Speech for Dart

Native Dart bindings for Apple Speech (SpeechAnalyzer pipeline) using Dart Build Hooks.

speech_kit exposes SpeechAnalyzer, SpeechTranscriber, DictationTranscriber, SpeechDetector, and AssetInventory from Dart with a small façade—not the legacy SFSpeechRecognizer / task-based API. Native code is compiled via hooks and bridged with FFI.

Features

• Permissions — Speech recognition and microphone status / requests (native checks for usage-description keys where applicable)
• Asset inventory — AssetInventory.status, download/install, and bestAvailableAudioFormat for configured modules
• Analysis sessions — analyzeFile, analyzePcm, analyzePcmStream with optional AnalysisContext, SpeechAnalyzerOptions, and prepareToAnalyze-style preparation with progress
• Custom language model (training data) — exportCustomLanguageModelData with phrase counts, custom pronunciations, and optional PhraseCountsFromTemplatesConfig (template / compound tree)
• Custom language model (compile) — prepareCustomLanguageModel from exported training data; supportedCustomLanguagePhonemes per locale
• Dictation integration — SpeechLanguageModelPaths for SFSpeechLanguageModel.Configuration on dictation modules
• Powered by Dart Build Hooks — No Flutter dependency (pure Dart on supported hosts)

Platform Support

Platform	Support
macOS	✅
Windows	❌
Linux	❌
iOS	❌
Android	❌

macOS is the supported target for the native dylib in this release (platforms in pubspec.yaml is macOS-only). The module APIs used here require recent Apple OS releases (e.g. macOS 26+ for the current Swift bridge, consistent with iOS 26+ / visionOS 26+ for the same Speech symbols). Callers need dart:io (VM or an embedder that provides dart.library.io); non-IO configurations use stubs that throw UnsupportedError.

Installation

Add the package:

dart pub add speech_kit

From a local path:

dependencies:
  speech_kit:
    path: ../speech-kit

Example

import 'package:speech_kit/speech_kit.dart';

Future<void> main() async {
  const kit = SpeechKit();
  final speech = await kit.speechRecognitionAuthorizationStatus();
  // Request permission if needed, build SpeechModuleConfiguration list,
  // ensure assets, then analyzeFile / analyzePcm / analyzePcmStream.
}

Usage flow

1. Check permissions — speechRecognitionAuthorizationStatus() / requestSpeechRecognitionPermission() and microphone APIs as needed.
2. Describe modules — Build a List<SpeechModuleConfiguration> (transcriber, dictation, optional speech detector).
3. Assets — assetInventoryStatus / ensureAssetsInstalled; use bestAvailableAudioFormat when preparing audio.
4. Analyze — analyzeFile (or PCM / stream) with optional AnalysisContext, SpeechAnalyzerOptions, and preparation callbacks.
5. Optional: custom LM — exportCustomLanguageModelData → prepareCustomLanguageModel; attach compiled paths via dictation configuration when required.
6. Retention — Call endSpeechModelRetention when you no longer need models kept after analysis (if you used lingering / process-lifetime retention options).

For CLI samples (permission_status, asset_inventory, analyze_file), see example/README.md.

Architecture

This package uses Dart Build Hooks to compile native code and bridge Apple’s Speech APIs to Dart.

Dart layers (inward dependencies): domain (value objects, errors) → application (SpeechKit façade, SpeechAnalysisSession) → infrastructure (macOS FFI + speech_kit_no_io stub). The public barrel exports domain types and the façade.

Dart (barrel + SpeechKit)
 │
 │  FFI
 ▼
Native bridge (Swift)
 │
 ▼
Speech (Apple framework)

This keeps the Dart surface small while delegating OS work to the native library.

Example app

See the example/ directory for command-line samples and macOS bundle notes. Full instructions: example/README.md.

Roadmap

Major capability areas implemented today are listed under Features: permissions, asset inventory, file/PCM/stream analysis with context and analyzer options, custom language model export (including template-based phrase counts), compile step, supported phonemes, and dictation LM paths.

Partially exposed or not exposed in the Dart API includes anything not covered in .cursor/skills/speech-kit-api-coverage/SKILL.md. iOS FFI is not wired in the current dart:io path; macOS is the supported desktop target for this release.

Additional documentation

• Contributing — setup, tests, and PR guidelines
• Domain model — layers and value objects
• Apple Speech — Speech framework, SpeechAnalyzer, SFCustomLanguageModelData, advanced speech-to-text sample