Universal Robust Speech Adaptation for Cross-Domain Speech Recognition and Enhancement

Chien-Chun Wang, Hung-Shin Lee · 2026

Read Universal Robust Speech Adaptation for Cross-Domain Speech Recognition and Enhancement by Chien-Chun Wang, Hung-Shin Lee (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-04307.txt.

Description

Speech systems trained in clean conditions collapse in new noise. Chien-Chun Wang and 3 co-authors present URSA-GAN, a single adaptation framework for recognition and enhancement, in a 2026 paper. Domain shifts in noise and channel conditions are the twin culprits.

Separate noise and channel encoders, pretrained on limited in-domain data, capture the target setting's signature. Those embeddings guide a generative model that builds speech matched in sound to the target domain while keeping what was said. A dynamic noise-based penalty adds robustness. One framework adapting both recognizers and enhancers beats retraining domain by domain, the everyday reality of deployment.

Catalog record

CREATORChien-Chun Wang, Hung-Shin Lee
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE69 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-04307

Browse the full library · About the Tower of Babel