the transfer function module, expressed in
the equation :
Yθ (f ) =
|H|dB (f ) =
=
M
−1
X
kl
1
X(f ) ·
ejω c sin θ
M
k=0
Y (f ) θ
20 log X(f ) −1
M
X jω klc sin θ 20 log e
(2.7)
[dB].
(2.8)
k=0
In picture 2.3 is shown the equation 2.8 in normalized form on polar diagram
in the specific case M = 4, l = 0.2 m, f = 1 kHz and c ∼ 343 m/s, varying the
microphone capsules number; this picture illustrates that the virtual orientation
θ = 0 of the array has not prevented the generation of secondary lobes on remarkable amplitude. Also, it can be noted that, fixed the sampling distance 1/l,
the acoustic focus - hence the array spatial resolution - increases together with
the employed capsules number, while it decreases with the frequency.3
2.1.2
Advanced beam forming techniques
As already written, the number of application fields for microphone arrays is
very large, but some limits of the traditionally adopted beam forming theory can
3
Concept are well and concisely expressed in [12], where are also present some C code
examples, but for a more deep and scientifically precise analysis, see [4].
2.1 The beam forming technique
11
Figure 2.3: Normalized beam patterns by varying the microphone capsules number.
discourage the use of this kind of probes in place of usual, well experienced technology; for example, the noise analysis inside vehicles, requires to consider a wide
frequency range, because the various kinds of stresses that normally affecting a
vehicle, and beam forming technique cannot have a large frequency range in such
small environments, as equation 2.3 states.
Some of the research efforts to surpass traditional beam forming theory limits are
aimed to develop combined techniques that is where another one is employed in
the application domains where beam forming is weak. An example for the noise
analysis field is the Spatial Transformation of the Sound Field (STSF4 ) that employs Near-field Acoustic Holography (NAH5 ) and its more recent modification,
the Statistically Optimized NAH (SONAH6 ), that, being able to do a more accurate analysis of the near field, it has been proved they are good and efficient
partners of beam forming.7
Basically, NAH is a technique that allows the study of spatial and frequency (or
4
See
See
6
See
7
See
5
[13].
[18].
[15].
[14].
12
Microphone arrays: the state of the art
time) dependence of acoustic fields. Rather than measure all points in 3D space
directly, NAH is used to reconstruct the quantities of acoustic field above and
below the measured hologram plane, which means that the data-acquisition process to be considerably shortened.
The theory of NAH is based on the Helmholtz integral equation and the twodimensional spatial Fourier transform. Complex pressure on a plane parallel to
the hologram plane is computed by inverse 2D FFT of product of hologram pressure spectra and modified Green’s functions. The acoustic velocity vector is then
acquired using Euler’s equation, active and reactive intensity vector as real and
imaginary parts of product of computed complex pressure and conjugate acoustic
velocity.
The STSF was developed for stationary sound fields: the technique was then extended to Non-Stationary STSF, becoming able to apply time domain averaging
techniques, like, for example, the transient analysis.
To be noted that some of the theories listed, like NAH, were known since 1985,
but the CPU computation power needed made them unenforceable at the time
of their formulations.
Another source locations technique, this time as an alternative to beam forming,
is the Inverse Boundary Elements Method, that connects together the Inverse
Method and, as the name said, the Boundary Elements Method; while the latter is a well-known numerical technique for solving general radiation problems,
the first is a method to solve the acoustic source problems that consist of the
determination of source descriptors - e.g. the surface vibration on a virtual surface wrapping the object under investigation - from a knowledge of the radiated
acoustic pressure as measured by a microphone array located around the source.
The major advantage of IBEM is the possible estimation of the vibration velocity and the sound intensity right on the surface of a totally irregularly shaped
source. Therefore, this allows a very precise source location and a very precise
noise source ranking.8
8
For a more complete review of spatial array techniques, see [11].
2.2 Array for noise maps elaborations
13
Figure 2.4: An example of two noise maps overlapped to the machinery analyzed.
2.2
Array for noise maps elaborations
An important field of application for microphone arrays is certainly the environmental acoustic, in particular, these probes are often employed as acoustic
cameras giving to the user a visual map of the pressure level of the space to
which the array is pointed. Usually this map is made partially transparent and
overlapped with a photo of the environment analyzed, so it’s really straightforward the localization of noise sources, as can be seen in figure 2.4.
Some commercially available microphone arrays are especially made for environmental application and described in the following sections.
2.2.1
Brüel & Kjær
Brüel & Kjær produces some arrays dedicated to different environmental applications; after research on microphone dispositions to find optimized random
distributions that correspond to array pattern minimization, the Wheel Array
represents a compromise solution between the difficulty setting the freedom degrees of the random configuration and the complexity of the physical structure
requested by a too random disposition.
The Wheel Array is composed of an odd number of line arrays inclined by a fixed
angle respect the wheel spokes, each of them has the same random configuration,
optimized with the array pattern. The array mechanical design is entirely modular: every spoke can be removed from the support rings - one in the interior and
one in the exterior - and it’s possible to reassemble themselves in different ways.
In figure 2.5 on the left, the version with 90 microphones can be seen: 15
identical spokes with 6 microphones with a total diameter of 2.43 m; it hasn’t an
14
Microphone arrays: the state of the art
Figure 2.5: The Brüel & Kjær Wheel Array.
embedded video camera, but the software allow the overlap and the alignment of
an optical image, captured separately by the user, with the acoustic one, finding
two common points in the two images. In the same picture, on the right, it’s
shown a smaller array, with 66 microphones, 11 spokes and diameter of, circa,
1 m; it’s provided of an embedded camera and it’s more suited for acquisitions
in the near field: the array diameter determines the main lobe amplitude and,
accordingly, the angular resolution.
The package includes the dedicated acquisition hardware, PULSE™ , that offers
some interesting utilities for multichannel calibration, ensures the synchronization between channels and a high quality A/D conversion.
A software, also named PULSE™ , is provided: it can produce both static and dynamic - the acoustic movie - maps. It’s important to mention that Brüel & Kjær
is the only array manufacturer that has released public documentation describing the algorithm employed,9 that is basically a beam forming with a particular
gains set,10 and it’s relevant to point out that, regardless the microphone calibration, the system cannot output absolute pressure levels because it’s impossible to
quantify the spatial noise in a map point.
An acoustical map produced by Brüel & Kjær’s PULSE™ and measured with
the Wheel Array is reported in figure 2.6: coloured spots appears only in corre9
10
See [4].
The documentation reports how these gains has been obtained.
2.2 Array for noise maps elaborations
15
Figure 2.6: Example of Brüel & Kjær’s PULSE™ acoustic map.
Figure 2.7: Brüel & Kjær spherical beam-former.
spondence with the noise sources, this mean that the software shows noise levels
only above a certain threshold, but it’s unknown if automatic or user-defined.
Another interesting model from the danish company is a spherical array (figure
2.7) with 12 video cameras on the surface, available in two versions, with 36 or
50 microphones. The data are acquired placing the sphere in the center of the
measuring area; the algorithm employed is the spherical beam forming, a technology that allows an omni-directional mapping independent from the acoustic
environment.
Brüel & Kjær is also the developer of the SONAH algorithm, that, combining
16
Microphone arrays: the state of the art
NAH with beam forming, tries to conciliate the capabilities of both technologies:
calibrated intensity, pressure and velocity maps of the particles near the source
from the first,11 and high performance with irregular disposition from the latter.
In particular, SONAH is used in the evaluation of holography and the scaling of
the sound intensity outputted by the beam forming algorithm, when collecting
measures with an ad-hoc designed array, the Sector Wheel Array, that maintains
the rotational symmetry of the Wheel Array, but is composed by identical sectors
containing 12 elements disposed following an irregular pattern to optimize the
Maximum Side lobe Level.
2.2.2
gfai tech
Gfai tech is the German company that developed and produced Acoustic Camera, the first microphone array for acoustic visualization appeared on the market.
This producer is experienced and engaged in research in many branches of media
technology and offers various models of arrays, each equipped with camera.
The first reported model is the Ring Array, visible in figure 2.8, it is a measure
system designed for applications in acoustics laboratories, suitable for near field
acquisitions, it is available in versions with 32, 36, 48 and 72 microphones, with
different sizes and characteristics, for example, Ring 36 has a diameter of 0.35 m
and a mapping frequency range between 1.0 to 20 kHz, but the Ring 48, with a
diameter of 0.75 m, maps frequencies between 400 Hz to 20 kHz; in every model
the ring is made in carbon fiber.
In figure 2.9 is shown the Star Array: it is composed of three spokes, the
maximum opening amplitude is 3.40 m, and is suitable for the far field, that
means long distances (from 3 to 300 m) and low frequencies (from 100 Hz to
7 kHz). The system can be folded and moved quickly, so it’s easy to carry it.
The non planar structure is patented and it guarantees the maximum attenuation
for signals arriving from behind that is a requirement for measures in disturbed
environments.
11
This implies a uniform measure grid, with spacing inferior to the half of the wave length,
that covers completely the sound source.
2.2 Array for noise maps elaborations
17
Figure 2.8: gfai tech Ring 48 (size: 0.75 m, Measured range: 0.7 ÷ 5 m, SPLmax :
130 dB, frange : 400 ÷ 20k Hz).
Figure 2.9: gfai tech Star 48 (size: max span 3.4 m, Measured range: 3 ÷ 300 m,
SPLmax : 130 dB, frange : 100 ÷ 7k Hz).
For measures of large objects and special applications like wind gallery, gfai tech
proposes the Spiral Array (figure 2.10) that has excellent acoustic and dynamic
ranges.
18
Microphone arrays: the state of the art
Figure 2.10: gfai tech Spiral 120 (size: 2 × 2 m, Measured range: 3 ÷ 150 m, SPLmax :
130 dB, frange : 200 ÷ 20k Hz).
For closed environment applications, where sound is typically more or less
diffused, so a good resolution on all solid angle is requested, the Sphere Arrays
(figure 2.11) are suggested, suitable for measures on all the space around, they
are built for very small spaces, also vehicle cabins, and high frequencies. As the
Ring model, the sphere array is available with 48 or 120 microphones, with different sizes and characteristics: the first has a diameter of 0.035 m and a frequency
range between 1 kHz to 10 kHz, the latter has a diameter of 0.6 m and measures
between 400 Hz to 10 kHz. The structure is in carbon fiber and guarantees the
maximum acoustic transparency, even the design minimizes aliasing and reflection effects.
Another options offered from gfai tech is that of microphones freely configurable (figure 2.12) presented as the integration of gfai tech beam forming hardware and NAH software from SenSound LLC (USA).
As can be seen in figure 2.13, like Brüel & Kjær PULSE™ , the Acoustic Camera software from gfai tech, plots colours only in correspondence of the sources, so
when it’s capable to reject lateral diffraction and aliasing lobes: it is not known
if this is done simply using a noise threshold, or applying more complex processing algorithms because no description of the techniques used is offered. Acoustic
2.2 Array for noise maps elaborations
19
Figure 2.11: gfai tech Sphere 48 (size: 0.35 m, Measured range: 0.3 ÷ 1.5 m,
SPLmax : 130 dB, frange : 1 ÷ 10 kHz).
Figure 2.12: gfai tech’s free configurable 24 channel microphone bundle (10-10k Hz).
Camera operates in post-processing, but allows a real-time preview with limited
screen resolution that probably uses a simplified algorithm.
2.2.3
Nittobo
The Japanese manufacturer Nittobo produces a system called Noise Vision™ (figure 2.14) that consists in a spherical transducer provided with 31 microphones and
12 video cameras (weights: 12 kg) and a notebook computer with software; the
system allow the generation of an acoustic map in every direction referred to the
spherical transducer. The map representation can be both static, to analyze stationary phenomenon, and animated, to analyze phenomena from non-stationary
or moving sources. The declared frequency range is 200 ÷ 5000 Hz.
20
Microphone arrays: the state of the art
Figure 2.13: Examples of acoustic maps produced by gfai tech Acoustic Camera.
Figure 2.14: Nittobo’s Noise Vision™ system.
The manufacturer doesn’t release specific technical documentation, but only
very generic informations.
2.3
Directive microphoning
Another particularly interesting field of applications for microphones arrays is the
so called directive microphoning which shares working principles and theory with
the system presented in the previous chapter, but for which some manufacturers
realized, in some cases very specific products, in some other more generic ones.
