Detection of Number of Words in Continuous Colloquial Bengali Speech
Word segmentation is a crucial part in any speech to text conversion. Many works has been done on popular languages, especially on English. But a very few work is conducted on Bengali language, especially on colloquial speech. In this paper, we present a simple technique to find the number of words in a Bengali speech. We extract the feature of the existence of words based on the pitch profile of a speech. To find the pitch profile, we have used the state phase technique. A simple deviation of a 20ms window is studied to fine the pitch. In order to reshape the pitch, power profile of the speech is used. Finally, we cluster the pitch profile. Silhouette index is used to select the number clusters present in the data. The algorithm is tested over continuous colloquial Bengali speech and the results are significantly good.
Keywords - Speech ToText, State Phase Analysis, Quasi-periodic, Quasi-random, Average Magnitude Difference function