Parakeet vs Whisper on Apple silicon: our benchmark

On Apple silicon, Parakeet v3 was both more accurate and about four times quicker than Whisper Large v3 Turbo in our benchmark: 0.46% word error rate against 0.83%, and a median 82 ms to transcribe a short dictation against 320 ms. Whisper's advantage is language coverage, 99 languages to Parakeet's 25, so it stays in Sayburst for everything Parakeet doesn't speak.

Sayburst runs speech recognition on your Mac, so the model you pick decides both how accurate it is and how long you wait after you let go of the key. We benchmarked the four models below using the same transcription code that runs in the app. Below are the numbers, exactly how we got them, and their limits.

#The results

Measured on a Mac Studio with an M4 Max and 36 GB of memory, macOS 15.6, on October 1, 2026.

What each column means:

• Word error rate: wrong, missing and extra words divided by the words in the reference text, across 1,082 words of LibriSpeech audio. Lower is better.
• Speed: seconds of LibriSpeech audio processed per second of compute. 96× means a 10-second clip takes about a tenth of a second.
• Median time to text: how long transcription took on eight short, dictation-length clips, from handing over the finished audio to getting the transcript back. It's the wait after you release the key.
• First load: how long the model took to load at the start of the run, before the first transcription.

#The same results as raw counts

Percentages hide how small the sample is. Here are the same figures as error counts, plus the slowest of the eight dictation clips:

Five errors against seven is a tie at this size. Five against eighteen is a real gap. Every model here turned a short clip into text in well under half a second.

#How we measured

Hardware and software. One Mac Studio, M4 Max, 36 GB, macOS 15.6. The models ran through Sayburst's own engines: Parakeet on Core ML, Whisper through whisper.cpp with a 5-bit quantized (q5) Large v3 Turbo. Each model got one untimed warm-up transcription before measurement.

Accuracy set. 40 utterances from LibriSpeech test-clean, a public corpus of English read from public-domain audiobooks: 10 consecutive utterances from each of four points in the test set, giving four speakers and 1,082 reference words. Every model heard the same 40 files.

Scoring. Before comparing, we lowercase both texts and strip punctuation. Numbers aren't spelled out, so "10" against "ten" counts as an error, and punctuation quality isn't scored at all. Word errors are substitutions plus deletions plus insertions, the standard definition described in Hugging Face's WER documentation.

Latency set. Eight dictation-style prompts (an email, a chat message, a reminder, an AI prompt, a sentence with product names, a sentence with times and dates, a long ramble and a two-word reply), generated with the macOS say command in two system voices. This is synthetic speech. We use it to time short clips and check formatting, never to claim accuracy.

#What the numbers don't tell you

• Real dictation is messier than audiobooks. Background noise, a laptop mic across the room, fast speech, half-finished sentences and unusual names all push error rates up for every model.
• LibriSpeech is familiar ground. NVIDIA's model card for Parakeet TDT 0.6B v3 lists LibriSpeech among the corpora in its training data. LibriSpeech keeps its test sets separate from its training sets, but clean audiobook English still plays to the model's strengths.
• Only English was tested. Parakeet v3 covers 25 European languages and Whisper 99; we haven't measured any of the others.
• We tested one Mac. We haven't measured an M1, M2 or M3, or any MacBook Air. Expect slower times on smaller chips.
• Not every model was tested. Sayburst also offers Parakeet v2 and Whisper Small, which aren't in this run.
• Formatting isn't measured. Real dictation also needs punctuation, capitals and fillers removed, which this score ignores.

To measure accuracy on your own voice, microphone and vocabulary, follow our guide to testing dictation accuracy. It uses the same scoring idea with a passage you can read aloud.

#What the models are

• Parakeet v3 is NVIDIA's 0.6-billion-parameter Parakeet TDT model, covering 25 European languages with built-in punctuation and capitalization.
• Parakeet Ultra is a larger Parakeet model post-trained from v3 by moondream, covering the same languages.
• Whisper Large v3 Turbo is OpenAI's pruned version of Whisper Large v3, with the decoder cut from 32 layers to 4. OpenAI's model card describes it as much faster with a minor drop in quality.
• Whisper Base (English) is a small English-only Whisper model: quick to load, quick to run, least accurate here.

#What we chose for Sayburst

Parakeet v3 is the default. It had the lowest error count in this test, transcribes a short dictation in well under a tenth of a second on this Mac, and runs on the Neural Engine, which leaves the GPU free for Sayburst Writer, the optional on-device AI that turns your speech into finished writing.

Whisper Large v3 Turbo is the one to pick for languages Parakeet doesn't cover, such as Chinese, Japanese, Korean, Hindi or Arabic. Its median of 320 ms is still a short wait.

Parakeet Ultra is there if you want the larger Parakeet model. On this clean read speech it was slightly quicker but didn't pull ahead on accuracy; two errors either way is noise at this sample size.

Here's what the default does with an ordinary, untidy sentence:

"can you, um, send the slides to Priya before the client call, the client call"

Inserted with Sayburst's Clean mode:

Can you send the slides to Priya before the client call?

Recognition with Parakeet v3, cleanup on-device, nothing uploaded. Models download once inside the app, are checksum-verified, and work offline afterwards; the models documentation lists each one's languages, download size and memory.

#Credits

Parakeet TDT 0.6B v3 and v2 by NVIDIA, CC BY 4.0, converted to Core ML by FluidInference. Parakeet Ultra by moondream, post-trained from NVIDIA Parakeet TDT 0.6B v3, CC BY 4.0, converted to Core ML by FluidInference. Whisper by OpenAI, MIT license; ggml conversions by the whisper.cpp project. LibriSpeech is CC BY 4.0.

Questions people ask

Is Parakeet more accurate than Whisper?

On our small English test, Parakeet v3 made 5 errors in 1,082 words (0.46% word error rate) and Whisper Large v3 Turbo made 9 (0.83%). That's a small sample of clean read speech, so it points the same way without proving a general winner. Whisper covers far more languages.

Is Parakeet faster than Whisper on a Mac?

In our test on an M4 Max, Parakeet v3 took a median 82 ms to transcribe a short dictation and Whisper Large v3 Turbo took 320 ms. Whisper Base was quickest at 39 ms but made the most errors.

How accurate is voice recognition?

It depends heavily on the audio. On clean audiobook recordings, the best model in our test got about 1 word in 200 wrong. Real dictation, with background noise, fast speech and unusual names, produces higher error rates for every model.

Which model does Sayburst use by default?

Parakeet v3, which covers English and 24 European languages and runs on the Neural Engine. For other languages, such as Chinese, Japanese, Korean, Hindi or Arabic, choose Whisper Large v3 Turbo in the app.