Detection of Alarm Sounds in Noisy Environments

2017 25th European Signal Processing Conference (EUSIPCO)
Detection of Alarm Sounds in Noisy Environments
Dean Carmel, Ariel Yeshurun, Yair Moshe
Signal and Image Processing Laboratory (SIPL)
Andrew and Erna Viterbi Faculty of Electrical Engineering, Technion – Israel Institute of Technology
Technion City, Haifa, Israel,
Abstract—Sirens and alarms play an important role in
everyday life since they warn people of hazardous situations, even
when these are out of sight. Automatic detection of this class of
sounds can help hearing impaired or distracted people, e.g., on the
road, and contribute to their independence and safety. In this
paper, we present a technique for the detection of alarm sounds in
noisy environments. The technique is not limited to particular
alarms and can detect most electronically generated alerting
sounds within 200 ms. We consider a set of acoustic features and
use the ReliefF algorithm to select only the ones that best
differentiate between alarms and other sounds. We use an SVM
classifier as the detector. On the tested dataset, consisting of
several dozen alarm sounds and several dozen background noises,
the proposed technique shows an accuracy of 98% per audio
frame. With a larger training dataset, this result is expected to
substantially improve.
Keywords- alarm detection; siren detection; acoustic event
detection; sound recognition; assistance for hearing impaired
The diversity of environmental sounds is vast and includes
the sounds generated in indoor and in outdoor environments.
These sounds convey information about human and social
activities. Examples include cheering of the audience in a sports
event, a gunshot in the street and hasty steps in a nursing home.
Such information is helpful in applications that analyze audio
and video content. Alerting sounds such as emergency vehicles,
smoke alarms and medical monitoring alarms are of a special
importance, as they are usually designed to warn people of
hazardous situations, even when these are out of sight.
Unfortunately, alerting sounds may be missed due to hearing
impairment or just simple distraction, leading to hazardous and
life-threatening situations. Automatic detection of such sounds
may have many applications, both for hearing aids, and for
intelligent systems that need to respond to their acoustic
environments. A strong need exists for hearing impaired people
driving a car. In such a situation, background noise is often very
loud, causing many people with a hearing loss to drive without
amplifications. Therefore, they are unable to hear the siren of
an approaching emergency vehicle. To support hearing
impaired people, different assistive devices are commercially
available, but no supporting devices for environmental sound
awareness are available for mobile use [1].
Humans can identify a particular sound as an alarm even
when they have never heard it before, and even with a
significant background noise. Since by their nature alarm
sounds are intended to be easily detected, we could expect
alarm sound detection to be a relatively simple task. However,
the distinctive characteristics of alarm sounds are not formally
ISBN 978-0-9928626-7-1 © EURASIP 2017
defined. Moreover, such sounds are varied, and it is not obvious
that they do indeed share common characteristics, rather than
being learned by listeners as the conjunction of a set of more
special purpose sound types [2]. The International Organization
for Standardization has defined a recommendation for “auditory
danger signals” (ISO 7731) [3]. However, this recommendation
only gives rough guidelines for alarm sounds and is not widely
used worldwide. Instead, most countries have their unique siren
standardization. Moreover, many alarm sounds are not
standardized at all, e.g., alarm clocks. In an attempt to define
common characteristics shared by alarm sounds, three types of
siren are defined in [4]:
 Pulsed alarms – Consist of a repetition of the same sound
with a silent gap between each instance.
 Sirens – Sounds in which the frequency continuously
 Alternating alarms – Consist of two different alternating
tones with no silence gap between them.
Spectrograms of examples of these three types of alarm sounds
are shown in Fig. 1.
In the last few decades, automatic processing of audio
signals drew great interest in both academia and industry.
