The ESC-50 dataset is a labeled collection of 2000 environmental audio recordings suitable for benchmarking methods of environmental sound classification.
The dataset consists of 5-second-long recordings organized into 50 semantical classes (with 40 examples per class) loosely arranged into 5 major categories:
| Animals | Natural soundscapes & water sounds | Human, non-speech sounds | Interior/domestic sounds | Exterior/urban noises |
|---|---|---|---|---|
| Dog | Rain | Crying baby | Door knock | Helicopter |
| Rooster | Sea waves | Sneezing | Mouse click | Chainsaw |
| Pig | Crackling fire | Clapping | Keyboard typing | Siren |
| Cow | Crickets | Breathing | Door, wood creaks | Car horn |
| Frog | Chirping birds | Coughing | Can opening | Engine |
| Cat | Water drops | Footsteps | Washing machine | Train |
| Hen | Wind | Laughing | Vacuum cleaner | Church bells |
| Insects (flying) | Pouring water | Brushing teeth | Clock alarm | Airplane |
| Sheep | Toilet flush | Snoring | Clock tick | Fireworks |
| Crow | Thunderstorm | Drinking, sipping | Glass breaking | Hand saw |
Clips in this dataset have been manually extracted from public field recordings gathered by the Freesound.org project. The dataset has been prearranged into 5 folds for comparable cross-validation, making sure that fragments from the same original source file are contained in a single fold.
A more thorough description of the dataset is available in the original paper with some supplementary materials on GitHub: ESC: Dataset for Environmental Sound Classification - paper replication data.
The dataset can be downloaded as a single .zip file (~600 MB):
Numerous machine learning & signal processing approaches have been evaluated on the ESC-50 dataset. Most of them are listed here. If you know of some other reference, you can message me or open a Pull Request directly.
• CNN - Convolutional Neural Network
• CRNN - Convolutional Recurrent Neural Network
• GMM - Gaussian Mixture Model
• GTCC - Gammatone Cepstral Coefficients
• GTSC - Gammatone Spectral Coefficients
• k-NN - k-Neareast Neighbors
• MFCC - Mel-Frequency Cepstral Coefficients
• MLP - Multi-Layer Perceptron
• RBM - Restricted Boltzmann Machine
• RNN - Recurrent Neural Network
• SVM - Support Vector Machine
• TEO - Teager Energy Operator
• ZCR - Zero-Crossing Rate
| Title | Notes | Accuracy | Paper | Code |
|---|---|---|---|---|
| Unsupervised Filterbank Learning Using Convolutional Restricted Boltzmann Machine for Environmental Sound Classification | CNN with filterbanks learned using convolutional RBM + fusion with GTSC and mel energies | 86.50% | sailor2017 | |
| Learning from Between-class Examples for Deep Sound Recognition | EnvNet-v2 (tokozume2017a) + data augmentation + Between-Class learning | 84.90% | tokozume2017b | |
| Novel Phase Encoded Mel Filterbank Energies for Environmental Sound Classification | CNN working with phase encoded mel filterbank energies (PEFBEs), fusion with Mel energies | 84.15% | tak2017 | |
| Knowledge Transfer from Weakly Labeled Audio using Convolutional Neural Network for Sound Events and Scenes | CNN pretrained on AudioSet | 83.50% | kumar2017 | |
| Unsupervised Filterbank Learning Using Convolutional Restricted Boltzmann Machine for Environmental Sound Classification | CNN with filterbanks learned using convolutional RBM + fusion with GTSC | 83.00% | sailor2017 | |
| Novel TEO-based Gammatone Features for Environmental Sound Classification | Fusion of GTSC & TEO-GTSC with CNN | 81.95% | agrawal2017 | |
| Learning from Between-class Examples for Deep Sound Recognition | EnvNet-v2 (tokozume2017a) + Between-Class learning | 81.80% | tokozume2017b | |
| 🎧 Human accuracy | Crowdsourcing experiment in classifying ESC-50 by human listeners | 81.30% | piczak2015a | 📜 |
| Look, Listen and Learn | 8-layer convolutional subnetwork pretrained on an audio-visual correspondence task | 79.30% | arandjelovic2017 | |
| Novel TEO-based Gammatone Features for Environmental Sound Classification | GTSC with CNN | 79.10% | agrawal2017 | |
| Learning from Between-class Examples for Deep Sound Recognition | EnvNet-v2 (tokozume2017a) + data augmentation | 78.80% | tokozume2017b | |
