Grapheme Segmentation of Tamil Speech Signals using Excitation Information with MFCC and LPCC Features

The major components of automatic Speech Recognition(ASR)are the pronunciation dictionary, language models, acoustic model and decoder. The Pronunciation dictionaries define the mapping between the words and basic sounds of a language and thus play a vital role in speech recognition systems. Construction of the pronunciation dictionary is expensive and time consuming since it requires the knowledge of the target language. Since phoneme is context dependent, a higher unit than phoneme is considered with the aim to develop a sophisticated tool for speech to text application. The proposed segmentation algorithm is tested on the continuous speech read by 4 native speakers. Energy and spectral centroid features are used to remove the silent portion and vowel onset point is used as the anchor point to find the beginning of the vowel. The proposed method is analysed with various time tolerances and the results are presented. Keywords— Speech Segmentation, Grapheme Segmentation, MFCC, LPCC.