Whisper

"Robust Speech Recognition via Large-Scale Weak Supervision" — an encoder-decoder Transformer trained on 680,000 hours of multilingual audio data collected from the web. The largest variant (large-v3) has 1.55B parameters. Supports transcription and translation across 99 languages.

Whisper demonstrated that scaling weakly supervised pre-training on diverse audio data approaches human-level robustness without domain-specific fine-tuning. Became the most widely-used open-source speech recognition model, integrated into thousands of applications. ICML 2023. By Radford, Kim, Xu et al. MIT License.