基于常数Q变换的构音障碍语音自动检测和评估  

Automatic Dysarthric Speech Detection and Severity Assessment with Constant Q Transform

作　　者：孙国伦 王丽 SUN Guolun;WANG Li(Key Laboratory of Speech Acoustics and Content Understanding,Institute of Acoustics,Chinese Academy of Sciences,Beijing,100190,China;University of Chinese Academy of Sciences,Beijing,100049,China)

机构地区：[1]中国科学院声学研究所语音与智能信息处理实验室,北京100190 [2]中国科学院大学,北京100049

出　　处：《网络新媒体技术》2025年第1期16-25,40,共11页Network New Media Technology

基　　金：国家重点研发计划(编号:2022YFF0608501)。

摘　　要：针对构音障碍语音的自动检测与严重程度评估问题,提出一种融合构音障碍语音病理特性的自动检测和评估方法。通过结合常数Q变换的频谱以及源-滤波器假设,将频谱分解,并使用常数Q变换频谱及其分量来捕捉构音障碍患者发音过程中的共振峰歪曲等发音特点。同时,采用密集连接网络对含有病理特性的声学特征建模实现构音障碍检测与严重程度评估。实验结果表明,该方法在英语、意大利语和中文等语种的语音自动检测任务上分别取得2%以上的准确率绝对值提升;在英语和中文语料的构音障碍严重程度评估任务上分别取得2%和10%以上的准确率绝对值提升。这表明该方法能够在不同语种和不同任务上一致提升构音障碍语音建模性能。This study addresses the issue of automatic detection and severity evaluation of dysarthric speech by proposing an integrated approach that incorporates the pathological features of dysarthric speech.Leveraging the spectral characteristics of the Constant-Q Transform(CQT)in conjunction with the source-filter hypothesis,the method decomposes the spectrogram components to capture the formant distortions and other articulatory traits exhibited by dysarthric speakers.Densenet encoder is then employed to model these pathological acoustic features,thereby establishing an automatic detection and assessment framework for dysarthric speech.The experimental results demonstrate significant improvements in accuracy,with an absolute increase of about 2%for the automatic detection tasks across English,Italian,and Chinese language datasets,and an absolute increase of more than 2%and 10%for the severity assessment tasks in English and Chinese datasets.These results suggest that the proposed method is effective in enhancing the performance of dysarthric speech modeling consistently across different languages and tasks.

关 键 词：构音障碍语音 语音共振峰 常数Q 变换 密集连接网络 源-滤波器 

分 类 号：TN912.3[电子电信—通信与信息系统] R767.92[电子电信—信息与通信工程]