Neural Network-Based Approach for Identification and Classification of Speech Disfluency: The Apraxia of Speech

Abstract

Over the past decade, the field of signal processing has witnessed remarkable growth, particularly in speech processing, with a substantial impact from the integration of Artificial Intelligence (AI) and Machine Learning (ML). The focus on AI/ML-based speech processing has notably advanced in the identification of voice disfluencies, particularly within biomedical applications. Given the critical nature of disfluency identification, the range of potential applications is extensive, as these inconsistencies pose challenges to effective human communication. This paper specifically delves into the examination of apraxia of speech, presenting an algorithm designed for its identification—a distinctive form of speech disfluency. The algorithm is built upon a Convolutional Neural Network (CNN) deep neural network, forming the cornerstone of its development for categorizing normal and apraxic speech. Feature extraction involves the utilization of Teager energy operators, encompassing fundamental frequency, Mel-Frequency Cepstral Coefficients (MFCC), short-term zero crossing rate (STZCR), and Teager energy (TEO). Notably, the incorporation of STZCR as a classification parameter significantly enhances the classifier's efficiency compared to TEO. The inclusion of STZCR results in an impressive 89 percent efficiency in speech categorization, whereas TEO alone yields an efficiency of 80 percent. This research underscores the pivotal role of AI/ML-based approaches in addressing speech disfluencies, particularly in the context of apraxia, contributing to advancements in early and accurate identification of communication disorders.