| Unsupervised Filterbank Learning Using Convolutional Restricted Boltzmann Machine for Environmental Sound Classification | CNN with filterbanks learned using convolutional RBM | 78.45% | sailor2017 | |
| Learning from Between-class Examples for Deep Sound Recognition | Baseline CNN (piczak2015b) + Batch Normalization + Between-Class learning | 76.90% | tokozume2017b | |
| Novel TEO-based Gammatone Features for Environmental Sound Classification | TEO-GTSC with CNN | 74.85% | agrawal2017 | |
| Learning from Between-class Examples for Deep Sound Recognition | EnvNet-v2 (tokozume2017a) | 74.40% | tokozume2017b | |
| Soundnet: Learning sound representations from unlabeled video | 8-layer CNN (raw audio) with transfer learning from unlabeled videos | 74.20% | aytar2016 | 📜 |
| Learning from Between-class Examples for Deep Sound Recognition | 18-layer CNN on raw waveforms (dai2016) + Between-Class learning | 73.30% | tokozume2017b | |
| Novel Phase Encoded Mel Filterbank Energies for Environmental Sound Classification | CNN working with phase encoded mel filterbank energies (PEFBEs) | 73.25% | tak2017 | |
| Classifying environmental sounds using image recognition networks | 16 kHz sampling rate, GoogLeNet on spectrograms (40 ms frame length) | 73.20% | boddapati2017 | 📜 |
| Learning from Between-class Examples for Deep Sound Recognition | Baseline CNN (piczak2015b) + Batch Normalization | 72.40% | tokozume2017b | |
| Novel TEO-based Gammatone Features for Environmental Sound Classification | Fusion of MFCC & TEO-GTCC with GMM | 72.25% | agrawal2017 | |
| Learning environmental sounds with end-to-end convolutional neural network (EnvNet) | Combination of spectrogram and raw waveform CNN | 71.00% | tokozume2017a | |
| Novel TEO-based Gammatone Features for Environmental Sound Classification | TEO-GTCC with GMM | 68.85% | agrawal2017 | |
| Classifying environmental sounds using image recognition networks | 16 kHz sampling rate, AlexNet on spectrograms (30 ms frame length) | 68.70% | boddapati2017 | 📜 |
| Very Deep Convolutional Neural Networks for Raw Waveforms | 18-layer CNN on raw waveforms | 68.50% | dai2016, tokozume2017b | 📜 |
| Classifying environmental sounds using image recognition networks | 32 kHz sampling rate, GoogLeNet on spectrograms (30 ms frame length) | 67.80% | boddapati2017 | 📜 |
| WSNet: Learning Compact and Efficient Networks with Weight Sampling | SoundNet 8-layer CNN architecture with 100x model compression | 66.25% | jin2017 | |
| Soundnet: Learning sound representations from unlabeled video | 5-layer CNN (raw audio) with transfer learning from unlabeled videos | 66.10% | aytar2016 | 📜 |
| WSNet: Learning Compact and Efficient Networks with Weight Sampling | SoundNet 8-layer CNN architecture with 180x model compression | 65.80% | jin2017 | |
| Soundnet: Learning sound representations from unlabeled video | 5-layer CNN trained on raw audio of ESC-50 only | 65.00% | aytar2016 | 📜 |
| 📊 Environmental Sound Classification with Convolutional Neural Networks - CNN baseline | CNN with 2 convolutional and 2 fully-connected layers, mel-spectrograms as input, vertical filters in the first layer | 64.50% | piczak2015b | 📜 |
| auDeep: Unsupervised Learning of Representations from Audio with Deep Recurrent Neural Networks | MLP classifier on features extracted with an RNN autoencoder | 64.30% | freitag2017 | 📜 |
| Classifying environmental sounds using image recognition networks | 32 kHz sampling rate, AlexNet on spectrograms (30 ms frame length) | 63.20% | boddapati2017 | 📜 |
| Classifying environmental sounds using image recognition networks | CRNN | 60.30% | boddapati2017 | 📜 |
| Comparison of Time-Frequency Representations for Environmental Sound Classification using Convolutional Neural Networks | 3-layer CNN with vertical filters on wideband mel-STFT (median accuracy) | 56.37% | huzaifah2017 | |
| Comparison of Time-Frequency Representations for Environmental Sound Classification using Convolutional Neural Networks | 3-layer CNN with square filters on wideband mel-STFT (median accuracy) | 54.00% | huzaifah2017 | |
| Soundnet: Learning sound representations from unlabeled video | 8-layer CNN trained on raw audio of ESC-50 only | 51.10% | aytar2016 | 📜 |
| Comparison of Time-Frequency Representations for Environmental Sound Classification using Convolutional Neural Networks | 5-layer CNN with square filters on wideband mel-STFT (median accuracy) | 50.87% | huzaifah2017 | |
