LSCodec: Low-Bitrate and Speaker-Decoupled Discrete Speech Codec
Yiwei Guo, Zhihan Li, Chenpeng Du, Hankun Wang, Xie Chen, Kai Yu
TL;DR
LSCodec tackles the need for low-bitrate, speaker-decoupled discrete speech tokens suitable for LM-based speech generation by introducing a three-stage, unsupervised training pipeline with explicit speaker perturbation. It combines a continuous-space VAE (Stage 1) with a single-codebook VQ-VAE (Stage 2) and a discrete-token vocoder (Stage 3) to produce compact tokens while reducing timbre leakage. Experiments on LibriTTS demonstrate strong reconstruction quality at 0.45 kbps (50Hz) and 0.25 kbps (25Hz), with favorable WER, SECS, MOS, and VC metrics, along with robust speaker-probing results indicating effective disentanglement. Ablation studies confirm the importance of perturbation, SSL-token prediction, and the multi-stage framework, highlighting LSCodec as a promising building block for efficient speech LMs and speaker-controlled generation.
Abstract
Although discrete speech tokens have exhibited strong potential for language model-based speech generation, their high bitrates and redundant timbre information restrict the development of such models. In this work, we propose LSCodec, a discrete speech codec that has both low bitrate and speaker decoupling ability. LSCodec adopts a multi-stage unsupervised training framework with a speaker perturbation technique. A continuous information bottleneck is first established, followed by vector quantization that produces a discrete speaker-decoupled space. A discrete token vocoder finally refines acoustic details from LSCodec. By reconstruction evaluations, LSCodec demonstrates superior intelligibility and audio quality with only a single codebook and smaller vocabulary size than baselines. Voice conversion and speaker probing experiments prove the excellent speaker disentanglement of LSCodec, and ablation study verifies the effectiveness of the proposed training framework.