Local Linear Wavelet Neural Network and RLS for Usable Speech Classification

While operating in a co -channel environment, the accuracy of the speech processing technique degrades. When more than one person is talking at same time, then there occurs the co-channel speech. The objective of usable speech segmentation is identification and extraction of those portions of co-channel speech that are degraded in a negligible range but still needed for various speech processing application like speaker identification. Some features like usable speech measures are extracted from the co-channel signal to differentiate between usable and unusable types of speech. The features are extracted recursively by this new method and variable length segmentation is carried out by making sequential decision on class assignment of LLWNN pattern classifier. The correct classification using this technique is 84.5% whereas the false classification is 15.5%. The result shows that the proposed classifier gives better classification and is robust.