Modified segmentation algorithm based on Short Term Energy & Zero Crossing Rate for Maithili speech signal

Converting the voice signals (uttered in Maithili language) to text has lots of applications including speaker identification, voice mode interaction, bio-metric identification etc. Most of the research papers reported in literature, till date, concentrated mainly in English. Some works are also available for the languages Hindi, Bengali and Tamil. But working with regional languages like Maithili, Bhojpuri and Magahi etc remains untouched. So, in this research work one of the most popular regional language of Bihar state under India territory, Maithili language is selected for our case study. In order to segment a voice signal firstly the word boundary is detected from the voice signals of sentences using Short Term Energy (STE) and Zero Crossing Rate (ZCR). Since there exist a large number of words for the Maithili language; converting the words directly to a text is a complex and tedious job. Hence words are further dived into syllables, which are mostly unique in terms of signal and in manageable in size. Detection of rough syllable boundary from the word boundary has been performed by group delay algorithm. To make these syllable boundaries more accurate group delay algorithm has been modified by considering the differences of consecutive peaks in the negative derivative phase. Accuracy is quantified through error rate measured by taking the differences between the ground truth boundary of the signals (determine manually) and system detected boundary of the syllable signals. Experiment using the data set of 25 different sentences from 10 different speakers have been performed and an accuracy rate of 85.62% is computed using above algorithm.