Адаптація архітектури глибинної згорткової моделі для емоційного аналізу в edge ai системах

Abstract

This work is devoted to the selection of neural network architectures for the task of speech-based emotion recognition (SER). The objective is to enhance recognition performance on devices with constrained computational resources. For model training, publicly available datasets were employed, from which log-mel spectrograms were extracted as input representations. An experimental evaluation was carried out to investigate the effectiveness of lightweight adaptations of several widely used convolutional neural architectures.