SpeechKit ASR

On-device speech recognition SDK for iOS 18+ / macOS 15+. Runs Audio8-ASR-0.1B locally: Apple Neural Engine for the audio encoder, CPU int4 for the language-model decoder. No network request is required by the SDK.

The public demo also exposes process memory footprint, peak footprint, CPU, thermal state, battery state, and a rough whole-device power estimate for device-side validation.

Installation

1. Add the package

dependencies: [
    .package(path: "../SpeechKit"),   // or the git URL + version once hosted
]
targets: [
    .target(name: "YourApp", dependencies: [
        .product(name: "SpeechKit", package: "SpeechKit"),
    ]),
]

2. Add the model bundle

Build it once (python3 Scripts/make_asset_bundle.py) -> dist/ASRModels.bundle (~437 MB, models pre-compiled to .mlmodelc). Drag the bundle into your app target as a folder reference (blue icon), or download it on first launch and pass its URL to ASRTranscriber(bundleURL:).

3. Info.plist

Only needed if you capture the microphone yourself: NSMicrophoneUsageDescription.

Usage

One-shot transcription

import SpeechKit

let transcriber = try ASRTranscriber()      // finds ASRModels.bundle in main bundle
transcriber.warmUp()                        // optional; off main thread

let result = try transcriber.transcribe(samples)   // [Float], 16 kHz mono
print(result.text, result.timings.total)

// Async, with cancellation:
let token = ASRTranscriber.CancellationToken()
let result = try await transcriber.transcribe(samples, cancellation: token)

// Any audio file:
let result = try transcriber.transcribeFile(url)

Streaming and VAD are not part of this ASR-only build. Capture an utterance, convert it to 16 kHz mono Float32, then call transcribe.

Error handling

Every API throws SpeechError (SpeechCore). Cases cover asset validation (assetCorrupted, assetSchemaUnsupported), audio input (audioTooShort), inference (inferenceFailed, promptTooLong), and cancelled.

Pass Options.verifyAssets = true on the first launch after install/update to SHA-256-check every asset (~1-2 s), then persist a flag and skip it.

Threading contract

ASRTranscriber methods are thread-safe; inference is serialized internally.
Synchronous transcribe blocks; never call from the main thread.
Async callbacks resume from an internal queue; dispatch to main for UI.

Constraints


Language	Model-defined
Audio window	30 s max per call
First launch	ANE compilation ~30-60 s (system-cached afterwards); `warmUp()` recommended
Memory	~600 MB resident with models loaded
Min OS	iOS 18 / macOS 15 (multifunction Core ML models)

Regression testing

swift run dev-check                              # lightweight local smoke check
swift test                                       # full Xcode recommended
swift run -c release asrkit-cli <workspace-dir>   # numeric regression vs Python refs
swift run -c release asrkit-cli <workspace-dir> --stress   # 100-pass memory check
swift run -c release asrkit-cli <workspace-dir> --file /path/to/audio.wav --repeat 10

dev-check does not require model assets and is the recommended first check on a fresh machine. Some Command Line Tools-only installations cannot expose the Apple Testing/XCTest frameworks cleanly; install and select full Xcode before using swift test.