stat_ort_plugin 0.0.2

On-device speech recognition (Vaani ASR by IISc Bangalore) for Flutter via dart:ffi — file and streaming transcription with Silero VAD and speaker diarization, powered by ONNX Runtime.

stat_ort_plugin

On-device speech recognition (Vaani ASR) for Flutter via dart:ffi, powered by ONNX Runtime. It provides both file transcription and real-time streaming transcription, with Silero VAD based segmentation and optional speaker diarization.

Status: early release (0.0.1). The native pipeline works; the API may still change.

Features

• 🎙️ Streaming transcription from a microphone PCM stream, segmented by voice activity.
• 📁 File transcription for 16 kHz mono int16 WAV files.
• 🧑‍🤝‍🧑 Optional speaker diarization (speaker-labelled segments).
• ⚡ Native C pipeline; heavy work runs off the UI isolate.

Supported platforms

Platform	Status	ONNX Runtime source
Android	✅	com.microsoft.onnxruntime:onnxruntime-android (Gradle)
iOS	✅	onnxruntime-c (CocoaPods)

macOS / Windows / Linux are not currently supported.

Models

This package does not ship any models. You supply ONNX model files and a vocabulary at runtime:

• ASR encoder (encoder-vaani.onnx)
• ASR decoder/joint (decoder_joint-vaani.onnx)
• Vocabulary (tokens.txt)
• (optional) Silero VAD (silero_vad.onnx) — required for streaming segmentation
• (optional) Speaker embedding (voxblink2_samresnet34_ft.onnx) — required for diarization

Audio must be 16 kHz, mono, 16-bit PCM. See NOTICE.md for model licensing. The example app loads models from its assets and copies them to a temp directory; for a real app, download them on first launch rather than bundling hundreds of MB into your binary.

Installation

Add the dependency (from a git ref or, once published, from pub.dev):

dependencies:
  stat_ort_plugin:
    git:
      url: https://github.com/suhrt/stat_ort_plugin

No extra setup is needed: ONNX Runtime is pulled in automatically by the Android Gradle dependency and the iOS CocoaPod.

Usage

File transcription

import 'package:stat_ort_plugin/stat_ort_plugin.dart';

final vaani = await Vaani.create(
  encoderPath, decoderPath, vocabPath, 4, // encoderThreads
  vadPath: vadPath,
  speakerPath: speakerPath,
);

final transcript = await vaani.transcribe(wavPath);
print(transcript);

vaani.dispose();

Streaming transcription

Feed raw 16 kHz mono int16 PCM as it arrives (e.g. from the record package). The plugin buffers internally into the 512-sample frames Silero VAD requires.

final stream = vaani.createStream();

await for (final Uint8List chunk in micPcmStream) {
  final segment = stream.pushChunk(chunk.buffer.asInt16List());
  if (segment != null) print(segment); // a finalised "[mm:ss - mm:ss] [Speaker N]: ..." line
}

// IMPORTANT: flush before closing to get the trailing segment.
final tail = stream.finish();
if (tail != null) print(tail);
stream.close();

A complete example (mic streaming + file transcription) is in example/.

Threading notes

• Vaani.create and transcribe run the native work on a background isolate.
• A single Vaani (pipeline) may back multiple VaaniStreams; VAD state is held per-stream, so independent streams don't interfere. Do not, however, drive the same pipeline's inference from multiple threads concurrently without your own serialization.
• vaani_pipeline_init resolves the ONNX Runtime API once and is safe to call from multiple isolates.

Regenerating the FFI bindings

lib/stat_ort_plugin_bindings_generated.dart is generated from src/stat_ort_plugin.h with package:ffigen. It is committed so the package builds without a generation step. Regenerate after changing the header:

dart run ffigen --config ffigen.yaml

Debug logging

The native code logs errors via LOGE always, and verbose progress via LOGD only when built with VAANI_DEBUG defined (CMake: -DVAANI_DEBUG=ON). Release builds carry no verbose-logging overhead.

License

MIT — see LICENSE. Third-party dependencies and models are listed in NOTICE.md; verify model licenses for your use case before shipping.