According to a report on arXiv cs.CL, researchers have released YODAS v3, described as a weakly-labeled speech corpus containing over 1.1 million hours of 48kHz multi-channel audio spanning 147 languages. The release is published under a CC BY 3.0 license and features high-fidelity stereo audio alongside language-balanced collection methodologies that yielded over 10,000 hours across 22 languages and 73 languages total.
Scale and Open Science Significance
Large open datasets remain critical for training modern automatic speech recognition and audio foundation models. By providing high-bandwidth, stereophonic, and multilingual data at this scale, corpora like YODAS v3 lower barriers for researchers studying acoustic modeling across diverse linguistic landscapes without relying on proprietary datasets.
Unknowns and Future Tracking
The announcement did not specify the exact hardware infrastructure used to crawl and process the dataset, nor did it detail the specific filtering mechanisms applied to the weak labels. Observers will need to monitor how downstream machine learning models perform when trained on this specific corpus compared to existing audio datasets.
Further technical documentation and benchmarks will determine how widely the dataset is adopted by the broader speech research community.
Lead image: Lukas Blazek / Pexels
