跳到正文
arXiv cs.CL· Ben Xiaolu Huang·· 9 小时前AI 评分31

ClearTalk:面向构音障碍儿童的中文发音分析与个性化沟通支持系统

ClearTalk: Data-Driven Mandarin Articulation Analysis and Personalized Communication Support for Children with Dysarthria

AI 导读

ClearTalk 是一款基于智能手表的中文构音障碍儿童发音分析系统,将语音识别差异转化为可解释的发音画像用于个性化练习与分阶段沟通。在 13 名构音障碍说话人上,系统音节级准确率与专家可懂度评分的 Pearson 相关系数为 0.908(R²=0.824,p<0.001),系统选出的优先训练目标与三位盲评言语治疗师的共识重合度达 86.7%。

正文 · 原文

View PDF HTML (experimental)

Abstract:Children with dysarthria may experience reduced intelligibility and difficulty expressing everyday needs. ClearTalk is a smartwatch-based system that converts Mandarin speech-recognition differences into interpretable articulatory profiles for personalized practice and staged communication. Target and recognized utterances are aligned at the syllable level, summarized across seven articulatory positions, and ranked by a count-sensitive asymmetric heuristic. For 13 dysarthric speakers from the Mandarin Dysarthria Speech Corpus, system syllable-level accuracy showed a Pearson correlation of 0.908 with expert intelligibility, with an R-squared value of 0.824 and a p-value below 0.001. The system-selected priority targets achieved an 86.7 percent overlap with the consensus of three blinded speech-language pathologists; across 210 target consonants, automated labels achieved 86.7 percent agreement, a Cohen's kappa of 0.76, and a macro-F1 score of 0.81. In a balanced single-session pilot with 15 children, mean pre/post gains were 14.4, 29.4, and 27.7 percentage points for mild, moderate, and severe groups, while communication-round completion rates were 81.3, 66.7, and 48.0 percent. These findings support the feasibility of using interpretable patterns of speech errors to guide wearable articulation practice and communication assistance.
Comments: 5 pages, 3 figures, 1 table. Accepted to the IEEE ICDM 2026 Teen Research Track
Subjects: Human-Computer Interaction (cs.HC); Computation and Language (cs.CL)
Cite as: arXiv:2611.21639 [cs.HC]
  (or arXiv:2611.21639v1 [cs.HC] for this version)
  https://doi.org/10.48550/arXiv.2611.21639

arXiv-issued DOI via DataCite (pending registration)

Submission history

From: Ben Xiaolu Huang [view email]
[v1] Fri, 25 Sep 2026 13:46:32 UTC (1,179 KB)

来源:arXiv cs.CL · arxiv.org