Dysarthric Speech Enhancement Based on Convolution Neural Network.

Area of Science:

• Speech processing
• Artificial intelligence
• Neurolinguistics

Background:

• Dysarthria often results in distorted speech with reduced intelligibility for humans and machines.
• Conventional speech enhancement (SE) methods primarily focus on noise reduction, which may not fully address severe signal distortion in dysarthric speech.

Purpose of the Study:

• To develop and evaluate a deep learning-based SE system for reconstructing severely distorted dysarthric speech signals.
• To enhance the intelligibility and automatic speech recognition (ASR) accuracy of dysarthric speech.

Main Methods:

• A convolutional neural network (CNN) model was trained using paired dysarthric and normal speech utterances.
• Dynamic time warping (DTW) was employed to align the speech utterances for effective model training.
• The trained CNN-based SE system was evaluated using the Google ASR system and a subjective listening test.

Main Results:

• The proposed SE system demonstrated a notable enhancement in recognition performance, exceeding 10% for both ASR and human recognition compared to unprocessed dysarthric speech.
• The system effectively reconstructed severely distorted speech signals, leading to improved intelligibility.

Conclusions:

• Deep learning-based speech enhancement, specifically using a CNN model, can significantly improve the intelligibility and ASR accuracy of dysarthric speech.
• This approach offers a promising method for clinical relevance, aiding communication for individuals with dysarthria.