2.3.1
Acoustic Magic
The Voice Tracker™ Array Microphone of Acoustic Magic contains an array if
8 elements and doesn’t present moving parts. It has a rugged structure with
2.3 Directive microphoning
21
Figure 2.15: Acoustic Magic Voice Tracker™ I Array Microphone.
Figure 2.16: Acoustic Magic Voice Tracker™ II Array Microphone.
reduced weight and is easy to use: no specific software is requested, it is not very
demanding for computer CPU and is compatible with different operating systems. It measures 45.7 cm (width) by 6.3 cm (height) and an image is shown in
figure 2.15, where it can be noticed the five optical indicators for chosen speaker
position and signal level.
In the product documentation it is asserted that the system can localize automatically the active speaker and that it can orient electronically the listening
beam around 360°, moving from a speaker to the neighbor in an amount of time in
the order of milliseconds. The speaker is also free to turn or move. The declared
range is about 9 m for meeting recording and some meters for automatic speaking
detecting; frequency response comes from 100 Hz to 11250 Hz.
In addition to focusing on the speaker, the Voice Tracker™ system realize a spatial filtering on noise coming from other directions, but there are no explications
on how this is done. The manufacturer mentions proprietary algorithms for noise
reduction, capable to filter background noise and environment reverberation, in
particular, always without explications, the Location Dependent Source (LDS)
technology is cited that can desensitize microphones for sounds coming from preselected directions.
The manufacturer has also released the version II of Voice Tracker™ , that features
6 microphones instead of 8, an Acoustic Echo cancellation filter, a quantitative
talker location capability, power through USB interface and more compact di-
22
Microphone arrays: the state of the art
mensions.
2.3.2
Yamaha
A representative model, suitable for medium-small environments is the PJP-50R
Conference Microphone System, visible in figure 2.17; the name stands for ProjectPhone: it is a videoconference system with 16 microphones 4 loudspeakers,
connected to, alternatively, a Local Area Network or, an analog phone line or an
audio device with analog line in/out connections.12 It is suitable for four speakers placed approximately around the device. The ring shaped microphone array
can detect automatically the sound source locations, by the analysis of signal
retards, but it is possible to specify some directions to exclude from the analysis
in advance, like known noise sources (for example, blowers). No details about the
employed algorithms on the documentation, but it is mentioned an automatic
system for echo cancellation that corrects filter coefficients accordingly with the
environment.
The loudspeakers reproduce the voice clearly with a volume sufficiently high per
large rooms, in addition there is the possibility (divide mode) to reproduce different voices with different speakers, accordingly to the speakers’ position, so it
can be easy to detect its placement.
A more interesting model from Yamaha was the PJP-100UH (figure 2.18), that
featured 32 microphones, 12 loudspeakers and could serve up to eight speakers,
four for every side of the table where the device was typically placed,13 but now
this product has been discontinued and substituted with a modular solution, the
YVC-1000 (figure 2.19), a system formed by a main unit to which are connected
one or more microphone-speaker units, each of them with 3 microphone capsules
installed: it seems that Yamaha, instead of developing a more complex technology
for a single dense microphone arrays, has preferred to rely to distributed, simpler,
devices, minimizing efforts and speeding up results.
12
13
It exists also a USB version, but equipped with 8 microphones and 4 loudspeakers.
With the optional camera PJP-CAM1, the system could manage a video signal too: it
could choose and bounds the image from the area in which active speaker was located.
2.3 Directive microphoning
23
Figure 2.17: The Yamaha PJP-50R Conference Microphone System.
Figure 2.18: The discontinued Yamaha PJP-100UH Conference Microphone System
with 16+16 microphones on long sides.
2.3.3
Squarehead Technology
The AudioScope (figure 2.20) represents on large scale what can be realized with
a planar array. It is a stand-alone audio system designed for sport and live entertainment shows substantially composed by three parts: the first is a strongly
directive array, realized with a 2 m carbon fiber disc where 315 microphones are
arranged in a regular pattern and with a video camera at the center; this is placed
above the area to shoot and it can cover a surface of the size of a basketball court.
Through a communication device, the array communicates with one or more control centers from which the operator manages the system simply moving the
cursor on the image reproduced on the monitor: in particular a trackball let to
follow the moving sources and with a knob it is possible to rewind the recording
for the replays. The system allows the selection of five audio sources at the same
24
Microphone arrays: the state of the art
Figure 2.19: The Yamaha YVC-1000.
Figure 2.20: The Squarehead Technology’s Audio Scope system: on the left the array,
on the right the control station.
time.
2.3.4
mh acoustics
The microphone array Eigenmike® EM32 (figure 2.21)of American company mh acoustic is since some years at the center of a collaboration between the IED/AIDA
and CRIT.
2.3 Directive microphoning
25
Figure 2.21: mh acoustic Eigenmike® EM32.
The microphone consists of an anodized aluminum sphere, with 43 mm radius, where 32 high quality electret microphones capsules are arranged on the
surface following a relatively uniform pattern. In the sphere interior it has been
placed the entire analog preamplifying circuitry, with software controlled gain,
and the analog to digital converters that can sample capsules signals with 24 bit
of resolution and 48 kHz of sampling rate. The signals are then transferred to the
Eigenmike® Microphone Interface Box (EMIB) with a fully digital connection,
using a common CAT-5 network cable that guarantees disturb immunity also for
long distances between probe and interface (up to 140 m).
The EMIB is connected to a computer through FireWire® interface, behaving
like a normal audio interface; drivers are provided for OS X® and Windows® , but
it works also on GNU/Linux systems using Ffado drivers.
The manufacturer bundles with the microphone a software, the EigenStudio,14
that employs the beam forming technique for sound sources pointing.
This device, starting from the same principles, is quite different from the previous listed, because it offers at the same time, high audio quality, due to the
Sennheiser® capsules, very compact dimensions and a smart connection to the
dedicated hardware, using only one small cable: these features have made the
Eigenmike® an ideal candidate for live audio broadcast applications, and is thanks
to this device that the virtual microphone technique, described in the next chapters, has been developed.
14
Available only for Microsoft® Windows® platform.
26
Microphone arrays: the state of the art
2.3.5
Customs
The requests made to the IED by the broadcast company RAI have conducted
to the development of other Eigenmike® -based arrays,15 more suitable for some
very specific applications.
2.3.5.1
Planar Array
The original Eigenmike® EM32 is a spherical probe and it can resolve sound
sources placed above, below and behind its placement, but in some real cases
there’s no needing to point virtual microphones in these directions, for example
when shooting a show on a stage from a frontal point of view: what happens
in other directions than the frontal one is a disturb, and more spatial resolution
is needed on the front side. A shape that can satisfies these specification is the
planar one, with capsules arranged with semi-random distances between them so
there aren’t privileged wavelength. To get a pointing visual reference, a video
camera - completely missing in the Eigenmike® - has been placed in the center of
the array. The realization is shown in figure 2.22.
2.3.5.2
Cylindrical Array
Similar specifications to which that led to the realization of the IED Planar Array,
gave the birth of the Cylindrical Array16 : this time the scene has to be shot 360°
around the probe, but not above or below, so the ideal shape is a cylinder, always
with the capsules placed following a semi-random pattern. Since in live recording
applications, together with the audio is always needed a video reference for virtual
microphone pointing a camera capable of 360° shooting, described in chapter 4,
has to be placed on the top of the device, as can be seen in the figure 2.23.
15
In these realizations, all the electronic of the original Eigenmike® EM32 has been used, but
placing the microphone capsules in different patterns.
16
Thanks to the hard work of Eng. Lorenzo Ebri and Eng. Lorenzo Chiesi.
2.3 Directive microphoning
Figure 2.22: The IED Planar array with the control laptop.
27
28
Microphone arrays: the state of the art
Figure 2.23: The IED Cilindrical array.
Chapter 3
Virtual microphones
In recent years a lot of research has been produced about technologies for capturing and reproducing the spatial properties of the sound. Most of the proposed
approaches employ massive arrays of microphones and loudspeakers, and process
the signals by means of very complex mathematical theories, based on various
modifications of the classical Huygens principle. These methods include decomposition of the sound field as a superposition of plane waves, working in cartesian
coordinates,1 or as a superposition of spherical harmonics, working in spherical
coordinates.2
Recently, even more complex methods have been proposed by Nelson and
Fazi.3
Whatever method is employed, it is possible always think to the whole processing as the synthesis of a number of virtual microphones, each of them feeding
a loudspeaker in the playback system.
Having realized that, we decided to remove the constraints inherent with any
of the previously-known techniques, and to generate directly the desired virtual
microphones as a direct transformation of the raw recorded signals, without relying on an intermediate layer (or kernel) of basic waveforms.
1
For example Wave Field Synthesis [2].
High Order Ambisonics [19].
3
Decomposing the sound field in complex Hankel functions [9].
2
30
Virtual microphones
Albeit this approach can work, in principle, with any geometrical layout of the
microphone array, we decided to develop our system around an high-quality 32capsules spherical microphone array, the Eigenmike® EM32 made by mh acoustic
fully described in section 2.3.4.
The 32 signals are filtered employing a massive convolution processor, capable of real-time synthesis and steering of up to 7 virtual directive microphones,
controlling their aiming and capture angle by means of a mouse/trackpad, and
employing a wide-angle panoramic video camera and a graphical view & point
interface for an easy-to-operate user’s interface.
This can be done in real-time and with small latency during a live broadcasting
event; alternatively, the raw signals from the 32 capsules can be recorded, together
with the panoramic video, for subsequent synthesis of the virtual microphone
signals in post-production.
3.1
A theoryless approach
The Virtual Microphone System (VMS) developed at IED is not based on any of
the previously known theories, such as High Order Ambisonics, WFS and similar.
There is no intermediate spatial kernel, such as a set of spherical harmonics or
Hankel functions.
The virtual microphones being synthesized can be highly directive,4 and are
intrinsically coincident, so the signals can be mixed without concerns of comb
filtering; it is possible to continuously move their aiming for following actors or
singers on scene, or for giving instantaneous miking to people in the audience.
Surround recording of a concert is just one of the possible scenarios for employing
this approach, which has also been successfully tested for dramas, sport events,
and TV shows in which there is systematic interaction of conductors and guests
with in-studio audience.
4
The polar pattern is, in fact, constant with frequency, and with a beam width much sharper
than a shotgun microphone.
3.1 A theoryless approach
31
y1 (t)
x2 (t)
SIGNAL
PROCESSOR
..
.
xM (t)
x1 (t)
..
.
yV (t)
y1 (t)
h1,1 (t)
x2 (t)
y2 (t)
V Outputs
M Inputs
x1 (t)
h2,1 (t)
..
.
xM (t)
hM,1 (t)
Figure 3.1: Scheme of signal processing.
A careful analysis of the performances of the new microphone system did show
that frequency response, signal-to-noise ratio and rejection of off-beam sounds are
better than those obtainable employing traditional processing algorithms applied
to the same input signals, or dedicated top-grade ultra-directive microphones.
3.1.1
Computation of the filters
Given an array of transducers, a set of digital filters can be employed for creating
the output signals. In our case the M signals coming from the capsules need to
be converted in V signals yielding the desired virtual directive microphones: so
we need a bank of M ×V filters. For a couple of reasons, FIR filters are preferred.
Assuming xm as the input signals of M microphones, yv v as the output signals
of V virtual microphones and hm,v the matrix of filters, the processed signals can
be expressed as:
yv (t) =
M
X
xm (t) ∗ hm,v (t)
(3.1)
m=1
Where ∗ denotes convolution, and hence each virtual microphone signal is
obtained summing the results of the convolutions of the M inputs with a set of
M proper FIR filters.
32
Virtual microphones
One of the most used techniques for deriving filter banks for generating virtual microphones with arbitrary directivity is the Ambisonics method: first the
M signals are processed deriving an equal or smaller number of spherical harmonics. Later these spherical harmonics signals are added together with proper
gains, for synthesizing the desired virtual microphones. This has the advantage
of allowing for the derivation of a large number of virtual microphones with a
small additional effort, as most of the effort is represented by the computation of
the spherical harmonic signals.
Furthermore, it is possible to change dynamically the aiming or the directivity
of every virtual microphone simply changing the gains employed when adding the
signals together.
In our approach, instead, every desired virtual microphone is derived directly
from the original M signals, avoiding the Ambisonics encoding and decoding: the
outputs of the processing system are directly the result of the digital filtering of
the input signals, with a different set of filters for every virtual microphone.
