Buchumschlag

AudioSet-EV: an AudioSet-derived distribution of Emergency Vehicle Siren sounds

Fuente: Zenodo

Gespeichert in:

Bibliographische Detailangaben
Hauptverfasser:	Giacomelli, Stefano, Rinaldi, Claudia
Format:	Recurso digital
Sprache:	Englisch
Veröffentlicht:	Zenodo 2025
Schlagworte:	Sound Sound Recordings Deep Learning Deep Learning/statistics & numerical data AudioSet Python PyTorch Datasets as Topic/statistics & numerical data
Online-Zugang:	Acceder al recurso
Tags:	Tag hinzufügen Keine Tags, Fügen Sie den ersten Tag hinzu!

_version_	1866902036185874432
author	Giacomelli, Stefano Rinaldi, Claudia
author_facet	Giacomelli, Stefano Rinaldi, Claudia
contents	<h1>AudioSet-EV</h1> <h2><br>Overview</h2> <blockquote> <div><strong>AudioSet-EV</strong>, is a case-study and tailored distribution of <a href="https://research.google.com/audioset/index.html" target="_blank" rel="noopener">AudioSet (©Google)</a> (AS) for acoustic Emergency Vehicle (EV) siren detection and recognition. By selectively grouping siren and non-siren urban sounds, enforcing taxonomy consistency, and mitigating class imbalances, AudioSet-EV offers a robust, large-scale resource for research in Machine Learning and Deep Learning acoustic modeling. <strong>AudioSet-EV v2</strong> is a refined and extended distribution of AudioSet-EV specifically derived from the PANNs (Pre-trained Audio Neural Networks) <a href="https://huggingface.co/datasets/confit/audioset-qiuqiangkong" target="_blank" rel="noopener">AudioSet 2020 release</a>, providing a larger amount of data (due to original contents availability). By leveraging selective grouping of Siren and Non-Siren urban sounds, enforcing taxonomy consistency, implementing negatives stratified class balancing, and supporting both binary and 4-way classification (Negative sounds VS Police, Ambulance, Fire Truck Sirens), AudioSet-EV v2 provides the most robust, and largest-scale resource for research in Machine Learning and Deep Learning acoustic EV modeling.</div> </blockquote> <h3><br>Key Improvements over <strong>AudioSet-EV v1</strong> (AudioSet 2025 availability)</h3> <ul> <li><strong>Enhanced Data Quality and Quantity</strong>: built from PANNs AS release with stabilized audio quality and larger files avilability</li> <li><strong>Stratified Negative Sampling</strong>: balanced negative samples across 39 sound categories (traffic, music, speech, alarms, etc.)</li> <li><strong>Extended Coverage</strong>: ~28,000 total samples (7,900 Positives, 20,916 Negatives)</li> <li><strong>Multi-Class Support:</strong> fine-grained annotations for police, ambulance, and fire truck sirens</li> <li><strong>Segment-Type Metadata</strong>: files organized by AudioSet <code>'balanced_train'</code>, <code>'eval'</code>, and <code>'unbalanced'</code> segments</li> <li><strong>Reproducible Processing</strong>: deterministic stratification for consistent experimental setups</li> </ul> <h2> </h2> <h2>Methodology</h2> <div> <p>Our design methodology encompasses a systematic selection and filtering of relevant AS samples, with <a href="https://github.com/StefanoGiacomelli/audioset-tools/tree/main" target="_blank" rel="noopener">AudioSet-Tools</a> and a <em>binary distinction</em> between True Positives (siren-related) and True Negatives (non-siren) samples, mitigating class imbalances and label contamination. We emphasize that, given the <a href="https://research.google.com/audioset/download.html" target="_blank" rel="noopener">original weak labeling nature</a>, total reliability of the label association process cannot be guaranteed.</p> <p>We structured AudioSet-EV into two primary groups:</p> <ol> <li><strong>Positives</strong>: including only EV-siren-related classes, specifically <code>'Police car (siren)', 'Ambulance (siren)', 'Fire engine, fire truck (siren)'</code>, and the ontology containers class <code>'Emergency vehicle', 'Siren'</code>, to account for any weakly labeled or meaningful sound.</li> <li><strong>Negatives</strong>: consisting of a diverse and challenging set, comprising: <ul> <li> <em>vehicle-related sounds</em> (<code>'Car', 'Car passing by', 'Power windows, electric windows', 'Tire squeal', 'Motor vehicle (road)', 'Truck', 'Air brake', 'Ice cream truck, ice cream van', 'Bus', 'Motorcycle', 'Skidding', 'Race car, auto racing', 'Bicycle', 'Train', 'Rail transport', 'Train wheels squealing', 'Railroad car, train wagon', 'Skateboard'</code>)</li> <li><em>alarm signals</em> (<code>'Car alarm', 'Vehicle horn, car horn, honking', 'Bicycle bell', 'Train horn', 'Train whistle', 'Foghorn', 'Toot', 'Reversing beeps', 'Beep, bleep', 'Civil defense siren', 'Alarm', 'Smoke detector, smoke alarm', 'Fire alarm', 'Buzzer'</code>),</li> <li><em>environmental noises</em> (<code>'Traffic noise, roadway noise', 'Outside, rural or natural', 'Outside, urban or manmade'</code>).