MuQ-MuLan audio tower, ONNX
The audio half of OpenMuQ/MuQ-MuLan-large, exported to ONNX and stored in float16, so it can run in a browser through ONNX Runtime Web.
Used by describesong.com to turn a track into the 512 numbers it is searched by. Matches the PyTorch original to cosine 1.000000.
Batch size must be 1. The input is declared [batch, 1001, 1025, 2], but a shape constant left behind when the STFT was removed — view_1/shape = [1, -1, 240000] — pins the batch to 1. A batched call returns correct output for the first item only; the rest come back at roughly cosine 0.85 to their true vectors, with no error and no warning. Feed one window per run.
The STFT is not in this graph. It is the one step ONNX Runtime Web has no kernel for, so the caller computes it and feeds the spectrogram in: n_fft 2048, hop 240, periodic Hann, reflect-padded 1024 either side, input shape [batch, frames, 1025, 2] at 24 kHz.
Weights and licence follow the original model.