continuous-representations

Proposes HybridCodec, a novel framework combining temporally compressed discrete tokens with continuous residuals to improve speaker characteristic retention in speech language models, reducing autoregressive steps while maintaining quality.