speech-to-speech
Full-duplex voice-to-voice in your browser.
Rust → WebAssembly + WebGPU. No server, no API key.
Your audio never leaves your machine.
The download button is disabled until the demo is usable.
A Q4_K-quantized PersonaPlex-7B, pruned to 24 layers with QLoRA recovery, running on Burn's wgpu backend compiled to WebAssembly. The pipeline is microphone → Mimi encoder → temporal and depth transformers → Mimi decoder → playback, end to end in your browser.
Walkie-talkie mode works, with a handful of voice presets. Generation runs about 3× slower than realtime, so it is turn-based for now. Audio quality is poor.
Full duplex mode, where the model can listen and speak at the same time, is not supported yet. The model runs in "walkie mode": it listens, then speaks.
Next: true duplex streaming, then performance work.
Quantized, pruned PersonaPlex-7B (24 layers, QLoRA-recovered) model (Mimi audio codec + temporal/depth transformers), custom Rust inference compiled to WebAssembly, GPU acceleration via WebGPU.