Anomalous Event Detection in Traffic Audio
Minakshee Shukla, Renu M Rameshan and Shikha Gupta
Vehant Research Lab, Vehant Technologies Pvt. Ltd., Noida, India
Keywords:
Sound Event Detection, Monophonic, Teacher-Student Strategy, ATST, CRNN.
Abstract:
This work focuses on detecting anomalous sound events from traffic audio. The audio used is recorded from
the microphone associated with a surveillance camera. We have defined six anomaly classes and generated
synthetic data using real background audio which corresponds to Indian traffic sound obtained from a surveil-
lance camera microphone. Using a teacher-student training strategy, we have obtained F1 score of 96.12%
and an error rate 0.06. We also show that even when the event occurs farther away from the microphone, the
performance is still impressive, with an F1 score of 92.55 and an error rate of 0.12.
1 INTRODUCTION
Human auditory system has evolved over millions
of years to have the ability to identify the different
sounds that exist in the environment even when they
are all mixed together. It understands when and from
which direction the sound comes under most circum-
stances. But this is not true for a man made system.
Systems that can understand the occurrence of a par-
ticular sound event in the presence of other signals are
still scarce. In sound event detection one needs to de-
tect the occurrence of a particular sound: when it oc-
curred, when it ended, as well as what it is. This work
focuses on detecting sound events which are anoma-
lous with respect to a traffic scene, where there is a
constant background of traffic noise which makes the
task challenging.
Unlike speech and music, where there is a statis-
tical structure, an audio event, more so an anoma-
lous event, doesn’t have any such structure. This
makes the conventional approaches based on Gaus-
sian Mixture Models (GMM) and Hidden Markov
Models (HMM) or those with SVM classifiers, inef-
ficient. The past few years has seen a prolific rise of
deep learning based solutions for sound event detec-
tion (SED) as well as for anomalous/rare sound event
detection.
In this preliminary study, we have chosen to work
with six anomaly classes. Our classes of interest are
car crash, tire skidding, glass breaking, gunshot, ex-
plosion, and screaming. Such a system can promptly
alert emergency response teams in the event of a
mishap to help those who are affected. This work
also can be useful for acoustic monitoring of natu-
ral or artificial environments including but not lim-
ited to smart cities, for noise monitoring for safe and
healthy living as well as to manage the noise pollution
for safe hearing. Sound is not a conventional signal
for anomaly detection in traffic surveillance which is
usually done using videos. A video based surveillance
system fails in the cases of fog, cloud, snow, smoke,
occlusions etc. and in such cases an audio based sys-
tem can take over the anomaly detection.
Considering the rare nature of the anomaly classes
that we have considered, we design a monophonic
sound event detection system. The main challenges
in designing a learning based solution is the lack of
strongly labelled data. Annotating real traffic data is
costly and time intensive, leading to the issue of in-
sufficient labeled data for the anomaly classes we se-
lected for our use case. Also lack of realistic data for
the classes of interest is another major issue. In order
to solve these issues for the classes that we selected,
we created a dataset for these six classes, using real
traffic background sound and also used a collection of
strongly labeled, weakly labeled and unlabeled sam-
ples.
A transformer based student-teacher network has
been used for Sound Event Detection (SED) in (Shao
et al., 2023). We adapt this method and retrain it for
our purpose. We achieve a 99.33% accuracy. In ad-
dition, we also show, via simulation that the perfor-
mance of the system does not degrade much with the
distance of occurrence of the event from the camera.
Shukla, M., Rameshan, R. M. and Gupta, S.
Anomalous Event Detection in Traffic Audio.
DOI: 10.5220/0013152100003905
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2025), pages 613-619
ISBN: 978-989-758-730-6; ISSN: 2184-4313
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
613

2 RELATED WORKS
A brief overview of related work is given in this sec-
tion. Prior research in audio processing has explored
various supervised, unsupervised and self-supervised
learning approaches. Supervised methods rely on la-
beled data, which can be expensive and time con-
suming to obtain. Unsupervised and self-supervised
learning methods are gaining popularity due to their
ability to learn from patterns or by generating rep-
resentations for audio samples using contrastive loss
(Khosla et al., 2020).
