Optimal Feature Learning for Speech Emotion Recognition – A DeepNet Approach

The identification of a speaker's state of mind via the use of algorithms that use machine learning is a significant challenge in the field of speech emotion recognition (SER). SER is essential for analyzing speaker' emotions in ERs as well as immediate uses such as evaluating how people act, communicating with robots, and utilizing augmented reality (AR). To enhance the precision of identification as well as complexity of the model, previous research in this field has focused primarily on explicitly constructed characteristics as well as traditional convolutional neural network, or CNN, methods for extracting high-level characteristics from speech spectrograms. In contrast, we provide an innovative model for SER that uses similarity measures between clusters in redial-based functional nets (RBFNs) to determine which parts of the sequence are most important. The selected sequences is converted into a spectrum using the STFT technique, whereby then serves as inputs to the CNN method, which extracts the salient as well as distinguishing features of the uttered spectrogram. We also normalized the CNN characteristics to provide reliable identification effectiveness. The extensive short-term and long-term memory bidirectional (BiLSTM) is then fed these traits to acquire the temporal data required to identify the ultimate emotional state. The proposed technique restores the characteristics of CNN before analyzing them, allowing it to readily detect the spatial information, while also reducing the computational difficulty of the entire models by analyzing just the relevant portions of the speech. The proposed system is evaluated on industry-standard databases including IEMOCAP, EMO-DB, as well as RAVDESS to improve recognition precision as well as decrease models time for processing. The results demonstrate the robustness and efficacy of the suggested SER approach, which achieves up to 72.25 percent reliability on the IEMOCAP data set, 85.5 percent reliability on the EMO-DB data set, as well as 77.02 percent reliability on the RAVDESS database when compared with cutting-edge SER techniques.