arXiv cs.CL· Hung-Yang Sung, Chien-Chun Wang, Tien-Hong Lo, Yu-Sheng Tsao, Yung-Chang Hsu, Berlin Chen·· 10 小时前AI 评分18
T-SANDHI：面向低资源台语（闽南语）语音识别的变调感知自适应网络
T-SANDHI: Tone Sandhi-aware Adaptive Network with Decoupled Hybrid Injection for Low-resource Taiwanese Hokkien Speech Recognition
AI 导读
针对台语（闽南语）ASR，研究提出 T-SANDHI，在冻结的 Whisper 骨干上显式解耦表层声学与底层词汇意图，用词典引导的多任务学习和文本伪标签驱动的轻量混合注入模块，通过动态门控整合本调与变调两条语音流。实验表明语音基础模型本身能有效处理变调，真正瓶颈是变调与保留本调之间的局部混淆；该方法在 TAT-MOE 语料和两个盲测集上以严格参数效率超越基线。
正文 · 原文
Abstract:In Taiwanese Hokkien automatic speech recognition (ASR), prior studies often treat tone sandhi as a major challenge under the assumption that models fail to process implicit phonological variations. However, our experiments on Taiwanese Hokkien reveal that speech foundation models actually handle tone sandhi variations effectively, and the real performance bottleneck stems from a localized confusion between these variations and retained citation tones. To address this, we propose T-SANDHI to explicitly decouple surface acoustics from underlying lexical intent on top of a frozen Whisper backbone. Using a lexicon-guided multi-task learning structure driven by text-derived pseudo labels, our lightweight hybrid injection module integrates independent citation and sandhi phonetic streams via dynamic gating. Extensive evaluation on the TAT-MOE corpus and two blind test sets demonstrates that this explicit disentanglement effectively resolves tonal mapping confusion, outperforming baselines with strict parameter efficiency.
| Comments: | Accepted to IEEE SLT 2026 |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2609.18194 [cs.CL] |
| (or arXiv:2609.18194v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2609.18194 arXiv-issued DOI via DataCite |
Submission history
From: Hung-Yang Sung [view email]
[v1]
Wed, 16 Sep 2026 06:31:31 UTC (1,317 KB)
[v2]
Tue, 10 Nov 2026 02:23:15 UTC (1,317 KB)
来源：arXiv cs.CL · arxiv.org