Methods which used traditional machine learning
worked with Gaussian Mixture Models (GMM) (Ito
et al., 2009), Hidden Markov Models (HMM) for fea-
ture extraction and classification (Zeiger, 2008) or
support vector machines (Foggia et al., 2016) with
features based on MFCC. (Giri et al., 2020) used
pool of support vector machines (SVM) to detect two
classes of hazardous road events tire skidding and car-
crash.
As mentioned in the previous section, these tradi-
tional methods do not perform well when it comes to
the anomalous sound event detection task.
Recent approaches rely on convolutional neural
networks, specifically models like auto encoders, and
variational auto encoders. Also, transformer based
and knowledge distillation based solutions exist for
sound event detection. Two of the auto encoder based
solutions are (Koizumi et al., 2019b) and (Wichern
et al., 2021). In the former an auto encoder was
trained to minimize the reconstruction error of ob-
served sounds, thus reducing the false positive rate.
They used synthetic data. (Wichern et al., 2021)
used an objective function based on Neyman Pearson
lemma (Neyman and Pearson, 1933) to train the auto
encoder to maximize the true positive rate. Wichern
et al. used attentive neural processes, a meta learning
approach, to train a masking based auto encoder for
audio anomaly detection. They have reported the re-
sults on publicly available anomaly sound detection
dataset for machine sounds, MIMII (Purohit et al.,
2019). MIMII is an open-source dataset for malfunc-
tioning industrial machine investigation and inspec-
tion.
Another class of approaches that we explored in-
clude large models specifically for feature generation.
(Kong et al., 2020) designed PANNs (pretrained au-
dio neural networks) by training a CNN model called
wavegram-logmel-CNN using both log-mel spectro-
grams and waveform as input feature, though, it gives
SOTA results for the audio tagging tasks for 527 au-
dioset (Gemmeke et al., 2017) classes of sound. Simi-
lar performance is achieved with another open source
model named YAMNet (YAM, ). YAMNet used Mo-
bileNetV1 (Howard et al., 2017) architecture and was
trained with audioset data (Gemmeke et al., 2017) for
providing inference for 521 predefined classes of au-
dioset. While both the models performed well for
audio tagging, the performance on anomalous sound
event detection was not good in the presence of back-
ground noise.
Though natural sounds occur in a polyphonic
manner, because of the way we model our problem,
we treat the sound samples as monophonic. The work
in (Radford et al., 2023) is a supervised approach
dealing with polyphonic sound event detection. This
work proposes a convolutional recurrent neural net-
work (CRNN) which combines the ability of CNNs
to extract high level spatio-temporal invariant features
and the ability of RNN to learn long term temporal
correlations.
In recent years, DCASE (DCA, ) has attracted
many researchers to the domain of acoustic event and
scene detection task which led to multiple novel so-
lutions to this problem. Some of the recent ones be-
ing DCASE 2023 task 4 (DCA, 2023) baseline pro-
vided by the challenge organizers, which uses BEATs
(Chen et al., 2023) (Bidirectional Encoder represen-
tation from Audio Transformers) and a CRNN net-
work in a teacher-student manner for sound event de-
tection task in domestic environment using DESED
dataset (DES, ). A paper on similar lines is, Shao et al.
(Shao et al., 2023) which replaced BEATs in (Chen
et al., 2023) with ATST-Frame model introduced in
(Li et al., 2023). We are using this idea and model
in our proposed anomalous sound event detection for
traffic audio.
Another recent work is SPecific anomaly IDen-
tifiER network called SPIDERnet (Koizumi et al.,
2020) which is a one-shot anomaly detection method
for anomalous sound. In this anomaly detec-
tion system they used a neural network-based fea-
ture extractor for measuring similarity in embedded
space and attention mechanisms for absorbing time-
frequency stretching. Although SPIDERnet outper-
forms conventional methods and robustly detects var-
ious anomalous sounds we decided not to adapt this
solution as the results are provided only for the ma-
chine datasets ToyADMOS (Koizumi et al., 2019a)
and MIMII (Purohit et al., 2019).
