A Study on Noisy Speech Recognition

The paper concentrates on a speech recognition algorithm to work with speech samples of nonhomogeneous quality. The speech samples are acquired using different microphones and are of different quality. Two algorithms of feature extraction are utilized,including the use of Toeplitz matrices and distances of feature points from the Cartesian origin point (0;0) as a reference point. For classification the Nearest Neighbor approach is used. The obtained results are promising.The paper also involves the description of the process in the preparation of speech samples. The approach to estimate the frequency range which contains enough information for proper speech recognition is undertaken. The studies in this paper show that cutting frequencies above 2200 Hz have rather low influence on the proper recognition but may rather lead to increase the error rate.