Robust Speech Recognition via Large-Scale Weak Supervision

This work studies the capabilities of speech processing systems trained simply to predict large amounts of transcripts of audio found on the internet. By scaling this weakly supervised approach to 680,000 hours of multilingual and multitask supervision, the authors train models that learn robust speech representations directly from naturally occurring transcript data rather than from curated, task-specific labels.

The resulting models generalize well to standard benchmarks and are often competitive with prior fully supervised results in a zero-shot transfer setting, without needing any fine-tuning, and they approach human accuracy and robustness. The authors release their models and inference code to serve as a foundation for further work on robust speech processing.