3 PROPOSED SOLUTION
The most general approach to anomaly detection is
to train a system to learn what is normal and anything
which falls out of the distribution of normal is marked
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
614

as anomalous. Anomalous events are usually rare,
and unknown apriori, making the learning process
quite challenging. In certain restricted cases where
we know the types of anomalies that could come up,
one can follow a classification based approach for
anomaly detection. Our focus being anomaly detec-
tion in Indian traffic, we select six possible classes
and train a system to raise an alarm when any of these
six events occur. The classes considered are 1. car
crash 2. tire skidding 3. gun shot 4. explosion 5. glass
breaking and 6. screaming .
In this paper we have followed the work of Shao et
al. (Shao et al., 2023) which finetunes the pre-trained
ATST model for sound event detection task. We use
the architecture from (Li et al., 2023) and apply the
finetuning strategy of (Shao et al., 2023) for training
using our dataset. Our focus is on monophonic multi-
class traffic anomaly detection along with recognizing
the start time i.e onset and end time i.e offset of the
anomaly.
The architecture consists of the pretrained ATST-
Frame (Li et al., 2023) model along with a CRNN net-
work. ATST-Frame is an audio teacher-student trans-
former based model architecture which is trained in
such a way that it can learn frame-wise representa-
tions by maximizing the similarity between student’s
and teacher’s frame-level embeddings. It is impor-
tant to note that, Shao et al. replaced pre-trained
BEATs (Chen et al., 2023) with pre-trained ATST-
frame model from the baseline solution of DCASE
2023 challenge (DCA, 2023) and the got SOTA re-
sults for sound event detection task using DESED
dataset (DES, ).
The idea of the paper (Shao et al., 2023) is to use
a trained CRNN to finetune the ATST model. Train-
ing progresses in two stages: in stage 1 the ATST
model is kept frozen and the CRNN is trained using
the labelled samples via BCE loss and the output of
ATST via the Mean Teacher (MT) loss. In the second
stage of training the output of CRNN acts as pseudo
label data for finetuning the ATST. In this step the
training loss is a combination of binary cross entropy
(BCE), mean-teacher (MT) and interpolation consis-
tency training (ICT). It may be noted that both the
models are updated in the latter stage.
The system has been trained using strongly la-
belled, weakly labelled and unlabelled data. The pro-
cess of data creation is given in the next section.
4 DATASET CREATION AND
TRAINING
We have prepared the dataset for sound event detec-
tion problem with explicitly collected background au-
dios from specified traffic sites and event only audios
collected from various open-source datasets includ-
ing SESA (Sound Events for Surveillance applica-
tions) (SES, ) MIVIA labs audio events dataset (MIV,
a), Mivia Labs ROAD Audio Events Dataset (MIV,
b) and DCASE 2017, TUT Rare Sound Events 2017
(DCA, 2017).
4.1 Real Background Audio Data
Creation
The background audios are recorded at a single site.
Nevertheless, a natural variability is introduced by the
varying congestion conditions that exists across the
day. In addition, we have introduced two different
SNR levels: 0dB and 5dB for the background audio.
For recording the background sound, we used in-
built microphone from an IP camera device DH-IPC-
HF5231EP-E 2MP WDR Box Network Camera (Dah,
) from Dahua Security. This camera is installed at a
traffic junction in a busy Indian city. Background data
is recorded in .AAC format at 16KHz sampling rate
and saved on an hourly basis. One hour long back-
ground data is then converted from .aac to .wav for-
mat and splitted into segments of duration 10 second
each.
4.2 Event Audio Collection
The anomalous event audios were extracted from
their respective datasets and the event alone part was
clipped out based on the onset and offset information
given in their respective meta data. These are, then
down-sampled to 16kHz audio segments.
