MioCodec 25 Hz · 44.1 kHz — in your browser

Neural audio codec Aratako/MioCodec-25Hz-44.1kHz-v2 exported to ONNX and running locally on WebGPU. 25 tokens per second · 341 bps · 44.1 kHz output · 128-d speaker embedding. No server, no upload — audio never leaves the page.

backend: checking… webgpu: ? decoder: not loaded encoder: not loaded

1 · Load the models — from the Hugging Face CDN, cached in your browser

Weights are fp16 with fp32 arithmetic, so the download is half the size. The decoder is bit-faithful; the encoder matches the fp32 model on ~95% of frames (a few tokens per clip can differ at the quantiser boundary). Playback is level-matched to −20 dBFS for fair A/B. The decoder is all you need for playback and voice conversion; the encoder adds “audio → tokens”.

Decoder 168 MB

tokens + speaker embedding → 44.1 kHz waveform

Encoder 248 MB

44.1 kHz audio → 25 Hz tokens + 128-d speaker embedding

2 · Try it — presets are pre-encoded, so the decoder alone is enough

Decode a preset clip from its 25 Hz token stream and compare with the original.