Graphic Representation Method and Neural Network Recognition of Time-Frequency Vectors of Speech Information

Currently, various time-frequency representations are often used for sound analysis. These representations, on the one hand, are convenient for visible sensation of sound by a human and, on the other hand, can be used for automatically analyzing sound pictures. In this paper, various methods for representation of sound as two-dimensional time-frequency vectors of a fixed dimension and their use for speech and speaker recognition problems are discussed. Probabilistic, distance-based, and neural-network methods for the recognition of these vectors by examples of separate words are considered. Numerical experiments showed that the best among them is the method based on a three-layer neural network, the short-time Fourier transform, and the two-dimensional wavelet transformation. For the speaker recognition problem, a distance-based recognition method employing the adaptive Hermite transform turned out the best among all.