4.3 Background and Event Mixing
Procedure
After the above mentioned steps, the resulting data
had a total of 22,225 background audio samples of 10
seconds each and a total of 670 event audio files. The
six event classes are Carcrash, Gunshot, Glassbreak-
ing, Screaming, Explosion and Tireskidding. These
audio events were then synthetically mixed with the
traffic environment background audios.
Based on the energy level of the audio clips the
background data is split into three categories of low,
medium and high energy audio files. These three
Anomalous Event Detection in Traffic Audio
615

Figure 1: t-SNE scatter plot for Carcrash and Screaming
using PANNs embeddings.
Table 1: Segment Based Performance Metrics.
Class
Stage 1 Stage 2
Precision Recall Precision Recall
Tireskidding 87.6 77.3 97.0 94.8
Carcrash 59.0 78.5 96.7 92.9
Screaming 52.3 91.9 96.7 98.9
Explosion 98.7 97.7 98.2 98.9
Glassbreaking 74.0 66.5 79.8 98.9
Gunshot 97.6 59.4 97.1 92.5
Overall 87.73 81.78 95.73 96.51
Overall Accuracy 97.45 99.33
Overall F-score 84.65 96.12
Overall Error-rate 0.22 0.06
Table 2: Event Based Performance Metrics.
Class
Stage 1 Stage 2
Precision Recall Precision Recall
Tireskidding 61.7 64.0 83.3 89.5
Carcrash 31.7 55.8 90.9 87.9
Screaming 41.2 76.5 100 100
Explosion 94.0 95.5 97.4 98.0
Glassbreaking 51.0 51.7 70.2 96.5
Gunshot 75.2 37.4 98.5 90.8
Overall 66.11 61.74 92.74 93.86
Overall F-score 63.85 93.30
Overall Error-rate 0.62 0.10
Table 3: Segment Based Performance Metrics for Stage 2
(SNR 0 dB).
Class
Event Scale 1 Event Scale 0.25
Precision Recall Precision Recall
Tireskidding 97.1 94.4 97.2 90.6
Carcrash 96.8 91.9 95.3 78.7
Screaming 96.5 98.9 97.8 97.8
Explosion 98.3 98.7 97.2 93.8
Glassbreaking 78.9 99.5 77.7 89.1
Gunshot 96.8 92.4 97.0 85.7
Overall 95.53 96.42 94.99 90.23
Overall Accuracy 99.30 98.74
Overall F-score 95.97 92.55
Overall Error-rate 0.06 0.12
background audio categories and each of the class-
wise event data are split in the ratio of 6:2:2 for train
set, validation set and test set, respectively. This en-
Table 4: Event Based Performance Metrics for Stage 2
(SNR 0 dB).
Class
Event Scale 1 Event Scale 0.25
Precision Recall Precision Recall
Tireskidding 83.7 88.1 87.4 87.1
Carcrash 91.6 87.5 75.9 70.9
Screaming 100 100 96.8 98.0
Explosion 96.6 97.6 60.7 76.1
Glassbreaking 68.7 95.6 63.0 84.4
Gunshot 98.8 90.8 88.3 82.0
Overall 92.48 93.49 75.59 81.26
Overall F-score 92.98 78.32
Overall Error-rate 0.11 0.41
Table 5: Performance Metrics for CRNN only Network.
Class
Segment Based Event Based
Precision Recall Precision Recall
Tireskidding 75.0 98.1 71.8 85.7
Carcrash 74.1 75.6 29.2 27.2
Screaming 81.6 91.7 69.9 82.7
Explosion 91.0 99.1 71.3 74.8
Glassbreaking 61.6 80.5 40.1 40.1
Gunshot 58.6 90.0 62.1 67.5
Overall 74.97 93.08 61.86 66.19
Overall F-score 83.05 63.95
Overall Error-rate 0.35 0.71
ergy based background audio splitting ensured that
all three levels of Indian traffic congestion conditions
- low, medium and highly noisy - are equally dis-
tributed across the training, validation and testing sets
of the data created.
