[A multiscale feature extraction algorithm for dysarthric speech recognition].

Area of Science:

• Speech Processing
• Machine Learning
• Biomedical Engineering

Context:

• Dysarthria significantly impairs speech intelligibility, posing challenges for accurate speech recognition.
• Existing speech recognition systems struggle with the unique acoustic characteristics of dysarthric speech.
• High computational complexity and feature loss are common issues in single-channel neural network training for speech tasks.

Purpose:

• To develop a multi-scale mel domain feature map extraction algorithm for improved dysarthria speech recognition.
• To propose an effective speech recognition network model addressing feature loss and computational complexity.
• To enhance the accuracy and efficiency of automatic speech recognition (ASR) for dysarthric speakers.

Summary:

• A novel algorithm decomposes speech signals using empirical mode decomposition, extracting Fbank features and first-order differences from effective components to create detailed frequency domain feature maps.
• A new speech recognition network model is proposed to mitigate effective feature loss and reduce computational complexity during training.
• The model was trained and validated on the UA-Speech dataset, demonstrating significant performance improvements.

Impact:

• The proposed algorithm and model achieved a 92.77% accuracy rate in recognizing dysarthric speech.
• This approach offers a promising solution for improving communication accessibility for individuals with dysarthria.
• The findings contribute to the advancement of assistive technologies for speech disorders.