muq-mulan-audio-onnx

MuQ-MuLan audio tower, post-mel ONNX export

The audio tower of OpenMuQ/MuQ-MuLan-large (Tencent AI Lab; CC-BY-NC 4.0) exported to ONNX after the mel spectrogram. MuQ's torchaudio STFT does not export, so the front end is computed by the caller.

• Input mel: (1, 128, frames) float32 — torchaudio MelSpectrogram(sample_rate=24000, n_fft=2048, hop_length=240, n_mels=128) (power 2, centred reflect padding, periodic Hann, HTK mel, no norm), AmplitudeToDB (10·log10(max(x, 1e-10))), the last frame dropped, then normalized as (x − 6.768444971712967) / 18.417922652295623. One clip is ≤ 10 s (1000 frames).
• Output embedding: (1, 512) float32, L2-normalized, in the MuQ-MuLan joint audio–text space. MuLan's convention for a track is the mean over consecutive 10 s clips (the tail completed with the start of the track), without renormalizing.
• Verified against the full PyTorch path at cosine 1.000000; opset 17, fp32, 1.27 GB. SHA-256 30d8b3bb55219830febd3840885519713f89a3c4c48cb36f888871eb621a7f4c.

Non-commercial use only, per the source weights' license; attribution to the MuQ authors: MuQ: Self-Supervised Music Representation Learning with Mel Residual Vector Quantization (Zhu et al., 2025).