vad

Diagnostics & Crash-Reporting Improvements

Adds an onLog callback to VadHandler.create for routing the package's production log lines into your app's crash reporter (Crashlytics, Sentry, Bugsnag, etc.), plus an always-on breadcrumb logged immediately before each ONNX
Runtime session creation on native platforms. Useful for diagnosing native crashes inside libonnxruntime.so on low-end Android devices.

Highlights:

• Add optional VadLogCallback? onLog parameter to VadHandler.create and matching VadLogCallback typedef exported from package:vad/vad.dart
• Emit always-on VadModel: creating OrtSession (model=... bytes=... sampleRate=... os=... osVersion=...) breadcrumb before each native session creation — the last line in logcat before any crash inside libonnxruntime.so
• Route all non-debug failures through onLog when set: session-creation errors (native + web), inference errors, frame-processing errors, audio-stream and microphone-permission failures
• README: add Reporting Issues troubleshooting section with adb logcat capture instructions and a Crashlytics wiring example

No breaking changes — onLog is optional and defaults to current behavior (print/logcat).

Full Changelog:
See CHANGELOG.md

• Add onLog callback to VadHandler.create for routing production log lines to consumer-side telemetry (Crashlytics, Sentry, Bugsnag, etc.) without zone or debugPrint overrides
  • API: Add optional VadLogCallback? onLog parameter to VadHandler.create
  • API: Add VadLogCallback typedef (void Function(String message)) exported from package:vad/vad.dart
  • Core: Thread onLog through VadIterator.create → VadInference.create → native/web inference factories → Silero v4/v5 models
  • Coverage: Route all non-debug failure log lines through the callback when provided — model session-creation errors (native + web), inference errors, frame-processing errors, audio-stream and microphone-permission failures
  • Behavior: When onLog is null, log lines continue to go to print/logcat as before. Verbose if (isDebug) print(...) state-transition logs are intentionally not routed through the callback
• Add always-on breadcrumb log immediately before each ONNX Runtime session creation on native platforms
  • Native: Emit VadModel: creating OrtSession (model=... path=... bytes=... sampleRate=... intraOp=... interOp=... os=... osVersion=...) from both SileroV4Model.create and SileroV5Model.create
  • Diagnostics: Lands in logcat as the last line before any native crash inside libonnxruntime.so, so crash reports always carry the context needed to diagnose device-specific failures (especially on low-end 32-bit armeabi-v7a Android devices)
  • Enrich the catch-block error message with the same context fields and stack trace
• Documentation
  • README: Document the new onLog parameter on VadHandler.create
  • README: Add a "Reporting Issues" troubleshooting section with adb logcat capture instructions and a Crashlytics/Sentry/Bugsnag wiring example for onLog

Major Update: Android 16KB Page Size Support & Custom Audio Streams

Highlights:

• ✅ Add Android 16KB page size support - Full compatibility with Android 15+ and Google Play requirements
• 🚀 Now using Microsoft ONNX Runtime 1.22.0 directly via FFI - Replaced third-party wrapper for better performance
• 🎤 Add support for custom audio streams - Provide your own audio stream for advanced use cases
• 🐛 Fix AudioRecorder disposal and recreation issue - Prevents "Recorder has already been disposed" error
• 🖥️ Add desktop platform support - Windows, macOS, and Linux now supported
• 🌐 Rewrite web VAD in pure Dart - Better browser compatibility and performance

Breaking Changes:

• ⚠️ Asset management: Models now loaded from CDN by default (can override for offline use)
• ⚠️ Model identifier renamed from 'legacy' to 'v4'
• ⚠️ onEmitChunk stream type changed to include isFinal flag

