Non-Streaming Speech Recognition from File

This app is a Tauri v2 desktop application that performs non-streaming speech recognition on audio and video files. It uses Silero VAD to detect speech segments and then runs an offline ASR model on each segment.

Note

This app works with files (e.g., .wav, .mp3, .mp4), not with microphone input. You can use it to transcribe audio/video files and generate subtitles in SRT format.

Features

62+ ASR models supported (SenseVoice, Paraformer, Whisper, Transducer, Moonshine, Parakeet, Canary, Qwen3 ASR, etc.)
Audio/video playback with waveform display
SRT subtitle export
Segment WAV export — save individual speech segments as WAV files
Progress tracking with cancellation support
Cross-platform — macOS (universal), Linux (x64/aarch64), Windows (x64)
Pure-Rust audio decoding via symphonia — no system dependencies

Supported Audio Formats

Any format supported by symphonia:

Audio: MP3, FLAC, AAC, OGG/Vorbis, WAV, AIFF, ADPCM
Video: MP4/M4A, MKV, WebM (audio track extracted)

Pre-built Apps

Pre-built apps for 62+ models and 4 platforms are available at:

https://k2-fsa.github.io/sherpa/onnx/tauri/pre-built-app.html#non-streaming-speech-recognition-from-file

The filename follows the pattern:

vad-asr-{version}-{lang}-{model}-{platform}.{ext}

where:

version: the current version, e.g., 1.12.39
lang: the language of the model, e.g., en for English, zh for Chinese
model: the name of the model used in the app
platform: one of linux-x64, linux-aarch64, universal.app (macOS), windows-x64
ext: .tar.gz for Linux, .tar.bz2 for macOS, .zip for Windows

You can download all supported models from https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models.

Building the App

Prerequisites

Rust (stable)
Node.js (for the Tauri CLI)
Tauri CLI prerequisites

Install npm dependencies:

npm install

Example 1: SenseVoice (multilingual, Chinese/English/Japanese/Korean/Cantonese)

This example uses model type 15 — sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2024-07-17. It supports Chinese, English, Japanese, Korean, and Cantonese.

Step 1: Download the model and Silero VAD

cd tauri-examples/non-streaming-speech-recognition-from-file

curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2024-07-17.tar.bz2
tar xvf sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2024-07-17.tar.bz2
rm sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2024-07-17.tar.bz2

curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/silero_vad.onnx

Step 2: Copy assets

Tauri bundles files from src-tauri/assets/. Place the model directory and silero_vad.onnx inside it:

mkdir -p src-tauri/assets
cp -a sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2024-07-17 src-tauri/assets/
cp -a silero_vad.onnx src-tauri/assets/

After copying, you can delete the original model directory to save disk space:

rm -rf sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2024-07-17

This gives:

src-tauri/assets/
├── silero_vad.onnx
└── sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2024-07-17/
    ├── model.int8.onnx
    └── tokens.txt

Tip

SenseVoice supports a homophone replacer. To enable it, also download lexicon.txt and replace.fst into src-tauri/assets/:

curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/hr-files/lexicon.txt
curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/hr-files/replace.fst
cp -a lexicon.txt replace.fst src-tauri/assets/

The app works fine without these files.

Step 3: Set model type

Edit src-tauri/src/lib.rs and make sure the constants match:

const MODEL_TYPE: u32 = 15;
const MODEL_NAME: &str = "sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2024-07-17";

Step 4: Build and run

npm run dev

To build a release binary:

npm run build

The output is in src-tauri/target/release/bundle/.

Example 2: Parakeet TDT 0.6b v2 (English)

This example uses model type 30 — sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8. It supports English only.

Step 1: Download the model and Silero VAD

cd tauri-examples/non-streaming-speech-recognition-from-file

curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8.tar.bz2
tar xvf sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8.tar.bz2
rm sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8.tar.bz2

curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/silero_vad.onnx

Step 2: Copy assets

mkdir -p src-tauri/assets
cp -a sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8 src-tauri/assets/
cp -a silero_vad.onnx src-tauri/assets/

After copying, you can delete the original model directory to save disk space:

rm -rf sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8

This gives:

src-tauri/assets/
├── silero_vad.onnx
└── sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8/
    ├── encoder.int8.onnx
    ├── decoder.int8.onnx
    ├── joiner.int8.onnx
    └── tokens.txt

Step 3: Set model type

Edit src-tauri/src/lib.rs:

const MODEL_TYPE: u32 = 30;
const MODEL_NAME: &str = "sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8";

Step 4: Build and run

npm run dev

# Or build a release binary
npm run build

Using a Different Model

The app supports 62+ models (types 0–61) defined in src-tauri/src/model_registry.rs. To use a different model:

Choose a model type from the registry (0–61).
Set MODEL_TYPE and MODEL_NAME in src-tauri/src/lib.rs.
Download the corresponding model and place it in src-tauri/assets/.
Always include silero_vad.onnx in src-tauri/assets/.

The model files required depend on the model family:

Paraformer: model.int8.onnx, tokens.txt
Transducer: encoder*.onnx, decoder*.onnx, joiner*.onnx, tokens.txt
Whisper: *-encoder*.onnx, *-decoder*.onnx, tokens.txt
SenseVoice: model.int8.onnx, tokens.txt
Moonshine: encode.onnx, decode.onnx, tokens.txt
NeMo CTC: model.onnx, tokens.txt

How It Works

Click Select Audio File to choose an audio/video file via native dialog.
The file is decoded by symphonia to mono f32 PCM samples.
Audio is resampled to 16 kHz (required by Silero VAD).
Audio is fed to the VAD in 512-sample (32 ms) chunks.
Each detected speech segment is decoded by the offline recognizer.
Results are displayed in a table with start/end timestamps and text.
You can export the results as an SRT subtitle file.

License

The code is licensed under Apache-2.0. Please check the license of your selected model separately.