Algorithm for Constructing and Analyzing Spectrograms of Audio Signals

Abstract

The paper describes one of the areas of sound signals research sound analysis using spectrograms as a means of visualizing dynamic changes in the intensity of the frequency components of the signal. Due to the fact that sound and, in particular, speech messages remain the most natural form of information exchange, this area is in demand in various technologies related to the processing of audio data. Spectrograms are used by recording studios to remove noise from musical works recorded on old analog media. In human speech recognition technologies, spectrograms are a promising source of data for analyzing the formant composition of speech sounds using neural networks focused on image analysis. Therefore, obtaining an image of high clarity and contrast, allowing stable identification of formants, both in music and in speech, seems to be an urgent task. Known algorithms for constructing spectrograms are based on the use of a discrete Fourier transform, which is due to the presence of a fast transformation algorithm (FFT), which significantly reduces computational costs. The paper points out the shortcomings of the FFT algorithm that may arise when studying the properties of speech signals and presents a new method for constructing spectrograms based on subband representations. The method is based on the use of subband matrices. The work demonstrates the effectiveness of the proposed approach, which consists in a clearer display of areas where the energy of the analyzed sound signal is concentrated, compared to known methods.