In principle this allows for synthesizing virtual microphones having an arbitrary
directivity pattern. In practice we decided, for now, to synthesize frequencyindependent high-order cardioid virtual microphones.
The directivity factor Q of a virtual cardioid microphone of order n is described,
in spherical coordinates θ, φ , by the expression:
Qn (θ, φ) = [Q1 (θ, φ)]n
(3.2)
Where Q1 (θ, φ) is the directivity factor of a first order cardioid microphone:
Q1 (θ, φ) = 0.5 + 0.5 · cos(θ) · cos(φ)
(3.3)
The processing filters h are usually computed following one of several complex mathematical theories, based on the solution of the wave equation, often
under certain simplifications, assuming the microphones are ideal and identical.
In some implementations the signal of each microphone is processed through a
digital filter for compensating its deviation, with a heavier computational load.
3.1 A theoryless approach
33
90
1
120
60
0.5
150
30
order 1
order 2
order 3
180
0
order 4
order 5
order 10
210
330
240
300
270
Figure 3.2: Polar plots of virtual cardioid microphones of various orders (target
patterns Qn ).
In this novel approach no theory is assumed: the set of filters h are derived
directly from a set of measurements, made inside an anechoic room. A matrix of measured impulse response coefficients c is formed and the matrix has to
be numerically inverted5 ; in this way the outputs of the microphone array are
maximally close to the ideal responses prescribed. This method also inherently
corrects for transducer deviations and acoustical artefacts (shielding, diffraction,
reflection, etc.).
The characterization of the array is based on a matrix of measured anechoic
impulse responses, obtained with the sound source placed at a large number D
of positions all around the probe, as shown in figure 3.3.
The processing filters h should transform the measured impulse responses c
into the prescribed theoretical impulse responses p:
pd =
M
X
cm,d ∗ hm
d = 1...D
(3.4)
m=1
5
usually employing some approximate techniques, such as Least Squares plus regularization
34
Virtual microphones
d = 1...D

Source positions
d(t)
C=
m = 1...M
microphones














c1,1
c2,1
..
.
cm,1
..
.
cM,1
c1,2
c2,2
..
.
cm,2
..
.
cM,2
...
...
...
...
...
...
c1,d
c2,d
..
.
cm,d
..
.
cM,d
...
...
...
...
...
...
c1,D
c2,D
..
.
cm,D
..
.
cM,D















Figure 3.3: Impulse response measurements from D source positions to the M microphones.
Please notice that in practice the target impulse responses pd are simply obtained applying a direction-dependent gain Q, given by equation 3.2, to a delayed
unit-amplitude Dirac’s delta function δ: pd = Qd · δ.
Computation is easier in frequency domain.6 Let’s call C, H and P the resulting complex spectra. This way, the convolution reduces to simple multiplication
between the corresponding spectral lines, performed at every frequency index k:
Pd =
M
X
Cm,d,k · Hm, k
m=1


d = 1...D

k = 0 . . . N/2.
(3.5)
Now we pack the values of C, H and P in proper matrices, taking into account
all the M input microphones, all the measured directions D and all the V outputs
to create:
[Hk ]M ×V =
[P ]M ×V
[Ck ]D×M
(3.6)
This over-determined system doesn’t admit an exact solution, but it is possible
to find an approximated solution with the least squares method, employing a
regularization technique for avoiding instabilities and excessive signal boost7 . The
block diagram of the least-squares method is shown in figure 3.4.
In this scheme we observe the delay block δ, required for producing causal
filters, and the resulting total modelling error e, which is being minimized by the
6
That is, computing the complex spectra, by applying the FFT algorithm to the N -points-
long impulse esponses c, h and p
7
See [16] and [17].
3.1 A theoryless approach
35
δk
Q
-
D
+
Ck
M
ed
Hk
Figure 3.4: Scheme of the Least Squared method with a delay in the upper branch.
least-squares approach.
In general, the frequency-domain representation of a Dirac’s delta delayed by
n0 samples is given by:
n0
δk = e−j2πk N
(3.7)
Albeit various theories have been proposed for defining the optimal value of
the causalisation delay n0 , we did take the easy approach, setting n0 = N/2.
Choosing N/2 samples is a safe choice, which creates inverse filters with their
main peak close to their centre, and going smoothly to zero at both ends.
Furthermore, a regularization parameter is required in the denominator of the
matrix computation formula, to avoid excessive emphasis at frequencies where the
signal is very low.
So the solution formula, which was first proposed in [17], becomes:
[Hk ]M ×V
[Ck ]∗M ×D · [Q]D×V · e−jπk
=
[Ck ]∗M ×D · [Ck ]D×M + βk · [I]M ×M
(3.8)
As shown in picture 3.5, the regularization parameter β should depend on frequency. A common choice for the spectral shape of the regularization parameter
is to specify it as a small, constant value inside the frequency range where the
probe is designed to work optimally, and as much larger values at very low and
very high frequencies, where conditioning problems are prone to cause numerical
instability of the solution.
36
Virtual microphones
β(f )
βoh
βol
βi
fl1
fl2
fh1
fh2
f
Figure 3.5: Regularization parameter in dependence of the frequency.
3.1.2
Experimental characterization of the array
Measurements of the microphone array were made employing the Exponential
Sine Sweep (ESS) method, in order to obtain 32 Impulse Responses for each direction of arrival of the test signal.
The ESS method was chosen due to its capability of removing unwanted artifacts due to nonlinearities in the loudspeaker, and because it provides significantly
better S/N than other methods based on periodic signals, such as MLS or linear
sine sweep (TDS), as the researches of A.Farina already proved.8 This made it
possible to get a good S/N ratio employing short test signals, speeding up the
measurement procedure, and consequently enhancing the time invariance of the
system during the measurement period.
These measurements were made inside an anechoic room, to avoid undesired reflections and to maximize the signal/noise ratio.
The test signal was pre-filtered with a suitable FIR in order to flatten perfectly
the frequency spectrum and to linearize the phase response of the full-range dualconcentric loudspeaker employed as sound source.
The array was rotated along azimuth (36 steps) and elevation (18 steps), using
a movable fixture for the azimuth rotation and a turntable for the elevation. In
this way we obtained 36 × 18 × 32 impulse responses, each 2048 samples long (at
48 kHz).
8
See [7].
3.1 A theoryless approach
37
Figure 3.6: Frequency response and polar patterns for capsule n.1.
The raw result of the measurement are the response of each capsule of the
array to the sound arriving by every direction. In picture 3.6 are shown the
results for the capsule n.1 (all the others are very similar):
From the graphs shown in figure 3.6 it is possible to verify that the directivity of the capsules, intrinsically omnidirectional at low frequency, is highly
influenced by the microphone structure (the aluminum sphere): above 1 kHz the
capsule becomes significantly directive. From the magnitude spectrum calculated
in the direction in which the capsule shows the higher sensitivity, it is possible to
verify that the transducers have a quite good frequency response from 30 Hz to
13 kHz, with a gentle roll off at higher frequencies.
A small phase error was introduced by this measurement technique in every
impulse response, due to the variation of the distance between the microphone
array and the sound source during the rotation. In particular the rotating table
produced a sinusoidal variation of the distance caused by the little eccentricity
of the support, which was accurately measured thanks to the absolute time-offlight of the sound from the loudspeaker to the microphone, as shown in figure 3.7.
With a mathematical calculation, based on the theoretical distance between
capsule and source, implemented in a Matlab program, the impulse responses
were carefully realigned in time, correcting for these positioning errors. Before
38
Virtual microphones
Figure 3.7: Variation of the distance during rotation.
Figure 3.8: Improvement of polar patterns of a 3rd order virtual cardioid due to delay
compensation.
this operation the worst case of distance error was 30 mm, after it was of 2 mm, a
length that corresponds to the wavelength of a 170 kHz signal. For this reason we
could assume that the phase coherence of impulse responses is reasonably good
in the spectral region of interest. As shown in the figure 3.8, in the case of a
3rd order virtual cardioid, the compensation improves the polar pattern at 4 and
8 kHz.
In the band of 16 kHz the signal is compromised in both cases, and this is probably
due to the spatial aliasing.
3.1 A theoryless approach
39
Figure 3.9: 1st order figure of eight.
3.1.3
Synthesis and test of virtual microphones
In order to derive the matrix of filters, a Matlab script was produced. This script
employs 2048 samples of each impulse response and it needs as inputs the number
of virtual microphones to synthesize, their directivity, their azimuth and elevation. From these inputs, according with the theory and the procedure described
in paragraph 3.1.1, it is possible to invert the matrix of impulse responses obtaining the matrix of filters to apply to the capsule signals.
In our specific case the number of incoming sound directions D was 648, the
number of microphones M was 32 and the number V of virtual microphones was
the one desired.
The convolution of the FIRs matrix with the 32 signals coming from the
capsules of the array should give as outputs the signals of virtual microphones
with the desired characteristics. In pictures 3.9, 3.10 and 3.11 are shown some
experimental results, showing some of the different directivity patterns obtained.
40
Virtual microphones
Figure 3.10: 3rd order cardioid.
Figure 3.11: 6th order cardioid.
Chapter 4
The optical part
The virtual microphone system, like the acoustic camera-like products listed in
chapter 2, needs always a visual reference, otherwise it would be impossible to
point correctly the virtual microphones to the desired sources. This is requested
in both post-processing and real-time applications, so a new problem was introduced during the development of the research project between IED and RAI
broadcasting company: the systems, to be suitable for live events, must be provided with a camera and the recording software must be capable to show the
frame as background in the virtual microphone pointing space, as fully described
in section 5.1.4.
But while a probe like the Planar Array requires a “planar” background image,
so every video camera can be suitable for this use, a non-planar array needs visual
reference that covers as much as possible the spatial range of this kind of probe:
for a spherical array, the shooting angles, in terms of width and elevation ranges,
are 360°/180° and for a cylindrical array they are 360°/∼ 100°. The solution that
has been successfully experimented in the realizations of IED is the coupling of
a camera with an hyperbolic mirror that outputs a full panoramic image to be
unwrapped with a proper software algorithm, as described in section 4.2.2.
4.1
The camera
In the first attempts to embed a camera in the array a low-cost webcam, was used,
connected to the computer through a USB port, and this kind of device revealed
42
The optical part
immediately a couple of drawbacks, the most important of them are:
• USB connections are limited to relatively short distances: in a broadcast
live set cable connections can be long tenths or hundreds of meters;
• webcams have cheap fixed lens and quite low resolutions, not suitable to
obtain good quality output frames after the unwrapping process.
To overcome these issues, the search of suitable devices was done between
cameras with IP technology (IP-cameras): in these cameras, typically, only a
standard network cable with RJ45 plugs is needed to connect them because power
also comes from this cable.1 This is indeed a great benefit for the system robustness and portability because with only two network cables - that, as known, can
be long tenths of meters - it is possible to feed, send and receive data from the
array and the camera. The quality target was achieved choosing two different
cameras, one for the Planar array where unwrapping process is not needed and
another for the applications where a full 360° image is necessary; the first camera is the model IPELA® SNC-CH210 from Sony® (figure 4.1), the latter is the
WS-M8P31-38B model (figure 4.2) from Chinese manufacturer Wision.
In table 4.1 some characteristics of the two cameras are reported.
Both cameras deliver the signal using the Real Time Streaming Protocol (RTSP),
widely adopted and well supported by the most known video player and recorders.
4.2
4.2.1
The panoramic frame
Mirror and lens
In order to obtain a good panoramic unwrapped frame it is absolutely necessary
to do a very precise optical and mechanical coupling between mirror and camera
lens; for the IED Cylindrical array was chosen a mirror built by 0-360° Panoramic
Optic™ (figure 4.3), that is a glass hyperbolic mirror with a horizontal visual field
1
The so called Power over Ethernet (PoE) technology; obviously a common network adapter
or switch cannot power up these devices: a PoE power unit is needed, but this has not to be
near the device.
