Whisper
An MIT-licensed speech recognition model for local transcription, multilingual audio and translation into English, with several model sizes.
Whisper is an open source speech recognition model for transcribing recordings on your own hardware. It's for developers building voice features and people working with multilingual audio.
- Speech transcription: Convert speech into text in its original language, with English-only models available for English recordings.
- English translation: Translate non-English speech into English using a multilingual model trained for translation.
- Language and timing: Identify the spoken language and produce phrase-level timestamps alongside the transcript.
- Model choice: Smaller models favor speed and lower memory use, while larger models offer a different accuracy tradeoff; turbo favors faster transcription.
OpenAI publishes the models and inference code under MIT. The Python implementation uses PyTorch and can process audio locally, so recordings can stay on the machine running inference.
Approximate GPU memory requirements range from 1 GB for tiny and base to 10 GB for large; turbo needs about 6 GB. Actual speed depends on hardware, language and the recording. The tiny.en and base.en models tend to outperform their multilingual counterparts on English.
Accuracy varies substantially by language. Whisper's training includes accents, background noise and technical language, but turbo isn't trained for translation.
Logotyper
