Peak Based Clustering of Continuous Bengali Speech for Medium Vocabulary Speech Recognition System

Throughout this research work, we have introduced the peak based clustering of continuous speech in Bengali language. We have done this to fill up the noble aim of using Bengali speech recognition in various areas such as text writing in the cell-phone and computer, establishing speech-based user interface, etc. To achieve this goal, original continuous speech was recorded by two different speakers in a room environment and stored as RIFF wave (.wav) file format. First, all speech blocks were segmented by our developed algorithm into words or segments. Then we incorporated moving average algorithm into the positive half cycle of the time spectrum of these segments to obtain smooth curve representation. From this curve, the peaks were traced and clustering was done through the property called number of peaks. This medium vocabulary speech database, containing a total of 355 words, was used to evaluate the system performance. The developed system achieved average segmentation accuracy at about 99.12% and clustering accuracy at about 96.53%.