[Prepublication] Abstract: In this paper we introduce a new framework and approach for convolutional neural network computation. By introducing layer functions that intelligently process complex-domain data in deep neural network architectures, we improve upon prior understanding and performance in complex-valued convolutional neural networks. Using novel derivations of convolutional, down-sampling, non-linear, and affine layers implemented in a complex-valued counterpart to Caffe, we proved results when evaluated against real-valued models in the context of environmental sound classification. Finally, we demonstrated these results are applicable to many other fields, including, but not limited to MRI imaging, signal processing, and LiDAR mapping.
Questions? Contact edmunds@ml.berkeley.edu. More information on my website.
Credits: the real-valued cNN model is based on Karol J. Piczak's paper, Environmental Sound Classification with Convolutional Neural Networks.
MIT License
Copyright (c) 2016-2017 Riley F. Edmunds
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
Download | A peek inside | License
The ESC-50 dataset is a public labeled set of 2000 environmental recordings (50 classes, 40 clips per class, 5 seconds per clip) suitable for environmental sound classification tasks.
See ESC: Dataset for Environmental Sound Classification - paper replication data for the full paper with a more thorough analysis.
The dataset consists of 50 classes of recordings in 5 loosely defined groups:
Animals
- 101 - Dog
- 102 - Rooster
- 103 - Pig
- 104 - Cow
- 105 - Frog
- 106 - Cat
- 107 - Hen
- 108 - Insects (flying)
- 109 - Sheep
- 110 - Crow
Natural soundscapes & water sounds
- 201 - Rain
- 202 - Sea waves
- 203 - Crackling fire
- 204 - Crickets
- 205 - Chirping birds
- 206 - Water drops
- 207 - Wind
- 208 - Pouring water
- 209 - Toilet flush
- 210 - Thunderstorm
Human, non-speech sounds
- 301 - Crying baby
- 302 - Sneezing
- 303 - Clapping
- 304 - Breathing
- 305 - Coughing
- 306 - Footsteps
- 307 - Laughing
- 308 - Brushing teeth
- 309 - Snoring
- 310 - Drinking - sipping
Interior/domestic sounds
- 401 - Door knock
- 402 - Mouse click
- 403 - Keyboard typing
- 404 - Door - wood creaks
- 405 - Can opening
- 406 - Washing machine
- 407 - Vacuum cleaner
- 408 - Clock alarm
- 409 - Clock tick
- 410 - Glass breaking
Exterior/urban noises
- 501 - Helicopter
- 502 - Chainsaw
- 503 - Siren
- 504 - Car horn
- 505 - Engine
- 506 - Train
- 507 - Church bells
- 508 - Airplane
- 509 - Fireworks
- 510 - Hand saw
Clips have been constructed from public field recordings gathered by the Freesound.org project. The dataset has been prearranged into 5 folds. Clips stemming from the same original source file are contained in a single fold.
File naming scheme:
category_id - category_name/fold_number-Freesound_clip_ID-take_letter.ogg
File details:
5-second-long recordings reconverted to a unified format:
- 44100 Hz,
- single channel (monophonic),
- Vorbis/Ogg compression @ 192 kbit/s.
The dataset can be downloaded as a single .zip file (~200 MB):
Waveforms and mel-spectrograms of ESC-50 dataset recordings:
The dataset is available under the terms of the Creative Commons license - Attribution-NonCommercial.
A smaller subset (ESC-10) is available under CC BY (Attribution).
In academic settings please cite:
K. J. Piczak. ESC: Dataset for Environmental Sound Classification. In Proceedings of the ACM International Conference on Multimedia, in press, ACM, 2015.
Due to GitHub limitations (README length limit) licensing details for individual clips are available in the plain text README.