The event only signal duration varies from 0.5s up
to 9.9s with varying duration for different classes. Es-
pecially, event segments like tire skidding have com-
paratively longer duration. As the background audio
duration is 10s, total duration of the mixture is fixed
to 10s. Mixing of the background and event are done
independently in all the three categories, thus ensur-
ing that no sample is common across the three sets:
train, validation and test.
The background and the event are mixed at two
SNR levels 0dB and 5dB. Let e[n] be the event and
b[n] be the background, the scaling factor a is calcu-
lated as:
a =
v
u
u
u
u
u
t
N−1
∑
i=0
e
2
[i]
10
(SNR)
10
M−1
∑
i=0
b
2
[i]
(1)
As N, M, the total length of event and background,
respectively, are different with the condition M > N,
we need to decide, where in the duration 0 to M − N,
we need to add the event. For this, we select an event
audio file and a corresponding background audio file
into which the event file is added at a random location
between 0 to M −N.
We mixed event and background data in such a
way that we could prepare the data according to these
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
616

segments: strongly labelled data, weakly labelled data
and unlabelled data.
In strongly labelled data, for each mixed audio file
event the start time (onset) and event end time (offset)
and event label (class to which it belongs) is given.
In weak labels, only the event class label is available,
and in unlabelled segments we have only mixed files
with no other information, and these form the self-
supervised data.
Total of 34716 files have been created, out of
which 21256 are used for training, 9001 for testing
and 4459 audio files for validation.
4.4 Training
Training was done in two different stages because
both the models required very different training
schedule and learning parameters.
In stage 1, ATST-Frame was kept frozen and only
CRNN network was trained with batch size 4,8,8
for strongly labelled, weakly labelled and unlabelled
data. The input audio clips were divided in frames
of duration 128ms with a hop length 16ms. 128-
dimensional Log-Mel features were extracted for each
frame.
Data augmentation is randomized with no aug-
mentation having a probability of 0.5 with mix-up and
frequency-warping, each having a probability of 0.25,
each. For stage 1, learning rate was set at 1e-3 for both
CNN and RNN.
Same batch size was used for stage 2 as well. For
stage 2, learning rate was set at 2e-4 and 2e-3 for CNN
and RNN, respectively.
Adam optimizer has been used. Binary cross en-
tropy loss and both mean teacher loss and interpola-
tion consistency training loss has been used for super-
vised and unsupervised loss calculation, respectively.
5 RESULTS
We evaluated the above system on a total of 9001
audio clips. The evaluation has been performed us-
ing two methods: event based and segment based, in-
spired from polyphonic sound detection score metric
calculation (Bilen et al., 2020), where event wise eval-
uation provides clip level evaluations, segment wise
results provide evaluation score based on frame length
of 64 ms. For segment based results, we have calcu-
lated the overall accuracy (calculated based on micro
average) for the system which is going from 97.45
percent to 99.33 percent in stage 1 and stage 2, re-
spectively.
Table 1 gives segment-wise precision and recall
for stage 1 and stage 2. It may be noted that both pre-
cision and recall have increased considerably for all
the 6 classes after stage 2 training. Table 2 gives the
event based precision and recall for both the stages.
Here also a remarkable improvement in the perfor-
mance is seen after stage 2 training.
In order to see the performance when the location,
at which the event occurred is far away from the mic,
we created a test data set where the event amplitudes
were scaled by a factor of 0.25. To compare the per-
formance between original and the scaled audio, we
created two test datasets, one with event scale 1 and
the other with event scale 0.25. Everything else, in-
cluding the onset and offset of the events is same be-
tween the two event scales for fair comparison. Table
3 gives the stage 2 results for segment-based evalua-
tion, when the event source moves away. It is seen
that, precision falls only slightly, indicating that the
true positives are still predicted correctly to a large
extent. Whereas the fall in recall, indicates that many
of the true cases are missed, being classified as false
negatives. This is also reflected in the error score.