</li> </ul> </li> </ol> <p>We also included some <code>Speech</code>, <code>Music</code> and <code>Engine</code>-related sounds to improve robustness against waveform pattern similarities and semantic taxonomy proximities.</p> </div> <div> </div> <div> <h2>Pre-Processing</h2> <p>For <em>Positives</em> category, segments processing followed these steps:</p> <ul> <li><em>Selection by Label</em>: <code>balanced</code>, <code>unbalanced</code> and <code>eval</code> AS segments were filtered according to our <strong>Positives</strong> label selection.</li> <li>Segments Differentiation <strong>(New in v2)</strong>: given the original tracking and results availability consistency, samples were differentiated across original segment origin in <code>.csv</code> files, while still belonging to the same Positives group.</li> <li><em>Blacklist Filtering</em>: to refine our selection, any <code>'Civil defense siren'</code> sample was removed to prevent contamination with non-emergency <em>Vehicle</em> sounds.</li> </ul> <p>For the <em>Negatives</em> category, datasets processing followed these steps:</p> <ul> <li><em>Selection by Label</em>: <code>balanced</code>, <code>unbalanced</code> and <code>eval</code> AS entries, matching our defined <strong>non-siren</strong> labels, were extracted.</li> <li>Segments Differentiation <strong>(New in v2)</strong>: same as for Positives.</li> <li><em>Partial Blacklist Filtering</em>: to avoid overlaps with the <strong>Positives</strong> category, samples containing at least one positive class label were removed, except for <code>'Civil defense siren'</code>, which is taxonomically included within the <code>'Siren'</code> container class.</li> <li><em>Stratified Class Re-Balancing</em><strong> (New in v2)</strong>: samples were stratified balanced across 39 ontology categories using random (label-aware) sampling, to preserve class representation while mitigating imbalance. This ensures diversity across alarm types, traffic sounds, music, speech, and environmental noises.</li> </ul> <p>Final <code>.csv</code> files were processed through two independent instances of our AudioSet-Tools downloader, configured to re-sample YouTube audio to 32KHz (without pitch shifting artifacts!), reduce files to mono, and avoid amplitude normalization. We stress the aspect that, given the large amount of <em>Negatives</em>, there actually exist multiple variants of this subset (due to the randomized class down-sampling process). </p> </div> <p> </p> <h2>Comparative POSITIVES Summary Statistics </h2> <table style="width: 100.063%;"> <tbody> <tr> <td style="width: 31.6959%;"> </td> <td style="width: 15.5263%;">Samples</td> <td style="width: 13.1313%;"><code>Police car (siren)</code></td> <td style="width: 11.9949%;"><code>Ambulance (siren)</code></td> <td style="width: 27.6515%;"><code>Fire engine, fire truck (siren)</code></td> </tr> <tr> <td style="width: 31.6959%;">Positives</td> <td style="width: 15.5263%;">8 409 </td> <td style="width: 13.1313%;">3 643 </td> <td style="width: 11.9949%;">1 931</td> <td style="width: 27.6515%;">3 187</td> </tr> <tr> <td style="width: 31.6959%;"><code>Downloaded</code> AudioSet v1</td> <td style="width: 15.5263%;">7 324</td> <td style="width: 13.1313%;">3 124</td> <td style="width: 11.9949%;">1 637</td> <td style="width: 27.6515%;">2 852</td> </tr> <tr> <td style="width: 31.6959%;"><code><strong>Downloaded</strong></code><strong> AudioSet v2</strong></td> <td style="width: 15.5263%;"><strong>7 900</strong></td> <td style="width: 13.1313%;"><strong>3 219</strong></td> <td style="width: 11.9949%;"><strong>1 732</strong></td> <td style="width: 27.6515%;"><strong>2 947</strong></td> </tr> </tbody> </table> <h2>NEW: AudioSet-EV v2 Multi-Label Awareness</h2> <table style="width: 100.063%;"> <tbody> <tr> <td style="width: 41.1616%;"> </td> <td style="width: 13.7626%;"><strong>Pure</strong> Samples</td> <td style="width: 21.8506%;"><strong>Multi-Positive</strong> Samples</td> <td style="width: 23.0989%;">Total Samples</td> </tr> <tr> <td style="width: 41.1616%;"><code>Police car (siren)</code></td> <td style="width: 13.7626%;">2 643</td> <td style="width: 21.8506%;">912</td> <td style="width: 23.0989%;"><strong>3 124</strong></td> </tr> <tr> <td style="width: 41.1616%;"><code>Ambulance (siren)</code></td> <td style="width: 13.7626%;">1 020</td> <td style="width: 21.8506%;">617</td> <td style="width: 23.0989%;"><strong>1 637</strong></td> </tr> <tr> <td style="width: 41.1616%;"><code>Fire engine, fire truck (siren)</code></td> <td style="width: 13.7626%;">2 713</td> <td style="width: 