However, most effort has been invested in speech and music
processing and significantly less in the processing of
environmental sounds, albeit hearing impairment is one of the
most widespread physical disabilities worldwide. The majority
of works dealing with environmental sounds are concentrated
on audio classification, e.g., explosions vs. door slams vs. dog
barks, or detection of abnormal audio events. The main research
problem these works deal with is choosing a set of suitable
audio features. Common features in use for these tasks are the
Mel-frequency cepstral coefficients (MFCC), wavelet-based
features and individual temporal and frequency features. As the
set of sound classes in these works is very limited, and the
features selected manually, these works are usually hard to
generalize. Thus, to create a reliable alarm sound detector, a
more specialized technique should be designed.
With this insight in mind, several previous works have tried
to deal specifically with the task of alarm sound detection. Since
building a general model of alarm sounds is difficult, most
works try to detect only particular alarms, usually sirens of
emergency vehicles of a specific country. Many of these works
do not perform well out of laboratory conditions since they do
not model well enough ambient background sounds and usually
do not consider shifts in frequency due to the Doppler effect.
For example, in the work presented in [5], the authors try to
detect a small set of pre-selected warning sounds in a simulated
environment by cross-correlation. In [6], an artificial neural
network was used to detect police vehicles in Italy. MFCC
2017 25th European Signal Processing Conference (EUSIPCO)
Fig. 1. Spectrograms of three types of alarm sounds. Red represents high values and blue represents low values. (a) Alarm clock: Pulsed alarm - Repetition of the
same sound with a silent gap between each instance. (b) Ambulance driving away: Siren - The frequency continuously changes. (c) Fire alarm: Alternating alarm - Two different alternating tones with no silence gap between them.
coefficients were extracted as features for classification.
Another work for detecting sirens of emergency vehicles in
Italy is presented in [7]. The system runs in real-time by
estimating the pitch frequency and comparing it to pre-defined
siren frequencies. Another simple system for siren detection
that runs in real-time is described in [8]. Based on the
periodicity of alarm sounds, the results of autocorrelation are
analyzed by a classifier tailored to acoustic emergency signals
in Germany. Sirens of German police cars are also being
detected in [9], where part-based models originally proposed in
computer vision are trained and used for classification. In [10],
a technique for detecting an ambulance siren sound in Taiwan
in presented. It uses frequency matching by finding the longest
common subsequence, a technique commonly applied for
pattern recognition.
One of the first works [2] to attempt to detect alarms in
general compared two different approaches – an artificial neural
network and a sinusoidal modeling. All tests were done at 0 dB
signal-to-noise ratio (SNR), and both approaches were stated to
perform poorly. A much simpler approach for automated
detection of alarm sounds is described in [11]. Four classes of
alarm sounds (bells/chimes, buzzers/beepers, horns/whistles,
and sirens) are detected in four noise backgrounds (cafeteria,
park, traffic, and music) based on the normalized
autocorrelation function and the overall sound level. The
authors report true positive rate of 80% with false positive rate
above 25%. The autocorrelation function is used for detection
of alarms also in [4]. For improved robustness, autocorrelation
is performed on the envelope of the signal after band-pass
filtering around the fundamental frequencies of the alarms to be
detected. The system was tested on alarms compliant to the ISO
7731 recommendation [3] with SNRs ranging from 0 to 15 dB.
An average true positive rate of 95% with an average latency of
1.27 sec are reported for pulsed alarms. For other types of
alarms, the reported true positive rate is about 50% and for
alarms not compliant to the ISO 7731 recommendation, even
poorer performance is reported. Recently, a smartphone
application for recognizing warning sounds was described in
[1]. The application can adapt to sounds defined by the user.
Warning sounds occurring in road traffic were chosen as a
demonstrator. For classification, the application uses 13 MFCC
coefficients and the zero crossing rate (ZCR) in a feed forward
ISBN 978-0-9928626-7-1 © EURASIP 2017
neural network. The authors report an average recall of 83%
with an average precision of 99%, but it is not clear what
background noises were used in the experiments.
