Grapheme Gaussian model and prosodic syllable based Tamil speech recognition system

Automatic Speech Recognition is an active field of research to identify speech patterns for providing the equivalent text. The challenges in automatic speech recognition start right from choosing the appropriate unit for the language being dealt with. Some of the units that could be used are word, phoneme, triphone, syllable, demisyllable, senone and morpheme. A syllable based Speech model is used which reduces the vocabulary size and also easier to align. This also suits Tamil, which is a syllable based language. In the model described in this paper, the connected word inputs are segmented into individual words using short term energy and the isolated words are further broken down into characters using Varied-Length Maximum Likelihood (VLML) algorithm. Gaussian Mixture Model (GMM), which is a speaker-independent model suitable for large sets of data, is used for classifying the characters for later pattern matching against the trained syllables. This paper introduces a new algorithm named VLML algorithm that is used for identifying the boundary of each character and explains the proposed process of speech recognition system for Tamil language.