Reproduce “YODAS: Youtube-Oriented Dataset for Audio and Speech” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xinjian Li, Shinnosuke Takamichi, Takaaki Saeki, William Chen, Sayaka Shiota, Shinji Watanabe
Carnegie Mellon University researchers introduce YODAS (YouTube-Oriented Dataset for Audio and Speech), a publicly available, large-scale multilingual speech dataset comprising over 560,000 hours of audio across 140 languages, addressing the data scale disparity between academic and industrial applications. Baseline experiments demonstrate strong performance on the manually transcribed subset and highlight the impact of subtitle quality and data alignment on speech recognition metrics.
Key insights
- Manual subtitles consistently yield better speech recognition performance due to higher quality and more accurate speech-text alignment compared to YouTube-generated automatic subtitles.
- Carefully selecting an alignment score threshold is essential; while stricter filtering removes noise, an overly strict threshold can inadvertently reduce data diversity and overall model performance.
- Increasing training data from 10,000 to 1 million utterances improves model performance, but benefits can plateau with simpler model architectures, suggesting that more complex models are needed to fully exploit extremely large datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.