Robustness of Localized Steering for Autoregressive Speech Generation

Determine whether localized logit-level contrastive interventions generalize robustly to autoregressive speech generation models or instead exacerbate autoregressive error accumulation.

Background

Phrase-Localized Language-Contrastive Guidance is designed for discrete diffusion LLM backbones and is evaluated using the OmniVoice discrete diffusion TTS model. Although the authors note that logit-level contrastive steering could in principle be applied to autoregressive speech generation models, they have not established whether localized interventions remain robust in that setting or introduce compounding generation errors.

References

While logit-level contrastive steering could empirically be applied to autoregressive (AR) speech generation models, whether such localized interventions would generalize robustly or conversely exacerbate autoregressive error accumulation remains unverified, thereby bounding the immediate extensibility of our method to alternative model paradigms.

— Phrase-Localized Language-Contrastive Guidance: Training-Free Localized Accent Control for Code-Switching Text-to-Speech  (2609.01016 - Lee et al., 1 Sep 2026) in Section “Limitations,” second paragraph