Articulatory feature-based gender factor minimization in automatic speech recognition

This paper presents articulatory feature-based automatic speech recognition for Japanese spoken language. Automatic speech recognition system suffers from some hidden factors, such as speaking style, gender effects, and noisy acoustic environment. These hidden factors degrade the performance of automatic speech recognizer. Therefore, the effect of these factors should be minimized for achieving better recognition performance. In this study, we have incorporated articulatory feature based gender effects normalization technique, where male- and female-dependent DPF extractors are firstly used to map LFs onto two DPF spaces corresponding to the gender type. Two DPF vectors extracted by each DPF extractor are called DPF-male and DPF-female, respectively. These DPF extractors are trained individually with a male speech and a female speech data set. In addition, a gender-independent (GI) DPF extractor is used to compensate errors of a DPF selector. GI-DPF extractor is trained with both the male and the female speech data set. After evaluating the Tohoku University and Matsushita Spoken Word Database it is observed that the proposed method improves word correct rate and word accuracies by a certain limit.