English

Speech Synthesis From Continuous Features Using Per-Token Latent Diffusion

Audio and Speech Processing 2025-11-25 v2

Abstract

We present SALAD, a zero-shot TTS autoregressive model operating over continuous speech representations. SALAD utilizes a per-token diffusion process to refine and predict continuous representations for the next time step. We compare our approach against a discrete variant of SALAD as well as publicly available zero-shot TTS systems, and conduct a comprehensive analysis of discrete versus continuous modeling techniques. Our results show that SALAD achieves superior intelligibility while matching the speech quality and speaker similarity of ground-truth audio.

Keywords

Cite

@article{arxiv.2410.16048,
  title  = {Speech Synthesis From Continuous Features Using Per-Token Latent Diffusion},
  author = {Arnon Turetzky and Avihu Dekel and Nimrod Shabtay and Slava Shechtman and David Haws and Hagai Aronowitz and Ron Hoory and Yossi Adi},
  journal= {arXiv preprint arXiv:2410.16048},
  year   = {2025}
}

Comments

ASRU 2025

R2 v1 2026-06-28T19:29:46.734Z