Rethinking Music Tokenization: A Semantic Codec toward High-Fidelity LLM Music Generation
Discrete audio tokenization has become the critical interface between raw waveforms and autoregressive modeling in recent music generation. As a result, music tokenizers must simultaneously support high-fidelity reconstruction and produce discrete sequences that remain amenable to language modeling. Existing reconstruc...