Phrase-Localized Language-Contrastive Guidance: Training-Free Localized Accent Control for Code-Switching Text-to-Speech
LCG targets the accent only where the code-switched phrase appears, without retraining the TTS model.
The paper says current systems often pronounce an inserted foreign phrase with the main language’s accent. Its method applies separate language guidance to different regions of the utterance, using self-attention probing to find phrase boundaries without external alignments. The authors report improved native accent on the switched phrase, less accent leakage, and preserved speaker identity and naturalness across language pairs. Accepted to EMNLP 2026 Main Conference. ArXiv · AI/CL/LG's note
The paper says current systems often pronounce an inserted foreign phrase with the main language’s accent. Its method applies separate language guidance to different regions of the utterance, using self-attention probing to find phrase boundaries without external alignments. The authors report improved native accent on the switched phrase, less accent leakage, and preserved speaker identity and naturalness across language pairs. Accepted to EMNLP 2026 Main Conference. ArXiv · AI/CL/LG's note
score 4