4.2 The panoramic frame
43
Sony IPELA SNC-CH210
Wision WS-M8P31-38B
1/2.8’ 3M Pixels CMOS
2.0 lux
1/2.5’ 5M Pixels CMOS
Color: 1.0 lux @ F1.2
B/W: 0.1 lux @ F1.2
Camera
Image Sensor
Minimum illumination
Shutter control
1 to 1/10000 s
fixed
Lens type
88◦
3.3 mm
Horizontal viewing angle
Focal length
-
2.8
-
0.5 m
-
F-number
Minimum object distance
1/5 to 1/50000 s
interchangeable with C/CS mount
-
Image
2048×1536
2592×1920
2048×1536 (JpegOnly),
1920×1080,1600×1200,
1680×1056,1440×912, 1280×1024,
1280×960,1376×768, 1280×800,
1024×768,1024×576, 768×576,
720×576,704×576, 800×480,
720×480,640×480, 640×368,
384×288,352×288, 320×240,
320×192, 176×144
2592×1920, 2048×1536,
1920×1080 (1080p),
1280×720 (720p)
Maximum size (H×V)
Codec image size (H×V)
Video compression format
Maximum frame rates
H.264, MPEG-4, M-JPEG
H.264:
H.264, M-JPEG
10 fps: 2592×1920
15 fps (1920×1080)/
30 fps (1280×720)
M-JPEG: 12
15
30
MPEG-4: 15
30
20 fps: 2048×1536
fps (2048×1536)/
fps (1920×1080)/
fps (1280×720)
30 fps: 1920×1080
30 fps: 1280×720
fps (1920×1080)/
fps (1280×720)
Network
Protocols
IPv4,IPv6,TCP,UDP,ARP,ICMP,
IGMP,HTTP,HTTPS,FTP (client),
SMTP,DHCP,DNS,NTP,RTP/
RTCP,RTSP,SNMP (MIB-2)
ONVIF software
HTTP,DHCP,UDP,RTP/RTSP,
NTP,FTP
Yes
Yes
10BASE-T/100BASE-T (RJ-45)
10BASE-T/100BASE-T (RJ-45)
100 g
-
44 × 93 mm
DC 12V/PoE
Interface
Ethernet
General
Weight
Dimensions
Power
PoE (IEEE802.3af compliant)
-
Table 4.1: Some characteristics of the Sony® and Wision cameras adopted for planar
and cylindrical array respectively.
of 360° and a vertical one of 115°,2 that was, in fact, a project constraint. The
hyperbolic shape conducts the reflected rays in the first focal point of the hyperboloid - “internal” to the mirror - hence the optical center of the camera lens
must be placed in the second focal point. A mirror with these characteristics is
2
Not symmetric: 52.5° above and 62.5° under the horizon.
44
The optical part
Figure 4.1: The Sony® IPELA® SNC-CH210 IP-camera with black case.
Figure 4.2: The Wision WS-M8P31-38B IP-camera in bulk version without lens.
defined as single point of view, and allows the correct estimation of distances and
moving objects as it were a natural perspective image. That was not true for
spherical mirrors, where rays doesn’t intersect in a single point, creating a spherical aberration; parabolic mirrors, however, require a telecentric lens in order to
obtain single point of view images.
Once the mirror has been chosen, the correct camera lens had to be found; the
target was to obtain a complete mirror image with minimum or no aberration at
all, minimizing at the same time the distance between mirror and lens, to make
the entire system as small as possible.
Different lenses was tested, evaluating the following parameters:
• the minimum distance D that allows the correct field angle on the minor
4.2 The panoramic frame
45
Figure 4.3: The 0-360° Panoramic Optic™ hyperbolic mirror.
Figure 4.4: The optical path in an omnidirectional vision system with hyperbolic
mirror.
CCD sensor dimension (usually, the vertical one);
• the minimum focus distance on vertical pattern Wr (D on figure 4.5);
• the field depth, or the correct focus within a defined range;
• the qualitative optical distortion;
• the lens physical size.
46
The optical part
Figure 4.5: Scheme of the field depth.
Figure 4.6: The mirror and the camera in test mount; the cover of the space around
the lens is missing and the body is not yet painted in the final black color.
Finally, it was chosen a 6 mm lens, with iris range from F1.2 to F16
The Camera and mirror has been mounted, perfectly coaxial, inside a transparent
plexiglass container, with a black needle that removes internal optical reflections.
The finished object is visible in figure 4.6.
Since, once the system is complete the camera is not anymore easily reachable,
before screwing the structure a lens setup has been done, finding the regulations
that allow the largest field of applications: lens focus was regulated to find a
4.2 The panoramic frame
47
compromise focus for near and far subjects in the mirror reflected image. The
field depth is, in this case, the really important parameter in order to focus
subjects placed at different distances. The chosen setup is slightly out from the
near field.
Then the iris diaphragm opening has to be regulated, determining the amount
of light that enters in the lens and heavily influencing the field depth obtainable.
Closed iris means long field depth, conversely is the iris is open, the field depth
will result very short. Obviously the desired behavior was to have an infinite
field depth, but when the environment is poor illuminated, the camera built-in
automatic image lightness control system intervenes, opening the iris to get more
light and reducing the field depth. So the iris opening was regulated in order
to obtain the maximum field depth and the minimum disturb allowed in poor
illuminated environments, about 2/3 of the fully opening.3
4.2.2
The unwrapping technique
There are a lot of commercial and free software to obtain the projection on a
rectangular surface of the 360° mirror reflected image - the unwrapping -, but all
of them does single images processing; in this case the camera video stream has
to be unwrapped in real-time, frame by frame, thus a software was developed
for this purpose, in the form of GStreamer (see section 5.1.4) plugin, because the
images processed are large (5 Mpixel) and the performance requested is high.
The core of the geometric transformation is the passage from the cartesian
coordinates system of the starting frame (rectangular, with the image of the mirror, as can be seen in figure 4.7) to a polar coordinates system centered on the
center of the image of the mirror, then returning to a new cartesian coordinates
system with only the pixel contained in the mirror image suitably translated so
as to form a rectangular image.
The processing algorithm to work properly needs some parameters that de3
The same light regulation system corrects the entire image lightness preferring an interme-
diate value between minimum and maximum: this mean that environments with concentrated
light sources and dark zones results, respectively, over and under-exposed.
48
The optical part
Figure 4.7: The image of the mirror seen by the camera.
x1
y
y1
Rmax
P
Rmin
r
θ
x
Figure 4.8: The polar coordinates system.
pends by the particular physical shape of the employed mirror and by the optical
coupling between mirror and camera; in figure 4.8 are highlighted the radius Rmax
and Rmin that can be determined with a calibration procedure, using a static test
frame, as can be seen also in figure 4.9.
The output reference system is shown in picture 4.10, where the scale factor of
both axis is user defined and it depends by the desired size of the output frame.
4.2 The panoramic frame
49
Figure 4.9: The optical calibration procedure.
x2
y2
x0
Ouput Frame Height
y0
Output Frame Width
Figure 4.10: The output (cartesian) coordinates system.
It is also necessary to know the curvature of the mirror shape, in this case
expressed by the experimentally obtained4 equation
y 0 = −0.4r2 + 1.4r,
(4.1)
that correlates a pixel final y 0 coordinate (figure 4.10) to the radius r of the
4
No informations about it were given from the producer.
50
The optical part
Figure 4.11: The unwrapped frame.
starting (θ, r) reference system (figure 4.8).
Now it is possible to obtain a law, based on the polar to cartesian conversion
equations


x = r cos θ

y = r sin θ,
(4.2)
that accomplish the desired unwrapping, as can be seen in figure 4.11.
This algorithm was firstly written as Matlab script, then implemented modifying
an existing GStreamer module5 and finally a brand new module, named VMSUnwrap, has been written and inserted in the main 3D VMS source code.
5
Work done mostly by Eng. L.Chiesi.
Chapter 5
The software
After the first successful experiments with the virtual microphones technique,
mainly executed with Matlab scripts, it was quickly clear that to obtain a product stable, reliable and easy installable in a live broadcast or post-production
setup, as requested by the partnership with RAI, a new custom software had to
be developed, and, through progressive and continuous enhancements and improvements, this software, called 3D Virtual Microphones System (3D VMS), is
currently always used as part of the standard RAI live broadcast equipment. To
reach this target, a number of issues had to be faced:
• provide high directivity in order to exclude the undesired noises that usually
affect audio recordings in broadcast and film production;
• provide the capability of changing orientation and directivity in real-time
by using a controller (joystick/mouse/trackpad);
• the capability of using this system in post-production, varying the virtual
microphones on recorded audio;
• the capability of seeing the subject on which the virtual microphone is
focusing;
• low latency;
• high audio quality;
• the control unit must run on Apple® OS X® .
52
The software
3D VMS
Main process
Control Interface
SMTPE/LTC
manager
Video acquisition
engine
(Gstreamer)
Other independent
processes
Playback/record
engine
(Ecasound)
Filtering
engine
(BruteFIR)
JACK Audio Connection Kit
CoreAudio
Kernel network
driver
Hardware
side
EMIB
IP Camera
Figure 5.1: The structure of 3D VMS.
5.1
3D Virtual Microphones System
How can be seen in figure 5.1, the 3D VMS is composed of different parts, in some
ways independents between them: roughly it is possible to individuate three main
blocks:
• the filtering engine
• the recording/playback engine
• the control interface.
In the really first implementation of the 3D VMS these blocks resides on different machines because of an attempt to make the entire system rack-mounted
and remote controlled, so the audio components were enclosed in a black box,
controlled by a laptop, where the graphical interface was installed: this layout influenced deeply the development of 3D VMS that, like sometimes happens in big
5.1 3D Virtual Microphones System
53
projects, in fact never leave this architecture; actually the system runs flawlessly
on a single, well boosted, laptop,1 but the software components still acts as they
are on different machines, so they are completely separated processes.
These blocks can communicate in bi-directional way between them to grant
full control on what is happening and all the audio is routed through Jack Audio
Connection Kit 2 (JACK).
As previously mentioned, a very important specification, needed in particular
for live processing, is the latency that has to be very low, typically inferior to
40 ms.
5.1.1
From Pyhton to OS X® App
Actually the 3D VMS is written completely in C/C++, but at the beginning of
the project it was not so: while BruteFIR and Ecasound were - and still they are
- stand alone programs, to obtain quickly a working prototype of the system,
Python programming language was employed to write the GUI, at the time indistinguishable from the control interface, built with wxWidgets3 libraries that
guarantee a visual family-feeling with other OS X® applications.
The very first versions of the software were written with a prototyping paradigm
in mind, so very little true object-oriented programming (OOP) was used, but
this, when the complexity begun to grow, soon had shown its limits: in 2011
a first complete rewrote of the code brought to a full OOP Python code, much
more readable and simpler to maintain.
Python is a wonderful tool that allows to write programs in a small amount
of time, forcing the author to use some “style”, but it is and remains a very high1
When the first prototype was built, in 2009, the CPU computational power requested by the
filtering engine was heavy, and a single portable computer cannot run the entire system without
glitches in the audio output. Nowadays, the audio engine is still demanding, but laptops are
much more powerful and for live events less boxes are needed, the better is.
2
http://www.jackaudio.org
3
http://www.wxwidgets.org
54
The software
level language that needs an interpret and a lot of installed libraries, plus, in this
case, the wxPyhon framework; one of the first big problem encountered was the
incompatibility between the default OS X® Python version4 and the binary wxPython distribution: this was resolved easily freezing the Python version to 2.6,
but it comes immediately clear the risk of a software hang after, for example, an
OS update that made unavailable the chosen version of Python.5 Again, Python
is a non-compiled language, this means that every kind of error comes out at
runtime, and it is possible to detect them only running the program from the
console, if this doesn’t happen, the application hangs and the user doesn’t know
why.
Finally the 3D VMS has to be simply installed and eventually uninstalled,
like a common OS X® application: the Python version of 3D VMS had really a
lot of dependencies, not easy to bundle in a reliable way.
For these reason, in the fall 2013, the 3D VMS was entirely converted to C++
and all its components bundled in a OS X® App; the only external dependency
remained is JACK that still have to be installed separately.
The code conversion was quite straightforward because most part of the
Python source had been written using wx-based classes6 and since wxWidgets
C++ namespace is very similar to its Python implementation, the editor search/replace
function was one of the most used.
As side-effect, the C++ version has significantly better performances than the
previous Python one, and this became really crucial when GStreamer entered in
3D VMS source mainstream (see paragraph 5.1.4).
4
In OS X® 10.5 and 10.6 the default Python version was 2.6; in OS X® 10.7, 2.6 was still
installed, but the default was 2.7.
5
Another considered solution was to embed in the 3D VMS installation package the entire
working Python 2.6 + wxPython distributions, but writer fully disagree with this kind of
approaches.
