Abstract
Employing stress in speech can transfer more information to a listener but makes more problems in speech recognition. The first step toward stressed speech recognition is the recognition of boundaries in stressed speech. In this research, the boundaries of prosodic stress were extracted in Farsi stressed sentences. The acoustic and prosodic features were used to train hidden Markov models for stress boundaries recognition. Using fast correlation-based filter (FCBF) method, the efficient features were selected for stress recognition. The influence of different feature sets on stress boundaries recognition performance was evaluated in this study. Based on this evaluation, a combined classifier scheme was proposed. Experimental results showed that the proposed combined model improved the stress boundaries detection performance by 12% as compared to the baseline model. So, the final recognition rate of the proposed classifier was 85% for prosodic stress boundaries recognition.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have