Transformer based Hybrid System for Speech Transcription and Disorder Detection

Stuttering is a speech disorder accompanied by disruptions in the fluency of speaking and is characterized by involuntary repetitions, prolongations, or blocks in speech. The early detection and recommendation for personalized therapies are considered to be very important in effective treatment. Deep learning and machine learning-based techniques like ANN's and CNN's have held much promise as solutions in the implementation of an automatic stuttering detection and therapy suggestion system. The paper is interested in a number of methods followed for detection of stuttering and focuses on the integration of wav2vec for learning speech representation and BERT for classification based on text for stuttering events. We review existing models and techniques; compare the performances, and an experimental evaluation on the approach presented, using wav2vec and BERT. Our results show that deep learning models, especially language and speech pre-trained models, significantly improve the stuttering detection accuracy.