6
wxWidgets library offers not only GUI related classes, but a very complete set of helpers
and accessories to cover all the needing of a generic application in a cross-platform way.
5.1 3D Virtual Microphones System
5.1.2
55
The filtering engine
One of the most critical parts of the system is obviously the filtering process, that
involves not only a non-negligible number of channels, but also it must have the
capability of real-time filter switch, when the user moves a virtual microphone in
the space.
To reach a good reliability as soon as possible, instead of writing a filtering engine from scratch, a ready-made solution was searched, finding an open source
tool of proved stability and reliability, BruteFIR7 written by Anders Torger8 , that
matches system requests.
BruteFIR permits the implementation of up to 256 filters and considering that
for one single microphone the system needs 32 filters - one for each input channel
- we can get 8 virtual microphones. The choice of synthesizing only 7 virtual
microphones is due to the need of having one free slot (32 filters) dedicated to
the double buffering, the dynamic switch from the old and the new virtual microphone every time the user needs to modify the characteristics of one of the virtual
microphones (figure 5.2). With BruteFIR, it is also possible to have a cross-fading
during the change of the coefficients with the benefit of a glitch-free transition
between two different positions of the virtual microphone.
BruteFIR, born for GNU/Linux, was never ported to OS X® systems, but its
intrinsic POSIX nature and the full support of JACK audio interface had permitted a relatively straightforward recompilation on OS X® environment, with only
some minor changes.
To keep the 3D VMS control interface informed about program status, an
error communication channel was added, implemented with a POSIX socket to
which is possible for another application to connect and receive status messages
from BruteFIR.
7
8
http://www.ludd.luth.se/ torger/brutefir.html
See [21].
56
The software
32 Inputs
224 Filters
x0
0
x1
1
..
.
..
.
x31
7 Outputs
V
y0
V
y1
31
32
33
..
.
63
..
..
.
192
193
..
.
V
y6
223
Unused filters
224
225
..
.
255
Figure 5.2: BruteFIR filtering scheme.
5.1.2.1
The logic module
Obviously BruteFIR doesn’t come with the implementation of the FIR’s synthesis
process described in the paragraph 3.1.3, but it’s fully modular and its open source
nature permits to write new modules9 starting from existing ones, so it had to
be possible to realize a 3D VMS own module implementing direction dependent
filter synthesis, and permitting the control of 7 independent virtual microphones.
The virtual microphone’s parameters can be modified in real-time via POSIX
9
In BruteFIR world a module is a plugin, loaded in runtime on request.
5.1 3D Virtual Microphones System
57
socket interface10 through a proper protocol. Every time the parameters are
modified, the system automatically recalculates the FIR filters and substitutes
them in the memory of the convolver through a double buffering mechanism.
The filter computation subroutine was rewritten in highly-optimized C, making
it possible to recalculate all the filters in a few milliseconds, so that the virtual
microphones can be smoothly moved around without artifacts.
5.1.3
The recording/playback engine
The recording of 32 channels at the same time also, is a task for which there are
plenty of affordable and reliable good solutions, so another time an open source
program fits perfectly the requests: Ecasound11 by Kai Vehmanen.
To be honest, this software is much more than a multichannel player and recorder,
having also advanced mixing and filtering capabilities, not needed for this project,
but the interactive mode and the ready-made Python and C++ interfaces come
immediately very useful when Ecasound has to be controlled from an external
process, and this is the case of 3D VMS: thanks to these features it is possible at
every instant to send commands and to have a feedback from the audio transport
without latency at all.
Unlike BruteFIR, Ecasound is distributed for various platforms, including OSX,
but the version included in the 3D VMS package is in many ways customized to
support BWF file format, needed to store correctly timecode data, as explained
in section 5.1.5.
5.1.4
The multimedia engine
It was completely described in chapter 4 the optical system needed to have a
coherent visual reference of what is shot by the camera, and also it was described
that video frames have to be processed to give out the correct unwrapped image,
but the first 3D VMS software version mentioned in section 5.1.1, in fact, aren’t
10
BruteFIR support various communication interfaces: when filtering and control were on two
separated machines, TCP/IP was used.
11
http://nosignal.fi/ecasound/
58
The software
Figure 5.3: The 3D VMS main dialog window.
able to do the elaboration in real-time: this was done only in post-processing,
because the wxPython wrap of the framework used to display the captured video
doesn’t allow the addition of code for frame processing.
Since the correct visualization of the video reference during real-time filtering
had to be an essential software feature, an extensible and embeddable tool had
been searched, and GStreamer12 was found.
GStreamer is indeed a really complex piece of software: it is open source and
what follows is only a subset of all its features, taken from the website:
• graph-based structure allows arbitrary pipeline construction;
• multi-threaded pipelines are trivial and transparent to construct;
• clean, simple and stable API for both plugin and application developers;
• extremely lightweight data passing means very high performance/low latency;
• dynamically loaded plugins provide elements and media types, demandloaded via a registry cache;
12
http://gstreamer.freedesktop.org
5.1 3D Virtual Microphones System
59
Figure 5.4: GStreamer architecture.
• GStreamer capabilities can be extended through new plugins.
• Some features available using the GStreamer own plugins, not counting any
3rd party offerings:
container formats : asf, avi, 3gp/mp4/mov, flv, mpeg-ps/ts, mkv/webm,
mxf, ogg;
streaming : http, mms, rtsp;
codecs : FFmpeg, various codec libraries, 3rd party codec packs;
metadata : native container formats with a common mapping between
them;
• gst-launch command-line tool for quick prototyping and testing, similar to
Ecasound;
• a lot of documentation, including partially completed manual and plugin
writer’s guide;
• access to GStreamer API with various programming languages.
60
The software
A scheme of how GStreamer works is reported in picture 5.4, where is pretty
clear its full modular structure, much similar to a studio with a lot of rack
mounted devices routed together to obtain the desired effects on the incoming
signal that can be video or audio or both at the same time. From the user
point of view, the modules are connected using the principle of the pipeline,
and the GStreamer distribution provide a command line tool named gst-launch
to which it is possible passing the desired pipeline as a string parameter with
some syntax conventions. Every module can have one or more inputs named
sinks and/or one or more outputs named sources; modules without sinks, normally start the pipeline, because they themselves are a source, conversely modules
without sources, are placed at the end of the pipeline. It is possible to split the
pipeline in two or more branches using, for example, a tee, a module with one
sink and two or more sources and, with the GStreamer API, it is possible to dynamically change the pipeline.
A source and a sink, when connected together must be compatible,13 if it’s not
the case, an appropriate conversion module must be placed between them.
A module behavior can be controlled by parameters that can be varied by the
user exactly like pushing buttons or moving knobs on a rack device. For example
gst-launch videotestsrc ! ximagesink
this really simple pipeline use the standard test video signal generator module,
named videotestsrc piped - the ’!’ symbol - to the X window server monitor, so
in a X based environment a window will display the piped video signal. Another
more complex example
gst-launch videotestsrc ! \
video/x-raw-yuv, framerate=20/1, width=640, height=480 ! \
x264enc bitrate=512 ! mpegsmux ! filesink location=test.ts
Here the same video test signal is piped to a x264 encoder, with bitrate set
to 512, through caps filtered 14 source/sink which set the image format and the
13
The tool gst-inspect called with a module name as parameter, gives to the user a full
information page about its parameters, sources and sinks.
14
That is the contraction of capabilities filter: it is not a module but it is a limitation imposed
to the source/sink of connected modules.
5.1 3D Virtual Microphones System
61
framerate; then the signal is multiplexed into a MPEG container and sent to a
file whose name is passed as location parameter to the filesink module.
Thanks to the develop and debug tools made available, the unwrapping plugin
prototype was quickly developed, then the challenge was to integrate GStreamer
with the existing 3D VMS mainstream that is wxWidgets based, so it hasn’t
nothing to do with the glib15 based GStreamer, that, like BruteFIR and Ecasound,
works as an independent process with its own threads, a messages’ bus, etc. The
junction point is the capability of GStreamer to send a video stream on various
destinations, one of them is the so-called appsink from which, defining a callback
function, it is possible to access the frame buffer and, for example, to copy it to
a window canvas: this is exactly what happens inside 3D VMS, and it becomes
clear how performance is important because the amount of data that has to be
copied from GStreamer buffer to the window canvas can be huge in the case of
5 Mpixel RGB images.
The scheme of the pipeline used in the Realtime activity of 3D VMS is fully
reported in picture 5.5: the RTSP signal coming from the IP-camera is collected
by the rtspsrc module that begin the pipeline, pass through some conversions,
then, if it is the case of a panoramic image, it is unwrapped, cropped and boxed.16
If the array is placed upside-down, the (unwrapped) frame is vertically flipped
and, after a proper scaling,17 the signal route is split: one branch goes to appsink
module, making the current frame available for the virtual microphones pointing
canvas, the other branch, after some additional conversions, writes the stream to
a file. The recording branch is fully connected only when the user press the record
button on the application.
15
glib is one of the core libraries of all GTK applications, containing the definitions of the
base classes from which the GTK objects are derived. The interesting - or crazy, it depends by
the point of view - thing is that all this was written in C plain, implementing from scratch a
complete object oriented superstructure.
16
The boxing process is the addition of black bands above and below to fit the desired frame
size.
17
This scaling can be requested by the zoom function, as explained in section 5.2.6.
rtspsrc
sink
tee
src
src 02
src 01
sink
rtpjpgdepay
is connected when the camera is placed upside-down.
sink
rec queue
sink
src
src
sink
videoconvert
monitor queue
src
sink
videorate
sink
appsink
src
src
sink
VMSUnwrap
sink
src
sink
videocrop
videoconvert
src
sink
jpegenc
src
src
sink
videobox
Modules for panoramic frame unwrapping
sink
avimux
src
src
sink
videoflip
sink
filesink
src
Flips upside-down
sink
videoscale
src
62
The software
Figure 5.5: The GStreamer pipeline implemented in the Recording activity of 3D VMS.
Green modules are connected only when a panoramic camera is used, and pink module
5.1 3D Virtual Microphones System
5.1.5
63
The timecode manager
When shooting broadcast events, it’s normal to have a lot of audio and video
signal coming from a certain number of cameras and microphones; so, in postproduction the amount of tracks to edit is very large and it is absolutely necessary
that all tracks are synchronized together. The ciak sound is the best known trick
to find sync between audio and video, but nowadays the LTC/SMPTE is the
standard signal to keep synced all the electronic equipment used in the set. No
exception for 3D VMS.
The system had to be able to:
• decode the LTC/SMPTE signal from an external source, maintain the reference during the recording session and store it in the output audio file;
• encode an LTC/SMPTE signal when playback an audio file with timecode
reference;
this became possible thanks to libltcsmpte 18 by Robin Gareus, now deprecated, but working perfectly as decoder/encoder for such kinds of signals.
Solving the decoding problem was not immediate because the EMIB interface,
that is normally connected to the laptop during a shooting session, has as the
only input the Eigenmike® and there aren’t other connectors: the only audio input available was the Line-in connector of the internal laptop audio card, but the
JACK server wasn’t able to manage more than one audio interface.
The solution found was to write the timecode codec as direct CoreAudio
thread, leaving JACK to manage the virtual microphones related audio connections, as shown in figure 5.1.
Since it is not possible to maintain physically connected the LTC/SMPTE
signal to whatever acquisition device during the entire recording session, the
3D VMS timecode subsystem, has to be autonomous after the synchronization
18
http://ltcsmpte.sourceforge.net/
64
The software
procedure and this was achieved by storing the timecode reference as offset between the laptop internal clock that has far than sufficient precision and stability.
So, the procedure to synchronize the 3D VMS with an external timecode
source is the following:
1. connect a 3.5 inches male jack terminated cable with LTC/SMPTE signal
on the mono/left channel to the Line-in connector of the laptop19 ;
2. press the Synchronize button on the Realtime or Recorder application’s GUI
and wait 5 seconds;
3. disconnect the cable. The system is now synchronized.
The LTC/SMPTE is then stored as audio channel on the output audio file
that is thus formed by 32+1 channels.20
In the Playback application, the timecode codec can be synchronized with
BWF file informations, but this is not a very used options because commonly
W64 with 32+1 channels are used, thus LTC/SMPTE signal is fully available
and directly routed to output port, together with virtual microphones signals.
5.1.6
The control interface
This part of the software has the task of managing the operations of the entire
application: it starts and stops processes and continuously checks if there are
errors.
