Dysarthric Speech Data Augmentation using Generative Adversarial Networks

Both the issue of data scarcity and the problem of class imbalance are factors that hinder the performance of automated dysarthric speech recognition (ADSR) systems. The use of data augmentation allows for the creation of synthetic speech samples, which helps to reduce the effects of data scarcity and class imbalance issues. Traditional statistically-based data augmentation approaches, on the other hand, have shown that ADSR systems have less than satisfactory speech intelligibility and accuracy. The purpose of this study is to provide the augmentation of dysarthric speech data via the use of a Generative Adversarial Network (GAN) that maintains the fluctuations in quality, prosodic variations, and intonation across the synthetic samples. The Mel Log Frequency Spectrogram (MFLS) features and the Deep Convolution Neural Network (DCNN) are used in order to verify the results of the system for the DSR. Mean square error (MSE) and peak signal-to-noise ratio (PSNR) are the metrics that are used to assess the efficacy of the system on the UASpeech dataset. After applying the GAN-based data augmentation, the total PSNR was determined to be 36.59, the MSE was 48.53, and the MAE was 18.24. An enhanced accuracy of 95.2%, recall of 95.2%, precision of 95.7%, net present value of 94.8%, and selectivity of 95.7% were all provided by the DCNN-GAN solution.