21.8506%;">139</td> <td style="width: 23.0989%;"><strong>2 852</strong></td> </tr> </tbody> </table> <h2>NEW: AudioSet-EV v2 Segment Awareness</h2> <table style="width: 100.063%; height: 140.657px;"> <tbody> <tr style="height: 20.0938px;"> <td style="width: 21.5599%; height: 20.0938px;"><strong>POSTIVES</strong></td> <td style="width: 30.2141%; height: 20.0938px;"><code>Downloaded</code><strong> </strong>Samples</td> <td style="width: 27.3636%; height: 20.0938px;">Total Samples</td> <td style="width: 21.3063%; height: 20.0938px;">Percentage (%)</td> </tr> <tr style="height: 20.0938px;"> <td style="width: 21.5599%; height: 20.0938px;"><code>balanced_train</code></td> <td style="width: 30.2141%; height: 20.0938px;"><strong>135</strong></td> <td style="width: 27.3636%; height: 20.0938px;">146</td> <td style="width: 21.3063%; height: 20.0938px;">92.5%</td> </tr> <tr style="height: 20.0938px;"> <td style="width: 21.5599%; height: 20.0938px;"><code>eval</code></td> <td style="width: 30.2141%; height: 20.0938px;"><strong>135 </strong></td> <td style="width: 27.3636%; height: 20.0938px;">148 </td> <td style="width: 21.3063%; height: 20.0938px;">91.2%</td> </tr> <tr style="height: 20.0938px;"> <td style="width: 21.5599%; height: 20.0938px;"><code>unbalanced</code></td> <td style="width: 30.2141%; height: 20.0938px;"><strong>7 630</strong></td> <td style="width: 27.3636%; height: 20.0938px;">8 154</td> <td style="width: 21.3063%; height: 20.0938px;">93.6%</td> </tr> <tr style="height: 20.0938px;"> <td style="width: 21.5599%; height: 20.0938px;"><strong>NEGATIVES</strong></td> <td style="width: 30.2141%; height: 20.0938px;"><code>Downloaded</code><strong> </strong>Samples</td> <td style="width: 27.3636%; height: 20.0938px;">Total Samples</td> <td style="width: 21.3063%; height: 20.0938px;"><strong> </strong></td> </tr> <tr style="height: 20.0938px;"> <td style="width: 21.5599%; height: 20.0938px;"><code>balanced_train</code></td> <td style="width: 30.2141%; height: 20.0938px;"><strong>10 963</strong></td> <td style="width: 27.3636%; height: 20.0938px;">11 871</td> <td style="width: 21.3063%; height: 20.0938px;">92.4%</td> </tr> <tr style="height: 20.0938px;"> <td style="width: 21.5599%; height: 20.0938px;"><code>eval</code></td> <td style="width: 30.2141%; height: 20.0938px;"><strong>9 953</strong></td> <td style="width: 27.3636%; height: 20.0938px;">10 808</td> <td style="width: 21.3063%; height: 20.0938px;">92.1%</td> </tr> </tbody> </table> <p><strong>Note</strong>: The <code>'unbalanced'</code> segment is <strong>not included</strong> in the Negative set to avoid excessive class imbalancing.</p> <h2>Comparative Summary Statistics</h2> <table style="width: 100.063%; height: 59.2814px;"> <tbody> <tr style="height: 19.5938px;"> <td style="width: 29.7983%; height: 19.5938px;"> </td> <td style="width: 35.6064%; height: 19.5938px;">AudioSet-EV v1</td> <td style="width: 34.5963%; height: 19.5938px;">AudioSet-EV v2</td> </tr> <tr style="height: 19.5938px;"> <td style="width: 29.7983%; height: 19.5938px;">Positives</td> <td style="width: 35.6064%; height: 19.5938px;">7 324 </td> <td style="width: 34.5963%; height: 19.5938px;">7 900<strong> (+7.9%)</strong></td> </tr> <tr style="height: 20.0938px;"> <td style="width: 29.7983%; height: 20.0938px;">Negatives</td> <td style="width: 35.6064%; height: 20.0938px;">6 702</td> <td style="width: 34.5963%; height: 20.0938px;">20 916 <strong>(+212%)</strong></td> </tr> <tr> <td style="width: 29.7983%;"><strong>Samples Difference</strong></td> <td style="width: 35.6064%;"><strong>+576</strong></td> <td style="width: 34.5963%;"><strong>+12 214 (+105%)</strong></td> </tr> </tbody> </table> <div> <div> <div> </div> </div> </div> <div> </div> <div> </div> <div> <h2>File Structure</h2> <p>Shared by both Version (backward compatibility between versions):</p> </div> <div> <div><code>AudioSet-EV_vX_main.zip (XX GB)</code></div> <div><code>        ├── Positive_files/</code></div> <div><code>        │    ├── balanced_train/          # AudioSet-EV v2 differentiation</code></div> <div><code>        │    ├── eval/</code></div> <div><code>        │    └── unbalanced/</code></div> <div><code>        ├── Negative_files/</code></div> <div><code>        │    ├── balanced_train/          # AudioSet-EV v2 differentiation</code></div> <div><code>        │    └── eval/</code></div> <div><code>        ├── EV_Positives.csv              # Metadata with MIDs and labels (vX compatibility)</code></div> <div><code>        └── EV_Negatives.csv</code></div> </div> <h2>Info</h2> <p>Each <code>.csv</code> file contains:</p> <ul> <li>- <code>'yt_id'</code>: YouTube video identifier</li> <li>- <code>'positive_labels'</code>: List of AudioSet MIDs (<em>for