On the top of the control interface there is a Graphical User Interface (GUI) that
19
With the latest MacBook Pro® machines, a new problem came out: the Line-in connector
disappeared, and all audio connections to the internal audio card pass through a single 3.5
inches combo jack connector for both input and output; the input is not a line-in but a mic and
is turned on after driver sensing, so a little harware interface has to be developed to connect
an external line signal.
20
An attempt was done to use the BWF (Broadcast Wave Format) audio file format to
store timecode data, but since the timecode management is not natively supported by audio
player/recorder Ecasound, a lot of modifications would be needed to add full support, especially
for playback activity.
5.1 3D Virtual Microphones System
65
is what the final user see and what he interacts with; the GUI sends commands to
the control interface, and receives information from it, giving full feedback to the
user; figure 5.1 describes well the interactions between 3D VMS components and,
as can be seen, the control interface resides in the main process and communicates
bidirectionally with everything else, to maintain the system in a known state and,
in case of errors, to inform instantly the user on what happened.
5.1.7
The actual structure
Currently the 3D VMS application is composed by three main activities:
• Playback
• Realtime
• Recording
that will be described in the next sections; they share some common modules,
modeled as objects, corresponding to the various resources and tasks used by the
application during its workflow. The resources include:
Configuration manager , used to store the global 3D VMS configuration: this
is a singleton always available to all instantiated objects;
Matrix manager , that made available to the 3D VMS the matrices installed
on the system; on application startup it always does a scan in the standard
directories looking for existing setups, and make a list that is available
through the Configuration manager;
IP-camera manager , that checks IP-camera connection and, if present, tries
to configure it with Configuration manager parameters.
The tasks, however, include:
JACK connections manager : a system that query the running JACK server
for existing connections, and, accordingly with other tasks setup or user
requests, make routing between running clients;
66
The software
BruteFIR manager , to start, stop and communicate with BruteFIR;
Ecasound manager , the same as the previous but with Ecasound;
Virtual Microphones manager : the system that collects user requests on
whatever concerns virtual microphones through its GUI and communicate
them to BruteFIR manager;
Timecode manager , fully described in section 5.1.5;
VU-meter display , window that shows real-time microphone capsules levels.
Once launched the application follows these steps:
• check for system shared memory configuration21
• check for connected audio cards and if these are suitable for recordings with
32 channels: if this is the case, the more 3D VMS compliant interface will
be selected then Recording and Realtime activities will be enabled;
• if a recording audio card was found, search for IP-camera and, if present,
configure it;
• do software parameter initialization, accordingly with previous checks, searching for existing matrices/setups;
• start JACK server for the selected audio interface;
• show main dialog, waiting for user actions.
Now the user can start one of the available activities, then, after the session
is terminated, it will return to this dialog.
21
This is requested by BruteFIR, that, being a multiprocess program, it needs a correctly
configured shared memory to work properly.
5.2 How to use the 3D VMS
5.2
67
How to use the 3D VMS
In order to give a more complete perspective of the software, the following paragraphs may be considered a sort of user manual, with detailed instruction about
software and hardware installation, and the description of all the current functionalities plus the more common operating procedures.
5.2.1
System requirements
3D VMS works on 64-bit OS X® platforms ≥ 10.7 with at least 4 GB of RAM
memory and a lot of space on disk for recordings, because they are really huge.22
A fast CPU is recommended23 to get a smooth processing of virtual microphones,
especially when video processing is also requested.
As previously said, the software needs JACK, that can be downloaded from its
website, but the ready-made JackOSX binary is usually the best choice for the
Mac® environments. Since 3D VMS is JACK-version agnostic there isn’t a version
recommended: choose the latest stable, but the compatibility with the installed
OS X® release has always to be verified.
5.2.2
Installation
The 3D VMS software package is composed by an application bundle named
3D vms.app24 whose icon is shown in picture 5.6 plus a bunch of matrix files that
may be organized in subdirectories.
The application bundle can be launched from everywhere, but the best place
to copy it is, as always, the Application directory; for all the 3D VMS related
resources, a directory named vms must be created in the current user home directory, then the matrix files must be copied in the subdirectory $HOME/vms/matrices25 ;
now, in order that 3D VMS can find correctly the matrix files, it may contain:
22
10 GB for a 30 minutes recording is a normal size.
Intel® Core™ i7 with four cores at the time of writing, late 2014.
24
In a normal OS X® installation, the extension .app should be invisible.
25
In all operating systems, of the Unix family, $HOME is an environment variable that points
23
to the current user home directory; for example, if there is a user named joe on a OS X® system,
68
The software
Figure 5.6: The 3D VMS application icon.
• single files with extension .dat, the very first version of matrix files, in which
are stored 684 IRs26 ;
• file pairs .dat/.dir, the first holds the impulse responses, the latter the azimuth and elevation angles in which they are measured;
• subdirectories containing a pair .dat/.dir plus a file named metadata.xml
that holds the setup parameters.
The latter is the recommended method to store 3D VMS setups,27 because
from the parameter stored in the metadata file, a complete and consistent state
for the system can be initialized.
Last thing to do to complete the installation is the configuration of the shared
memory. This is due to the multiprocess nature of BruteFIR, since processes in a
UNIX® system normally have separated memory addressing spaces and can share
datas using Inter-Process Communication (pipes, sockets, etc.); for a realtime
application this is a largely inefficient method, thus a shared memory addressing
space has to be configured to allow fast data sharing between BruteFIR filtering
processes. The shared memory configuration consist of parameters that has to
be passed to the operating system kernel and stored in the file /etc/sysctl.conf:
it is a simply text28 file and if it doesn’t exist, it must be created. Here the lines
to add:
# brutefir settings: 1GB of shared memory
then $HOME will point to /Users/joe.
26
This is exactly the total number of steps of the array characterization described in section
3.1.2.
27
The file formats are fully described in appendix A.
28
A plain text editor must be used to create or modify this file: don’t use a word processor!
5.2 How to use the 3D VMS
69
kern.sysv.shmall=262144 #shmmax / 4096
kern.sysv.shmseg=8192
#shm segments per user
kern.sysv.shmmni=8192
#shm id availables
kern.sysv.shmmin=1
#min size of shm segment
kern.sysv.shmmax=1073741824 #max size of shm segment
the symbol # precedes comments. After editing, the kernel must be reinitialized, therefore a computer restart is needed.
In a future, all these operations will be executed by installer scripts.
5.2.3
Things to do before launching the application
As described in section 5.1.7, once launched, the program checks for connected
hardware and never do it when running, thus every device must be connected
before (see pictures 5.7 or 5.8), in particular: the audio card and, if it is intended
to do a real-time or recording session, the IP-camera.
Keep in mind that if there aren’t connected audio cards with at least 32 inputs,
Realtime and Recording activities will never be available.
For what concerns the IP-camera, currently only the two models described in
section 4.1 are supported: in both cases they have to be configured with the fixed
IP address 192.168.0.202,29 without changing the default access credentials, all
the other parameters can be set directly by 3D VMS.30 At this point31 a new
connection has to be set up in the OS X® network preferences:
1. open Network from System Preferences window;
2. from Locations menu, choose Edit locations...;
3. press the small ’+’ button and give a name to the new location (ex. IPcamera);
29
30
This can be done with the camera bundled software.
There is an exception: the Sony® IPELA® camera frame ratio setting must to be 4:3, and
this actually can be done only using its web interface. It has been experienced that if this
parameter is accidentally set to 16:9, 3D VMS cannot detect correctly the device.
31
It is assumed that the IP-camera is already connected to the laptop with the network cable
and through the PoE power supply.
70
The software
FireWire cable
CAT-5 network cable
EMIB
CAT-5 network cable
signal only
PoE
power
injector
CAT-5 network cable
signal + power
Figure 5.7: Hardware connections layout: Eigenmike® + external IP-camera.
4. from left panel, chose Ethernet device and in the right panel select Manually
in the Configure IPv4 menu, give a consistent fixed IP address, for example
192.168.0.200, and the subnet mask 255.255.255.0
5. press Apply button at the lower right;
since now, every time is needed the connection with the IP-camera, simply
select the corresponding voice from Locations menu.
If the 3D VMS doesn’t detect the camera, check the connection by issuing the
ping command from a terminal:
ping 192.168.0.202
if the answer is no route to host, the connection has to be checked, otherwise,
if the answer is positive, update camera firmware to the latest version.32 If still
3D VMS doesn’t recognize the camera, last attempt is to set manually - with
cameras’ web interface - the parameters as follows:
Sony® IPELA® : Single stream JPEG, frame rate 4:3, resolution 1024x768;
Wision: Single stream MJPEG, resolution 5 MPixel.
32
This is especially recommended for Sony® IPELA® camera.
5.2 How to use the 3D VMS
FireWire cable
71
CAT-5 network cable
EMIB
CAT-5 network cable
signal only
PoE
power
injector
CAT-5 network cable
signal + power
Figure 5.8: Hardware connections layout: IED Cylindrical Array (with embedded
IP-camera).
5.2.4
The main dialog
When double-clicking on the application icon, after some waiting for matrix setups scanning and hardware checking, the main dialog (figure 5.3) should appear.
In the right panel, below the logo picture, a status log can be found: here is
reported if the EMIB, preferred audio interface, was found and the same for the
IP camera; the controls are all placed on the left of which follows description,
from top to bottom:
Activity buttons start the corresponding activity, Playback, Realtime, or Recorder;
to be noted that last two can be deactivated if proper audio interface is not
present;
Array selection menu from which the user can select the connected array in
case of Realtime or Recorder activity or the array used for the recording that
user wants to Playback 33 ; the entries of this menu are the various matrix
files/setups the system find in the default matrix directories;
Interfaces selection the system can use different interfaces for playing and
recording activities: in these menus the more suitable interfaces should
be listed after initial scan;
33
At the time of writing, the system cannot detect which kind of array was used for a certain
recording, thus if the wrong setup is chosen, then an incorrect filtering will have place during
Playback activity and virtual microphones will not work as expected.
72
The software
Setup playback ports this button opens a dialog that can be used to setup the
default JACK routing for audio software components. Actually there is some
work in progress on it and it will appear as shown in figure 5.9, allowing
the user to set the routing for both playing and recording activities;
Timecode by checking the Use timecode flag, the timecode subsystem will be
activated and the timecode display (figure 5.17) will be shown in Realtime
and Recorder activities. Here it is possible to select the FPS used by the
external synchronization system that will be connected to 3D VMS;
Opening angle using this radiobutton it is possible to force the acoustic horizontal range to a different, predefined value: usually this value is detected
from the selected matrix/setup and as consequence of this selection some
values can be disabled34 ;
Unwrap video stream if checked, force the video frame unwrap. This option
is checked and unmodifiable if a cylindrical array setup is selected;
Upside-down placement when checked, horizontal and vertical angles of virtual microphones will be inverted: as the name say, this option must used
when the array is placed upside-down.
5.2.5
The Playback activity
The Playback activity is intended for post-processing of recorded files; as shown
in figure 5.10 it’s divided in two windows, one is the control panel, the other is
the multimedia canvas.
Basically the user has to:
1. load the audio file and, optionally (but recommended, a static or dynamic
(movie) background
2. press play;
34
For example: a planar array cannot resolve a full 360° azimuth range.
5.2 How to use the 3D VMS
73
Figure 5.9: The 3D VMS default JACK routing dialog.
3. enjoy virtual microphones
Files can be loaded in some way:
• using the menus File/Load Audio... and/or File/Load Media...;
• using the menus Load Audio and/or Load Media buttons placed on the right
of the control panel;
• dragging file icons on the media canvas or on the transport display;
in every case if audio and media file have the same name - and obviously different extension! - loading the media file, movie or photo, will load automatically
the audio file.
Supported media files for static images are ISO Portable Network Graphics (.png)
and ISO JPEG format (.jpg,.jpeg); for movies supported containers are
• Apple® QuickTime® File Format (.mov,.qt).
• Microsoft® Advanced Streaming Format (.asf, .wmv),
74
The software
• Microsoft® Audio Video Interleave (.avi),
• ISO MPEG-PS (.mpg, .mpeg),
• ISO MPEG-TS (.ts, .tsv),
• ISO MPEG-4 (.mp4, .m4v)
with some of the most diffused codec (MPEG-2, M-JPEG and all codecs supported by the embedded libav 35 ).
Audio file supported are W64, the default output for 3D VMS recordings,
CAF, the Apple® CoreAudio format, and WAV but this is not very useful for its
limits on size (see paragraph 5.2.7.1).
