Paper Title
Speech Emotion Recognition Through Federated Learning for Quality Assurance in Call Centers
Abstract
For call centers, emotion recognition (ER) is a helpful application since it helps prevent customer abuse, match
clients with qualified agents, provide agent support and training, and raise client happiness. Federated learning makes it
possible to train an extremely accurate ER model while protecting privacy. FL makes sure that private information is kept on
local servers and devices and isn't shared with a centralized server. We propose a thorough approach to emotion analysis in
our project, which uses a dual encoder architecture made up of a CNN and a Bi-LSTM network. The call center management
can then receive an analytical report showing the agent's emotion trend and utilize the trained ER model to categorize the
emotions of their customers. Our method shows 88% accuracy on training data and 82% accuracy on test data in recognizing
and comprehending emotional subtleties in customer-agent discussions. Furthermore, as it can assist in identifying areas
where agents require additional training or support, ER through FL can also be utilized for quality assurance. In conclusion,
ER via FL is a potential method for leveraging machine learning in a privacy-preserving manner to improve the client
experience in call centers.
Keywords - Federated learning, neural networks, call centers, privacy-preserving, speech emotion recognition, quality
assurance, dual encoder