Distil-Whisper is an open-source speech recognition model from Hugging Face, a distilled version of OpenAI's Whisper for English only. Knowledge distillation trains a smaller student model to copy the outputs of a larger teacher model. Distil-Whisper is about 6 times faster than Whisper and about half its size, and stays within 1% word error rate (WER) on test sets unlike its training data.
The authors kept Whisper's whole encoder unchanged and only two of its decoder layers, then trained on 22,000 hours of pseudo-labelled audio, that is, audio transcribed by Whisper itself, from 10 domains. A developer runs the model with the Hugging Face Transformers library: audio goes in and text comes out, with segment-level timestamps on request; long recordings are split into chunks. The checkpoints also run in whisper.cpp, Transformers.js and Candle. As an assistant model for speculative decoding, it gives exactly Whisper's output about 2 times faster.
Distil-Whisper is free under the MIT license, which allows commercial use.