Datasets:
File size: 2,788 Bytes
154b008 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 |
# VocalSound Dataset
## Overview
VocalSound is a crowdsourced dataset of human non-speech vocal sounds for audio classification tasks. It was created to support research on building robust and accurate vocal sound recognition systems, addressing limitations in existing datasets that have relatively small numbers of samples or noisy labels.
**Source:** https://github.com/YuanGongND/vocalsound
**Paper:** ICASSP 2022 - IEEE International Conference on Acoustics, Speech and Signal Processing
**License:** Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0)
## Dataset Statistics
| Metric | Value |
|--------|-------|
| Total Samples | 21,024 |
| Unique Speakers | 3,365 |
| Countries Represented | 60 |
| Classes | 6 |
### Data Splits
| Split | Samples | Percentage |
|-------|---------|------------|
| Training | 15,570 | 74% |
| Validation | 1,860 | 9% |
| Evaluation | 3,594 | 17% |
All three sets are speaker-independent. The evaluation set has been manually verified for quality.
### Classes (Balanced)
Each class contains 3,504 samples:
1. Laughter
2. Sigh
3. Cough
4. Throat clearing
5. Sneeze
6. Sniff
## Speaker Demographics
### Gender
- Male: 55%
- Female: 45%
### Age Distribution
- Range: 18-80 years
- Majority: 20-40 years
- Subjects over 50: 321
### Geographic Distribution
- United States: 60.3%
- India: 10.8%
- Brazil: 8.3%
- Other countries: 20.6%
### Native Languages
- English: 67.2%
- Portuguese: 8.7%
- Italian: 6.8%
- Other languages: 17.3%
### Health Status
- 4% of subjects reported symptoms affecting speech
## Audio Specifications
### Original Format
- Sample Rate: 44.1 kHz
- Format: WAV
- Mean Duration: 4.18 seconds
- Median Duration: 3.72 seconds
- Standard Deviation: 1.81 seconds
### WebDataset Format (This Collection)
- Sample Rate: 48 kHz
- Bit Depth: 16-bit
- Channels: Mono
- Format: FLAC
- Total Tar Files: 22
## Baseline Performance
Using an EfficientNet-B0 based classifier:
- Overall Accuracy: 90.5%
- Male Speakers: 89.2%
- Female Speakers: 91.9%
Adding VocalSound to existing training data improves vocal sound recognition performance by 41.9%.
## Authors
- Yuan Gong (MIT CSAIL)
- Jin Yu (Signify)
- James Glass (MIT CSAIL)
## Citation
```bibtex
@INPROCEEDINGS{gong_vocalsound,
author={Gong, Yuan and Yu, Jin and Glass, James},
booktitle={ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
title={Vocalsound: A Dataset for Improving Human Vocal Sounds Recognition},
year={2022},
pages={151-155},
doi={10.1109/ICASSP43922.2022.9746828}
}
```
## References
- GitHub Repository: https://github.com/YuanGongND/vocalsound
- arXiv Paper: https://arxiv.org/abs/2205.03433
- IEEE Xplore: https://ieeexplore.ieee.org/document/9746828
|