Once the files are loaded successfully, the green display on the right will be
consequently updated with file names and play button will be enabled: now the
user can start the audio (and movie) reproduction pressing play button and other
transport controls will work as usual. Main transport commands are also available on menu Transport.
It is important to point out that the movies playback is implemented with
GStreamer engine and since there aren’t synchronization “hooks” to be used in
conjunction with Ecasound, audio and video playback are not strictly synchronized, and the play/pause/stop commands are simply issued to the two subsystem
one after the other. This is acceptable because the video part is only a reference
for virtual microphones pointing and the delay between audio and video - if
present - is really negligible for this kind of application.
During the playback, the multimedia canvas shows the loaded image or movie
of the recording environment and overlaps the virtual microphones, represented
by circles, to it: the user can drag the virtual microphones over the image, and
can modify the microphone order using the mouse wheel or the corresponding
gesture on the touchpad. It is also possible to modify with more accuracy the
35
http://www.libav.org
5.2 How to use the 3D VMS
75
position and order of virtual microphones acting on text controls of the control
panel: write the desired angle, then press Enter on laptop keyboard to issue the
command.
The gains of virtual microphones can be modified using the mixer controls on the
control panel.
When there are all virtual microphones on the canvas and it get quite confusing,
it is possible to activate the Select by buttons function by checking the square
on control panel and use the radio-button at the side to select the corresponding
virtual microphone: now every click on the canvas will correspond to a position
change.
It is possible to save and load the virtual microphone spatial configuration
to a text file36 using the menus Preset/Save microphones preset and Preset/Load
microphones preset respectively. It if also possible to load a microphone preset
simply dragging its file icon on the media canvas.
At every time, the user can select the number of virtual microphones using
the control at the left of the control panel.
5.2.5.1
JACK routing
Even without the help of the handy JackPilot application from JackOSX in every activity of 3D VMS it is always possible to modify the routing using the
JACK connections dialog, available from menu JACK and shown in picture 5.11.
Like the most known JACK routing applications, from this dialog it is possible to
check current connections and to connect or disconnect ports by selecting both
ends and pressing the button corresponding to desired action.
Every activity also is provided with a shortcut, accessible from menu Headphones, useful to quickly connect headphone ports to an array capsule or to a
virtual microphone.
36
Description in appendix A.2
76
The software
Figure 5.10: The 3D VMS Playback activity.
5.2.6
The Realtime activity
The main windows layout of the Realtime activity is similar to that of Playback:
a control panel, plus a multimedia canvas; but here the input doesn’t come from
a file: audio and video - if a camera is present - are real-time captured by the
connected hardware devices and processed by the filtering engine, so it is given
to the user the feeling of moving a real microphone in the pointing space having
immediately the feedback of what happens. When shooting a show, for example,
it is possible to follow an actor moving on the stage, like a spot-light, without
losing the acoustic focus on his voice and without the use of radio microphones
or similar.37
37
For this kind of application the latency time is absolutely critical and it has to be the
smallest possible.
5.2 How to use the 3D VMS
77
Figure 5.11: The JACK connections dialog. Here are exploded the 3D VMS filtering engine input ports (on the playback side) and the outputs, also known as virtual
microphones, on the capture side.
The control panel gives to the user roughly the same functions of the Playback
activity with the substitution of the play/pause buttons with a record button; the
user can also decide whether to record the 32 array channels row, or the 3D VMS
filtered signals, the so-called virtual microphones.
Being an acquisition activity, a slider to control the gain of Eigenmike® “family” of array is provided; to be noted that the gain updating of these devices is
a slow process because it is implemented with a MIDI protocol, thus a delay can
be noticed from slider movement to effective gain change.
If a camera is connected, pressing Record will record not only the audio, but
78
The software
Figure 5.12: Using a spherical array with a non panoramic camera system there is
a big difference between acoustic and optical angles: the visual reference is very small
and is difficult to point virtual microphones on the speakers.
the video stream too; if not differently specified using the change folder button,
all files will be stored in the default directory $HOME/EigenmikeRecordings, soon
deprecated in favor of $HOME/vms/recordings.
A zooming function for the multimedia canvas had also been recently implemented: this was needed because often it happens that the acoustical angular
ranges of the microphone array are different from the optical ones of the connected
camera; to leave to the user the full acoustical ranges for virtual microphone operations and since these are often much larger than optical aperture, the movie
frame appears small to the user and surrounded by a black area (figure 5.12).
With the zoom feature, the user can enlarge the movie frame, reducing the visible acoustic range, but making easier to point virtual microphones, as can be
seen in figure 5.13 or 5.14.
As accessory window, to check the array capsules level a Level Meter frame
is also present (figure 5.15): it is a JACK client connected directly to the capture
audio interface ports.
If chosen from the main application dialog, it is possible to do a timecodesynchronized recording session: a small dialog window that shows (figure 5.17)
5.2 How to use the 3D VMS
79
Figure 5.13: Increasing the zoom level make easier to point virtual microphone, but
the acoustic space available for visual pointing has been reduced, as can be seen from
the angular bounds reported on the sides of the canvas.
Figure 5.14: Zooming further, the virtual microphones circles don’t overlap and the
canvas appears less confusing, but the acoustic space is far smaller than the full angular
range of a spherical microphone.
timecode status will appear when starting the activity, and the Sync button of
the control panel will be enabled. The timecode synchronization has to be done
every time the user opens a Realtime activity session from the main application
80
The software
Figure 5.15: The 3D VMS meter monitoring array capsules output levels.
Figure 5.16: A 3D VMS Realtime session with VU-meter and timecode display.
dialog following the procedure described at the paragraph 5.1.5.
5.2.7
The Recorder activity
In professional application, when is preferred to do the virtual microphones steering in post-processing, a simple ad robust application that can record the raw 32
channels of the microphone array with no filtering at all, can be a better choice
than the complex Realtime activity; the Recorder activity was written with this
kind of task in mind, a rec-button only application that records 32 tracks (figure
5.19).
By users choice, the application can also record the video incoming from the ar-
5.2 How to use the 3D VMS
81
Figure 5.17: The timecode display.
Figure 5.18: A 3D VMS Recorder session with camera monitor, VU-meter and timecode display.
ray camera; in this case an additional window is shown that works as camera
monitor.
Like the Realtime activity, it is possible to do a timecode-synchronized recording
with the same modalities.
5.2.7.1
Output file formats
The default output file format for audio recording is the Sony® Sound Forge™
Wave 64, with extension W64: it is an extension of the Microsoft® WAV format
that overcomes its intrinsic 4 GB file length limit, having a full 64 bit header.
82
The software
Figure 5.19: The 3D VMS Recorder activity.
The system can also output WAV files, but for a non-timecoded recording, the
length will be limited to about 10 minutes.
Another output options, without any kind of limitation, is the Apple® CoreAudio
Format, with extension CAF.
Chapter 6
Endings
It is since 2009 that RAI broadcasting company employs a 3D VMS system in
show production, and if the first times the task was, in fact, an experimental
application of a research project, now the system has gained in stability and reliability, becoming a sort of acoustic swiss knife that really can save some tens
of meters of cables and the placement of other real microphones in non movable
positions and with static polar shapes. Now, when, for example, the audio editing
team has available the 32 raw recorded signals from array capsules, they can place
the virtual microphones and decide the better polar pattern in post-production,
and if something doesn’t sound good, they simply have to modify virtual microphone position and order.
Or also it is possible to take full advantage of the realtime features for a live
broadcast production, using 3D VMS virtual microphone in place of 7 physical
microphones: the first RAI production that employed this technology was the
opera Lucia di Lammermoor broadcast live by Rai-Radio3 from Theater Regio
of Turin the 21th of March 2011.
To achieve these results, the software part requested an accurate and long
work, as described in chapter 5 because if it’s certainly true that BruteFIR, Ecasound, GStreamer are excellent and intrinsically robust programs, to manage the
concurrent execution of them and to maintain the entire system in a consistent
state in case of failure of one of the subprocesses - that are always independent
84
Endings
Figure 6.1: Example of post-processing with 7 fixed microphones (La Bohème - Theater Regio, Turin, 20 May 2010 ).
- it is absolutely not easy: still today some issues are not completely resolved,
mainly in cases of incompatible startup setup.1 Currently the system watches all
running processes and when an error status is detected, the activity is paused
or stopped and the user is warned, then a rearm is tried, but there are cases in
which the behavior of a process in error status is not completely predictable and
the safer way to restore a consistent state is to restart the main process.
In general terms, the program restart as an error recover procedure is a bad solution, plus 3D VMS can have a long initialization phase on startup: what has been
done to avoid these uncomfortable status is a careful software architecture design
aimed to minimize runtime error status and numerous controls on Graphical User
Interface options to prevent status with unpredictable behavior.
The feedback from RAI technicians are always very precious in this sense,2 showing that the road taken was right, thus now it is possible to consider 3D VMS
maybe still beta, but reliable, also in very chaotic situations like a mobile live
broadcast set.
1
For example: the user tries to start a cylindrical array setup when a camera with different
resolution than expected is connected to the laptop. Actually the system warns the user about
the incompatibility, and if one activity is launched, not always the video frame adaptation is
successful.
2
Especially when some time is given to work on bugs. . .
6.1 Future developments
6.1
85
Future developments
In the next future, there are a lot of feature requests for 3D VMS, here a partial
list:
• addition of other polar shapes for virtual microphones, for example the
figure-of-eight;
• addition of recording system related information in the output files as metadata;
• increase the number of available virtual microphones;
• VST version of the system;
• version for iOS® /Android™ based portable devices of the playback application.
Whilst the first and the second are already intrinsically available in the filtering engine and the audio file manager library respectively,3 the others require
a much more radical action: as described in section 5.1.2, the limit of 7 virtual
microphones is due to a limit of the software BruteFIR, therefore increase this
number implies the complete replacement of the filtering engine. The extension
of BruteFIR engine is not taken in account because its multiprocess based architecture is to be considered obsolete, too much memory consuming and slow to
setup4 compared to modern multithreaded applications.
An attempt was done in 2012 with the software Jconvolver5 by Fons Adriaensen,
a JACK based multithreaded library that features up to 64 inputs and 64 outputs,
quite easily embeddable in the 3D VMS architecture, but the needing of a more
portable solution and the needing, at the time, to strengthen the mainstream
system rather than writing new modules stopped the tests.
An apparently good alternative is represented by the filtering engine of XVolver6
3
For W64 file format the library used is libsndfile by Eric de Castro Lopo (http://www.mega-
nerd.com/libsndfile).
4
The time needed to start a new process is effectively longer than the amount of time needed
to start a new thread.
5
http://kokkinizita.linuxaudio.org
6
http://www.desknotes.it/siten/Software.html
86
Endings
software by Carlo Chiari, already written with portability in mind and for this
reason can be a valuable base for a VST version of the software. The cons is the
absence of a dynamic filter change feature that has to be implemented.
With portable devices, the computational power available and storage space are
limited, for some kinds of devices very limited: in this case the challenge is to
individuate a good trade-off between resources available on-device and online,
for example pre-filtered tracks, balancing streaming and CPU loads in order to
obtain a responsive feedback for the user.
Appendix A
File format descriptions
In the following sections a complete descriptions of file formats used by 3D VMS
software.
A.1
Matrix files
Since the very first versions of the software the measured impulse responses of
the array were stored in a dat binary format used by Matlab® to store its data.
First array1 measurements have a fixed, high, number of positions or pairs azimuth/elevation, 684, as described in section 3.1.2. Now this number is not
anymore fixed, but the dat has been paired (= having the same name) with a dir
in which are stored the measured positions.
So, actually if the system find a not-paired dat file, it assumes that the file contains all the 684 “standard” positions.
An additional extension was done in order to store not only the raw impulse responses, but other array related metadata, contained in an xml file that mus be
placed in the same directory where dat and dir files reside. This file must have
the name metadata.xml and it can contain a complete 3D VMS configuration.
1
Essentially the Eigenmike® , a work done by Eng. L.Chiesi.
88
Appendix A.
