Continuous diffusion shows scaling laws for spoken language models
July 5, 2026
Continuous diffusion (CD) spoken language models exhibit scaling laws for validation loss and phoneme Jensen-Shannon divergence (pJSD) comparable to discrete autoregressive models. This approach bypasses the discretization bottlenecks inherent in traditional speech-to-token methods.
HOW THIS AFFECTS YOU
●
researcherThe discovery of scaling laws in CD SLMs suggests a viable path toward high-performance speech-only models without tokenization bottlenecks.