Browse Teach Curate

MuQ: Music SSL via Mel Residual Vector Quantization

Trains a self-supervised music representation model to predict tokens from Mel Residual Vector Quantization (Mel-RVQ), a residual linear-projection quantizer of the Mel spectrum that stabilizes SSL target extraction, beating prior music SSL with only 0.9K hours of data (scaling to 160K with iterative training) and yielding MuQ-MuLan, a contrastive music-text model with SOTA zero-shot music tagging.

Rosie Text

MuQ: Self-Supervised Music Representation Learning with Mel Residual Vector Quantization Haina

MuQ is a self-supervised music representation model trained to predict tokens produced by Mel Residual Vector Quantization (Mel-RVQ). Unlike prior music SSL that uses random projection or off-the-she…