Neural Network Based Curve Fitting to Enhance the Intelligibility of Dysarthric Speech

• Published in: Speech and Computer Vol. 13721; pp. 545 - 553
• Main Authors: Pavithra, K. S., Chandrashekar, H. M., Karjigi, Veena
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2022; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Curve fitting; Dysarthria; Intelligibility; Neural network; Synthesis
• ISBN: 3031209796; 9783031209796
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-031-20980-2_46

Dysarthria is a motor speech disorder resulting from disturbance in neuromuscular control. The speech produced by people with dysarthria is distorted speech, whose intelligibility is poor compared to the normal speakers. This work attempts to increase the intelligibility of dysarthric speech by using a fitting function neural network transformation model created by Levenberg-Marquardt algorithm and Bayesian Regularization algorithm. The Linear Predictive coefficients from dysarthric speech signal and normal speech signal are taken as input and target for the fitting model respectively. The modified LP coefficients are obtained for the test dysarthric speech signal using transformation model and modified speech signal is reconstructed using LP synthesis followed by Overlap and Add method. It is observed that the mean opinion score is increased from 1.24 to 1.37 and 1.32 after modification for given set of dysarthric speech signals with Levenberg-Marquardt algorithm and Bayesian Regularization algorithm respectively.