In this paper, we present a technique for detecting alarm
sounds in general in noisy environments that is able to
generalize away from the specific examples used in
development to cover most real-world alarm sounds. The
technique is primarily intended for a future assistive device in
road traffic for hearing impaired persons, but can also be used
in other application domains, and was tested accordingly. We
use a machine learning approach where an SVM classifier is
trained beforehand on features extracted from labeled sound
samples and is used for classification. As the classification
results heavily depend on the representation power of the audio
features, we consider a large set of audio features, both in the
time and frequency domain, and use only the ones that best
differentiate between alarms and no-alarm sounds. Feature
selection is performed using the ReliefF algorithm [12].
We perform alarm sound detection using a supervised
machine learning approach, as shown in Fig. 2. In the offline
training phase, audio frames labeled as alarm or no-alarm are
used to train an SVM classifier. This trained classifier is used
in the online classification phase to classify new unlabeled
audio frames as alarms or no-alarms. In both the training and
classification phases, audio samples are first band-pass filtered
and their time envelope is extracted. Then, features
characterizing the signal are extracted and used for training or
classification. To make classification results more robust, a
median filter of size 3 is applied to the binary SVM
classification results. We use audio frames of size 100 ms with
a 50% overlap. These short time frames contain enough
information for accurate alarm sound detection and they allow
us to do so with a short delay of 200 ms. Moreover, using short
time frames enables detection of short or intermittent alarm
Following [4], to remove noise and irrelevant information,
pre-processing is performed before feature extraction. First, the
signal is band-pass filtered. As alarm sounds tend to be periodic
with fundamental frequencies usually in the range of
2017 25th European Signal Processing Conference (EUSIPCO)
Offline Training
Online Classification
Fig. 2. Proposed alarm sound detection technique. In the offline training phase, audio frames labeled as alarm or no-alarm are used to train an SVM classifier. This
trained classifier is used in the online classification phase to classify new unlabeled audio frames as alarms or no-alarms.
500-1500 Hz, a band-pass filter in this frequency range is used.
Then, we extract the time envelope of the filtered signal using
the magnitude of the analytic signal computed using the Hilbert
transform [13]. The extracted envelope is proportional to the
amplitude of the signal and captures the slowly varying features
of the signal. Therefore, it allows detection of an alarm’s energy
variation in time.
Careful selection of features is a key issue when designing
an effective audio signal detector. For each time frame, we seek
features that comprise the essence of information relevant to
classification, are robust to variations in the signal and to noise,
and are efficient in terms of computational and space
complexity. To find such features, we have considered a large
set of features that were previously used for different tasks of
audio processing, especially the features described in [14]. We
will now describe these features briefly and then describe a
technique we have used to leave only a subset of these features
that contribute to the accuracy of alarm sound detection. We
divide the features into three categories:
1) Time domain features
 Pitch - The fundamental frequency computed using the YIN
algorithm [15].
 Short time energy – Alarm sounds tend to have high energy
in the given spectral band compared to the energy attributed
to noise.
 Zero crossing rate (ZCR) - The number of times the sign of
a time series changes within a frame. It roughly indicates the
frequency that dominates during that frame.
For the short-time energy and the zero crossing rate, we divide
each frame into sub-frames of size 20 ms, with a 50% overlap.
For each feature, we compute 18 statistics of the sub-frames:
 maximum (max)
 minimum (min)
 mean
 median
 standard deviation (std)
 max/mean
 max/median
 std2/mean2
 2nd smallest value
 3rd smallest value
 5th smallest value
smallest value
 mean of 4 smallest values
2) Frequency domain features
 MFCC – We use 13 Mel-frequency cepstral coefficients to
describe the spectral shape of the signal. These coefficients
are extracted by applying the discrete cosine transform
(DCT) to the log-energy outputs of the nonlinear mel-scale
 Spectral flux, spectral roll-off, spectral centroid, and
spectral flatness – Spectral flux is a measure of how quickly
the power spectrum of a signal is changing, defined as the
Euclidean distance between the spectra of two adjacent
frames. The spectral flux can be used to determine the
timbre of an audio signal, or for onset detection. Spectral
roll-off is defined as the Nth percentile of the power spectral
distribution, where N is usually 85% or 95%. The roll-off
point is the frequency below which the N% of the
magnitude distribution is concentrated. This measure is
useful in distinguishing voiced speech from unvoiced.
