Whisper models aren't the only option for offline transcription. Whisper Notes has six speech models that run offline:
- Parakeet (built in)
- Whisper Large V3 Turbo
- Whisper Small
- Qwen3-ASR 1.7B (Mac)
- SenseVoice Small
- Apple Speech (macOS 26 or later on Mac, iOS 26 on iPhone)
Which one is most accurate depends on the language you speak.
Which model is most accurate?
We tested every model on the same MacBook Air (M5). Error rates (lower is better; the lowest in each row is in green):
| Language | Parakeet | Whisper Large V3 Turbo1 | Whisper Small | Qwen3-ASR 1.7B1 | SenseVoice Small | Apple Speech |
|---|---|---|---|---|---|---|
| English | 5.89% | 5.49% | 7.04% | 4.49% | 8.46% | 8.80% |
| German | 3.72% | 4.05% | 8.67% | 2.85% | — | 6.26% |
| Dutch | 6.91% | 5.69% | 16.75% | 7.36% | — | — |
| Japanese | — | 5.30% | 11.37% | 5.74% | 6.78% | 6.36% |
| Polish | 6.34% | 5.45% | 15.81% | 12.59% | — | — |
| Russian | 5.72% | 5.13% | 11.37% | 5.65% | — | — |
| French | 4.95% | 5.97% | 13.24% | 4.57% | — | 7.61% |
| Spanish | 3.77% | 3.62% | 6.22% | 3.38% | — | 5.41% |
| Korean | — | 4.25% | 7.30% | 3.54% | 7.85% | 4.31% |
| Ukrainian | 5.11% | 6.33% | 17.48% | — | — | — |
| Chinese | — | 7.97% | 20.50% | 6.49% | 7.69% | 7.97% |
| Cantonese | — | —2 | —2 | 6.44% | —2 | 8.69% |
| Italian | 2.48% | 2.58% | 7.64% | 2.58% | — | 4.39% |
| Portuguese | 5.69% | 5.44% | 8.61% | 5.17% | — | 9.35% |
| Swedish | 13.12% | 8.78% | 21.66% | 19.80% | — | — |
| Turkish | — | 5.95% | 14.24% | 8.81% | — | — |
| Czech | 10.85% | 11.94% | 38.71% | 24.59% | — | — |
| Norwegian | — | 8.99% | 24.50% | — | — | — |
| Finnish | 10.03% | 7.11% | 25.91% | 27.71% | — | — |
| Arabic | — | 15.72% | 32.81% | 14.32% | — | — |
| Slovak | 8.58% | 9.59% | 32.43% | — | — | — |
| Danish | 15.68% | 13.63% | 34.14% | 22.24% | — | — |
| Hindi | — | 29.64% | 61.26% | 13.19% | — | —3 |
| Thai | — | 13.40% | 23.52% | 6.53% | — | — |
| Indonesian | — | 6.31% | 16.31% | 4.91% | — | — |
| Hungarian | 13.63% | 14.59% | 40.65% | 36.63% | — | — |
| Vietnamese | — | 8.10% | 18.12% | 4.91% | — | — |
| Malay | — | 9.22% | 19.72% | 10.62% | — | — |
| Greek | 20.96% | 13.94% | 31.64% | 30.22% | — | — |
| Romanian | 9.54% | 8.36% | 29.45% | 20.20% | — | — |
| Filipino | — | 9.72% | 25.60% | 23.31% | — | — |
| Bulgarian | 12.28% | 14.54% | 39.68% | — | — | — |
| Slovenian | 18.49% | 19.15% | 48.54% | — | — | — |
| Estonian | 15.40% | 19.08% | 50.84% | — | — | — |
| Croatian | 12.49% | 11.49% | 33.76% | — | — | — |
| Lithuanian | 18.00% | 23.28% | 67.23% | — | — | — |
| Maltese | 19.19% | 74.35% | 95.68% | — | — | — |
| Icelandic | — | 19.75% | 73.96% | — | — | — |
| Catalan | — | 5.80% | 15.51% | — | — | — |
| Galician | — | 12.23% | 31.09% | — | — | — |
| Macedonian | — | 17.46% | 37.11% | 18.57% | — | — |
| Bosnian | — | 14.85% | 39.60% | — | — | — |
Character error rate for Chinese, Japanese, Korean, Cantonese and Thai, word error rate for the rest. Languages not in the table: see Supported Languages.
