Dysarthric Speech Data Augmentation Using Generative Adversarial Networks

The data scarcity problem and class imbalance problems limit the performance of automatic Dysarthric speech recognition (ADSR) systems. Data augmentation helps to create synthetic voice samples to minimize data scarcity and class imbalance problems. However, traditional statistical-based data augmentation techniques have shown poor voice intelligibility and accuracy of ADSR systems. This paper presents the Dysarthric speech data augmentation using a Generative Adversarial Network (GAN) that retains the quality, prosodic, and intonation variations over the synthetic samples. The outcomes of the system are validated for the DSR using Mel Log Frequency Spectrogram (MFLS) features and Deep Convolution Neural network (DCNN). The effectiveness of the system is validated on the UASpeech dataset based on mean square error (MSE) and peak signal-to-noise ratio (PSNR). The GAN based data augmentation resulted in overall PSNR of 36.59, MSE of 48.53 and MAE of 18.24. The DCNN-GAN offered improved accuracy of 95.2%, recall of 95.2%, precision of 95.7%, NPV of 94.8%, and selectivity of 95.7%.