Computer-based stuttered speech detection system using Hidden Markov Model

Abstract

Stuttering has attracted extensive research interests over the past decades. Most of the available stuttering diagnostics and assessment technique uses human perceptual judgment to overt stuttered speech characteristics. Conventionally, the stuttering severity is diagnosed by manual counting the number of occurrences of disfluencies of pre-recorded therapist-patient conversation. It is a time-consuming task, subjective, inconsistent and easily prone to error across clinics. Therefore, this thesis proposes a computerized system by deploying HMM-based speech recognition technique to detect the stuttered speech disfluency. The continuous Malay digit string has been used as the training and testing set for fluency detection. Hidden Markov Model (HMM) is a robust and powerful statistical-based acoustic modeling technique. With their efficient training algorithm (Forward-backward, Baum-Welch algorithms) and recognition algorithm, as well as its modeling flexibility in model topology and other knowledge sources, HMM has been successfully applied in solving various tasks. In this thesis, a set of normal voice for digit string as database is used for training HMM. Then, the pseudo stuttering voice was collected as testing set for proposed system. The generated experimental results were compared with the results made by Speech Language Pathologist (SLP) from Clinic of Audiology and Speech Sciences of Universiti Kebangsaan Malaysia (UKM). As a result, the proposed system is proven to be capable to achieve 100% average syllable repetition detection accuracy with 86.605% average sound prolongation detection accuracy. The SLP agreed with the result generated by the software. This system can be further enhanced for detecting stuttering disorder for daily speaking words where Microsoft Visual C++ 6.0 and Goldwave have been used for developing the software which can be executed under the window-based environment.