Full Changelog:
For detailed changes, see CHANGELOG.md

• Add Android 16KB page size support
  • Android: Bump ONNX Runtime to 1.22.0 which includes native 16KB page size support
  • Android: All native libraries (libonnxruntime.so, libonnxruntime4j_jni.so) are now properly aligned for 16KB page sizes
  • Android: Plugin is now fully compatible with Android 15+ devices and Google Play's 16KB page size requirement
• BREAKING CHANGE: Asset management changes - models now loaded from CDN by default
  • Core: Remove bundled assets from package to reduce package size
  • Core: Introduce companion NPM package @keyurmaru/vad to host ONNX model files via jsDelivr CDN
  • Core: Update baseAssetPath and onnxWASMBasePath parameters to default to CDN URLs (can be overridden for offline/self-hosted use)
  • Core: Delete lib/assets directory and its contents
  • Migration: For offline support, download model files and set baseAssetPath parameter in startListening()
• BREAKING CHANGE: Model identifier renamed from 'legacy' to 'v4'
  • API: Rename model parameter value from 'legacy' to 'v4' for clarity
  • Migration: Update model: 'legacy' to model: 'v4' in startListening() calls
• Internal architectural refactor to unify cross-platform implementation
  • Internal: Remove platform-specific internal classes (VadHandlerWeb, VadHandlerNonWeb, VadIteratorWeb, VadIteratorNonWeb)
  • Internal: Remove internal abstract base classes (VadHandlerBase, VadIteratorBase)
  • Internal: Introduce VadInference abstraction layer as platform split point
  • Core: Use record package for cross-platform audio capture, replacing custom web audio implementation
  • Example: Rename RecordingModel enum to VadModel in example app
• Internal native implementation overhauled with FFI for better performance
  • Internal: Replace onnxruntime package dependency with direct FFI-based implementation
  • Internal: Add ONNX Runtime C API headers and generate Dart bindings using ffigen
  • Internal: Create Dart wrappers for ONNX structs (OrtEnv, OrtSession, OrtValue)
  • Internal: Implement OrtIsolateSession to run native inference in separate isolate
  • Internal: Restructure lib/ into core and platform directories
  • Internal: Introduce abstract VadModel class to unify inference logic
• Add desktop platform support (Windows, macOS, Linux)
  • Platform: Add build configurations for Windows, macOS, and Linux desktop applications
  • Platform: Bundle pre-compiled x64 and arm64 ONNX Runtime binaries for Windows and Linux
  • Platform: Add CMakeLists.txt for each desktop platform to handle binary packaging
  • Platform: Add macOS podspec with dependency on onnxruntime-objc pod
  • Platform: Update Dart FFI bindings to dynamically detect OS and CPU architecture at runtime
  • Example: Add macOS example app
  • Example: Add Linux example app
• Rewrite web VAD implementation in pure Dart
  • Web: Remove pre-compiled JavaScript bridge (vad_web.js, bundle.min.js)
  • Web: Implement pure Dart web support using dart:js_interop to directly communicate with onnxruntime-web
  • Web: Add MicVAD class in lib/src/web/audio_node_vad.dart to manage audio pipeline using AudioContext and dynamically generated AudioWorkletProcessor
  • Web: Add typed wrappers for onnxruntime-web in lib/src/web/onnx_runtime_web.dart
  • Web: Add ScriptProcessorNode fallback from AudioWorklet for better browser compatibility (older browsers and non-secure contexts)
• BREAKING CHANGE: onEmitChunk stream type changed to include isFinal flag
  • API: Change onEmitChunk from Stream<List<double>> to Stream<({List<double> samples, bool isFinal})>
  • API: Add onEmitChunk stream for real-time audio chunk emission during active speech
  • API: Add isFinal flag to mark the last chunk of a speech utterance
  • API: Add numFramesToEmit parameter to startListening() to enable chunking (default: 0, disabled)
  • API: Add endSpeechPadFrames parameter to startListening() to control audio padding at speech end (default: 1 for v4, 3 for v5)
  • Core: Update VadIterator implementations to manage frame buffers and emit chunk events periodically and at speech end
  • Example: Update example app to demonstrate chunk emission feature with playback UI for individual chunks
  • Migration: Update listeners from vadHandler.onEmitChunk.listen((samples) { ... }) to vadHandler.onEmitChunk.listen((chunk) { final samples = chunk.samples; final isFinal = chunk.isFinal; ... })
• Network model loading support
  • Core: Update native implementation to support loading models from network URLs using HttpClient
• Fix Android audio playback interference with VAD detection
  • Example: Change audio player configuration to use AndroidUsageType.media instead of voiceCommunication
  • Example: Add AndroidAudioFocus.gainTransientMayDuck to prevent recorder from receiving AUDIOFOCUS_LOSS when playing back recordings
  • Example: Allow continuous speech detection during playback without requiring manual stop/start
• Fix iOS minimum version requirement
  • Platform: Correct minimum iOS version from 16.0 to 15.1 in vad.podspec to align with onnxruntime-objc dependency support
  • Example: Update iOS example app configuration to match 15.1 minimum deployment target
• Use forked record dependency for macOS echo cancellation fix
  • Dependencies: Temporarily override record and record_macos packages to point to git repository with echo cancellation fix
  • Dependencies: Will be removed once fix is merged into official release
• Modernize example app build configurations
  • Example: Migrate Android Gradle scripts from Groovy to Kotlin DSL (.kts)
  • Example: Upgrade Gradle wrapper from 8.3 to 8.12
  • Example: Bump Java compatibility to version 11
  • Example: Update Android package name to com.example.vad_example
  • Example: Clean up iOS Podfile, removing obsolete settings
  • Example: Update Xcode project files to match new dependencies
• Update package metadata
  • Pubspec: Add homepage and issue_tracker fields
• Example: Expose RecordConfig from record package for detailed audio input configuration
• Example: Update dependencies
  • Bump permission_handler to latest version
  • Bump audioplayers to latest version
• Add support for custom audio streams
  • API: Add optional Stream<Uint8List>? audioStream parameter to startListening() method
  • Core: Allow users to provide their own audio stream instead of using the built-in recorder
  • Core: When custom stream is provided, VadHandler bypasses internal AudioRecorder setup
  • Core: Custom stream should provide PCM16 audio data at 16kHz sample rate, mono channel
  • Example: Add CustomAudioStreamProvider demonstration class using the record library
  • Example: Add "Use Custom Audio Stream" toggle in settings dialog
  • Example: Automatically configure manageAudioSession: false when custom stream is used
  • Use case: Enables advanced scenarios like custom recording configurations, audio from non-microphone sources, or integration with existing audio pipelines
• Fix AudioRecorder disposal and recreation issue
  • Core: Change _audioRecorder from final to nullable field to allow recreation after disposal
  • Core: Add logic to recreate AudioRecorder instance in startListening() if it was previously disposed
  • Core: Prevent "Recorder has already been disposed" error when restarting after stop
  • Core: Properly set _audioRecorder to null after disposal in both stopListening() and dispose() methods