1 Mac only. 2 We don't list SenseVoice Small's or Whisper's Cantonese scores yet: we haven't confirmed that this test scores them fairly. 3 Apple Speech wrote all of its Hindi in English letters, so it can't be scored.
Which is fastest? How long does an hour of audio take?
| Model | Speed (× real time) | 1 hour of audio | Download | Memory |
|---|---|---|---|---|
| SenseVoice Small | 115–131× | about 30 s | about 450 MB | about 700 MB |
| Parakeet | 75.8× | about 50 s | Built in (about 630 MB) | about 800 MB |
| Apple Speech | 19.7× | about 3 min | Managed by macOS | — |
| Qwen3-ASR 1.7B | 7.6× | about 8 min | about 2.5 GB | about 2.6 GB |
| Whisper Small | 6.5× | about 9 min | about 400 MB | about 1 GB |
| Whisper Large V3 Turbo | 2.7× | about 22 min | about 2 GB | about 2.4 GB |
Speed, download and memory as shown on each model's card in the Mac app. Speed counts model processing time only; the hour figures are worked out from it.
The most accurate models are often the slowest, and the longer the recording, the more that matters.
Which models can I use on iPhone and Mac?
| Model | iPhone | Mac |
|---|---|---|
| Parakeet | Yes, built in | Yes, built in |
| Whisper Large V3 Turbo | No | Yes |
| Whisper Small | Yes | Yes |
| Qwen3-ASR 1.7B | No | Yes (Beta) |
| SenseVoice Small | Yes | Yes |
| Apple Speech | Yes (iOS 26) | Yes (macOS 26 or later) |
Green means available, amber means not available.
We want more models on iPhone too. The most requested one, Whisper Large V3, only runs on the newest few iPhones, whether full-size or compressed; most phones can't run it. The same goes for Qwen3-ASR 1.7B. We keep testing new models and smaller versions of the ones we have. Most recently, we switched the default model from Parakeet V3 to Parakeet Ultra, which makes about 24% fewer mistakes across 25 languages. See Parakeet Ultra vs Parakeet V3.
How do I switch models?
On Mac, open Settings and click a card under Transcription Model to download or switch to it; the Supported Languages link on each card lists that model's error rate for every language.
Settings → Transcription Model in the Mac app
How did we test?
We tested every model on the same MacBook Air (M5) with FLEURS read speech: 83 language sets and 12,090 sentences. Our Whisper results closely match the numbers OpenAI publishes. This is clean read speech; meetings and noisy recordings will have higher error rates.
Frequently Asked Questions
What is the best Whisper model?
Whisper Large V3 Turbo. It is more accurate than Whisper Small in every language in the table, for example 5.49% against 7.04% in English, but it is much slower, and the iPhone app only has Whisper Small.
What is the best model for Japanese?
Whisper Large V3 Turbo, at a 5.30% character error rate, followed by Qwen3-ASR 1.7B at 5.74%. On iPhone, use SenseVoice Small (6.78%).
What is the best model for Mandarin Chinese?
Qwen3-ASR 1.7B on Mac, at a 6.49% character error rate. Next is SenseVoice Small at 7.69%, which also works on iPhone.
Which Parakeet model is best?
Parakeet Ultra. Across 25 languages it makes about 24% fewer mistakes than Parakeet V3 (10.63% against 13.99%), and it is built into the iPhone app and both Mac versions.
Detailed benchmarks