Whisper, Parakeet or SenseVoice: numbers, not adjectives
By Andrey ChmerevI build Kekoso, which ships all five of these models, so I had to answer this question to build the model picker. The speed numbers are my own measurements on one machine with one clip; the word error rates are other people's, each linked to its publication — I later worked out how to measure word error rate on my own audio, which is the honest way to check any of them. Measured 23 August 2026.
Ask how to transcribe something locally and the answer is “Whisper”, the way the answer to how to edit a photo used to be “Photoshop”. Whisper transcription is a good default. It is also, for a lot of jobs, the wrong one, and the way to show that is with numbers rather than opinions.
I had to settle this to build a model picker, so here is what came out of it.
Why most comparisons you will find are broken
Whisper vs Parakeet is the comparison people arrive looking for, and the numbers they find will not answer it. Vendor numbers are not comparable with each other. One model’s page reports 96.5 times real time on an M2, another reports 19.3 on an M4 Max. Put those two figures in a table side by side and you have measured two laptops, not two models.
So the speed column below is a single run of my own: five models, one machine, one 28-second clip of continuous English speech, warm Neural Engine cache, best of two consecutive runs. MacBook Pro with an M4 Max, macOS 26.5.2, measured on 23 August 2026.
The accuracy column is not mine — measuring word error rate needs a labelled dataset, not one recording. Those figures come from published benchmarks on LibriSpeech test-clean, each linked in the table.
The word error rates below are other people’s, each linked to its publication. If you want to check any of them against your own recordings rather than trust them, the method is here and it is thirty lines of Python.
The five models
Three of them are Whisper builds, which is why “whisper transcribe” covers more than one thing: OpenAI Whisper transcription comes in sizes, and whisper audio transcription on a Mac usually means one of the two below.
| Model | WER | Speed | Size | Languages | Licence |
|---|---|---|---|---|---|
| Whisper large-v3-turbo | 1.96% | 20.1× | 616 MB | 99 | MIT |
| Parakeet TDT v3 | 2.5% | 122.3× | 240 MB | 25 | CC BY 4.0 |
| Parakeet TDT-CTC 110M | 3.01% | 165.4× | 217 MB | English only | CC BY 4.0 |
| SenseVoiceSmall | 3.25% | 247.5× | 226 MB | 5 | FunASR |
| Whisper small | 3.95% | 19.3× | 207 MB | 99 | MIT |
Word error rates: FluidAudio’s benchmark tables for both Parakeet models and SenseVoice, the WhisperKit paper for the compressed large-v3-turbo, and argmax’s LibriSpeech runs for small. One caveat on that last one: the published figure is for uncompressed small, and the 207 MB build is compressed, so treat 3.95% as the optimistic end.
Speed is RTFx — audio duration divided by transcription time. 122× means a minute of audio takes about half a second.
What those numbers look like on harder audio
Published word error rates are measured on LibriSpeech — read audiobook prose, recorded cleanly, without a proper noun in sight. I later ran three of these models over my own material: 200 words of synthesised meeting speech stuffed with the things real meetings are made of, including product names, two surnames and half a dozen pieces of jargon.
| Model | Published (LibriSpeech) | On my 200 words |
|---|---|---|
| Whisper large-v3-turbo | 1.96% | 11.0% |
| Whisper medium | not published here | 9.0% |
| Whisper small | 3.95% | 15.0% |
| Parakeet TDT v3 | 2.5% | 12.5% |
Five times worse across the board, and the reason is not that the published
figures are wrong. They measure a different thing. Nearly every error in my run
was a name or a term — Kekoso heard as Kakuzo, Grafana board as Gryphon
abort, rollback as robux — and audiobook prose contains none of those.
Two things survive the change of material and one does not. The ordering between Whisper large-v3-turbo and Parakeet TDT v3 holds, and Whisper small stays last. But the gap between the leaders shrinks from a clear 0.5 points to noise: on 200 words a single error moves the figure half a point, so 11.0% against 12.5% is three errors, and medium’s apparent win over turbo is four. Treat the top of that column as a tie.
The practical reading is that on your own vocabulary, all of these models are roughly as good as each other and all of them are much worse than their benchmark. What closes that gap is not a bigger model but a custom vocabulary, which on the same recordings took errors from 12.5% to 4.0% — a bigger improvement than any choice in the table above.
What the table actually says
One caveat before the reading: these ratios are measured on a 28-second clip, and for dictation-length phrases they mislead — Whisper’s encoder costs the same whether you say one word or twenty-six seconds of them.
Whisper is the most accurate, and it is not close on quality. 1.96% against 2.5% for the best Parakeet is a meaningful gap on hard material.
Whisper is also six times slower than that Parakeet, and twelve times slower than SenseVoice. For dictating a sentence into a text field, six times slower means the difference between text appearing as you finish speaking and text appearing a beat later. For a two-hour recording it is the difference between four minutes and twenty-five.
The compact Parakeet beats Whisper small on both axes. 3.01% against 3.95%, at eight times the speed, in the same disk space. If you are English-only and were about to download Whisper small, that is the trade you are making.
Three things the table hides
These are the parts I did not expect, and they matter more than the ranking.
Making Whisper smaller does not make it faster
The two Whisper models landed at 20.1× and 19.3×. Essentially identical, despite one being three times the size of the other.
