Distil-Whisper

Hugging Face researchers produced these English speech recognition checkpoints by distilling OpenAI Whisper: the student keeps Whisper's encoder, held fixed during training, while its decoder shrinks to just two layers that learn from transcripts the larger model generated. The payoff is transcription several times faster than Whisper large with word error rates that stay close to it. The family runs from distil-small.en, at 166M parameters for constrained devices, up to distil-large-v3 and the newest distil-large-v3.5, published in March 2025 after training on about 98,000 hours of public audio. Checkpoints load in the Transformers ASR pipeline and have ports for faster-whisper, whisper.cpp, Transformers.js and Candle, and the large variants double as draft models for speculative decoding against Whisper large-v3. A GPU helps for long-form batch jobs, though CPU inference works. Code and weights are MIT, so commercial deployment is fine. The GitHub training repo has been quiet since early 2025, and for non-English audio OpenAI's Whisper turbo is the better pick. Suited to English transcription at lower latency or cost.