Whisper large-v3-turbo

Whisper large-v3-turbo is a speech recognition model with 809 million parameters. It is a pruned version of Whisper large-v3: the number of decoding layers is cut from 32 to 4, so it is much faster, at the expense of a minor quality degradation (the authors' words). The Whisper family was trained on more than 5 million hours of labeled audio. Per the model card, the model recognizes speech in 99 languages, detects the language automatically, transcribes audio of arbitrary length and returns sentence-level and word-level timestamps. It can be fine-tuned, and it is supported in Hugging Face Transformers and in the authors' command-line tool. The authors' repository lists about 6 GB of VRAM for the turbo model and notes that it is not trained for translation: use the larger Whisper models to translate speech into English. There is no speaker separation and no streaming mode in the card. On the Open ASR Leaderboard (English short-form speech, nine test sets, checked 2026-10-20) it has an average word error rate of 8.39%, higher than most newer models in our list, and it is the oldest of them (released 2024-10-01). The weights are public without any acceptance step under the MIT license, which allows commercial use. Code and instructions are in the authors' repository on GitHub. Date of check: 2026-10-20.