Direct answer — Should you choose Whisper large-v3 Turbo or Distil-Whisper in 2026? Choose Whisper large-v3 Turbo if you dictate in more than one language: it keeps all 99 languages in 809M parameters. Choose Distil-Whisper (distil-large-v3, 756M) only for English, where its model card reports a relative latency of 6.3x versus large-v3. Both can run in whisper.cpp. Accuracy differences between the two families are small; language coverage is not.
Picking the best Whisper model in 2026 starts with a local speech model comparison, not a size chart. OpenAI’s large-v3 Turbo and Hugging Face’s Distil-Whisper both compress the 1,550M-parameter large-v3 to roughly half its size, and both exist for the same reason: decoding was too slow for comfortable real-time dictation.
They took different routes to get there, and those routes produce genuinely different trade-offs. This comparison covers what each model changes architecturally, what the published word error rates actually measure, how both behave inside whisper.cpp, and which belongs in a dictation workflow.
If you are still assembling the runtime, start with our whisper.cpp setup guide.
What is Whisper large-v3 Turbo?
Whisper large-v3 Turbo is a pruned and fine-tuned version of large-v3 carrying 809M parameters. OpenAI’s model card describes it as “a finetuned version of a pruned Whisper large-v3” in which “the number of decoding layers have reduced from 32 to 4”.
Pruning means removing whole components from a trained network — here, 28 of the 32 decoder layers — and then fine-tuning what remains so it recovers most of the lost behaviour. The encoder, which turns audio into an internal representation, is untouched.
That asymmetry is the whole design. In Whisper, the encoder runs once per 30-second window while the decoder runs once per generated token, so the decoder dominates latency on long transcriptions.
Critically, Turbo keeps the multilingual head. The model card lists the same 99 languages as large-v3, which is what makes it a drop-in replacement rather than a specialised variant.
What is Distil-Whisper, and why is it English-only?
Distil-Whisper is a family of English-only models trained by knowledge distillation from Whisper large-v2 and large-v3. Knowledge distillation trains a smaller “student” model to reproduce the outputs of a larger “teacher” model, rather than learning from ground-truth labels alone.
The Distil-Whisper paper (arXiv:2311.00430) reports a student with “51% fewer parameters” that is “5.8 times faster” while staying “within 1% WER on out-of-distribution test data” in a zero-shot transfer setting. Those figures describe distil-large-v2.
Like Turbo, Distil-Whisper keeps the encoder: the model card notes that it “is entirely copied from the teacher to the student and frozen during training”. The savings again come from a drastically shortened decoder.
The language limit is not an oversight but a training-data constraint. The Distil-Whisper repository states plainly that the series “is only available for English speech recognition” and recommends “using the Whisper Turbo checkpoint” for multilingual work.
The published Distil-Whisper family
| Model | Parameters | Relative latency | Short-form WER | Long-form WER |
|---|---|---|---|---|
| distil-large-v3 | 756M | 6.3x | 9.7% | 10.8% |
| distil-large-v2 | 756M | 5.8x | 10.1% | 11.6% |
| distil-medium.en | 394M | 6.8x | 11.1% | 12.4% |
| distil-small.en | 166M | 5.6x | 12.1% | 12.8% |
Figures as published in the Distil-Whisper repository. All four are English-only; latency is relative to Whisper large-v3 at 1.0.
Whisper V3 Turbo vs Distil-Whisper: specifications side by side
The two models are close on size and far apart on scope. Here is what the official model cards publish.
| Attribute | Whisper large-v3 | large-v3 Turbo | distil-large-v3 |
|---|---|---|---|
| Parameters | 1,550M | 809M | 756M |
| Languages | 99 | 99 | English only |
| Decoder layers | 32 | 4 | Shortened (encoder frozen) |
| Published speed factor | baseline | none published | 6.3x relative latency |
| Method | — | pruning + fine-tuning | knowledge distillation |
| whisper.cpp download target | ✅ | ✅ | ❌ conversion required |
One row deserves attention: OpenAI publishes no numeric speed factor for Turbo. The model card says only that “the model is way faster, at the expense of a minor quality degradation”. Any specific multiplier you see quoted for Turbo comes from third-party testing, not from OpenAI.
Which model is more accurate in 2026?
Both give up accuracy against large-v3, and both give up less than their size reduction would suggest. But the two families publish their numbers on different benchmarks, so the headline figures are not directly comparable.
