DYSARTHRIC SPEECH RECOGNITION USING WAVENET CONVOLUTIONAL NEURAL NETWORK WITH PROSODY CONSISTENCY LOSS FUNCTION

Dysarthric speech recognition concentrates on understanding speech impairments caused by neurological disorders. Speech recognition increases communication by enhancing adaptability and clarity for individuals. However, dysarthric speech recognition struggles to transcribe speech accurately due to variations in pitch, articulation, and rhythm that significantly differ from typical speech and processing dysarthric across different speakers. In this research, the WaveNet Convolutional Neural Network with Prosody Consistency Loss Function (WNCNN-PCLF) is proposed to recognise and classify dysarthric speech accurately. In CNN, WaveNet is incorporated for capturing local speech features that enhance the model’s ability to determine intricate variations and patterns in distorted speech. The PCLF assist in preserving natural speech patterns, which makes for more accurate rhythm and tone representation. Therefore, this integration enables better adaptation to dysarthric speech, which addresses both prosody and articulation issues effectively. Hence, the proposed WNCNN-PCLF achieves a high accuracy of 99.92% and 98.34% using UA-Speech and Kannada datasets compared to existing methods like Densely Squeezed and excitation Attention-gated Network (DySARNet).