Scaling Properties of Continuous Diffusion Spoken Language Models

TL;DR

This paper investigates the scaling properties of continuous diffusion spoken language models, demonstrating their potential for high-quality multilingual speech generation and analyzing their behavior as models grow larger.

Contribution

It introduces the phoneme Jensen-Shannon divergence metric and explores the scaling laws and efficiency of continuous diffusion SLMs compared to autoregressive models.

Findings

CD SLMs exhibit predictable scaling laws for validation loss and pJSD.

Optimal token-to-parameter ratios decrease with increased compute.

Scaling to 16B parameters enables diverse, multilingual speech generation.

Abstract

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.