Scholarlink

<p>Speech-based detection of cognitive impairment offers a non-invasive alternative to traditional diagnostic methods. This study explores the effectiveness of Wav2Vec, a self-supervised model that extracts features directly from raw audio, compared to conventional handcrafted features such as Mel-Frequency Cepstral Coefficients (MFCCs). Using the ADReSSo dataset, we evaluate three classifiers: Support Vector Machine (SVM), Time-Delay Neural Network (TDNN), and Bidirectional Long Short-Term Memory (BiLSTM). Results show that Wav2Vec combined with SVM achieves 71% accuracy and 74% F1-score, outperforming MFCC-based models, which reach 61% accuracy and 67% F1-score. Layer-wise analysis reveals that lower Wav2Vec layers (e.g. Layer 4) capture acoustic features crucial for identifying cognitive decline, while deeper layers overfit and degrade performance. These findings highlight Wav2Vec’s superiority in extracting meaningful speech representations and support its use in scalable, accurate systems for early detection and continuous monitoring of cognitive impairment.</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

Layer-wise analysis of Wav2Vec for early detection of cognitive decline

  • 期刊论文

  • 发表日期：

    2026年1月31日

查看全文
  • Sri Lakshmi Talasila,
  • Vijayakumari Rodda

摘要

Speech-based detection of cognitive impairment offers a non-invasive alternative to traditional diagnostic methods. This study explores the effectiveness of Wav2Vec, a self-supervised model that extracts features directly from raw audio, compared to conventional handcrafted features such as Mel-Frequency Cepstral Coefficients (MFCCs). Using the ADReSSo dataset, we evaluate three classifiers: Support Vector Machine (SVM), Time-Delay Neural Network (TDNN), and Bidirectional Long Short-Term Memory (BiLSTM). Results show that Wav2Vec combined with SVM achieves 71% accuracy and 74% F1-score, outperforming MFCC-based models, which reach 61% accuracy and 67% F1-score. Layer-wise analysis reveals that lower Wav2Vec layers (e.g. Layer 4) capture acoustic features crucial for identifying cognitive decline, while deeper layers overfit and degrade performance. These findings highlight Wav2Vec’s superiority in extracting meaningful speech representations and support its use in scalable, accurate systems for early detection and continuous monitoring of cognitive impairment.