The cost is not in the weights. Whisper decodes in fixed 30-second windows, and that structure dominates the runtime. Downloading Whisper small instead of large-v3-turbo to make transcription faster does not work — you give up accuracy and get nothing back. It saves disk space, and that is all it saves.
Clip length changes the ranking
Everything above is one 28-second clip. On a 94-second clip the picture moved: both Whisper models fell to roughly 7.6×, while Parakeet TDT v3 rose to 176×.
That is a gap widening from six times to over twenty, purely because the material got longer. Any single benchmark number — mine included — describes one length of audio. Short dictation and long recordings are different questions.
The first run of a model costs minutes, not seconds
The first ever load of Whisper large-v3-turbo took 2 minutes 35 seconds. The second took 1.5 seconds. For SenseVoiceSmall the first compile took around ten minutes.
That is CoreML compiling the model for the Neural Engine, cached by the system afterwards. It happens once per model per machine and never appears in a benchmark, because benchmarks report steady state. It is, however, exactly what a person experiences the first time they try a model, and it is why an app that downloads a model in the background should say what it is doing rather than showing a spinner.
The differences that are not about speed
Timings. Whisper reports word-level times. Parakeet reports token times, close enough for most uses — though I later timed the same file with both and they disagree by 130 ms at the median, which is close enough to click a transcript and not close enough for subtitles. SenseVoice reports none at all, so they have to be estimated by spreading words evenly across the window — which is why subtitles made from SenseVoice are visibly coarser. If timing matters, this is a bigger deal than the WER column.
Silence. Whisper invents text on it. Long pauses come back captioned with sign-off lines nobody said — an artefact of how it was trained on subtitled video. Any serious use of Whisper needs a filter that drops segments with no audio behind them. I measured this on four model sizes afterwards: both large models thank you for thirty seconds of digital zero, and none of the obvious settings stop them.
Languages. Whisper covers 99, Parakeet TDT v3 covers 25 European ones, the compact Parakeet is English-only, and SenseVoice covers Chinese, Cantonese, Japanese, Korean and English. This decides the choice more often than accuracy does: a 1.96% word error rate is irrelevant if the model does not speak your language. Which model covers which language is its own article, including why SenseVoice’s advertised count does not survive a look at its vocabulary.
Licence. MIT for Whisper, CC BY 4.0 for both Parakeet models, the FunASR model licence for SenseVoice. If you are shipping something built on these, read them rather than assuming — I did, and what those three licences do and do not cover has one surprise in it: not one of them mentions the transcript.
How to choose
Dictating into a text field. Speed wins, because latency is felt and a 0.5% accuracy difference is not. Parakeet if your language is covered, SenseVoice for Chinese, Japanese, Korean or Cantonese.
A long recording where accuracy matters. Whisper large-v3-turbo, and plan for it to take a while. This is the case where Whisper transcription earns its accuracy premium.
Subtitles. Whisper, for the word-level timings.
A language outside the European set. Whisper is often the only option, and that settles it.
English-only, tight on disk. The compact Parakeet, not Whisper small.
The reason Kekoso ships all five rather than picking one is that the table above has no winner. Every model on it is the right answer to some question and the wrong answer to others, and a product that hides the choice is deciding for you which question you are asking.
Methodology, the full run log and the two bugs the measurement uncovered are in the project’s own benchmark notes. Speed figures: one machine, 23 August 2026. Accuracy figures: other people’s published benchmarks, linked above.
Questions people ask
Is Whisper the most accurate local speech model?
On LibriSpeech test-clean, the compressed large-v3-turbo has the lowest word error rate of the five models compared here at 1.96%, ahead of Parakeet TDT v3 at 2.5% and SenseVoiceSmall at 3.25%. But it is also six to twelve times slower on the same machine and clip, so 'most accurate' and 'best choice' are not the same sentence.
Why is Whisper small not faster than Whisper large-v3-turbo?
Because what costs time is not the weights but the decoding window. Whisper processes audio in 30-second windows regardless of model size, and in a same-machine run the small model managed 19.3 times real time against 20.1 for large-v3-turbo — effectively identical. Downloading the smaller Whisper to go faster does not work.
How much faster is Parakeet than Whisper?
On the same Mac and the same 28-second clip, Parakeet TDT v3 ran at 122.3 times real time against 20.1 for Whisper large-v3-turbo — about six times faster, at a word error rate of 2.5% against 1.96%. SenseVoiceSmall was fastest at 247.5 times real time.
Does clip length change which model is fastest?
Yes, and it is the reason to distrust a single benchmark number. On a 94-second clip both Whisper models dropped to roughly 7.6 times real time while Parakeet TDT v3 rose to 176. Long files widen the gap; short dictation narrows it.
Why does the first run of a local model take minutes?
That is CoreML compiling the model for the Neural Engine, and the system caches the result. In my run the first ever prepare of Whisper large-v3-turbo took 2 minutes 35 seconds and the second took 1.5 seconds. SenseVoiceSmall's first compile took around ten minutes. It happens once per model per machine, and no benchmark that reports steady-state speed will warn you about it.
Which model should I use for subtitles?
Whisper, if timing matters. Whisper reports word-level times, Parakeet reports token times which are close enough for most work, and SenseVoice reports no timings at all — they have to be estimated by spreading words across the window, which makes its subtitles visibly coarser.