Spectral centroid is a measure indicating where the "center
of mass" of the spectrum is. Perceptually, it has a robust
connection with the impression of "brightness" of a sound.
Spectral flatness measures the variance of the frequency
spectrum of the signal. High spectral flatness indicates
similar amount of energy at all frequencies, e.g., white
noise. Low spectral flatness indicates that the signal energy
is concentrated mainly around a few frequency bands, e.g.,
alarm sound. As with the short-time energy and the zero
crossing rate, these features are computed by dividing each
frame into sub-frames and calculating 18 statistics of these
sub-frame features. This results in a feature vector of size
3) Wavelet-based features
The wavelet coefficients capture time and frequency localized
information about the audio waveform. The wavelet
transform can be viewed as a multi-level process of filtering
the signal using a low-pass (scaling) filter and a high-pass
(wavelet) filter. The first layer of the transform decomposition
of a signal splits it into two bands giving low-pass
approximation coefficients and high-pass detail coefficients.
 Discrete Wavelet transform (DWT) – In the DWT, each
level is calculated by passing only the previous wavelet
approximation coefficients through discrete-time low and
high pass filters. Thus, DWT results in a binary tree like
structure which is left recursive (where the left child
represents the lower frequency band). We decompose each
frame using the DWT with 10 decomposition levels. For the
 mean of values > 10-6
This results in a vector of size 1+2*18=37.
ISBN 978-0-9928626-7-1 © EURASIP 2017
2017 25th European Signal Processing Conference (EUSIPCO)
Fig. 3. Feature ranks and weights computed by the ReliefF algorithm. All feature types contribute to classification accuracy. The highest weights are assigned to
features based on the spectral flux (leftmost green bars), features based on the short time energy and the pitch. Wavelet-based features also have a significant
contribution to classification accuracy. Weight ≤ 0 indicates a feature that does not contribute to a higher classification accuracy.
resulting coefficients, we compute 4 features - energy,
variance, standard deviation, and waveform length [16].
This results in a vector of size 15*4=60.
 Wavelet packet transform (WPT) - In the WPT, both the
detail and approximation coefficients are decomposed to
form a full binary tree. The WPT results in equal-width subband filtering of signals as opposed to the coarser octave
band filtering found in the DWT. This makes wavelet
packets an attractive alternative to the DWT in certain
applications such as audio classification. We decompose
each frame using the WPT with 4 decomposition levels.
This results in a vector of size 31.
Feature Selection
The set of features we have described for representing the
audio data consists of many features but only some of them may
be useful for alarm sound detection. Feature selection is the
process of selecting a subset of the features by removing
redundant and irrelevant ones. This process reduces the
dimensionality of the dataset, which in turn reduces the
computational and storage complexity of the classification. We
select a subset of the features by using the ReliefF algorithm
[12]. ReliefF receives as input a dataset with  instances of 
features, belonging to two known classes. The key idea of
ReliefF is to estimate features importance according to how
well their values distinguish among neighboring instances. An
instance  is denoted by a -dimensional vector (1 , 2 , … ,  )
where  denotes the value of feature  of . Given  , ReliefF
searches for its nearest two neighbors, one from the same class
(called the nearest hit) and the other from a different class
(called the nearest miss). The weight vector  of the instance
 is then updated by:
 =  − | −  | + | −  |
Thus the weight of any feature decreases if it differs from that
feature in nearby instances of the same class more than nearby
instances of the other class, and increases in the reverse case.