A.1.1
File format descriptions
The file metadata.xml
All 3D VMS related tags must be enclosed in a ThreeDvmsConfigurationMetadata
tag; inside the following tags are recognized:
Label (string) the configuration label: this is the name that will appear as main
dialog array selection menu entry;
MatrixFileName (string) dat file name without extension2 ; it must be placed
in the same directory where metadata.xml is;
Device (string) the array type; recognized strings are: Eigenmike, Planar Array
and Cylindrical Array. It is mainly used to guess the correct camera setup;
AngularWidthDegrees (float) the horizontal (azimuth) angular range in degrees;
AngularHeightDegrees (float) the vertical (elevation) angular range in degrees;
CanvasWidthPixel (int) the media canvas width in pixels;
CanvasHeightPixel (int) the media canvas height in pixels;
MaxOrder (float) the maximum virtual microphones order;
CameraStream (string) the full address of rtsp camera stream;
UseDirections (string: yes/no) use or not the directions file. If ’yes’ a file
vith the same name of that specified in the MatrixFileName field, but with
extension dir must be present in the same directory of metadata.xml;
UnwrapFrame (string: yes/no) unwrap or not the frames coming from camera.
If ’yes’ the unwrap parameters of GStreamer module can be specified as tag
attributes (referencing to picture 4.8:
• outWidth (int, 0 ÷ 65535) output image width
• outHeight (int, 0 ÷ 65535) output image height
2
Spaces are not allowed
A.2 Microphones presets
89
• t0 (float −6.28 ÷ 6.28) wrapped image center angle θ0 in radians
• rMin (float, 0 ÷ 1) minimum wrapped image normalized radius Rmin
• rMax (float, 0 ÷ 1) maximum wrapped image normalized radius Rmax
• a (float, −10 ÷ 10) 2nd order vertical axis unwrapping function coefficient
• b (float −10÷10) 1st order vertical axis unwrapping function coefficient
• c (float −10 ÷ 10) 0th order vertical axis unwrapping function coefficient
• xc (float, 0 ÷ 1) wrapped image center x normalized coordinate
• yc (float, 0 ÷ 1) wrapped image center y normalized coordinate
Example:
<ThreeDvmsConfigurationMetadata>
<Label>Cylindrical Array</Label>
<MatrixFileName>Cyl_R_70HZ_2048</MatrixFileName>
<Device>Cylindrical Array</Device>
<AngularWidthDegrees>360.0</AngularWidthDegrees>
<AngularHeightDegrees>100.0</AngularHeightDegrees>
<UseDirections>yes</UseDirections>
<UnwrapFrame outWidth="1280" outHeight="400" t0="-2.3">yes\
</UnwrapFrame>
</ThreeDvmsConfigurationMetadata>
A.2
Microphones presets
Microphones presets can have the extension txt or prs, but in any case they are
plain text files containing the configuration of all current virtual microphones as
tab-separated values table.
Every line is referred to a single virtual microphones and contains the following
parameters:
• azimuth angle in degrees
90
Appendix A.
File format descriptions
• elevation angle in degrees
• order
• gain in deciBels.
To be noted that the file will have as many lines as the number of virtual
microphones currently available in the application.
If the user loads a preset with a smaller number of lines than virtual microphones,
the parameters of microphones in excess will not be updated.
Appendix B
LTC/SMPTE
Generally speaking, SMPTE timecode is a set of cooperating standards to label
individual frames of video or film with a time code defined by the Society of
Motion Picture and Television Engineers in the SMPTE 12M specification.
Timecodes are added to film, video or audio material, and have also been
adapted to synchronize music. They provide a time reference for editing, synchronization and identification and, practically, is a form of media metadata. The
invention of timecode made modern videotape editing possible, and led eventually to the creation of non-linear editing system.1
32
36
40
USER MINS USER MINS
BITS 4 UNITS BITS 5 TENS
44
48
52
56
USER HOURS USER
BITS 6 UNITS BITS 7
60
BGFB
USER SECS USER SECS
BITS 2 UNITS BITS 3 TENS
28
TENS
24
HOURS
20
Reserved
16
BGFB
FRAME USER
UNITS BITS 1
12
BPMPC
8
CF bit
Col.Frm
4
FRAME
TENS
bit 0
64
USER
BITS 8
68
72
76
79
SYNC WORD
Figure B.1: An example of SMPTE frame.
SMPTE timecodes contain binary coded decimal hour:minute:second:frame
identification and 32 bits for use by users; an example of the signal is shown in
figure B.1 but, in this form, is not used for timecode distribution and storage.
Instead the Linear (or Longitudinal) Time Code (LTC) encoding is used, that,
transforming the binary SMPTE signal in an audio signal, can be stored without
distortion in the old, fully analog, VTRs.
1
See [22].
92
Appendix B.
LTC/SMPTE
For backward compatibility and for reliability reasons, the analog LTC/SMPTE
signal is still today the standard for timecode distribution in every broadcast set.
Bibliography
[1] S. Berge and N. Barrett,
High Angular Resolution Planewave Ex-
pansion , Proc. of the 2nd International Symposium on Ambisonics and
Spherical Acoustics , (2010).
[2] A. J. Berkhout, D. de Vries, and P. Vogel, Acoustic control by wave
field synthesis , J. Acoust. Soc. Am., 93(5) (1993), pp. 2764–2778.
[3] M. M. Boone, E. N. G. Verheijen, and P. F. van Tol, Spatial
Sound-Field Reproduction by Wave-Field Synthesis , J. Audio Eng. Soc., 43
(1995), pp. 1003–1012.
[4] J. J. Christensen and J. Hald, Beamforming , Bruel & Kjær Technical
Review, 1 (2004).
[5] J. Daniel, S. Moreau, and R. Nicol, Further Investigations of HighOrder Ambisonics and Wavefield Synthesis for Holophonic Sound Imaging ,
in 114th AES Convention , mar. 2003.
[6] W. L. Dooley and R. D. Streicher, M-S Stereo: a Powerful Technique
for Working in Stereo , in 69th AES Convention , may 1981.
[7] A. Farina, Simultaneous measurement of impulse response and distortion
with a swept-sine technique , in 110th AES Convention , Paris, feb. 2000.
[8] A. Farina, A. Amendola, L. Chiesi, A. Capra, and S. Campanini,
Spatial PCM Sampling: a new method for sound recording and playback , in
52th AES Conference, Guildford, sep. 2013.
[9] F. M. Fazi and P. A. Nelson, The ill-conditioning problem in Sound
Field Reconstruction , in 123th AES Conference , New York, oct. 2007.
94
BIBLIOGRAPHY
[10] M. Frigo and S. G. Johnson, FFTW: An Adaptive Software Architecture for the FFT , Proceedings of the 1998 ICASSP conference , 3 (1998),
pp. 1381–1384.
[11] B. Ginn, J. J. Christensen, J. Hald, J. M. rkholt, A. Schuhmacher, and C. Blaabjerg,
A review of array techniques for noise
source location , in ICSV 10 , Stockholm, jul. 2003.
[12] A. Greensted, Delay sum beamforming — the lab books pages,
2012. http://www.labbookpages.co.uk/audio/beamforming/delaySum.
html, [Online; accessed 13-January-2015].
[13] J. Hald, STSF - a unique technique for scan-based Near-field Acoustic
Holography without restrictions on coherence , in Bruel & Kjær Technical
Review, vol. 1, 1989.
[14] J. Hald, Combined NAH and Beamforming Using the Same Array , Bruel
& Kjær Technical Review, 1 (2005), pp. 15–43.
[15] J. Hald, Patch Near-field Acoustical Holography Using a New Statistically
Optimal Method , Bruel & Kjær Technical Review, 1 (2005), pp. 44–54.
[16] O. Kirkeby and P. A. Nelson,
Digital Filter Design for Inversion
Problems in Sound Reproduction , J. Audio Eng. Soc., 47 (1999).
[17] O. Kirkeby, P. Rubak, and A. Farina, Analysis of ill-conditioning of
multi-channel deconvolution problems , in 106th AES Convention , Munich,
may 1999.
[18] J. Maynard, E. Williams, and Y. Lee, Nearfield acoustic holography:
I. Theory of generalized holography and the development of NAH , J. Acoust.
Soc. Am., 78 (1985), pp. 1395–1413.
[19] S. Moreau, J. Daniel, and S. Bertet, 3D sound field recording with
High Order Ambisonics -objective measurements and validation of a 4th order
spherical microphone , in 120th AES Conference , Paris, may 2006.
[20] V. Pulkki, Spatial Sound Reproduction with Directional Audio Coding ,
J. Audio Eng. Soc., 55 (2007), pp. 503–516.
[21] A. Torger and A. Farina, Real-Time Partitioned Convolution for Ambiophonics Surround Sound , in 2001 IEEE Workshop on Applications of
Signal Processing to Audio and Acoustics , New York, oct. 2001.
[22] Wikipedia,
Smpte timecode — wikipedia,
the free encyclopedia,
2014. http://en.wikipedia.org/w/index.php?title=SMPTE_timecode&
oldid=639775488, [Online; accessed 13-January-2015].
[23] W. Woszczyk, B. Leonard, and D. Ko, Space Builder: An Impulse
Response-Based Tool for Immersive 22.2 Channel Ambiance Design , in 40th
AES Conference , Tokyo, oct. 2010.
96
BIBLIOGRAPHY
Questa la scrivo in italiano.
È normale: quando capita fra le mani una qualsiasi tesi, ed ora come non mai se
ne producono in quantità, si d‘a una scorsa all’indice, poi molto frettolosa al testo,
ma alla fine si capita sempre lì, alla pagina dei “ringraziamenti”. Dopo pagine
e pagine di contenuti dalle ambizioni scientifiche dalle quali l’autore poco lascia
trapelare di sé, si va in cerca di calore umano, di quei nomi, di quelle esperienze
che hanno accompagnato la vita di colui che nella prima pagina si presenta come
“candidato”. Magari pensando di vedersi citati e provando una piccola emozione
quando, scorrendo l’elenco di nomi, si trova il proprio.
La pagina dei ringraziamenti è, nella sua essenza, un luogo romantico, dove si
può addirittura trovare conforto e serenità nello scoprire di aver avuto un ruolo,
talora inaspettatamente importante, nella vita di qualcun altro.
Non mi sottrarrò al rito: dopotutto ringraziare è una delle prime regole di buona
educazione che ci viene insegnata ed è indiscutibilmente bello sentirsi dire “grazie”, non ha grande importanza il perché. Ciò che conta è dirne tanti di “grazie”,
poiché a nessuno è dovuto proprio nulla e tutto, in fondo, è un dono.
Dunque grazie, prima di tutto, ad Angelo Farina ed al suo geniale entusiasmo
al di fuori di qualsiasi controllo, poi a tutti coloro con cui ho collaborato in questi
anni di assegni Unipr e di cui non mancano i contributi nelle pagine precedenti:
Enrico, Christian, Alberto, Andrea (ora in RCF, ma il prototipo del software l’ha
fatto lui), Lorenzo C. (difficile elencare tutti i suoi contributi in questo progetto
di ricerca), Lorenzo E. e Luca N.
Grazie anche al CRIT-RAI nella persona di Leonardo Scopece, senza cui questa
ricerca non sarebbe cominciata e grazie al quale c’è tutta l’intenzione di proseguirla: delle telefonate ne avrei fatto volentieri a meno, ma se il software ha
raggiunto certi obiettivi una gran parte del merito va a chi lo utilizza e ci crede.
L’occasione è propizia per dire grazie agli amici, che non solo è d’uso, ma esprime ciò che realmente avviene nella vita delle persone: quello che si fa è sempre
frutto di un contesto e di esperienze negli ambiti più vari: io sono musicista e
dirigo un coro. Non fossi stato un musicista, probabilmente non sarei mai entrato
in contatto con il gruppo di acustica dell’Università. Dunque voglio ringraziare
tutti i miei coristi, perché amici. Grazie anche a loro il legame con la musica in
questi anni è rimasto forte e, nonostante le inevitabili difficoltà, le soddisfazioni
ed il divertimento non sono mai mancati.
Mi piace ringraziare Michele, da diversi anni negli USA, non ci sentiamo
praticamente mai, ma la sua “lezione” e soprattutto la sua passione sono presenti
in ogni istante. Indimenticabili.
Non posso certo dimenticare i miei genitori, co-dedicatari di questo testo,
senza i quali non sarebbe iniziato proprio nulla, ma l’ultimo, il più importante
“grazie” va a Giovanna, che più di ogni altro mi ha spinto nel conseguire questo
dottorato nonostante i miei piagnistei: è lei il primo ed il più energico motore di
tutto ciò che faccio ed è grazie a lei se la nostra vita è una continua “ricerca” di
musica.
Typesetted on 2015 january 16 with LATEX 2ε
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement

Languages