positives only</em>)</li> <li>- <code>'segment_type'</code>: AudioSet segment (balanced_train, eval, unbalanced_train)</li> <li>- <code>'downloaded'</code>: Boolean flag indicating successful download (w.r.t. original 2017 AudioSet metadata release)</li> </ul> <div><strong>Use cases</strong></div> <ul> <li><strong>Binary Classification</strong>: Emergency Vehicle siren recognition (Presence VS Absence)</li> <li><strong>Multi-Class Classification</strong>: Fine-grained siren type recognition (Police VS Ambulance VS Firemans)</li> <li><strong>Transfer Learning</strong>: Pre-training on AudioSet-EV v2 for downstream Events detection tasks</li> <li><strong>Robust</strong> <strong>Testing</strong>: Evaluation against diverse confounding sounds (alarms, urban traffic, music)</li> <li><strong>Cross-Dataset Benchmarking:</strong> Comparison with sireNNet, LSSiren, FSD50K ESC-50 andUrbanSound8K pre-processed datasets.</li> </ul> <div><strong>Technical Notes</strong></div> <ul> <li> <div><strong>Audio Format</strong>: <code>.wav</code> (mono, 32 kHz, ~10 seconds)</div> </li> <li> <div><strong>Total Size</strong>: ~16 GB (.zip compressed), ~28 GB (uncompressed)</div> </li> <li> <div><strong>Label Reliability</strong>: <em>Weak labels</em> inherited from AudioSet (manual verification recommended for <em>critical</em> applications)</div> </li> <li> <div><strong>Reproducibility</strong>: Seed-controlled stratified sampling and groups selection ensures deterministic dataset splits</div> </li> </ul> <h1> </h1> <h1>References</h1> <div><strong>AudioSet-Tools Framework</strong></div> <div>S. Giacomelli <em>et al</em>.(2026) "AudioSet-tools: a Python framework for taxonomy-aware AudioSet curation and reproducible audio research" in EURASIP <em>Journal of Audio, Speech and Music Processing.</em> 2026, 2. DOI: <a href="https://doi.org/10.1186/s13636-025-00436-z" target="_blank" rel="noopener">10.1186/s13636-025-00436-z</a></div> <div> </div> <div> </div> <div><strong>E2PANNs Framework</strong></div> <div> <div>S. Giacomelli <em>et al</em>. (2025) "From Large-scale Audio Tagging to Real-Time Explainable Emergency Vehicle Sirens Detection", <strong><em>Under Review for IEEE TASLP</em></strong>, arXiv preprint DOI: <a href="https://arxiv.org/abs/2506.23437" target="_blank" rel="noopener">arXiv:2506.23437</a>, <strong>GitHub</strong>: <a href="https://github.com/StefanoGiacomelli/e2panns" target="_blank" rel="noopener">https://github.com/StefanoGiacomelli/e2panns</a></div> <div> </div> <div>M. Giordano <em>et al</em>. (2025) "Real-Time Emergency Vehicle Siren Detection with Efficient CNNs on Embedded Hardware," <em>2025 IEEE 6th International Symposium on the Internet of Sounds (IS2)</em>, L'Aquila, Italy, 2025, pp. 1-10, DOI: <a href="https://ieeexplore.ieee.org/abstract/document/11284671" target="_blank" rel="noopener">10.1109/IS264627.2025.11284671</a>.</div> </div> <div> </div> <div> </div> <div><strong>Related to...</strong></div> <div>J. F. Gemmeke <em>et al</em>. (2017) "Audio Set: An ontology and human-labeled dataset for audio events," <em>2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)</em>, New Orleans, LA, USA, 2017, pp. 776-780, doi: <a href="https://ieeexplore.ieee.org/document/7952261" target="_blank" rel="noopener">10.1109/ICASSP.2017.7952261</a>.</div> <div> </div> <div>Q. Kong <em>et al</em>. (2020) "PANNs: Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition," in <em>IEEE/ACM Transactions on Audio, Speech, and Language Processing</em>, vol. 28, pp. 2880-2894, 2020, DOI: <a href="https://ieeexplore.ieee.org/document/9229505" target="_blank" rel="noopener">10.1109/TASLP.2020.3030497</a>.</div> <h1> </h1> <h1>License & Credits</h1> <div>This dataset is derived from AudioSet, which is licensed under CC BY 4.0. Users must comply with YouTube's Terms of Service and AudioSet's license terms. Individual audio clips remain under their original licenses as specified by content creators.</div> <div> </div> <div>If you use <strong>AudioSet-EV v2</strong> in your research, please cite:</div> <div> </div> <div><code>```bibtex</code></div> <div><code>@dataset{giacomelli2025audiosetev_v2,</code></div> <div><code>author = {Giacomelli, Stefano and Rinaldi, Claudia},</code></div> <div><code>title = {AudioSet-EV v2: a refined AudioSet-derived distribution of Emergency Vehicle Siren sounds},</code></div> <div><code>month = feb,</code></div> <div><code>year = 2025,</code></div> <div><code>publisher = {Zenodo},</code></div> <div><code>version = {v2.0},</code></div> <div><code>doi = {10.5281/zenodo.18668076},</code></div> <div><code>url = {https://zenodo.org/uploads/18668076}</code></div> <div><code>}</code></div> <div><code>```</code></div>
format	Recurso digital
