About this document

Duration-Penalized Units in Spoken Language Models by brainx is a document available to read on EtoBox.

The paper investigates the performance of spoken language models (SLMs) by varying codebook size and unit coarseness using duration-penalized dynamic programming (DPDP). Results indicate that while coarser units provide little benefit at the phone and word levels, they improve performance in whole-sentence resynthesis and lexical and syntactic tasks, particularly at higher codebook sizes. The study highlights the nuanced relationship between unit granularity and task requirements, suggesting that DPDP is an

Author

brainx

Language

EN