Scaling Properties of Continuous Diffusion Spoken Language Models
This work provides scaling insights for continuous diffusion SLMs, addressing the computational and data demands of spoken language modeling for researchers aiming to match text model performance.
The paper investigates whether continuous diffusion (CD) spoken language models (SLMs) are more viable than discrete autoregressive (AR) SLMs, finding that CD SLMs exhibit scaling laws for validation loss and phoneme Jensen-Shannon divergence (pJSD), with optimal token-to-parameter ratios decreasing as compute scales. Scaling to 16B parameters and tens of millions of hours of data enables emotive, prosodic, multi-speaker, multilingual speech, but long-form coherence remains a challenge.
Speech-only spoken language models (SLMs) lag behind text and text-speech models in performance, with recent discrete autoregressive (AR) SLMs indicating significant computational and data demands to match text models. Since discretizing continuous speech for AR creates bottlenecks, we explore whether continuous diffusion (CD) SLM is more viable. To quantify the SLMs linguistic quality, we introduce the phoneme Jensen-Shannon divergence (pJSD) metric. Our analysis reveals CD SLMs, mirroring AR behavior, exhibit scaling laws for validation loss and pJSD, and show optimal token-to-parameter ratios decreasing as compute scales. However, for the latter, loss becomes insensitive to choice of data and model sizes, showing potential for fast inference. Scaling CD SLMs to 16B parameters with tens of millions of hours of conversational data enables generation of emotive, prosodic, multi-speaker, multilingual speech, though achieving long-form coherence remains a significant challenge.