id	zenodo_https___doi_org_10_5281_zenodo_18668076
institution	Zenodo
language	eng
publishDate	2025
publisher	Zenodo
record_format	zenodo
spellingShingle	AudioSet-EV: an AudioSet-derived distribution of Emergency Vehicle Siren sounds Giacomelli, Stefano Rinaldi, Claudia Sound Sound Recordings Deep Learning Deep Learning/statistics & numerical data AudioSet Python PyTorch Datasets as Topic/statistics & numerical data <h1>AudioSet-EV</h1> <h2><br>Overview</h2> <blockquote> <div><strong>AudioSet-EV</strong>, is a case-study and tailored distribution of <a href="https://research.google.com/audioset/index.html" target="_blank" rel="noopener">AudioSet (©Google)</a> (AS) for acoustic Emergency Vehicle (EV) siren detection and recognition. By selectively grouping siren and non-siren urban sounds, enforcing taxonomy consistency, and mitigating class imbalances, AudioSet-EV offers a robust, large-scale resource for research in Machine Learning and Deep Learning acoustic modeling. <strong>AudioSet-EV v2</strong> is a refined and extended distribution of AudioSet-EV specifically derived from the PANNs (Pre-trained Audio Neural Networks) <a href="https://huggingface.co/datasets/confit/audioset-qiuqiangkong" target="_blank" rel="noopener">AudioSet 2020 release</a>, providing a larger amount of data (due to original contents availability). By leveraging selective grouping of Siren and Non-Siren urban sounds, enforcing taxonomy consistency, implementing negatives stratified class balancing, and supporting both binary and 4-way classification (Negative sounds VS Police, Ambulance, Fire Truck Sirens), AudioSet-EV v2 provides the most robust, and largest-scale resource for research in Machine Learning and Deep Learning acoustic EV modeling.</div> </blockquote> <h3><br>Key Improvements over <strong>AudioSet-EV v1</strong> (AudioSet 2025 availability)</h3> <ul> <li><strong>Enhanced Data Quality and Quantity</strong>: built from PANNs AS release with stabilized audio quality and larger files avilability</li> <li><strong>Stratified Negative Sampling</strong>: balanced negative samples across 39 sound categories (traffic, music, speech, alarms, etc.)</li> <li><strong>Extended Coverage</strong>: ~28,000 total samples (7,900 Positives, 20,916 Negatives)</li> <li><strong>Multi-Class Support:</strong> fine-grained annotations for police, ambulance, and fire truck sirens</li> <li><strong>Segment-Type Metadata</strong>: files organized by AudioSet <code>'balanced_train'</code>, <code>'eval'</code>, and <code>'unbalanced'</code> segments</li> <li><strong>Reproducible Processing</strong>: deterministic stratification for consistent experimental setups</li> </ul> <h2> </h2> <h2>Methodology</h2> <div> <p>Our design methodology encompasses a systematic selection and filtering of relevant AS samples, with <a href="https://github.com/StefanoGiacomelli/audioset-tools/tree/main" target="_blank" rel="noopener">AudioSet-Tools</a> and a <em>binary distinction</em> between True Positives (siren-related) and True Negatives (non-siren) samples, mitigating class imbalances and label contamination. We emphasize that, given the <a href="https://research.google.com/audioset/download.html" target="_blank" rel="noopener">original weak labeling nature</a>, total reliability of the label association process cannot be guaranteed.</p> <p>We structured AudioSet-EV into two primary groups:</p> <ol> <li><strong>Positives</strong>: including only EV-siren-related classes, specifically <code>'Police car (siren)', 'Ambulance (siren)', 'Fire engine, fire truck (siren)'</code>, and the ontology containers class <code>'Emergency vehicle', 'Siren'</code>, to account for any weakly labeled or meaningful sound.</li> <li><strong>Negatives</strong>: consisting of a diverse and challenging set, comprising: <ul> <li> <em>vehicle-related sounds</em> (<code>'Car', 'Car passing by', 'Power windows, electric windows', 'Tire squeal', 'Motor vehicle (road)', 'Truck', 'Air brake', 'Ice cream truck, ice cream van', 'Bus', 'Motorcycle', 'Skidding', 'Race car, auto racing', 'Bicycle', 'Train', 'Rail transport', 'Train wheels squealing', 'Railroad car, train wagon', 'Skateboard'</code>)</li> <li><em>alarm signals</em> (<code>'Car alarm', 'Vehicle horn, car horn, honking', 'Bicycle bell', 'Train horn', 'Train whistle', 'Foghorn', 'Toot', 'Reversing beeps', 'Beep, bleep', 'Civil defense siren', 'Alarm', 'Smoke detector, smoke alarm', 'Fire alarm', 'Buzzer'</code>),</li> <li><em>environmental noises</em> (<code>'Traffic noise, roadway noise', 'Outside, rural or natural', 'Outside, urban or manmade'</code>).</li> </ul> </li> </ol> <p>We also included some <code>Speech</code>, <code>Music</code> and <code>Engine</code>-related sounds to improve robustness against waveform pattern similarities and semantic taxonomy proximities.