Smart VAD

Voice Activity Detection (VAD) slicing tool powered by Silero VAD. Detects speech segments in WAV audio, splits into separate files, and generates an interactive HTML report.

Features

• VAD Detection — Silero VAD via ONNX Runtime (CGO)
• Adaptive VAD — Dynamic noise floor estimation, auto-adjusting threshold/min_speech, RMS energy post-filter (--adaptive)
• Audio Slicing — Split WAV by speech segments
• HTML Report — Interactive waveform, VAD confidence chart, per-segment audio players with playback speed control
• Streaming API — Feed audio chunks incrementally via Process()/Flush()
• CLI + HTTP Server — Command-line tool and web upload interface

Quick Start

Make (recommended)

make init            # 安装 onnxruntime + 下载 Silero VAD 模型
make run-server      # 构建并启动 HTTP 服务 (localhost:8080)
make run-demo INPUT=audio.wav   # 构建并运行 CLI 分析
make test            # 运行测试
make clean           # 清理编译产物

Manual

# Install ONNX Runtime (macOS)
brew install onnxruntime

# Download Silero VAD model
pip3 install silero-vad
cp $(python3 -c "import silero_vad; import os; print(os.path.join(os.path.dirname(silero_vad.__file__),'data','silero_vad.onnx'))") .

# Run CLI (standard mode)
CGO_CFLAGS="-I/opt/homebrew/include/onnxruntime" \
CGO_LDFLAGS="-L/opt/homebrew/lib" \
go run ./cmd/demo --input audio.wav --model silero_vad.onnx

# Run CLI (adaptive mode — auto-adjusts to background noise)
CGO_CFLAGS="-I/opt/homebrew/include/onnxruntime" \
CGO_LDFLAGS="-L/opt/homebrew/lib" \
go run ./cmd/demo --input audio.wav --model silero_vad.onnx --adaptive

# Or run HTTP server
CGO_CFLAGS="-I/opt/homebrew/include/onnxruntime" \
CGO_LDFLAGS="-L/opt/homebrew/lib" \
go run ./cmd/server --model silero_vad.onnx

Adaptive VAD

In noisy environments (e.g. shopping malls, exhibitions), standard VAD with a fixed threshold can't balance sensitivity vs. false positives:

• Fixed low threshold → captures background chatter as speech
• Fixed high threshold → misses soft-spoken users in quiet moments

The --adaptive flag enables runtime adaptation with three mechanisms:

1. Noise Floor Estimation

Computes the average RMS energy of the quietest 10% of audio frames. This captures the background noise level rather than speech energy, ensuring accurate baseline in mixed speech/noise audio.

2. Dynamic Parameter Mapping

Maps the noise floor to VAD parameters:

Noise Floor	Threshold	Min Speech
≤ -50 dB	0.5	250 ms
-50 ~ -40	0.5 → 0.7	250 → 400
-40 ~ -35	0.7 → 0.8	400 → 500
> -35 dB	0.85	600 ms

3. RMS Energy Post-Filter

Discards detected segments whose average energy is below noiseFloor + 6 dB, filtering out distant/unintended speech.

SDK Usage

import "github.com/LiusCraft/smart-vad/vad"

// Batch mode
adaptCfg := vad.AdaptiveConfig{
    DetectorConfig: vad.Config{
        ModelPath:  "silero_vad.onnx",
        SampleRate: 16000,
    },
}
adaptDetector, _ := vad.NewAdaptiveDetector(adaptCfg)
result, _ := adaptDetector.Detect(pcm)

// Streaming mode
adaptDetector.Reset()
adaptDetector.Process(chunk1)
adaptDetector.Process(chunk2)
result := adaptDetector.Flush()

Project Structure

├── vad/            # VAD detection package (streaming + batch)
├── slice/          # Audio slicing and WAV export
├── html/           # HTML report generation (embedded templates)
├── template/       # HTML templates (embedded via //go:embed)
├── cmd/demo/       # CLI entry point
└── cmd/server/     # HTTP server

SDK Usage

import "github.com/LiusCraft/smart-vad/vad"

// Batch mode
result, err := detector.Detect(pcm)

// Streaming mode
detector.Reset()
detector.Process(chunk1)
detector.Process(chunk2)
result := detector.Flush()