Formatting-Only Release

This is a patch release to meet pub.dev static analysis requirements.

Changes:

• Apply dart format to all files to meet pub.dev static analysis requirements

No functional changes from v0.0.7.

Full Changelog:
See CHANGELOG.md

• Apply dart format to all files to meet pub.dev static analysis requirements

Full Changelog: v0.0.5...v0.0.6

• BREAKING CHANGE: Convert all VAD APIs to async Future-based methods for better async/await support
  • API: Convert startListening(), stopListening(), pauseListening(), and dispose() methods in VadHandlerBase to return Future<void>
  • Web: Update VadHandlerWeb implementation to use async method signatures
  • Non-Web: Update VadHandlerNonWeb implementation to use async method signatures and properly await internal async operations
  • Example: Update example app to use async/await pattern when calling VAD methods
• introduce pauseListening feature
  • API: Add pauseListening() to VadHandlerBase.
  • Web: implement pauseListeningImpl() in vad_web.js and expose via JS bindings.
  • Non-Web: add _isPaused flag in VadHandlerNonWeb; ignore incoming frames when paused; if submitUserSpeechOnPause is true, call forceEndSpeech().
  • Start/Stop: reset _isPaused in startListening(); guard vadInstance in stopListeningImpl() with null-check and log.
• Add pause/resume UI functionality to example app
  • Example: Add dynamic pause button that appears only while actively listening
  • Example: Transform start button to "Resume" when paused, calling startListening() to resume
  • Example: Hide pause button when paused state is active
  • Example: Add separate stop button (red) available in both listening and paused states
  • Example: Implement proper state management for isListening and isPaused tracking
• Add support for custom RecordConfig parameter in startListening() for non-web platforms
  • API: Add optional RecordConfig? recordConfig parameter to startListening() in VadHandlerBase.
  • Non-Web: Use custom RecordConfig if provided, otherwise fall back to default configuration with 16kHz sample rate, PCM16 encoding, echo cancellation, auto gain, and noise suppression.
  • Web: Accept the parameter for compatibility but ignore it (not applicable for web platform).
• Bump record package to version 6.0.0
• Example: Bump permission_handler package to version 12.0.0+1
• Example: Bump audioplayers package to version 6.5.0

Full Changelog: v0.0.2...v0.0.5

• Add support for Silero VAD v5 model. (Default model is set to v4)
• Automatically upsample audio to 16kHz if the input audio is not 16kHz (fixes model load failures due to lower sample rates).
• Expose onRealSpeechStart event to notify when the number of speech positive frames exceeds the minimum speech frames (i.e. not a misfire event).
• Expose onFrameProcessed event to track VAD decisions by exposing speech probabilities and frame data for real-time processing.
• Update example app to show the onRealSpeechStart callback in action and introduce VAD Settings dialog to change the VAD model and other settings at runtime.
• For web platform, bundle the required files within the package to avoid download failures when fetching from CDNs and to ensure offline support.
• Update example app to log onFrameProcessed details for debugging.

• Fixed a bug where default modelPath was not picked up, resulting in silent failure if modelPath was not provided.
• Export VadIterator class for manual control over the VAD process for non-streaming use cases. Only available on iOS/Android.
• Added comments for all public methods and classes.

• Switch to onnxruntime package for inference on a separate isolate on iOS and Android to avoid using a full browser in the background, overall reducing the app size and improving performance.
• Example app will show audio track slider with controls while speech segment is being played and it will reflect a misfire event on the UI if occurred.

Full Changelog: https://github.com/keyur2maru/vad/commits/v0.0.2

• Fix broken LICENSE hyperlink in README.md and add topics to pubspec.yaml

• Initial release