Scholarlink

<p>Dysarthria, a motor speech disorder resulting from neurological conditions, significantly impairs clear communication. It can arise from stroke, Parkinson’s disease, cerebral palsy, or traumatic brain injury, affecting approximately 2% of the global population. Traditional diagnostic methods, which rely on subjective evaluations by speech-language pathologists, are often time-consuming and susceptible to inaccuracies. This study addresses these challenges by integrating Generalized Representation of Audio and Music (GRAM) features with Kolmogorov-Arnold Networks (KANs) to form a GRAM-KAN model. The proposed system employs spectrograms, chromagrams, and spectral analysis for audio feature extraction, effectively capturing the nuances of dysarthric speech. Unlike multilayer perceptrons (MLPs), KANs utilize learnable activation functions on network edges, enhancing flexibility and interpretability. The model was trained and tested on the Easy Call corpus, comprising 21,386 audio recordings from 55 speakers (31 with dysarthria and 24 without). It achieved an accuracy of 99.3%, surpassing existing methods such as convolutional neural networks (97% accuracy) and artificial neural networks (98% accuracy), with minimal false positives and false negatives. The system demonstrated a precision of 1.00, a recall of 0.99, and an F1-score of 0.99. SHAP (Shapley Additive Explanations) analysis identified spectrogram formant frequencies and spectral filtering as the most influential features. This study offers three primary contributions: (1) integrating GRAM features with KANs for improved dysarthria detection, (2) developing a comprehensive audio feature extraction pipeline, and (3) validating the system on a large-scale dataset. This work establishes a robust framework for early diagnosis, personalized treatment planning, and clinical decision support for individuals with dysarthria.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

The GRAM-KAN approach: dysarthria speech identification with easy call corpus

  • 期刊论文

  • 发表日期：

    2025年6月16日

查看全文
  • Jothieswari Jayaprakash,
  • Suguna Sangaiah

摘要

Dysarthria, a motor speech disorder resulting from neurological conditions, significantly impairs clear communication. It can arise from stroke, Parkinson’s disease, cerebral palsy, or traumatic brain injury, affecting approximately 2% of the global population. Traditional diagnostic methods, which rely on subjective evaluations by speech-language pathologists, are often time-consuming and susceptible to inaccuracies. This study addresses these challenges by integrating Generalized Representation of Audio and Music (GRAM) features with Kolmogorov-Arnold Networks (KANs) to form a GRAM-KAN model. The proposed system employs spectrograms, chromagrams, and spectral analysis for audio feature extraction, effectively capturing the nuances of dysarthric speech. Unlike multilayer perceptrons (MLPs), KANs utilize learnable activation functions on network edges, enhancing flexibility and interpretability. The model was trained and tested on the Easy Call corpus, comprising 21,386 audio recordings from 55 speakers (31 with dysarthria and 24 without). It achieved an accuracy of 99.3%, surpassing existing methods such as convolutional neural networks (97% accuracy) and artificial neural networks (98% accuracy), with minimal false positives and false negatives. The system demonstrated a precision of 1.00, a recall of 0.99, and an F1-score of 0.99. SHAP (Shapley Additive Explanations) analysis identified spectrogram formant frequencies and spectral filtering as the most influential features. This study offers three primary contributions: (1) integrating GRAM features with KANs for improved dysarthria detection, (2) developing a comprehensive audio feature extraction pipeline, and (3) validating the system on a large-scale dataset. This work establishes a robust framework for early diagnosis, personalized treatment planning, and clinical decision support for individuals with dysarthria.