The short version
- In English, Parakeet matched or beat Whisper. On 432 clips of real people reading, Parakeet v2 made 2.14% word errors and Whisper large-v3 turbo 2.44%. With noise added: 2.78% vs 3.75%. On an accented speaker full of product names: 21.2% vs 28.4%.
- Parakeet was about 13× faster. After a clip ended it had the text in a median 0.10 s; Whisper took 1.37 s, and up to 2 s on longer dictations.
- Whisper wins on reach. Parakeet v3 handled Spanish, French, German and Russian as well as Whisper, but turned Turkish into nonsense. Whisper covers 99 languages.
- So the right answer is both: Parakeet for the languages it knows, Whisper for the rest.
A year ago, “local speech to text” meant Whisper. Now many local Mac dictation apps, Murmur included, also offer NVIDIA’s Parakeet, and the settings screen asks you to choose. The advice online is a mix of vendor blog posts and a single widely quoted test that put Parakeet at 12% word errors. That didn’t match what we saw in our own app, so we measured it properly.
We cared about two numbers: how many words come out wrong, and how long you wait after you stop talking. For dictation, the second is the one you feel every time.
What we tested, and how
Everything ran on one 2021 MacBook Pro with an M1 Pro and 16 GB of memory, on macOS 26.6.2. A newer chip will be faster across the board; the gaps between models are what to look at.
- Parakeet TDT 0.6B v2 (English only) and v3 (25 European languages), the 600-million-parameter models NVIDIA released in 2025, converted to Core ML and run through FluidAudio 0.17.7 on the Neural Engine.
- Whisper large-v3 turbo, OpenAI’s pruned large-v3 with 4 decoder layers instead of 32, run through Argmax’s WhisperKit 1.1.1 (the 626 MB Core ML build). We told it the audio was English for the English sets.
- Apple’s SpeechTranscriber, built into macOS 26, as a reference point.
- The same code for every model: the speech engines inside Murmur 1.1, driven by its built-in test harness. No dictionary, cleanup or AI rewrite was counted, only what each model heard.
- Timing: model already loaded, whole clip handed over when it ends, stopwatch until the final text. That is the wait after you let go of the dictation key.
And four sets of audio:
| Set | What it is | Clips |
|---|---|---|
| Clean | LibriSpeech test-clean: real people reading books aloud, the standard speech-recognition test. 63 minutes, 9,814 words, clips of 2 to 34 seconds. | 432 |
| Noisy | The same clips with steady background hiss mixed in at 10 dB signal-to-noise: clearly audible under the voice. | 432 |
| Accented | One of us, whose first language is Azerbaijani, dictating sentences full of product names (Claude, Supabase, kubectl) into a MacBook’s built-in mic. | 67 |
| Six languages | Twelve everyday sentences each in Spanish, German, French, Russian, Ukrainian and Turkish, read by macOS voices. | 72 |
Result 1: accuracy in English
Source: Murmur benchmark, October 29, 2026. MacBook Pro M1 Pro, macOS 26.6.2.
Show the numbers
| Word error rate | |
|---|---|
| Parakeet v3 (25 languages) | 20.9% |
| Parakeet v2 (English) | 21.2% |
| Whisper large-v3 turbo | 28.4% |
| Apple SpeechTranscriber | 32.3% |
This is the set that looks most like real dictation: a fast talker with an accent, a laptop mic, and a sentence like “Run kubectl rollout restart on the staging deployment.” Every model struggled with the jargon, which is why the error rates are high. Parakeet made about a quarter fewer errors than Whisper and a third fewer than Apple’s engine.
Whisper also did something Parakeet never did on any of our sets: on one clip it heard a clear spoken sentence and wrote “*sad music*”. Inventing text, usually from subtitles in its training data, is a known Whisper habit.
On clean and noisy read speech the field was closer:
| Model | Clean | Noisy (10 dB) | Accented |
|---|---|---|---|
| Parakeet v2 (English) | 2.14% | 2.78% | 21.2% |
| Parakeet v3 (25 languages) | 2.54% | 3.17% | 20.9% |
| Whisper large-v3 turbo | 2.44% | 3.75% | 28.4% |
| Apple SpeechTranscriber | 2.24% | 3.86% | 32.3% |
On clean audio, everything is within a few tenths of a point: about one wrong word in every 40 to 47. Background noise is where Parakeet v2 pulled away, holding at 2.78% while Whisper rose to 3.75%. That’s roughly one extra wrong word every 100, which you notice in a long email.
