Coarser duration-penalized speech units hurt phone/word discrimination but help resynthesis and lexical/syntactic modeling when the codebook is large enough.
In contrast to models that combine speech and text [2, 3], SLMs do not use text data in any of their components
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
Spoken Language Modeling with Duration-Penalized Self-Supervised Units
Coarser duration-penalized speech units hurt phone/word discrimination but help resynthesis and lexical/syntactic modeling when the codebook is large enough.