Technical explainer · Verified 2026-09-05

Distil-Whisper
English speech distillation

Distil-Whisper is a distilled version of Whisper for English speech recognition. Its student model uses only two decoder layers initialized from the teacher's first and last decoder layers, while the encoder is kept frozen. The project reports faster inference and a smaller model with word error rates close to its teacher on the stated evaluation sets.

What distillation changes

Distil-Whisper keeps the encoder frozen and uses a student decoder initialized from the first and last decoder layers of Whisper. The student has only two decoder layers. The project is explicitly for English speech recognition. Distillation creates an English-only student model with a different decoder. Its Whisper ancestry gives it neither multilingual coverage nor the identity of an unchanged checkpoint in a smaller file format.

Read the speed and error claims together

The repository describes Distil-Whisper as 6 times faster, 49% smaller and within 1% word error rate on out-of-distribution evaluation sets. The distil-large-v3 card separately reports 6.3x faster inference and within 1% WER of large-v3 on long-form audio. These results apply to the stated evaluations. Actual error rates and speed depend on the audio, runtime and application; the WER comparisons describe differences from the teacher rather than fixed error rates.

The checkpoint and runtime still matter

The repository's distil-large-v3 row lists 756M parameters, relative latency 6.3, short-form WER 9.7 and long-form WER 10.8. Keep the row's labels and evaluation context attached to those numbers. The model card provides converted weights for faster-whisper, whisper.cpp and the reference Whisper implementation. Distil-Whisper is MIT licensed, but available runtime support does not determine which models a desktop product actually includes.

How WhisperJot uses this

WhisperJot does not ship Distil-Whisper. Its macOS Jot Local Pro defaults to large-v3-turbo through WhisperKit, with a compressed turbo option and small.en also selectable. Windows and Linux use the product's GGML model choices. Distilled checkpoints are available in compatible independent runtimes; WhisperJot's selectable models come from its own model catalog. Jot Local is a separate Parakeet-based engine, and Auto chooses between the local engines per utterance.

Explore WhisperJot →

Frequently asked questions

Does Distil-Whisper support other languages?

Distil-Whisper is only available for English speech recognition. Its distilled checkpoints retain that English-only scope despite their Whisper ancestry. For multilingual transcription, compare the multilingual Whisper checkpoints and confirm that your chosen runtime supports them. Check the checkpoint's stated language scope.

Is Distil-Whisper the same as turbo?

No. Distil-Whisper is an English speech-recognition distillation project. Large-v3-turbo is a pruned and fine-tuned multilingual Whisper checkpoint. They are different models with distinct training methods and language coverage, even though both derive from Whisper. Choose around your language requirements and transcription workload.

Can faster-whisper load distil-large-v3?

Yes. The distil-large-v3 card provides pre-converted weights for faster-whisper as well as whisper.cpp and the reference implementation. That compatibility is a runtime option; WhisperJot does not ship Distil-Whisper as one of its selectable local models. Check product model selections separately from library compatibility.

Sources

Primary documentation checked 2026-09-05. Specifications and runtime requirements can change.

Related explainers

faster-whisper

Read faster-whisper's CTranslate2 speed and memory benchmarks with their model and precision settings, Python requirements, and when this runtime fits.

Whisper large-v3-turbo

Understand Whisper large-v3-turbo: fewer decoder layers, reference speed, translation limits, compressed variants and its role in WhisperJot Jot Local Pro.

Whisper model sizes

Compare Whisper model sizes using the official parameters, VRAM and relative speed table. Learn what English-only variants mean and how to choose.

See also: Whisper glossary definition · System requirements · Pricing