ankit-pn/MioCodec-25Hz-44.1kHz-v2-ONNX

MioCodec-25Hz-44.1kHz-v2 — ONNX (WebGPU-ready)

Browser-runnable ONNX export of Aratako/MioCodec-25Hz-44.1kHz-v2, a 25 Hz / 341 bps neural audio codec (WavLM-base+ encoder, iSTFTHead decoder, 44.1 kHz output).

Two graphs, split so the browser can load only what it needs:

file	size	inputs	outputs
encoder.fp16.onnx	~248 MB	waveform f32 (1, N) @ 44.1 kHz	tokens i32 (1, T) @ 25 Hz, global f32 (1, 128)
decoder.fp16.onnx	~168 MB	tokens i32 (1, T), global f32 (1, 128)	waveform f32 (1, L) @ 44.1 kHz

Weights are stored as fp16 with a Cast to fp32, so the files are half the size while all arithmetic still happens in float32. The FSQ quantiser and the iSTFT tables stay fp32 so tokenisation and synthesis remain stable. Both graphs have a dynamic sequence length.

Export notes

The reference implementation needed four rewrites to be ONNX/WebGPU-friendly; each is numerically equivalent (verified against the PyTorch model: identical tokens, identical speaker embedding, waveform max abs diff ~4e-6):

• RoPE without view_as_complex/polar (interleaved pairs, computed from cos/sin tables stored as graph constants and sliced by the runtime length).
• Windowed attention without triu/tril/FlashAttention: an additive bias built from a position vector, then explicit matmul + softmax.
• iSTFT without torch.fft.irfft/F.fold: one ConvTranspose1d whose kernel is the real inverse-DFT matrix folded with the Hann window, plus the same convolution of an all-ones frame for the window envelope.
• FSQ code<->index arithmetic in float32 instead of float64/int64 (exact for levels 8/8/8/5/5).

The encoder expects the waveform to carry the symmetric padding that MioCodecModel._calculate_waveform_padding applies (111 samples per side for a 3 s clip); the demo computes it in JS.

MIT, inherited from the upstream model.