The English-only v2 was a little better than the multilingual v3 on LibriSpeech; on our accented speaker, v3 edged ahead. Either is a safe pick for English.
Result 2: the wait after you stop talking
Source: Murmur benchmark, October 29, 2026. MacBook Pro M1 Pro, macOS 26.6.2.
Show the numbers
| Median wait | |
|---|---|
| Parakeet v2 (English) | 103 ms |
| Parakeet v3 (25 languages) | 106 ms |
| Apple SpeechTranscriber | 240 ms |
| Whisper large-v3 turbo | 1.37 s |
Parakeet answered in about a tenth of a second: faster than you can move your hand from the key to the mouse. Whisper large-v3 turbo took 1.37 seconds for the same clips. That’s 13 times longer, and long enough that you watch for the text to arrive.
The gap grows with what you say, because Parakeet runs through a long clip almost as fast as a short one:
Groups: 0–5 s (150 clips), 5–10 s (142), 10–20 s (116), 20–34 s (24). Source: Murmur benchmark, October 29, 2026.
Show the numbers
| 2.5 s | 7.5 s | 15 s | 27 s | |
|---|---|---|---|---|
| Whisper turbo | 1.2 s | 1.4 s | 1.7 s | 2.0 s |
| Apple | 0.13 s | 0.25 s | 0.47 s | 0.78 s |
| Parakeet v3 | 0.08 s | 0.11 s | 0.15 s | 0.24 s |
| Parakeet v2 | 0.08 s | 0.10 s | 0.14 s | 0.17 s |
For a 20 to 34 second dictation, about a paragraph, Parakeet v2 had the text ready in a median 0.17 s and Whisper in 1.99 s. Over a day of dictating that adds up to minutes of waiting.
The 95th-percentile waits tell the same story: 0.18 s for Parakeet v2, 0.25 s for v3, 2.0 s for Whisper. With noise added, none of the medians moved by more than 20 ms.
Result 3: other languages
Here the trade-off flips. Both models were given the audio with no language set and had to work out what they were hearing:
| Language | Parakeet v3 | Whisper large-v3 turbo |
|---|---|---|
| Spanish | 0.0% | 0.0% |
| French | 0.0% | 0.8% |
| German | 0.9% | 1.8% |
| Russian | 2.3% | 2.3% |
| Ukrainian | 4.5% | 8.0% |
| Turkish | Not supported: unusable | 5.1% |
| Median wait | 0.08 s | 1.27 s |
Within its 25 languages, Parakeet v3 was as good as Whisper or better here, and still about 15 times faster. Outside them it doesn’t fail politely. Given Turkish, it wrote English-looking sound-alikes: “Eve dönerken süt ve ekmek almayı unutma” (don’t forget to buy milk and bread on the way home) came out as “Eve don’t sootweek make alma yulnutma.” Whisper got it right.
A caution on this table: these are 12 synthetic sentences per language, read by macOS voices. That’s enough to show which languages work at all, not to rank two models a point apart. For a fuller look at speaking more than one language, see Mac dictation in multiple languages.
Why other tests disagree
The most quoted comparison, a 9to5Mac test reported by heise, transcribed one 7.5-minute file on an M2 Pro. Parakeet finished in 2 seconds and Whisper Large V3 Turbo in 40, but Whisper scored a 1% word error rate and Parakeet 12%.
We can’t check that result without their audio. A single long file is where implementations differ most: how the audio is cut into pieces, which Parakeet build is used, and whether punctuation and numbers are scored. Our numbers come from 432 separate clips of human speech, run through the same app code, and we kept every clip’s output, so each error can be traced.
The speed findings agree everywhere: Parakeet is an order of magnitude faster than Whisper on a Mac. Our earlier Apple vs Whisper benchmark found the same about Whisper’s wait growing with clip length.
Which one should you use?
- You dictate in English: Parakeet v2. Most accurate in our clean and noisy tests, answers in about 0.1 s.
- You switch between English and other European languages (Spanish, French, German, Italian, Portuguese, Polish, Russian, Ukrainian and more): Parakeet v3. It detects the language each time.
- You speak a language outside those 25 (Turkish, Azerbaijani, Arabic, Hindi, Japanese, Chinese…): Whisper large-v3 turbo, and accept a second or two of wait.
- You transcribe hour-long recordings after the fact: the wait matters less, and you should test both on your own audio.
