On-device transcription
Meetly transcribes on your device. There is no transcription server: the audio never leaves your iPhone, iPad, or Mac. Three engines cover the world's languages, and the app picks the right one for your device so you never have to.
The three engines
Settings shows one choice: which engine to use. Everything below it, including which Whisper model your device gets, is decided by the app.
Fast · recommended
NVIDIA's Parakeet model for 25 European languages including English. Quick on every supported device, with speaker labels.
Global · 99 languages
OpenAI's Whisper for everything else, from Hindi to Japanese to Arabic. Slower than Fast. Your device gets one model: Small on 8 GB iPhones, iPads and every Mac, Base on the rest.
Built-in · English
The recognizer already on your device. English only, nothing to download, ready the moment you allow Speech Recognition.
Speaker labels come from a separate 21.6 MB on-device model (FluidAudio 0.15.7, CC BY 4.0) that installs with Fast or Global. It labels voices as Speaker 1, Speaker 2, and so on, and can recognise you from an optional My Voice enrollment.
One model per device
The Global engine runs the largest Whisper model your device handles comfortably. Meetly reads the device's memory class and picks once; there is no variant picker to get wrong.
| Device | Memory | Global model | Default engine |
|---|---|---|---|
| iPhone 15 Pro and later, iPhone 16 and 17, iPad Pro and iPad Air with M-series chips | 8 GB or more | Whisper Small (491 MB) | Fast |
| iPhone 12 to 15, iPhone 16e, recent iPads | 6 GB | Whisper Base (151 MB) | Fast |
| iPhone SE, iPhone 11, iPhone 13 mini, older iPads | 4 GB or less | Whisper Base (151 MB) | Built-in |
| Every Apple-silicon Mac | 8 GB or more | Whisper Small (491 MB) | Fast, or Global on 16 GB Macs |
Not yet tested on your device? New model and device combinations are marked "Device testing is pending" in the app until we have measured them. You can still download with confirmation; an installation that already works is never replaced.
Languages
No language setting to choose. Fast covers the 25 outlined languages natively. Global hears all 99.
afsqamarhyasazbnbaeubebsbrbgmycazhhrcsdanlenetfofifrglkadeelguhthahawhehihuisiditjajwknkkkmkololalvlnltlbmkmgmsmlmtmrmnminenonnocpsfaplptparorusasrsnsdsiskslsoessuswsvtltgtatttethbotrtkukuruzvicyyiyoGlobal output is English for now. When the speech is not English, the Whisper models Meetly ships write the transcript in English (Whisper's translate mode). We measured the alternatives and native-script output in this model class was not reliable enough to ship; transcribing Hindi, Japanese, Arabic and the rest in their own script is the next transcription pass. English speech stays English on every engine.
How it works
| Where transcription runs | On the device, with the Neural Engine. Recording audio is never uploaded for transcription. See where your data goes. |
|---|---|
| Downloading a model | Once, after you confirm the size. Files are verified against a pinned manifest before use; a download that fails verification is discarded, never loaded. |
| Storage | Settings → Storage & Models lists every installed model with its size and lets you remove any of them. A recording that needs a removed model waits until it is installed again. |
| Long recordings | Whisper decodes a recording in chunks with two decoder workers on iPhone, and retries a chunk if the Neural Engine times out, so an hour-long recording finishes instead of failing at the end. |
| Changing engines | Settings → Transcription. Existing transcripts are kept; Recompute on a meeting re-runs it with the current engine. |
Model inventory from Meetly's pinned transcription catalog, version 2026-09-13.1. Whisper models are the WhisperKit Core ML conversions (MIT); Parakeet TDT 0.6B v3 and the speaker model are FluidAudio Core ML conversions (CC BY 4.0). Model attributions ship inside the app.
Questions about a language or device? Contact us.