Scaling Properties of Speech Language Models

arXiv:2404.00685v1 Announce Type: cross
Abstract: Speech Language Models (SLMs) aim to learn language from raw audio, without textual resources. Despite significant advances, our current models exhibit weak syntax and semantic abilities. However, if the scaling properties of neural language models hold for the speech modality, these abilities will improve as the amount of compute used for training increases. In this paper, we use models of this scaling behavior to estimate the scale at which our current methods will yield a …

abstract advances aim arxiv audio compute cs.ai cs.cl cs.ne current eess.as however language language models learn raw resources scaling semantic slms speech syntax textual type will