Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Modeling Music as a Time-Frequency Image: A 2D Tokenizer for Music Generation

About

Autoregressive music generation depends strongly on the audio tokenizer. Existing high-fidelity codecs often use residual multi-codebook quantization, which preserves reconstruction quality but complicates language modeling after sequence flattening, as the residual hierarchy imposes strong sequential dependencies and can amplify error accumulation. We propose BandTok, a generation-oriented 2D Mel-spectrogram tokenizer that represents each frame with Mel-frequency band tokens from a single shared codebook. This design yields a physically interpretable time-frequency token grid with a more independent token structure, making it better suited for autoregressive modeling. BandTok improves reconstruction with a multi-scale PatchGAN objective and EMA codebook updates. We further introduce an autoregressive language model with 2D Rotary Position Embedding (2D RoPE) to preserve temporal and frequency-band structure during generation. Experiments show that BandTok improves over residual-codebook tokenizers and achieves strong results in a data-limited setting. The source code and generation demos for this work are publicly available.

Yuqing Cheng, Xingyu Ma, Guochen Yu, Xiaotao Gu• 2026

Related benchmarks

TaskDatasetResultRank
Music GenerationICME contest (test)
FADCLAP0.482
8
Audio ReconstructionSongDescriber
Mel Score0.642
7
Showing 2 of 2 rows

Other info

Follow for update