SPEECH RECOGNITION FRONT-END FOR SEGMENTING AND CLUSTERING CONTINUOUS BANGLA SPEECH

Abstract

This research is concerned with the development of speech recognition front-end for segmenting and clustering continuous Bangla speech sentence to some predefined clusters. From the study of different previous research works it was observed that the front-end is an important part of any speech recognition system. In our work, the original speech sentences were recorded and stored as RIFF (.wav) file format. Then a segmentation approach was used to segment the continuous speech into uniquely identifiable and meaningful units. Among the different techniques, the word/sub-word segmentation is simple and produces very good results. This is why this technique was selected for speech segmentation to obtain improved performance. After segmentation, the segmented words were clustered into different clusters according to the number of syllables and the sizes of the segmented words. The test database contained 758 words/sub-words segmented from 120 sentences. Each sentence was recorded from six different speakers and saved as a different wave file. The developed system achieved the segmentation accuracy rate at about 95%.