voice_forge_speech #

Pure-Dart FFI bindings for sherpa-onnx (k2-fsa), the next-gen-Kaldi speech toolkit — server-side edition.

This is a Flutter-free, web-free build of the official sherpa_onnx Dart bindings (v1.13.5, Apache-2.0) with one small patch. It is used by voice_forge for VAD, speech-to-text, and text-to-speech in a pure-Dart voice agent.

What you get #

• Silero VAD (VoiceActivityDetector)
• ASR (offline + streaming): Whisper, SenseVoice, NeMo, Zipformer transducer, and more (OfflineRecognizer, OnlineRecognizer)
• TTS (Piper/VITS and others) (OfflineTts)
• Speaker identification / diarization, punctuation restoration, audio tagging, spoken language identification, speech denoising
• WAV read/write helpers

Why not just depend on the official sherpa_onnx from pub.flutter-io.cn? #

The official package cannot be used from a pure-Dart process, which is the whole point of voice_forge (dart compile exe, no Flutter engine):

• its pubspec declares flutter: sdk: flutter and imports package:flutter/foundation.dart — dart pub get refuses to resolve it, and importing package:flutter crashes in a plain Dart VM;
• it is a Flutter plugin: it bundles the native libraries through nine platform sub-packages (sherpa_onnx_macos, sherpa_onnx_linux, …) that are only wired up by the Flutter plugin loader.

So this package is the same upstream bindings (v1.13.5, Apache-2.0), adapted for a server process that downloads its own libsherpa-onnx-c-api.

Exact changes vs the official v1.13.5 #

Everything except the items below is byte-identical to the official package, including all of lib/src/* (the 87 KB sherpa_onnx_bindings.dart FFI surface and every recognizer/TTS/VAD/config file):

Keeping upstream attribution: LICENSE is untouched (Apache-2.0, Xiaomi Corporation), and the upstreaming of the setSherpaLibrary patch is encouraged so this fork can shrink over time.

Getting started #

Using the voice_forge framework? SherpaKit.load() downloads this library automatically on first run — nothing below is needed.

1. Get the native library — this package has no bundled natives; the host process must be able to load libsherpa-onnx-c-api. Download the prebuilt library from the official sherpa-onnx releases (https://github.com/k2-fsa/sherpa-onnx/releases):

   Platform	Asset (v1.13.5)
   macOS arm64	sherpa-onnx-v1.13.5-osx-arm64-shared.tar.bz2
   macOS x64	sherpa-onnx-v1.13.5-osx-x64-shared.tar.bz2
   Linux x64	sherpa-onnx-v1.13.5-linux-x64-shared.tar.bz2
   Windows / others	see the release page for *-shared builds

   Each archive contains lib/libsherpa-onnx-c-api.dylib (macOS), lib/libsherpa-onnx-c-api.so (Linux) or bin/sherpa-onnx-c-api.dll (Windows), plus the libonnxruntime it links against. Place both next to your binary (the current working directory), on the system library search path, or pass the directory to initBindings(...).

2. Add the dependency:

dependencies:
  voice_forge_speech: ^1.13.5

3. Initialize once before creating any recognizer/TTS/VAD object:

import 'package:voice_forge_speech/voice_forge_speech.dart';

void main() {
  initBindings(); // or initBindings('/path/to/native/lib')
  // ... create VoiceActivityDetector / OfflineRecognizer / OfflineTts
}

Note: initBindings() must be called in every isolate that uses sherpa-onnx APIs — each isolate has its own FFI binding state.

For concrete end-to-end usage see the dart-api-examples/ folder in the upstream repository. For a complete voice-agent built on this package, see the voice_forge project.

License #

Apache-2.0, copyright Xiaomi Corporation (see LICENSE). This is a vendored derivative of the official bindings; keep upstream attribution intact when redistributing.