Parakeet or Whisper for dictation on a Mac?
For live dictation in English or one of 25 European languages, Parakeet V3 wins: lower word error rate (6.3% against 7.4% to 8%) and several times the speed on the Neural Engine. Whisper wins when you need one of its 99 languages — Hebrew, Arabic, Hindi, CJK — or translation into English. Most people run both, one mode each.
The two families
Whisper is OpenAI's encoder–decoder speech model, released in 2022 and trained on 680,000 hours of audio in 99 languages; the decoder writes the transcript one token at a time. Parakeet is NVIDIA's family of smaller models built as token-and-duration transducers (TDT), which emit tokens and their durations together and skip most of the step-by-step decoding. Both run on the Mac as CoreML models — Whisper through WhisperKit, Parakeet through FluidAudio — and FreeFlow ships three of each.
| Model | Download | Languages | Avg English WER | Speed | Translate |
|---|---|---|---|---|---|
| Parakeet V3 | 484 MB | 25 European | 6.3% | Near-instant | No |
| Parakeet V2 | 465 MB | English | 6.1% | Near-instant | No |
| Whisper Small | 970 MB | 99 | ~8% | Fast | Yes |
| Whisper Large V3 Turbo | 1.6 GB | 99 | 7.8% | Medium | Yes |
| Whisper Large V3 | 3.1 GB | 99 | 7.4% | Slowest | Yes |
Speed
This is the axis dictation cares about most, and it is not close. Parakeet's transducer design finishes a sentence in a fraction of the time Whisper's autoregressive decoder needs, and it does so on the Neural Engine. On an M-series Mac a sentence comes back from Parakeet V3 in well under a second; Whisper Large V3 Turbo takes a second or two; the full Large can take several. Hacker News threads on local dictation apps converge on the same experience — Parakeet described as “very fast”, Whisper Large as the slower, more careful one.
Accuracy
On English, Parakeet V3 beats every Whisper size here: 6.3% against 7.4% for the full Large and 7.8% for Turbo. The gap narrows on clean, well-spoken English and widens on fast or accented speech. Outside English the picture flips by language: within Parakeet's 25 it stays competitive; beyond them it has nothing, and Whisper's coverage of 99 languages is the reason it exists. Commenters who dictate in non-English languages regularly report Whisper Large V3 as the more accurate of the two for their language, which matches the design.
Failure modes
- Whisper hallucinates on silence, famously producing “Thank you.” or “Subtitles by…” from a second of room noise. FreeFlow's silence gate — a check for speech-level sound before transcribing — exists because of this.
- Parakeet repeats: on noisy or long takes it can emit a phrase twice. A formatting pass or filler removal catches most of it; shorter takes avoid it.
- Both mishear names and jargon the same way every time, which is what the dictionary is for.
Translation
Whisper can translate into English as it transcribes; Parakeet cannot. If you dictate in one language and want English typed, that decides it.
What to do
- Default to Parakeet V3. Fastest, most accurate for English and its 25 languages, 484 MB.
- Add a Whisper Large V3 Turbo mode for any language Parakeet lacks, or for translate-to-English.
- Add Cohere Transcribe if you dictate long, careful text in one of its 14 languages and can wait a couple of seconds a take.
- Skip Whisper Small unless disk is tight; Parakeet is smaller and better in every language they share.
Questions people ask
Is Parakeet more accurate than Whisper?
For English, yes: 6.3% against 7.4% to 8% average WER. For languages outside Parakeet's 25, Whisper is the only option of the two.
Why is Parakeet so much faster?
It is a token-and-duration transducer rather than an autoregressive decoder, and it runs on the Neural Engine. It does less work per word.
Can I use both?
Yes. In FreeFlow each mode has its own model, so a Parakeet mode and a Whisper mode sit a menu-bar click apart.
Does either need the internet?
No. Both run as CoreML models on your Mac after a one-time download.