The evaluation results on the OpenAI model cards report a mean word error rate of 7.83 for large-v3-turbo against 7.44 for large-v3 on the English Open ASR Leaderboard suite — a gap of roughly 0.4 points.
The Distil-Whisper team evaluates on its own English suite, where distil-large-v3 records 9.7% short-form WER against large-v3’s 8.4%, and 10.8% on sequential long-form audio against large-v3’s 10.0%. On chunked long-form the distilled model actually edges ahead: 10.9% versus 11.0%.
Do not subtract 7.83 from 9.7. Different datasets, different decoding algorithms and different audio lengths produce different absolute WER for the same model. Each number is only meaningful against its own baseline.
The honest summary: relative to large-v3, Turbo loses about half a point of WER on the English leaderboard, and distil-large-v3 loses about 1.3 points short-form and 0.8 points on sequential long-form on the Distil-Whisper suite. For dictation, both sit well inside the range where microphone quality and background noise dominate the outcome.
We covered how to read speech benchmark numbers without over-reading them in our Voxtral and Whisper benchmark comparison.
Which is the fastest Whisper model in practice?
For English-only workloads on a local machine, Distil-Whisper has the stronger published evidence. Its model card includes a whisper.cpp-specific measurement: “In a provisional benchmark on Mac M1, distil-large-v3 is over 5x faster than Whisper large-v3.”
That is a scoped, reproducible claim — a named runtime, named hardware, named comparison — and it is the sort of figure worth trusting. It is also explicitly labelled provisional.
Turbo’s advantage is harder to quantify because OpenAI declined to publish one. Structurally, cutting the decoder from 32 layers to 4 removes the same bottleneck Distil-Whisper targets, so the two should land in a broadly similar range on the same hardware.
Three variables move real-world throughput more than the checkpoint does:
- Quantisation. whisper.cpp supports quantised ggml files (Q5_0 among others), which shrink memory footprint and speed up inference at some accuracy cost.
- Hardware acceleration. On Apple Silicon, whisper.cpp can push encoder inference to the Neural Engine via Core ML, and run inference on the GPU via Metal.
- Audio length and algorithm. Sequential and chunked long-form strategies produce different speed and accuracy profiles for the same model.
How do both models behave inside whisper.cpp?
Turbo is easier to adopt. The whisper.cpp repository lists large-v3-turbo among its downloadable models alongside the other official checkpoints, so fetching the ggml file is a single command.
Distil-Whisper is not one of those download targets. The distil-large-v3 model card documents running it “with the Whisper.cpp package with the original sequential long-form transcription algorithm”, which means a conversion step before you have a usable ggml file.
That difference matters more than it looks. A model you can pull directly stays current with upstream releases; a model you convert yourself is a step you maintain.
Memory is the other practical constraint. whisper.cpp documents roughly 3.9 GB for a large-class model, so a half-size checkpoint is what makes local dictation viable on an 8 GB machine at all.
Which model should you choose?
Work down this list and stop at the first line that matches.
- You dictate in more than one language. Take large-v3 Turbo. Distil-Whisper cannot serve you, and the Distil-Whisper project says so itself.
- English only, and latency is your binding constraint. Take distil-large-v3, accept the conversion step, and validate on your own audio.
- English only, and you want the simplest maintenance path. Take large-v3 Turbo anyway. The multilingual head costs you nothing at inference time.
- Accuracy above all, speed secondary. Stay on full large-v3.
- You want dictation, not a model pipeline. Use an application that manages the checkpoint for you.
That last line is the one most people land on. Compiling a runtime, converting checkpoints and tuning quantisation is a genuinely interesting afternoon — and a poor weekly commitment.
Weesper Neon Flow is built on whisper.cpp, runs entirely offline with no audio leaving the device, supports 55+ languages, and uses Metal acceleration on Apple Silicon. Try Weesper free for 15 days if you would rather dictate than benchmark.
The bottom line on Whisper V3 Turbo vs Distil-Whisper
These two models solve the same problem — a decoder too slow for real-time use — with different tools, and they land in a similar size class: 809M against 756M parameters, both roughly half of large-v3.
The accuracy gap between them is small enough that your microphone matters more. The language gap is absolute: 99 languages against one. For anyone whose work is not exclusively in English, that single row decides the comparison before speed enters the discussion.
If you want to see how these local models compare against the engines built into your operating system, read our breakdown of Apple Dictation and OpenAI Whisper.
Ready to dictate offline? Install Weesper Neon Flow on macOS or Windows, or browse the Help Center for setup and language configuration.