Similarly, Table 4 gives the stage 2 results for
event-based evaluations for event scale 0.25. Here,
the performance degrades, significantly. Both the
false positives and false negatives have increased con-
siderably as is evident from the lower precision and
recall as compared to the corresponding event scale 1
values. The error has risen to 0.41 in line with there
observations.
It may be noted that all the overall metric scores
mentioned in the tables (1-5) have been calculated on
the basis of micro-averaging.
To establish that the student-teacher based model
is far superior, we also show the results for a sim-
ple CRNN in Table 5. Comparison with other mod-
els are not made available as the embeddings learnt
using those (PANNs, YAMNet) were not discrimina-
tive enough. It was observed from the t-SNE plot that
these anomalous classes overlapped with each other.
In Fig. 1 we show the tSNE of PANNs (Kong et al.,
2020) embeddings for two classes: screaming and
carcrash. It is evident from the plot that classifica-
tion is nearly impossible. From table 5, it is seen that,
segment-based as well as event-based scores for fine-
tuned ATST-SED model on traffic anomaly classes
are far superior to the base CRNN results.
6 CONCLUSIONS
In this paper, we re-trained a transformer based
student-teacher network for Sound Event Detection
Anomalous Event Detection in Traffic Audio
617

task following Shao et al. (Shao et al., 2023), us-
ing synthesized dataset for our purpose. Based on
our model, unknown acoustic patterns are identified
into six different anomaly classes. The use of student-
teacher transformer allows the learning of long-term
temporal dependencies. When trained and fine-tuned
with on the synthetic dataset generated using real
traffic audio, the model gave an overall accuracy of
99.33% when tested on unseen audio. The model
performance degrades gracefully with distance of the
source of anomaly which is an added advantage.
REFERENCES
Dahua security network cameras. https://zenodo.org/
records/3519845. Accessed: October 22, 2024.
Dahua security network cameras. https://www.
dahuasecurity.com/in/products/All-Products/
Network-Cameras/WizMind-Series/5-Series/2MP/
DH-IPC-HF5231EP-E. Accessed: October 22, 2024.
DCASE community. https://dcase.community/. Accessed:
October 22, 2024.
DESED: Domestic environment sound event detection.
https://project.inria.fr/desed/. Accessed: October 22,
2024.
MIVIA audio events dataset. https://mivia.unisa.it/datasets/
audio-analysis/mivia-audio-events. Accessed: Octo-
ber 22, 2024.
MIVIA road audio events dataset. https:
//mivia.unisa.it/datasets/audio-analysis/
mivia-road-audio-events-data-set. Accessed:
October 22, 2024.
YAMNet. https://www.tensorflow.org/hub/tutorials/
yamnet. Accessed: October 22, 2024.
(2017). Dcase 2017 challenge: Rare sound event
detection. https://dcase.community/challenge2017/
task-rare-sound-event-detection. Accessed: October
22, 2024.
(2023). DCASE 2023 challenge: Sound
event detection with weak and soft la-
bels. https://dcase.community/challenge2023/
task-sound-event-detection-with-weak-and-soft-labels.
Accessed: October 22, 2024.
Bilen, C., Ferroni, G., Tuveri, F., Azcarreta, J., and
Krstulovic, S. (2020). A framework for the robust
evaluation of sound event detection. In ICASSP 2020-
2020 IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP), pages 61–
65.
Chen, S., Wu, Y., Wang, C., Liu, S., Tompkins, D., Chen,
Z., Che, W., Yu, X., and Wei, F. (2023). Beats: Au-
dio pre-training with acoustic tokenizers. In Inter-
national Conference on Machine Learning (ICML),
pages 5178–5193.
Foggia, P., Petkov, N., Saggese, A., Strisciuglio, N., and
Vento, M. (2016). Audio surveillance of roads:
A system for detecting anomalous sounds. IEEE
Transactions on Intelligent Transportation Systems,
17(1):279–288.
Gemmeke, J. F., Ellis, D. P. W., Freedman, D., Jansen,
A., Lawrence, W., Moore, R. C., Plakal, M., and Rit-
ter, M. (2017). Audio set: An ontology and human-
labeled dataset for audio events. In IEEE International
Conference on Acoustics, Speech and Signal Process-
ing (ICASSP), pages 776–780.
