A web-based experimentation tool for testing and comparing WaveKat library backends — VAD, turn detection, and more — side by side in real time.
Warning
Early development. Things may change.
- Live recording — capture audio from your microphone server-side, stream results to the browser in real time
- File analysis — upload a WAV file and run multiple configs against it at full speed
- Side-by-side comparison — fan out audio to N configurations simultaneously and compare outputs
- Preprocessing exploration — apply high-pass filters, RNNoise denoising, or normalization per-config
- Interactive visualization — waveform, spectrogram, and probability timelines with synchronized zoom, pan, and hover
make setup # one-time: install dependencies
make dev-frontend # Terminal 1: frontend (http://localhost:5173)
make dev-backend # Terminal 2: backend with auto-rebuild (http://localhost:3000)--host <HOST> Bind address (default: 127.0.0.1)
--port <PORT> Listen port (default: 3000)
| Backend | Description | Key Parameters |
|---|---|---|
| webrtc-vad | Google's WebRTC VAD — fast, low latency | Mode: quality, low-bitrate, aggressive, very-aggressive |
| silero-vad | Neural network VAD via ONNX Runtime — higher accuracy | Threshold: 0.0–1.0 |
| ten-vad | TEN framework VAD | Threshold: 0.0–1.0 |
| firered-vad | Xiaohongshu's FireRedVAD using DFSMN architecture | Threshold: 0.0–1.0 |
Each config can also enable per-config preprocessing: high-pass filter, RNNoise denoising, normalization.
| Backend | Description | Input |
|---|---|---|
| pipecat | Pipecat Smart Turn v3 — audio-based EOU detection | 16 kHz PCM audio |
| livekit | LiveKit Turn Detector — transcript-based EOU detection | ASR transcript text |
The Rust backend handles all audio capture and processing; the React frontend is embedded in the binary and handles visualization only.
┌─────────────────────────────────┐
│ Browser (React) │
│ Waveform + Spectrogram + │
│ Timelines + Config Panel │
└──────────┬──────────────────────┘
│ WebSocket
┌──────────▼──────────────────────┐
│ Server (Rust / Axum) │
│ ┌────────────┐ ┌────────────┐ │
│ │ Mic Capture │ │ WAV Loader │ │
│ │ (cpal) │ │ (hound) │ │
│ └─────┬──────┘ └─────┬──────┘ │
│ └──────┬─────────┘ │
│ ┌──────▼──────┐ │
│ │ Audio Frames │ │
│ └──────┬──────┘ │
│ ┌─────────┼─────────┐ │
│ ▼ ▼ ▼ │
│ Config 1 Config 2 Config N │
│ │ │ │ │
│ └─────────┼─────────┘ │
│ ┌────▼────┐ │
│ │ Results │ │
│ └────┬────┘ │
└───────────────┼──────────────────┘
▼
Browser UI
A web app for browsing and playing audio clips from the Mozilla Common Voice dataset. Filter by locale, split, demographics, and search sentences — with waveform playback powered by WaveSurfer.js. Source
https://commonvoice-explorer.wavekat.com/
Built on Cloudflare Workers + D1 + R2, with GitHub OAuth and a React 19 frontend.
| Video | Description |
|---|---|
 |
Testing Pipecat Smart Turn with WaveKat Lab Visual test of Pipecat Smart Turn v3 — live recording and VAD-gated pipeline mode simulating production workflows. |
 |
Adding FireRedVAD as the 4th backend Benchmarking Xiaohongshu's FireRedVAD against Silero, TEN VAD, and WebRTC across accuracy and latency. |
 |
VAD Lab: Real-time multi-backend comparison Live demo of VAD Lab comparing WebRTC, Silero, and TEN VAD side by side with real-time waveform visualization. |
Licensed under Apache 2.0.
Copyright 2026 WaveKat.