Scholarlink

<p>Dysarthria is a motor speech disorder caused by neurological impairments that affect speech intelligibility. Extraction of acoustic features specific to dysarthria is very challenging. The acoustic models, along with dysarthria, also capture speaker-specific information. Consequently, automatic dysarthria severity classification systems exhibit significantly reduced performance in speaker-independent (SI) test scenarios. Motivated by these observations, this paper investigates advanced deep learning architectures to develop an effective dysarthria severity classification system in SI test mode. The paper first examines the performance of the time-delay neural network (TDNN) using log Mel filterbank energies (LMFBE) and Mel-frequency cepstral coefficients (MFCC). Next, channel attention, propagation, and aggregation in TDNN (ECAPA-TDNN) and the residual network 34-squeeze-and-excitation (ResNet34-SE) are studied to develop dysarthria severity classification systems. The experimental results on the UA (Universal Access) speech corpus show that in SI test scenarios, the ECAPA-TDNN with MFCC achieves the highest accuracy of 71.28%. This improvement is mainly due to the effectiveness of ECAPA-TDNN in capturing long-range temporal dependencies in speech signals. The integration of the self-supervised learning-based Wav2Vec2.0 further enhances the accuracy of ECAPA-TDNN to 72.77%, which is one of the best results reported on this database. The generalization capability of the proposed framework is further validated on the TORGO database. These findings underscore the potential of advanced deep learning models and self-supervised learning in developing automated dysarthria severity classification systems.</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

Improving Performance of the Automatic Dysarthria Severity Classification System using Self-Supervised Learning and Advanced Deep Learning Techniques

  • 期刊论文

  • 发表日期：

    2026年5月6日

查看全文
  • Md. Talib Ahmad,
  • Gayadhar Pradhan,
  • Jyoti Prakash Singh

摘要

Dysarthria is a motor speech disorder caused by neurological impairments that affect speech intelligibility. Extraction of acoustic features specific to dysarthria is very challenging. The acoustic models, along with dysarthria, also capture speaker-specific information. Consequently, automatic dysarthria severity classification systems exhibit significantly reduced performance in speaker-independent (SI) test scenarios. Motivated by these observations, this paper investigates advanced deep learning architectures to develop an effective dysarthria severity classification system in SI test mode. The paper first examines the performance of the time-delay neural network (TDNN) using log Mel filterbank energies (LMFBE) and Mel-frequency cepstral coefficients (MFCC). Next, channel attention, propagation, and aggregation in TDNN (ECAPA-TDNN) and the residual network 34-squeeze-and-excitation (ResNet34-SE) are studied to develop dysarthria severity classification systems. The experimental results on the UA (Universal Access) speech corpus show that in SI test scenarios, the ECAPA-TDNN with MFCC achieves the highest accuracy of 71.28%. This improvement is mainly due to the effectiveness of ECAPA-TDNN in capturing long-range temporal dependencies in speech signals. The integration of the self-supervised learning-based Wav2Vec2.0 further enhances the accuracy of ECAPA-TDNN to 72.77%, which is one of the best results reported on this database. The generalization capability of the proposed framework is further validated on the TORGO database. These findings underscore the potential of advanced deep learning models and self-supervised learning in developing automated dysarthria severity classification systems.