Investigating the Recognition of Non-articulatory Sounds by Using Statistical Tests and Support Vector Machine

Abstract

People with articulation and phonological disorders need training to plan and to execute sounds of speech. Compared to other children, children with Down Syndrome have significantly delayed speech development because they present developmental disabilities, mainly apraxia of speech. In practice, speech therapists plan and perform trainings of articulatory and non-articulatory sounds such as blow production and popping lips in order to assist speech production. Mobile applications can be integrated into the clinical treatment to transcend the boundaries of clinics and schedules and therefore reach more people at any time. The use of artificial intelligence and machine learning techniques can improve this kind of application. The aim of this pilot study is to assess speech recognition methods prioritizing the training of sounds for speech production, particularly the non-articulatory sounds. These methods apply Mel-Frequency Cepstrum Coefficients and Laplace transform to extract features, as well as traditional statistical tests and Support Vector Machine (SVM) to recognize sounds. This study also reports experimental results regarding the effectiveness of the methods on a set of 197 sounds. Overall, SVM provides higher accuracy.