Giri, R., Tenneti, S. V., Cheng, F., Helwani, K., Isik, U.,
and Krishnaswamy, A. (2020). Self-supervised clas-
sification for detecting anomalous sounds. In Proc.
DCASE, pages 46–50.
Howard, A., Zhu, M., Chen, B., Kalenichenko, D., Wang,
W., Weyand, T., Andreetto, M., and Adam, H.
(2017). Mobilenets: Efficient convolutional neu-
ral networks for mobile vision applications. arXiv
preprint arXiv:1704.04861.
Ito, A., Aiba, A., Ito, M., and Makino, S. (2009). Detection
of abnormal sound using multi-stage gmm for surveil-
lance microphone. In Proc. IAS, volume 1, pages 733–
736.
Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y.,
Isola, P., Maschinot, A., Liu, C., and Krishnan, D.
(2020). Supervised contrastive learning. In Proceed-
ings of the Advances in Neural Information Process-
ing Systems.
Koizumi, Y., Saito, S., Uematsu, H., Harada, N., and
Imoto, K. (2019a). Toyadmos: A dataset of miniature-
machine operating sounds for anomalous sound detec-
tion. In Proc. of the Workshop on Applications of Sig-
nal Processing to Audio and Acoustics (WASPAA).
Koizumi, Y., Saito, S., Uematsu, H., Kawachi, Y., and
Harada, N. (2019b). Unsupervised detection of
anomalous sound based on deep learning and the ney-
man–pearson lemma. IEEE/ACM Transactions on Au-
dio, Speech, and Language Processing, 27(1):212–
224.
Koizumi, Y., Yasuda, M., Murata, S., Saito, S., Uematsu,
H., and Harada, N. (2020). Spidernet: Attention
network for one-shot anomaly detection in sounds.
In ICASSP 2020 - 2020 IEEE International Con-
ference on Acoustics, Speech and Signal Processing
(ICASSP), pages 281–285, Barcelona, Spain.
Kong, Q., Cao, Y., Iqbal, T., Wang, Y., Wang, W., and
Plumbley, M. D. (2020). Panns: Large-scale pre-
trained audio neural networks for audio pattern recog-
nition. arXiv:1912.10211.
Li, X., Shao, N., and Li, X. (2023). Self-supervised au-
dio teacher-student transformer for both clip-level and
frame-level tasks. arXiv preprint arXiv:2306.04186.
Neyman, J. and Pearson, E. S. (1933). On the problem of the
most efficient tests of statistical hypotheses. Philos.
Trans. Roy. Soc. London, 231:289–337.
Purohit, H., Tanabe, R., Ichige, K., Endo, T., Nikaido,
Y., Suefusa, K., and Kawaguchi, Y. (2019). Mimii
dataset: Sound dataset for malfunctioning industrial
machine investigation and inspection. In Proc. Detec-
tion Classification Acoustic Scenes Events Workshop
(DCASE), page 209.
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
618

Radford, A., Narasimhan, K., Jeong, D., Chowdhery,
J., Butler, J., Shuster, L., Parmar, N., Clark, D.,
and Elibol, J. (2023). Robust speech encoding
via large-scale weak supervision. arXiv preprint
arXiv:2309.06864v2.
Shao, N., Li, X., and Li, X. (2023). Fine-tune the pretrained
atst model for sound event detection. arXiv preprint
arXiv:2309.08153v2.
Wichern, G., Chakrabarty, A., Wang, Z. Q., and Roux, J. L.
(2021). Anomalous sound detection using attentive
neural processes. In 2021 IEEE Workshop on Appli-
cations of Signal Processing to Audio and Acoustics
(WASPAA), pages 186–190, New Paltz, NY, USA.
Zeiger, C. (2008). An HMM based system for acoustic
event detection. In Multimodal Technologies for Per-
ception of Humans.
Anomalous Event Detection in Traffic Audio
619