Spectro-temporal directional derivative based automatic speech recognition for a serious game scenario

Abstract

Speech is one of the important modalities in a serious game platform. Serious game can be very useful for the rehabilitation of individuals with voice disorders. Therefore, we need an efficient and high-performance automatic speech recognition (ASR) system. In this paper, we propose a spectro-temporal directional derivative (STDD) feature that requires less number of computations in the modeling and yet gives high recognition accuracy in the ASR system. The proposed STDD feature is achieved by applying different directional derivative filters in the spectro-temporal domain. The feature dimension is then compressed by discrete cosine transform. The experiments are performed with voice samples of Arabic numerals spoken by persons with and without voice pathology. The experimental results show that the STDD feature outperforms the conventional mel-frequency cepstral coefficients both in clean and noisy environments.