Choosing engines and models
How the "AI (Apple Intelligence)" and "Standard" transcription engines, Apple Intelligence for summaries, and the external WhisperKit and MLX models Pro differ, and which to pick. Downloading and deleting models. Everything runs on the device.
What you start with
Out of the box, Otonota uses these engines. Both are Apple’s own and free.
| Role | Default engine | Notes |
|---|---|---|
| Transcription | AI (Apple Intelligence) | The first time you record in a language, iOS downloads the speech model for it |
| Summaries | Apple Intelligence | No download. Uses the device’s Apple Intelligence |
Choose the transcription engine with Recognition Engine in Settings, and the summary engine with Summary Engine in Settings or Summary Settings in the summary column. Summary Engine appears only with Pro (without it, Apple Intelligence is the only one).
Transcription engines
| Engine | Edition | What it’s like |
|---|---|---|
| AI (Apple Intelligence) | Free | The default. Apple’s newer speech recognition (SpeechAnalyzer). Stable even in long meetings. Downloads a model per language the first time. Supports Mask Offensive Words |
| Standard | Free | Apple’s earlier speech recognition. No download, ready right away. Only languages the device can recognize on its own. Asks for Speech Recognition permission the first time |
| WhisperKit large-v3 (quantized) | Pro | High accuracy. Multilingual. About a 0.6 GB download. Devices with 6 GB of memory or more |
| WhisperKit small | Pro | Light and fast. Multilingual. About a 0.5 GB download. Devices with 4 GB of memory or more |
- The Language list in Settings shows only the languages the selected engine supports. For the two built-in engines, that depends on the device and iOS
- WhisperKit offers Japanese, English, Chinese, Korean, Spanish, French, German, Italian, Portuguese, Russian, Dutch, Swedish, Polish, Turkish, Arabic, Hindi, Vietnamese, Thai and Indonesian
- Unlike on the Mac, “Standard” never falls back to Apple’s servers for a language the device can’t recognize on its own. In that case it doesn’t transcribe and tells you why
- A new transcription engine takes effect from the next recording
Summary engines
| Engine | Edition | What it’s like |
|---|---|---|
| Apple Intelligence | Free | The default. No download. Lets you choose Safety Guardrails |
| MLX Qwen3.5 2B | Pro | Lightweight summary model. About a 1.6 GB download. Devices with 6 GB of memory or more |
| MLX Qwen3.5 4B | Pro | More thorough summaries. About a 3 GB download. Devices with 8 GB of memory or more |
- Whatever the engine, summaries are written only while Otonota is in the foreground (iOS doesn’t let background apps run that kind of work)
- An MLX model is loaded into memory when recording starts, which takes from a few seconds to over ten
- You can change the summary engine mid-recording; it applies from the next summary
- The meeting wrap-up (Pro) is always written by Apple Intelligence, whatever the summary engine
Which to pick
- Start with the defaults — “AI (Apple Intelligence)” and Apple Intelligence. The only download is the speech model for your language
- Record right away without downloading a language model — set transcription to “Standard”. It covers fewer languages
- Transcribe a language the built-in engines don’t support Pro — WhisperKit. If a language isn’t in the Language list, switch the engine to WhisperKit first, then choose it
- More accurate transcripts Pro — WhisperKit large-v3 (quantized). On devices with less memory, small
- Summaries in a language Apple Intelligence doesn’t support Pro — MLX. 2B is lighter; 4B is more thorough
- Ordinary conversation gets refused — before switching engines, check that Safety Guardrails in Settings is “Relaxed (for summaries and translation)”
External models use a lot of memory. Otonota uses the setting that raises the memory limit iOS gives an app, but a model is left out of the list on devices without enough memory for it.
| Device memory | External models you can choose |
|---|---|
| 4 GB | WhisperKit small |
| 6 GB | The above, plus WhisperKit large-v3 (quantized) and MLX Qwen3.5 2B |
| 8 GB or more | The above, plus MLX Qwen3.5 4B |
Downloading a model Pro
External Models in Settings lists the models that work on your device.
- Tap the ↓ button next to a model. On cellular, Otonota first asks “Download over cellular?” and shows the size. Wi-Fi is recommended
- A bar shows progress. To stop, tap the × button
- When it’s done, the size on the device appears, and the model can be chosen under Recognition Engine or Summary Engine in Settings
- Downloads run one at a time. Others you tap wait as “Waiting”
- If there isn’t enough free space, Otonota says “Not enough free space.” before starting, with how much is needed (a little more than the model itself)
- You can’t download while recording. Starting a recording stops a download in progress; tap it again afterwards
- Models are downloaded from where they’re published (Hugging Face). On networks that block it, such as some company networks, the download fails
- The speech models for “AI (Apple Intelligence)” are separate from external models; iOS downloads them when you start recording
Deleting a model
- In External Models, tap the trash button next to a model, then Delete. To use it again you’ll need to download it again
- A model being used for recording or summaries shows “In Use” and can’t be deleted
- If you delete the model you had selected, transcription goes back to “AI (Apple Intelligence)” and summaries to Apple Intelligence
- If Pro lapses, downloaded models can still be deleted from External Models
Models are kept inside the app and aren’t included in device backups.
What leaves the device
Transcription and summaries are processed on the device with every engine. No engine sends audio or text to a server. Otonota goes online only to download models (speech models, external models, the speaker separation model) and to buy or restore through the App Store.
