Technology & privacy
Whisper vs Parakeet: latency is the axis nobody compares on
Both run locally and both are good. The interesting difference for dictation is not which is more accurate on a benchmark — it is which returns your sentence before you lose patience.
Last updated
Why the usual comparison is the wrong one
Model comparisons are almost always about word error rate on a benchmark corpus. That is the right measure for transcription, where you feed in an hour of recording and wait.
Dictation is an input method. You hold a key, speak eight seconds, release, and the text needs to be there. A model that is half a percentage point more accurate and takes three seconds to return is worse for this job, because three seconds of staring at nothing breaks the flow in a way one wrong word does not.
So the axis that matters is: how fast does it return, at what accuracy, on the hardware the customer actually owns?
The two families
Whisper
Published openly by OpenAI, which is the single most consequential thing about it — it turned speech recognition from a rented service into a downloadable file, and most of the local dictation market exists because of that decision.
- Strengths: very broad language coverage, robust on messy audio, extremely well understood with mature optimised implementations such as whisper.cpp.
- Sizes: tiny through large, trading accuracy against memory and speed. The size choice is most of the engineering decision.
- For dictation: the smaller variants are fast enough; the larger ones are better suited to transcribing files where latency does not matter.
Parakeet
A family from NVIDIA, designed with throughput and latency as first-class goals rather than as a consequence of size.
- Strengths: notably fast on Apple Silicon for its accuracy, strong on English, compact.
- Trade-off: narrower language coverage than Whisper's larger variants.
- For dictation: well suited, which is why several Mac dictation apps default to it for English.
| Whisper family | Parakeet family | |
|---|---|---|
| Published by | OpenAI | NVIDIA |
| Design emphasis | Breadth and robustness | Speed at good accuracy |
| Language coverage | Very broad | Narrower |
| Sizes | tiny → large | Generally compact |
| Runs locally | Yes | Yes |
| Typical dictation fit | Smaller variants | Very good |
| Typical transcription fit | Larger variants | Usable |
A general orientation rather than a benchmark. Both families change with each release, and neither vendor publishes numbers directly comparable on Apple Silicon.
What actually decides your accuracy
In descending order of effect, for a real person dictating real work:
- Microphone distance. Bigger than any model choice. A close microphone beats a better model on a laptop's built-in one every time.
- Room acoustics. Echo is worse than noise, and hard-surfaced rooms send the model your voice twice.
- Whether the app knows your vocabulary. The words you correct daily are proper nouns no general model has seen. This is fixable and the model cannot fix it.
- How you speak. Whole phrases at a natural pace, not word by word.
- The model. Genuinely last, once the four above are handled.
This ordering is why "which model?" is usually the wrong question from someone whose dictation feels inaccurate. See fixing dictation mistakes.
Other families worth knowing about
- Moonshine. Built for low-latency streaming on modest hardware. Interesting for real-time use.
- SenseVoice. Fast, multilingual, with strengths in several Asian languages.
- Apple's own on-device recogniser. Not published as a model you can run, but it is what macOS uses for built-in dictation on Apple silicon and it is good.
The field moves quickly. A page confidently declaring a winner would be out of date within months, which is itself a reason not to choose a dictation app on which model it happens to ship today.
Should you be choosing at all?
Questions
Which model is more accurate?
Nobody can answer that honestly for your voice without a published method, and we are not going to pretend to. Larger Whisper variants are generally more robust on difficult audio; Parakeet is fast at good accuracy on English. On clean audio the difference is smaller than the difference your microphone makes.
Which is better for dictation specifically?
Whichever returns your sentence quickly enough that you do not notice waiting. That usually favours the faster, more compact models — which is why Parakeet and the smaller Whisper variants dominate live dictation while larger Whisper models dominate file transcription.
Can I pick the model in VV?
VV ships both families and selects for your Mac rather than exposing a general model picker in the standard interface. That is a deliberate narrowing. If choosing the model yourself matters to you, Superwhisper is more flexible and that is a legitimate reason to prefer it.
Do these models work in languages other than English?
Whisper's broader variants cover a great many languages. Parakeet's coverage is narrower. If you dictate in a language outside a local app's support, Apple's built-in dictation or a cloud service will serve you better.