Development of vocal tract length normalized phonetic engine for Gujarati and Marathi languages

摘要

Phonetic engine (PE) is a system that converts speech sound units into symbols without any higher-level information (such as semantic or linguistic details). This paper presents the development of PE in two Indian languages, viz., Gujarati and Marathi. To investigate the performance of PE, speech recorded in three different modes, viz., read, spontaneous and lecture is considered. Database consists of a large number of speakers in each mode for these languages. In order to reduce the effects of speaker differences in the databases, Vocal Tract Length Normalization (VTLN) using Lee-Rose method is incorporated. Here, performances of PEs are tested using state-of-the-art Mel frequency cepstral coefficients (MFCC) and vocal tract length normalized features. Hidden Markov model (HMM)-based approach is used for modeling the phonetic units. On an average, improvement of 3.12 % and 1.32 % is achieved using vocal tract length normalized PE over MFCCs for Gujarati and Marathi, respectively.