A SPEAKER VERIFICATION SYSTEM USING PROSODIC FEATURES

This paper describes a technique for automatic speaker verification based on prosodic knowledge in Hindi using neural networks. Properties of intonation patterns (changes in F0 as a function of time) and duration were exploited to extract speaker specific information from natural speech utterances, which were used for fixed text speaker verification task. A set of twenty-three features (fifteen pitch features and eight durational features) were extracted from a fixed natural utterance, using a word boundary hypothesization algorithm. A neural network model based on adaptive resonance theory (ART2) was used to verify speaker from the input feature set. The system was trained for twenty five speakers and tested with twenty seven impostors. The results show that the overall percentage of correct acceptance and correct rejection was found to be about 98%.