</p> </div> <div> </div> <div> <h2>Pre-Processing</h2> <p>For <em>Positives</em> category, segments processing followed these steps:</p> <ul> <li><em>Selection by Label</em>: <code>balanced</code>, <code>unbalanced</code> and <code>eval</code> AS segments were filtered according to our <strong>Positives</strong> label selection.</li> <li>Segments Differentiation <strong>(New in v2)</strong>: given the original tracking and results availability consistency, samples were differentiated across original segment origin in <code>.csv</code> files, while still belonging to the same Positives group.</li> <li><em>Blacklist Filtering</em>: to refine our selection, any <code>'Civil defense siren'</code> sample was removed to prevent contamination with non-emergency <em>Vehicle</em> sounds.</li> </ul> <p>For the <em>Negatives</em> category, datasets processing followed these steps:</p> <ul> <li><em>Selection by Label</em>: <code>balanced</code>, <code>unbalanced</code> and <code>eval</code> AS entries, matching our defined <strong>non-siren</strong> labels, were extracted.</li> <li>Segments Differentiation <strong>(New in v2)</strong>: same as for Positives.</li> <li><em>Partial Blacklist Filtering</em>: to avoid overlaps with the <strong>Positives</strong> category, samples containing at least one positive class label were removed, except for <code>'Civil defense siren'</code>, which is taxonomically included within the <code>'Siren'</code> container class.</li> <li><em>Stratified Class Re-Balancing</em><strong> (New in v2)</strong>: samples were stratified balanced across 39 ontology categories using random (label-aware) sampling, to preserve class representation while mitigating imbalance. This ensures diversity across alarm types, traffic sounds, music, speech, and environmental noises.</li> </ul> <p>Final <code>.csv</code> files were processed through two independent instances of our AudioSet-Tools downloader, configured to re-sample YouTube audio to 32KHz (without pitch shifting artifacts!), reduce files to mono, and avoid amplitude normalization. We stress the aspect that, given the large amount of <em>Negatives</em>, there actually exist multiple variants of this subset (due to the randomized class down-sampling process). </p> </div> <p> </p> <h2>Comparative POSITIVES Summary Statistics </h2> <table style="width: 100.063%;"> <tbody> <tr> <td style="width: 31.6959%;"> </td> <td style="width: 15.5263%;">Samples</td> <td style="width: 13.1313%;"><code>Police car (siren)</code></td> <td style="width: 11.9949%;"><code>Ambulance (siren)</code></td> <td style="width: 27.6515%;"><code>Fire engine, fire truck (siren)</code></td> </tr> <tr> <td style="width: 31.6959%;">Positives</td> <td style="width: 15.5263%;">8 409 </td> <td style="width: 13.1313%;">3 643 </td> <td style="width: 11.9949%;">1 931</td> <td style="width: 27.6515%;">3 187</td> </tr> <tr> <td style="width: 31.6959%;"><code>Downloaded</code> AudioSet v1</td> <td style="width: 15.5263%;">7 324</td> <td style="width: 13.1313%;">3 124</td> <td style="width: 11.9949%;">1 637</td> <td style="width: 27.6515%;">2 852</td> </tr> <tr> <td style="width: 31.6959%;"><code><strong>Downloaded</strong></code><strong> AudioSet v2</strong></td> <td style="width: 15.5263%;"><strong>7 900</strong></td> <td style="width: 13.1313%;"><strong>3 219</strong></td> <td style="width: 11.9949%;"><strong>1 732</strong></td> <td style="width: 27.6515%;"><strong>2 947</strong></td> </tr> </tbody> </table> <h2>NEW: AudioSet-EV v2 Multi-Label Awareness</h2> <table style="width: 100.063%;"> <tbody> <tr> <td style="width: 41.1616%;"> </td> <td style="width: 13.7626%;"><strong>Pure</strong> Samples</td> <td style="width: 21.8506%;"><strong>Multi-Positive</strong> Samples</td> <td style="width: 23.0989%;">Total Samples</td> </tr> <tr> <td style="width: 41.1616%;"><code>Police car (siren)</code></td> <td style="width: 13.7626%;">2 643</td> <td style="width: 21.8506%;">912</td> <td style="width: 23.0989%;"><strong>3 124</strong></td> </tr> <tr> <td style="width: 41.1616%;"><code>Ambulance (siren)</code></td> <td style="width: 13.7626%;">1 020</td> <td style="width: 21.8506%;">617</td> <td style="width: 23.0989%;"><strong>1 637</strong></td> </tr> <tr> <td style="width: 41.1616%;"><code>Fire engine, fire truck (siren)</code></td> <td style="width: 13.7626%;">2 713</td> <td style="width: 21.8506%;">139</td> <td style="width: 23.0989%;"><strong>2 852</strong></td> </tr> </tbody> </table> <h2>NEW: AudioSet-EV v2 Segment Awareness</h2> <table style="width: 100.063%; height: 140.657px;"> <tbody> <tr style="height: 20.0938px;"> <td style="width: 