ADAPTIVE FILTERING AND MACHINE LEARNING METHODS IN NOISE SUPPRESSION SYSTEMS, IMPLEMENTED ON THE SoC

Context. Modern video conferencing systems work in different noise environments, so preservation of speech clarity and provision of quick adaptation to changes in this environment are relevant tasks. During the development of embedded systems, finding a balance between resource consumption, performance, and signal quality obtained after noise suppression is necessary. Systems on a chip allow us to use the power of both processor cores available on the hardware platform and FPGAs to perform complex calculations, which contributes to increasing the speed or reducing the load on the central SoC cores. Objective. To conduct a comparative analysis of the noise suppression quality in audio signals by an adaptive filtering algorithm and a filtering algorithm using machine learning based on the RNNoise neural network in noise suppression devices on the technological platform SoC. Method. Evaluation using objective metrics and spectrogram analysis using the Librosa library in Python. Neural network training and model design are performed on the basis of Python and Torch tools. The Vitis IDE package was used for the neural network implementation on the platform SoC. Results. The analysis of two noise suppression methods using the adaptive Wiener filter and the RNNoise neural network was performed. In the considered scenarios, it was determined that the neural network shows better noise suppression results according to the analysis of spectrograms and objective metrics. Conclusions. A comparative analysis of the effectiveness of noise suppression algorithms based on adaptive filters and a neural network was performed for scenarios with different noise environments. The results of objective SIGMOS metrics were obtained to evaluate the quality of the received audio signal. In addition, the possibility of running the RNNoise neural network on the technological platform SoC ZYNQ 7000 was verified.