METHODS OF PROCESSING AUDIO SIGNALS

Abstract

The article examines the principal methods of acoustic signal analysis employed in contemporary research and practical applications. Special attention is devoted to three key approaches: spectral analysis, wavelet transformation, and machine learning methods, particularly neural networks. The author provides a detailed description of spectral analysis principles, which are based on Fourier transformation and its modifications (Discrete Fourier Transform, Fast Fourier Transform). The article emphasizes that spectral analysis is particularly effective for studying stationary processes, enabling precise characterization of energy distribution across frequencies. However, for analyzing non-stationary signals (such as automobile noise, musical and speech signals), time-dependent Fourier transform (STFT) is applied, which has certain limitations regarding simultaneous resolution in time and frequency domains.

Wavelet transformation is presented as an alternative mathematical tool that provides simultaneous representation of signals in both time and frequency domains. Unlike classical Fourier transformation, this method allows for the localization of spectral components in time, which is especially important for non-stationary acoustic signals. The principle of wavelet transformation involves decomposing signals into basis functions—wavelets—obtained through scaling and shifting of a mother wavelet. This approach enables the detection of signal features at different scales and at different moments in time, and also effectively reduces noise levels without significant loss of useful information.

The article also explores contemporary machine learning methods and neural networks for acoustic signal analysis. It emphasizes that over the past two decades, the use of machine learning for audio signal processing has grown substantially, and today these methods dominate new approaches to sound signal processing. Particular attention is paid to deep neural networks, which often outperform traditional signal processing methods. The author notes that despite borrowing many deep learning methods from image processing, there are important differences between these fields that require specialized approaches to audio analysis. Audio signals form one-dimensional time series that fundamentally differ from two-dimensional images and must be studied sequentially in chronological order. These properties have given rise to audio-specific solutions in the field of signal processing. The article concludes that the integration of these diverse methods allows for more comprehensive analysis of complex acoustic phenomena in various applications.