Distil-Whisper

Distil-Whisper is a distilled version of OpenAI’s Whisper large-v3 speech-recognition model that runs 6× faster and is 50% smaller while staying within 1% word-error rate. Created by the Hugging Face team, it keeps Whisper’s robustness to accents, background noise and technical language at a fraction of the serving cost. The project has about 4,100 stars on GitHub.

Overview

The trick is knowledge distillation done carefully: the student model keeps Whisper’s full encoder but a drastically reduced decoder, and is trained on 22,000 hours of pseudo-labelled audio generated by the teacher, with WER-based filtering to keep only trustworthy labels. The result, distil-large-v3, transcribes long-form audio with timestamp support and is a drop-in replacement in Transformers, faster-whisper and other Whisper stacks — same API, same language (English for the distilled checkpoints), much lower latency and memory. The repository documents the full recipe — data generation, filtering, training and evaluation across out-of-distribution test sets — so teams can distill their own variants. For production ASR serving, voice interfaces and bulk transcription, Distil-Whisper became the default way to cut Whisper’s cost without giving up its accuracy.

The checkpoints are released in several sizes — distil-small.en through distil-large-v3 — so teams can pick their own point on the speed/accuracy curve, and they work with speculative-decoding setups where Distil-Whisper drafts and full Whisper verifies. Evaluations in the repository cover ten out-of-distribution test sets to demonstrate the distilled models do not just memorize the teacher’s benchmark behaviour. For many production systems this halved GPU serving cost while keeping transcription quality effectively unchanged.

Key Features

• 6× faster inference and 49% fewer parameters than Whisper large-v3
• Within 1% WER of the teacher on standard benchmarks
• Drop-in compatible with Transformers and faster-whisper pipelines
• Long-form transcription with timestamps
• Full open distillation recipe: data, filtering and training code
• Robust to accents, noise and domain-specific vocabulary

💳 Pricing

Distil-Whisper is free and open source under the MIT license; weights are on the Hugging Face Hub for self-hosting.

Who Is It For

Distil-Whisper is for developers serving speech recognition at scale — transcription products, voice assistants, meeting tools — who want Whisper-level accuracy at half the compute bill.