VibeVoice TTS - ONNX for Browser
Microsoft's VibeVoice-Realtime-0.5B converted to ONNX for browser inference.
Architecture
VibeVoice uses a two-LLM architecture with diffusion-based vocoding:
- Text LM (187 MB INT8) - Processes input text to hidden states
- TTS LM (417 MB INT8) - Generates speech latents autoregressively
- Diffusion Head (25 MB INT4) - Denoises latents
- Vocoder (339 MB INT4) - Converts latents to 24kHz audio
- Acoustic Connector (0.5 MB INT4) - Feeds speech back to TTS LM
Total: ~969 MB quantized (down from 3.8 GB)
Files
text_lm_int8.onnx - Text encodertts_lm_int8.onnx - Speech token generatordiffusion_head_int4.onnx - Latent denoiservocoder_int4.onnx - Audio synthesizeracoustic_connector_int4.onnx - Speech feedbacktokenizer.json - Qwen2 tokenizerconfig.json - Model configuration
Usage
See demo.html for browser usage example.
License
Apache 2.0 (same as original VibeVoice model)