21.5599%; height: 20.0938px;"><strong>POSTIVES</strong></td> <td style="width: 30.2141%; height: 20.0938px;"><code>Downloaded</code><strong> </strong>Samples</td> <td style="width: 27.3636%; height: 20.0938px;">Total Samples</td> <td style="width: 21.3063%; height: 20.0938px;">Percentage (%)</td> </tr> <tr style="height: 20.0938px;"> <td style="width: 21.5599%; height: 20.0938px;"><code>balanced_train</code></td> <td style="width: 30.2141%; height: 20.0938px;"><strong>135</strong></td> <td style="width: 27.3636%; height: 20.0938px;">146</td> <td style="width: 21.3063%; height: 20.0938px;">92.5%</td> </tr> <tr style="height: 20.0938px;"> <td style="width: 21.5599%; height: 20.0938px;"><code>eval</code></td> <td style="width: 30.2141%; height: 20.0938px;"><strong>135 </strong></td> <td style="width: 27.3636%; height: 20.0938px;">148 </td> <td style="width: 21.3063%; height: 20.0938px;">91.2%</td> </tr> <tr style="height: 20.0938px;"> <td style="width: 21.5599%; height: 20.0938px;"><code>unbalanced</code></td> <td style="width: 30.2141%; height: 20.0938px;"><strong>7 630</strong></td> <td style="width: 27.3636%; height: 20.0938px;">8 154</td> <td style="width: 21.3063%; height: 20.0938px;">93.6%</td> </tr> <tr style="height: 20.0938px;"> <td style="width: 21.5599%; height: 20.0938px;"><strong>NEGATIVES</strong></td> <td style="width: 30.2141%; height: 20.0938px;"><code>Downloaded</code><strong> </strong>Samples</td> <td style="width: 27.3636%; height: 20.0938px;">Total Samples</td> <td style="width: 21.3063%; height: 20.0938px;"><strong> </strong></td> </tr> <tr style="height: 20.0938px;"> <td style="width: 21.5599%; height: 20.0938px;"><code>balanced_train</code></td> <td style="width: 30.2141%; height: 20.0938px;"><strong>10 963</strong></td> <td style="width: 27.3636%; height: 20.0938px;">11 871</td> <td style="width: 21.3063%; height: 20.0938px;">92.4%</td> </tr> <tr style="height: 20.0938px;"> <td style="width: 21.5599%; height: 20.0938px;"><code>eval</code></td> <td style="width: 30.2141%; height: 20.0938px;"><strong>9 953</strong></td> <td style="width: 27.3636%; height: 20.0938px;">10 808</td> <td style="width: 21.3063%; height: 20.0938px;">92.1%</td> </tr> </tbody> </table> <p><strong>Note</strong>: The <code>'unbalanced'</code> segment is <strong>not included</strong> in the Negative set to avoid excessive class imbalancing.</p> <h2>Comparative Summary Statistics</h2> <table style="width: 100.063%; height: 59.2814px;"> <tbody> <tr style="height: 19.5938px;"> <td style="width: 29.7983%; height: 19.5938px;"> </td> <td style="width: 35.6064%; height: 19.5938px;">AudioSet-EV v1</td> <td style="width: 34.5963%; height: 19.5938px;">AudioSet-EV v2</td> </tr> <tr style="height: 19.5938px;"> <td style="width: 29.7983%; height: 19.5938px;">Positives</td> <td style="width: 35.6064%; height: 19.5938px;">7 324 </td> <td style="width: 34.5963%; height: 19.5938px;">7 900<strong> (+7.9%)</strong></td> </tr> <tr style="height: 20.0938px;"> <td style="width: 29.7983%; height: 20.0938px;">Negatives</td> <td style="width: 35.6064%; height: 20.0938px;">6 702</td> <td style="width: 34.5963%; height: 20.0938px;">20 916 <strong>(+212%)</strong></td> </tr> <tr> <td style="width: 29.7983%;"><strong>Samples Difference</strong></td> <td style="width: 35.6064%;"><strong>+576</strong></td> <td style="width: 34.5963%;"><strong>+12 214 (+105%)</strong></td> </tr> </tbody> </table> <div> <div> <div> </div> </div> </div> <div> </div> <div> </div> <div> <h2>File Structure</h2> <p>Shared by both Version (backward compatibility between versions):</p> </div> <div> <div><code>AudioSet-EV_vX_main.zip (XX GB)</code></div> <div><code>        ├── Positive_files/</code></div> <div><code>        │    ├── balanced_train/          # AudioSet-EV v2 differentiation</code></div> <div><code>        │    ├── eval/</code></div> <div><code>        │    └── unbalanced/</code></div> <div><code>        ├── Negative_files/</code></div> <div><code>        │    ├── balanced_train/          # AudioSet-EV v2 differentiation</code></div> <div><code>        │    └── eval/</code></div> <div><code>        ├── EV_Positives.csv              # Metadata with MIDs and labels (vX compatibility)</code></div> <div><code>        └── EV_Negatives.csv</code></div> </div> <h2>Info</h2> <p>Each <code>.csv</code> file contains:</p> <ul> <li>- <code>'yt_id'</code>: YouTube video identifier</li> <li>- <code>'positive_labels'</code>: List of AudioSet MIDs (<em>for positives only</em>)</li> <li>- <code>'segment_type'</code>: AudioSet segment (balanced_train, eval, unbalanced_train)</li> <li>- <code>'downloaded'</code>: Boolean flag indicating successful download (w.r.t. original 2017 AudioSet metadata release)</li> </ul> <div><strong>Use cases</strong></div> <ul> <li><strong>Binary Classification</strong>: Emergency Vehicle siren recognition (Presence VS Absence)</li> <li><strong>Multi-Class Classification</strong>: Fine-grained siren type recognition (Police VS Ambulance VS Firemans)</li> <li><strong>Transfer Learning</strong>: Pre-training on AudioSet-EV v2 for downstream Events detection tasks</li> <li><strong>Robust</strong> <strong>Testing</strong>: Evaluation against diverse confounding sounds (alarms, urban traffic, music)</li> <li><strong>Cross-Dataset Benchmarking:</strong> Comparison with sireNNet, LSSiren, FSD50K ESC-50 andUrbanSound8K pre-processed datasets.