Application of Convolutional Neural Networks Using Multitaper Processing to Acoustic Signal Classification Problems for River Navigation

Abstract

Sound source identification based on the temporal realization of the signal is a pressing problem for various practical applications, ranging from speech recognition to vibration defect diagnostics. This paper proposes a classification algorithm based on modern convolutional neural networks. The input data is a log-mel spectrogram obtained by multitaper processing—a power spectrum estimation method that uses orthogonal windows (tapers) to suppress noise components and improve the information content of features. Multitaper processing allows one to estimate the power spectral density of a stationary ergodic random process using a finite segment of its realization. The mel scale is a psychophysical pitch scale that simulates human perception. This paper presents a description of the algorithm and its validation using river navigation data collected using a directional sound source—a microphone located in the focal plane of a parabolic reflector. Model training is evaluated depending on the amount of data, and the performance of various convolutional networks is compared.