After  iterations, the algorithm averages the contribution of
the nearest hits and misses to the weight vector to obtain the
relevance vector. Features are selected if their relevance is
greater than a threshold value. In practice, to increase the
algorithm robustness to noise, ReliefF does not search only for
nearest two neighbors but for  nearest hits and misses and
ISBN 978-0-9928626-7-1 © EURASIP 2017
averages their contribution to the weights of each feature. We
use =35. The feature ranks and weights computed by ReliefF
on our training dataset are depicted in Fig. 3. We can see in the
figure that all feature types contribute to classification accuracy.
The highest weights are assigned to features based on the
spectral flux (leftmost green bars), to features based on the short
time energy, and to the pitch. Wavelet-based features also have
a significant contribution to classification accuracy. We have
decided to put a threshold of 0 on the weight, removing about
12% of the features that do not contribute to a higher
classification accuracy.
To simulate real-life conditions, we assembled a dataset of
diverse alarm sounds and ambient noises. The dataset contains
70 audio signals 30 seconds long, 35 alarm sounds and 35
noises. Sounds were collected by searching the web and by
making recordings around the home and office. Of the 35 alarm
sounds, 20 are clean and 15 are real-life recordings with
significant ambient noise. Of the 35 noises, some are typical
everyday noises and some were carefully selected according to
their similarity to alarm sounds, e.g., busy roads noises, a truck
motor, and a helicopter.
To evaluate the performance of the proposed technique, we
used -fold cross-validation with  = 10. We obtained 98%
accuracy per audio frame (for both alarms and noises). In a reallife application that constantly monitors the environment, the
number of no-alarm audio frames is orders of magnitude higher
than the number of alarm audio frames. To make such an
application feasible, the false positive rate (percentage of noalarm frames miss classified as alarm frames) should be
negligible. We have found the false positive rate of the
proposed technique to be 0.4%. This result is encouraging
compared with previous works but is not small enough to make
an audio monitoring application feasible. Further analysis
revealed that false positives resulted from two noises in the
dataset – a loud sound of birds chirping and the sound of a
waterfall. The miss classification of frames of those two noises
is due to the fact that those noises differ greatly from other
sounds in the dataset. Therefore, in the -fold cross-validation
procedure, when one of those sounds is used for classification,
no similar sounds are available for training and it is difficult for
2017 25th European Signal Processing Conference (EUSIPCO)
Fig. 4. (a) A spectrogram of a police car operating its siren in intermittent bursts
and (b) alarm/no-alarm classification results of its audio frames. Classification
results are accurate (accuracy of 100%) although the siren operates in short
bursts and is interleaved with noise.
the classifier to generalize to this sound. This problem can be
solved by an expanding the dataset with more diverse ambient
noises. Note that the false positive rate can be further reduced
by allowing a longer delay and incorporating the results from
more consecutive frames, e.g., using a median filter of a larger
We have also performed several tests with real-life
unlabeled data to verify the robustness of our alarm detection
technique to noise, low amplitude, Doppler effect (e.g., moving
emergency vehicles) and short alarms with intermittent bursts.
For all those tests, satisfactory results were obtained. For
example, Fig. 4 demonstrates the ability of the proposed
technique to detect short alarms. It shows a spectrogram of a
police car operating its siren intermittently and alarm/no-alarm
classification results of its audio frames. Although the siren
operates in short bursts and is interleaved with noise, the
proposed alarm detection technique is able to detect the siren
In this paper, we present a detection technique for alarm
sounds in noisy environments. The technique is not restricted to
the recognition of a specific set of pre-defined alarm sounds and
can generalize to most electronically generated alerting sounds.
Acoustic features, which were verified to contribute to this task,
are used with an SVM classifier. Alarm sounds can be detected
with a short delay of 200 ms. On a dataset of several dozen
audio samples, we achieve an accuracy of 98% per 100 ms
audio frame. The false positive rate of the proposed technique
is 0.4% and can be further improved by future expansion of the
dataset. The proposed technique is robust to noise and to other
real-life conditions such as low alarm amplitude, Doppler effect
and short alarms with intermittent bursts.
