Models

Loading ahead of timeCopy Anchor Link

Models load on first use. To download and load them before the user needs them, for example during onboarding:

try await Scribe.shared.prepare(diarization: true) { progress in
    let percent = progress.fraction.map { " \(Int($0 * 100))%" } ?? ""
    print("\(progress.model.rawValue):\(percent) \(progress.detail)")
}

• diarization: false loads only the speech model.
• ModelProgress.model is .speechRecognition or .speakerDiarization.
• fraction is 0 to 1 when known. The speaker model reports nil while loading and 1 when ready.
• detail is a short status, such as "Downloading model (3/7 files)…".

prepare is safe to call repeatedly, including while a load is running: concurrent callers wait for the same load.

Check the state with isSpeechModelLoaded and isDiarizationModelLoaded.

Freeing memoryCopy Anchor Link

await Scribe.shared.unload()

The next transcription loads the models again from the disk cache.

Speaker countsCopy Anchor Link

Passing speakerCount configures a separate diarizer for that count. It is created on first use and kept, so later calls with the same count are fast.

CacheCopy Anchor Link

In a sandboxed app the path is inside the app's container. Delete the folders to force a new download.

HardwareCopy Anchor Link

The speech encoder runs in 8-bit precision on the Apple Neural Engine. On Apple Silicon, transcription runs many times faster than real time.