Abstract
Generative speech-to-speech models operating in real-time conversational systems may produce deviations from ground-truth data. A system may address this using a dual-layer architecture for factual correction. A preventative layer can utilize dynamic factual-logit masking to guide a model's token generation towards verified data by applying a mask derived from a token prefix trie. A subsequent curative layer can function as an in-flight acoustic self-repair engine, which may use a lookahead audio buffer to audit synthesized speech and correct discrepancies before transmission to a user. This dual-layer method can improve the factual and mathematical integrity of generative voice systems while operating within low-latency constraints to support conversational fluency.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Raman, Sundarrajan, "Dual-Layer Factual Correction for Generative Speech via Logit Masking and Acoustic Repair", Technical Disclosure Commons, (October 12, 2026)
https://www.tdcommons.org/dpubs_series/12097