Abstract

Discrete speech tokens have emerged as a fundamental representation for various downstream speech processing tasks, particularly in speech generation. However, most existing tokens encode dense, fixed-rate acoustic information, which introduces substantial redundancy and limits their efficiency. In this talk, I will first provide a brief review on the taxonomy of current discrete speech tokens, then present our works exploring the reduction of this information redundancy in two critical directions: (1) Speaker timbre disentanglement, introducing a low-bitrate, single-codebook and speaker-decoupled codec for speech. (2) Variable-rate temporal compression, exploring methods to dynamically adjust the frame rate of discrete tokens for better compactness and bitrate-performance tradeoff. Together, these efforts highlight pathways toward more efficient and controllable discrete speech representations, paving the way for the next generation of speech technologies.