Whisper: multilingual speech recognition

An encoder-decoder speech model that transcribes many languages, translates speech into English and detects the spoken language, all from one network.

• Input audio is converted to log-Mel spectrograms
• Special decoder tokens pick the task: transcribe, translate or identify the language
• English-only .en variants exist for the smaller sizes
• turbo is a faster large-v3 variant, not trained for translation

Summary written by SolxLab. Weights and code belong to their authors and are distributed under the license shown; always check the official page before use.