Automatic speech recognition system for people with speech disorders

Abstract

The conversion of speech to text is essential for communication between speech and visually impaired people. The focus of this study was to develop and evaluate an ASR baseline system designed for normal speech to correct speech disorders. Normal and disordered speech data were sourced from Lwazi project and UCLASS, respectively. The normal speech data was used to train the ASR system. Disordered speech was used to evaluate performance of the system. Features were extracted using the Mel-frequency cepstral coefficients (MFCCs) method in the processing stage. The cepstral mean combined variance normalization (CMVN) was applied to normalise the features. A third-order language model was trained using the SRI Language Modelling (SRILM) toolkit. A recognition accuracy of 65.58% was obtained. The refinement approach is then applied in the recognised utterance to remove the repetitions from stuttered speech. The approach showed that 86% of repeated words in stutter can be removed to yield an improved hypothesized text output. Further refinement of the post-processing module ASR is likely to achieve a near 100% correction of stuttering speech Keywords: Automatic speech recognition (ASR), speech disorder, stuttering