The authors would like to thank Prof. David Malah, head of
SIPL, for his advice and helpful comments.
ISBN 978-0-9928626-7-1 © EURASIP 2017
[1] M. Mielke and R. Brueck, "Design and evaluation of a
smartphone application for non-speech sound awareness for
people with hearing loss," in Engineering in Medicine and
Biology Society (EMBC), 2015 37th Annual International
Conference of the IEEE, 2015, pp. 5008-5011.
[2] D. Ellis, "Detecting alarm sounds," in Proc. Workshop on
Consistent and Reliable Acoustic Cues CRAC-2000, 2001.
[3] "ISO 7731: Ergonomics - Danger signals for public and work
areas - Auditory danger signals," International Organization for
Standardization, 2013.
[4] M.-A. Carbonneau, N. Lezzoum, J. Voix, and G. Gagnon,
"Detection of alarms and warning signals on an digital in-ear
device," International Journal of Industrial Ergonomics, vol. 43,
pp. 503-511, 2013.
[5] E. R. Bernstein, A. J. Brammer, and G. Yu, "Augmented warning
sound detection for hearing protectors," The Journal of the
Acoustical Society of America, vol. 135, pp. EL29-EL34, 2014.
[6] F. Beritelli, S. Casale, A. Russo, and S. Serrano, "An automatic
emergency signal recognition system for the hearing impaired," in
Digital Signal Processing Workshop, 12th-Signal Processing
Education Workshop, 4th, 2006, pp. 179-182.
[7] F. Meucci, L. Pierucci, E. Del Re, L. Lastrucci, and P. Desii, "A
real-time siren detector to improve safety of guide in traffic
environment," in Signal Processing Conference, 2008 16th
European, 2008, pp. 1-5.
[8] M. Mielke, A. Schäfer, and R. Brück, "Integrated circuit for
detection of acoustic emergency signals in road traffic," in Mixed
Design of Integrated Circuits and Systems (MIXDES), 2010
Proceedings of the 17th International Conference, 2010, pp. 562565.
[9] J. Schröder, S. Goetze, V. Grutzmacher, and J. Anemüller,
"Automatic acoustic siren detection in traffic noise by part-based
models," in ICASSP, 2013, pp. 493-497.
[10] J.-J. Liaw, W.-S. Wang, H.-C. Chu, M.-S. Huang, and C.-P. Lu,
"Recognition of the ambulance siren sound in Taiwan by the
Longest Common Subsequence," in Systems, Man, and
Cybernetics (SMC), 2013 IEEE International Conference on,
2013, pp. 3825-3828.
[11] R. A. Lutfi and I. Heo, "Automated detection of alarm sounds,"
The Journal of the Acoustical Society of America, vol. 132, pp.
EL125-EL128, 2012.
[12] I. Kononenko, E. Šimec, and M. Robnik-Šikonja, "Overcoming
the myopia of inductive learning algorithms with RELIEFF,"
Applied Intelligence, vol. 7, pp. 39-55, 1997.
[13] B. Boashash, "Estimating and interpreting the instantaneous
frequency of a signal - Part I: Fundamentals," Proceedings of the
IEEE, vol. 80, pp. 520-538, 1992.
[14] A. Rabaoui, M. Davy, S. Rossignol, and N. Ellouze, "Using oneclass SVMs and wavelets for audio surveillance," IEEE
Transactions on information forensics and security, vol. 3, pp.
763-775, 2008.
[15] A. De Cheveigné and H. Kawahara, "YIN, a fundamental
frequency estimator for speech and music," The Journal of the
Acoustical Society of America, vol. 111, pp. 1917-1930, 2002.
[16] R. Khushaba, "Feature Extraction Using Multisignal Wavelet
Transform Decomposition,
matlabcentral/fileexchange/37950-feature-extraction-usingmultisignal-wavelet-transform-decomposition," 2012.
Download PDF
Similar pages