Speech Models
Wispen can use several speech recognition models. They all run on your device. Apple's engines are built in; the others are optional downloads.
Choose a model in Settings › Speech Models › Speech Model. Tap Use on a downloaded model, or tap the size button to download one. The model in use shows a checkmark (iPhone and iPad) or In Use (Mac). Each model is tagged English only or Multilingual, with its engine.
Available models
| Model | Size | Languages | Notes |
|---|---|---|---|
| Apple Speech (SpeechAnalyzer) | Built-in | Multilingual | The default where your device supports it. Apple downloads your language the first time you use it. |
| Apple Neural Speech Engine | Built-in | Multilingual | Apple's classic engine. Used as the default on devices without SpeechAnalyzer. |
| NVIDIA Parakeet v2 (English) | 570 MB | English | Very accurate and fast enough to keep up with live dictation. |
| NVIDIA Parakeet Ultra (25 Languages) | 720 MB | English and 24 European languages | Detects the language automatically. Doesn't cover Japanese, Chinese or Korean. |
| Whisper Large-v3 Turbo | 626 MB | Multilingual | OpenAI's most accurate Whisper model here. Uses the most memory. |
| Whisper Large-v3 Turbo (ANE Optimized) | 632 MB | Multilingual | Large-v3 Turbo, compressed to use less memory. |
| Whisper Distil-Large-v3 | 756 MB | English | A smaller, faster version of Large-v3. |
| Whisper Distil-Large-v3 (Compact 594MB) | 594 MB | English | Distil-Large-v3, compressed to use less memory. |
| Whisper Small | 240 MB | Multilingual | A balance of speed and accuracy. |
| Whisper Base | 145 MB | Multilingual | Fast and light, but less accurate. |
| Whisper Tiny | 75 MB | English | The smallest and fastest, and the least accurate. |
Downloads come from Hugging Face and stay on your device until you delete them. You need an internet connection to download a model, but not to use it.
TIP
If you dictate in a language other than English, pick a multilingual model. Settings › Dictation warns you when your model doesn't support the language you picked.
Parakeet and your dictionary
When you download a Parakeet model, Wispen also downloads a small helper model (NVIDIA Parakeet TDT-CTC 110M) so Parakeet can spell the vocabulary in your dictionary correctly. If that extra download fails, Parakeet still works, and Wispen tries again the next time you use the dictionary.
Transcript Cleanup
Settings › Speech Models › Transcript Cleanup offers S1-mini by Superwhisper, an optional English cleanup model.
- Tap Download (462 MB). You can Cancel while it downloads.
- When it's done, Use for English Dictation turns on. It adds a moment after you stop speaking.
S1-mini cleans up finished English dictations in the Wispen app: it removes fillers, resolves corrections like "no wait, make that Thursday", and writes numbers, dates and punctuation properly. Your replacements, snippets and dictation mode still apply after it. With it off, Wispen uses its built-in cleanup rules.
On iPhone and iPad, S1-mini isn't used for Wispen Keyboard dictations, because iOS doesn't let it run in the background. Those use the built-in cleanup.
To remove it, tap Delete and confirm Delete Model. This frees 462 MB.
Advanced settings
| Setting | Default | What it does |
|---|---|---|
| Load Model When Wispen Opens | On | For Whisper and Parakeet: the first dictation starts faster, at the cost of memory. |
| Fast Whisper Decoding | On | Whisper only: text appears sooner when you pause. |
| Speech Detection | 0.14 | How loud you need to be to count as speaking, from 0.05 to 0.30. Raise it in noisy rooms. |
| Downloaded Models | Shows how much space your downloaded speech models use. Delete All removes them. Apple's built-in speech recognition keeps working. |
Delete one model
On iPhone and iPad, touch and hold a downloaded model. On the Mac, right-click it. Then choose Delete Download. You can't delete the model that's in use; switch to another one first.
Switch models quickly
- On the Mac or iPad with a keyboard, press ⌘K and pick a speech model.
- On the Mac, right-click the Wispen menu bar icon and use Dictation Model.