Implementation of Multiscale Conv-GRU-based Efficient Emotion Recognition Framework using Speech Signals based on Federated Learning

In real-life communication, nonverbal elements including laughter, crying, and other emotional interjections play a key role in conveying feelings. Emotions are essential for human interactions, aiding in mutual understanding. To enhance human-computer interactions and advance speech-based emotion recognition technologies, it is crucial to accurately identify emotions. Extracting emotional context from speech remains a significant challenge in the domain of AI. Building datasets for speech emotion recognition is a demanding and expensive task, resulting in limited data availability and uneven emotion representation. This imbalanced dataset can affect the performance and dependability of emotion detection systems. Innovative approaches are necessary to address the limitations of traditional emotion recognition methods. In this study, a novel deep learning-based speech-emotion recognition framework incorporating the principles of federated learning is proposed. In the initial phase, required speech signals for the validations are sourced from the benchmark resources. Next, from the collected speech signals, ensemble features like Mel-Frequency Cepstral Coefficients (MFCC), spectral features and sparse autoencoder-based deep features are extracted. Therefore, the extracted ensemble features are subjected to a Multiscale Convolutional Gated Recurrent Unit (MS-ConvGRU) to execute the emotion recognition procedure. Here, the ensemble features are offered in multiple scales on CGRU. Finally, the emotion recognition outcomes are obtained from MSC-GRU mode based on Federated Learning (FL). Later, various experiments are executed in the developed framework over classical techniques to verify its overall efficiency in recognizing emotions.