An Image-Based System for Spoken-Letter Recognition

摘要

A new trial on speech recognition from graphical point of view is introduced. Isolated spoken-letters and color-names words are considered. After recording, the speech signal is processed as an image by Power Spectrum Estimation. For feature extraction, classification and hence recognition, the algorithm of minimal eigenvalues of Toeplitz matrices together with other methods of speech processing and recognition are used. A number of examples on applications and comparisons are presented in the work. The efficiency of the method is very high in the case of the six Polish vowels and English color-names, and the results are encouraging to extend the algorithm to cover more word classes.