Two-stage algorithm of spectral analysis for the automatic speech recognition systems

The problem of the spectral analysis of speech signals in automatic speech recognition systems is considered within the framework of a dynamically developed direction of investigations in the field of acoustic measurements. We indicate that efficiency of the analyzed systems under unfavorable conditions of speech production (noise and insufficient intelligibility of speech sounds) is low as compared with human perception of oral speech. To improve the efficiency of automatic speech recognition systems, we propose to use a two-stage algorithm of spectral analysis of the speech signals. The first stage of processing of speech signals is their parametric spectral analysis performed by using an autoregressive model of the vocal tract of a conventional speaker. The second stage of processing is the transformation (modification) of the obtained spectral estimate according to the principle of frequency-selective amplification of the amplitude of main formants of the intraperiod power spectrum. The software implementation of the proposed algorithm is described on the basis of the computational procedure of fast Fourier transform. By using the software developed by the authors, we performed full-scale experiments and studied an additive mixture of vowel sounds in the speech of a control speaker with white Gaussian noise. The obtained experimental results enable us to conclude that the amplitudes of the main formants of speech signals are amplified by 10–20 dB and, hence, the intelligibility of speech sounds substantially improves. The developed algorithm can be used in the automatic speech recognition systems based on processing of the speech signals in the frequency domain, including the use of artificial neural networks.