| Comparison of Time-Frequency Representations for Environmental Sound Classification using Convolutional Neural Networks | 5-layer CNN with vertical filters on wideband mel-STFT (median accuracy) | 46.25% | huzaifah2017 | |
| 📊 Baseline - random forest | Baseline ML approach (MFCC & ZCR + random forest) | 44.30% | piczak2015a | 📜 |
| Soundnet: Learning sound representations from unlabeled video | Convolutional autoencoder trained on unlabeled videos | 39.90% | aytar2016 | 📜 |
| 📊 Baseline - SVM | Baseline ML approach (MFCC & ZCR + SVM) | 39.60% | piczak2015a | 📜 |
| 📊 Baseline - k-NN | Baseline ML approach (MFCC & ZCR + k-NN) | 32.20% | piczak2015a | 📜 |
| A mixture model-based real-time audio sources classification method | Dictionary of sound models used for classification (accuracy is computed on segments instead of files) | 94.00% | baelde2017 | |
| Utilizing Domain Knowledge in End-to-End Audio Processing | End-to-end CNN with learned mel-spectrogram transformation | N/A | tax2017 | 📜 |
| Deep Neural Network based learning and transferring mid-level audio features for acoustic scene classification | Transfer learning from various datasets, including ESC-50 | N/A | mun2017 | |
| Features and Kernels for Audio Event Recognition | MFCC, GMM, SVM | N/A | kumar2016b | |
| A real-time environmental sound recognition system for the Android OS | Real-time sound recognition for Android evaluated on ESC-10 | N/A | pillos2016 | |
| Comparing Time and Frequency Domain for Audio Event Recognition Using Deep Learning | Discriminatory effectiveness of different signal representations compared on ESC-10 and Freiburg-106 | N/A | hertel2016 | |
| Audio Event and Scene Recognition: A Unified Approach using Strongly and Weakly Labeled Data | Combination of weakly labeled data (YouTube) with strong labeling (ESC-10) for Acoustic Event Detection | N/A | kumar2016a |
-
2000 audio recordings in WAV format (5 seconds, 44.1 kHz, mono) with the following naming convention:
{FOLD}-{CLIP_ID}-{TAKE}-{TARGET}.wav{FOLD}- index of the cross-validation fold,{CLIP_ID}- ID of the original Freesound clip,{TAKE}- letter disambiguating between different fragments from the same Freesound clip,{TARGET}- class in numeric format [0, 49].
-
CSV file with the following structure:
filename fold target category esc10 src_file take The
esc10column indicates if a given file belongs to the ESC-10 subset (10 selected classes, CC BY license). -
Additional data pertaining to the crowdsourcing experiment (human classification accuracy).
The dataset is available under the terms of the Creative Commons Attribution Non-Commercial license.
A smaller subset (clips tagged as ESC-10) is distributed under CC BY (Attribution).
Attributions for each clip are available in the LICENSE file.
If you find this dataset useful in an academic setting please cite:
K. J. Piczak. ESC: Dataset for Environmental Sound Classification. Proceedings of the 23rd Annual ACM Conference on Multimedia, Brisbane, Australia, 2015.
@inproceedings{piczak2015dataset,
title = {{ESC}: {Dataset} for {Environmental Sound Classification}},
author = {Piczak, Karol J.},
booktitle = {Proceedings of the 23rd {Annual ACM Conference} on {Multimedia}},
date = {2015-10-13},
url = {http://dl.acm.org/citation.cfm?doid=2733373.2806390},
doi = {10.1145/2733373.2806390},
location = {{Brisbane, Australia}},
isbn = {978-1-4503-3459-4},
publisher = {{ACM Press}},
pages = {1015--1018}
}
Please be aware of potential information leakage while training models on ESC-50, as some of the original Freesound recordings were already preprocessed in a manner that might be class dependent (mostly bandlimiting). Unfortunately, this issue went unnoticed when creating the original version of the dataset. Due to the number of methods already evaluated on ESC-50, no changes rectifying this issue will be made in order to preserve comparability.
• Change to WAV version as default.
• Replace OGG recordings with cropped WAV files for easier loading and frame-level precision (some of the OGG recordings had a slightly different length when loaded).
• Move recordings to a one directory structure with a meta CSV file.
• Initial version of the dataset (OGG format).