Or don’t choose: what we built into Murmur
We ran this benchmark because Murmur ships all of these models and we had to decide what to recommend. The result is the setup above, built in:
- Pick the languages you speak, and Murmur suggests the model. English only gets Parakeet v2. Only Parakeet’s European languages gets v3. Anything else gets Whisper large-v3 turbo, which covers 99.
- Both models detect the language. With Parakeet v3 or Whisper you don’t switch anything by hand; you just start talking.
- Whisper gets Parakeet’s speed where it can. If you speak Azerbaijani and English, Murmur runs Whisper and checks the language of each dictation. English and Russian go to Parakeet, which answers in a tenth of the time and made fewer English errors in our tests, and Azerbaijani stays with Whisper.
- Everything runs on your Mac. Each model downloads once (480 to 630 MB), then works offline, even in Airplane Mode. No account, no audio leaves the Mac.
- The rest of Murmur sits on top: hold right ⌘, talk, let go, and polished text is typed wherever your cursor is, with your dictionary, per-app writing styles and “scratch that”. Calls get meeting notes with no bot joining.
Murmur costs $19 once, runs on Apple silicon Macs with macOS 14 Sonoma or later, and your first 2,000 words are free with no account. If privacy is why you want local models in the first place, our guide to offline dictation apps for Mac compares the others.
Repeat the test yourself
Behind this page sit a script that builds the noisy and multilingual sets, a run script, a scorer, and one line per clip with what each model heard and how long it took. Ask at the address below and we’ll send them. The English error rates use the harness’s scorer (lowercase, no punctuation, spelled numbers matched to digits). The multilingual ones use a Unicode-aware version of the same rule.
What we did not test: other Macs, streaming previews, whisper.cpp or MLX builds of Whisper, and recorded human speech in the five non-English languages. One honest footnote: the Mac went to sleep during the first Whisper run, so six clips logged waits of up to 85 minutes. We re-ran those six (same text, 1.3 to 1.7 s) and used the re-run timings.
Got different numbers on an M3 or M4? Write to [email protected] and we’ll add them here.
Frequently asked questions
Is Parakeet more accurate than Whisper?
In English, in our test, yes or tied. On 432 LibriSpeech clips Parakeet TDT 0.6B v2 made 2.14% word errors and Whisper large-v3 turbo 2.44%. With background noise it was 2.78% vs 3.75%, and on one accented speaker reading jargon 21.2% vs 28.4%. Whisper is the one to use for languages outside Parakeet’s 25, such as Turkish.
How much faster is Parakeet than Whisper on a Mac?
About 13 times faster in our test on an M1 Pro. After a clip ended, Parakeet v2 had the text ready in a median 0.10 seconds and Whisper large-v3 turbo in 1.37 seconds. On 20 to 34 second dictations the medians were 0.17 s and 1.99 s.
Which languages does Parakeet support?
Parakeet TDT 0.6B v2 is English only. Parakeet TDT 0.6B v3 covers 25 European languages, including English, Spanish, French, German, Italian, Portuguese, Polish, Russian and Ukrainian, and detects which one you are speaking. Whisper large-v3 turbo covers 99 languages, including Turkish, Azerbaijani, Arabic, Hindi, Japanese and Chinese.
Should I use Parakeet or Whisper for dictation on a Mac?
Use Parakeet if you dictate in English or another of its 25 European languages: in our test it was at least as accurate and answered in about a tenth of a second. Use Whisper large-v3 turbo for languages Parakeet doesn’t cover, and expect a wait of a second or two after each dictation.
Why do some tests say Whisper is much more accurate than Parakeet?
9to5Mac’s widely cited test found Whisper at 1% word error rate and Parakeet at 12% on one 7.5-minute file. Results like that depend on the audio, the Parakeet build and how the long file is split. On our 432 short clips of real speech, run through the same app code, the two were within half a point in English.
Sources
- NVIDIA Parakeet TDT 0.6B v2 model card
- NVIDIA Parakeet TDT 0.6B v3 model card (25 languages)
- OpenAI Whisper large-v3-turbo model card
- FluidAudio (Parakeet on Core ML), version 0.17.7
- Argmax WhisperKit (argmax-oss-swift), version 1.1.1
- heise online: Apple’s new speech APIs outperform Whisper on speed (July 4, 2025)
- LibriSpeech ASR corpus (OpenSLR 12)
Telman Gadimov builds Murmur, a dictation app for Mac that runs entirely on-device. We make Murmur, which ships both Parakeet and Whisper large-v3 turbo, so we have a stake in how this reads. Every run went through Murmur’s own engine code, and this page shows the setup, every result and the cases where Whisper won.
