Source: Robust Speech Recognition via Large-Scale Weak Supervision (Whisper)

Radford et al. 2022 paper introducing Whisper, a speech recognizer trained on ~680k hours of weakly supervised multilingual audio that generalizes zero-shot.

Contents
  • What: research paper (arXiv:2212.04356)
  • Author/publisher: Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, Ilya Sutskever (OpenAI)
  • Date: 2022-12-06
  • Original: arXiv:2212.04356

Summary

Shows that scaling weakly supervised training (internet audio paired with transcripts) to about 680,000 hours of multilingual, multitask data yields a speech model that transfers to standard benchmarks without fine-tuning.

Key claims

  • Training data: ~680,000 hours of multilingual and multitask (transcription + translation) audio.
  • Zero-shot performance is often competitive with prior fully supervised systems.
  • The models approach human accuracy and robustness.

Feeds

Multimodal models · Voice agents

Suggest changes →