</li> </ul> <div><strong>Technical Notes</strong></div> <ul> <li> <div><strong>Audio Format</strong>: <code>.wav</code> (mono, 32 kHz, ~10 seconds)</div> </li> <li> <div><strong>Total Size</strong>: ~16 GB (.zip compressed), ~28 GB (uncompressed)</div> </li> <li> <div><strong>Label Reliability</strong>: <em>Weak labels</em> inherited from AudioSet (manual verification recommended for <em>critical</em> applications)</div> </li> <li> <div><strong>Reproducibility</strong>: Seed-controlled stratified sampling and groups selection ensures deterministic dataset splits</div> </li> </ul> <h1> </h1> <h1>References</h1> <div><strong>AudioSet-Tools Framework</strong></div> <div>S. Giacomelli <em>et al</em>.(2026) "AudioSet-tools: a Python framework for taxonomy-aware AudioSet curation and reproducible audio research" in EURASIP <em>Journal of Audio, Speech and Music Processing.</em> 2026, 2. DOI: <a href="https://doi.org/10.1186/s13636-025-00436-z" target="_blank" rel="noopener">10.1186/s13636-025-00436-z</a></div> <div> </div> <div> </div> <div><strong>E2PANNs Framework</strong></div> <div> <div>S. Giacomelli <em>et al</em>. (2025) "From Large-scale Audio Tagging to Real-Time Explainable Emergency Vehicle Sirens Detection", <strong><em>Under Review for IEEE TASLP</em></strong>, arXiv preprint DOI: <a href="https://arxiv.org/abs/2506.23437" target="_blank" rel="noopener">arXiv:2506.23437</a>, <strong>GitHub</strong>: <a href="https://github.com/StefanoGiacomelli/e2panns" target="_blank" rel="noopener">https://github.com/StefanoGiacomelli/e2panns</a></div> <div> </div> <div>M. Giordano <em>et al</em>. (2025) "Real-Time Emergency Vehicle Siren Detection with Efficient CNNs on Embedded Hardware," <em>2025 IEEE 6th International Symposium on the Internet of Sounds (IS2)</em>, L'Aquila, Italy, 2025, pp. 1-10, DOI: <a href="https://ieeexplore.ieee.org/abstract/document/11284671" target="_blank" rel="noopener">10.1109/IS264627.2025.11284671</a>.</div> </div> <div> </div> <div> </div> <div><strong>Related to...</strong></div> <div>J. F. Gemmeke <em>et al</em>. (2017) "Audio Set: An ontology and human-labeled dataset for audio events," <em>2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)</em>, New Orleans, LA, USA, 2017, pp. 776-780, doi: <a href="https://ieeexplore.ieee.org/document/7952261" target="_blank" rel="noopener">10.1109/ICASSP.2017.7952261</a>.</div> <div> </div> <div>Q. Kong <em>et al</em>. (2020) "PANNs: Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition," in <em>IEEE/ACM Transactions on Audio, Speech, and Language Processing</em>, vol. 28, pp. 2880-2894, 2020, DOI: <a href="https://ieeexplore.ieee.org/document/9229505" target="_blank" rel="noopener">10.1109/TASLP.2020.3030497</a>.</div> <h1> </h1> <h1>License & Credits</h1> <div>This dataset is derived from AudioSet, which is licensed under CC BY 4.0. Users must comply with YouTube's Terms of Service and AudioSet's license terms. Individual audio clips remain under their original licenses as specified by content creators.</div> <div> </div> <div>If you use <strong>AudioSet-EV v2</strong> in your research, please cite:</div> <div> </div> <div><code>```bibtex</code></div> <div><code>@dataset{giacomelli2025audiosetev_v2,</code></div> <div><code>author = {Giacomelli, Stefano and Rinaldi, Claudia},</code></div> <div><code>title = {AudioSet-EV v2: a refined AudioSet-derived distribution of Emergency Vehicle Siren sounds},</code></div> <div><code>month = feb,</code></div> <div><code>year = 2025,</code></div> <div><code>publisher = {Zenodo},</code></div> <div><code>version = {v2.0},</code></div> <div><code>doi = {10.5281/zenodo.18668076},</code></div> <div><code>url = {https://zenodo.org/uploads/18668076}</code></div> <div><code>}</code></div> <div><code>```</code></div>
title	AudioSet-EV: an AudioSet-derived distribution of Emergency Vehicle Siren sounds
topic	Sound Sound Recordings Deep Learning Deep Learning/statistics & numerical data AudioSet Python PyTorch Datasets as Topic/statistics & numerical data
url	https://doi.org/10.5281/zenodo.18668076