Scholarlink

<p>Dysarthric speech recognition (DSR) plays a vital role in voice pathology, enabling the detection and analysis of impaired speech. However, challenges such as reduced speech intelligibility, high variability in speech patterns, increased reverberation, and noise degrade DSR performance. Traditional DSR systems often suffer from limitations, including weak feature representation, high complexity in deep learning architectures, and limited generalization. To address these issues, this paper presents an optimized DSRNet that considers the Multiple Acoustic Features (MAFs), the 2D representation of wavelet packet decomposition (WPT), and Mel spectrograms as inputs to enhance the feature representation of dysarthric voice. The 2D Deep Convolutional Neural Network (2D DCNN) is used to improve the local temporal and spectral variations of Mel spectrograms and the spatial correlations across different WPT subbands to localize abnormalities in dysarthric voice. The 1D DCNN helps to enhance the correlation between distinct local and global MAFs of the voice. The hyperparameters, such as learning rate, momentum, batch size, and dropout rate of the DSRNet, are optimized using a hybrid particle swarm based on the Archimedes Optimization algorithm (PSAOA). The DSRNet provides enhanced precision (0.98), recall (0.99), F1-score (0.99), accuracy (98.75%), and WER (1.25%) compared to existing state-of-the-art systems for DSR on the TORGO dataset.</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

Enhancing Dysarthric Speech Recognition Using Multi-acoustic Features and Wavelet–Spectrogram Fusion in Optimized DSRNet

  • 期刊论文

  • 发表日期：

    2026年7月8日

查看全文
  • Chandarani D. Pophale,
  • Shankar D. Chavan

摘要

Dysarthric speech recognition (DSR) plays a vital role in voice pathology, enabling the detection and analysis of impaired speech. However, challenges such as reduced speech intelligibility, high variability in speech patterns, increased reverberation, and noise degrade DSR performance. Traditional DSR systems often suffer from limitations, including weak feature representation, high complexity in deep learning architectures, and limited generalization. To address these issues, this paper presents an optimized DSRNet that considers the Multiple Acoustic Features (MAFs), the 2D representation of wavelet packet decomposition (WPT), and Mel spectrograms as inputs to enhance the feature representation of dysarthric voice. The 2D Deep Convolutional Neural Network (2D DCNN) is used to improve the local temporal and spectral variations of Mel spectrograms and the spatial correlations across different WPT subbands to localize abnormalities in dysarthric voice. The 1D DCNN helps to enhance the correlation between distinct local and global MAFs of the voice. The hyperparameters, such as learning rate, momentum, batch size, and dropout rate of the DSRNet, are optimized using a hybrid particle swarm based on the Archimedes Optimization algorithm (PSAOA). The DSRNet provides enhanced precision (0.98), recall (0.99), F1-score (0.99), accuracy (98.75%), and WER (1.25%) compared to existing state-of-the-art systems for DSR on the TORGO dataset.