CosyAccent: Duration-Controllable Accent Normalization Using Source-Synthesis Training Data

Qibing Bai, Shuhao Shi, Shuai Wang, Yukai Ju, Yannan Wang, Haizhou Li · 2026

Read CosyAccent: Duration-Controllable Accent Normalization Using Source-Synthesis Training Data by Qibing Bai, Shuhao Shi, Shuai Wang, Yukai Ju, Yannan Wang, Haizhou Li (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-2602-19166.txt.

Description

Accent normalisation often sounds unnatural or distorts content. Qibing Bai, with five co-authors, fixes both flaws in 2026 with CosyAccent. Poor training data and rigid duration modelling cause the trouble. Their source-synthesis method generates source second-language speech and uses authentic native speech as target. No real second-language data is needed for training, and text-to-speech artefacts are avoided.

CosyAccent itself is non-autoregressive. It models rhythm implicitly for flexibility, yet controls total output duration explicitly. That resolves the naturalness-versus-control tradeoff. Despite training without real second-language speech, it preserves content better and sounds more natural than strong baselines trained on real-world data. Accent normalisation thus gains a data-efficient path.

Catalog record

CREATORQibing Bai, Shuhao Shi, Shuai Wang, Yukai Ju, Yannan Wang, Haizhou Li
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE23 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-19166

Browse the full library · About the Tower of Babel