The paper introduces PhonoVoxtLM, a speech language model built on the existing VoxtLM framework. The central aim is to investigate whether phonological tokens can support both speech understanding and speech generation within a single model.

The model is described as using a single-codebook representation for these tokens, as indicated by the name. The authors position this as a way to explore how speech token representations can be shared across understanding and generation tasks.

The abstract as provided focuses on the motivation and design, without detailing experimental outcomes. The work appears to be an early exploration of unified speech modeling with phonological tokens.