Speaker Recognition Using Constrained Convolutional Neural Networks in Emotional Speech.

Area of Science:

• Machine Learning
• Speech Processing
• Artificial Intelligence

Background:

• Speaker recognition is crucial for classification but struggles with emotional speech and interference.
• Current deep models are parameter-heavy, limiting real-time application on edge devices.
• Need for constrained, efficient speaker recognition solutions for Internet of Things (IoT) systems.

Purpose of the Study:

• To propose a simple and constrained convolutional neural network (CNN) for speaker recognition.
• To evaluate the robustness of this CNN for speaker recognition in emotional speech conditions.
• To explore quantization techniques for developing a constrained network suitable for edge devices.

Main Methods:

• Development of a simple, constrained convolutional neural network architecture.
• Examination of three quantization methods: floating-point 8-bit, ternary scalar quantization, and binary scalar quantization.
• Evaluation of model performance on the SEAC dataset, focusing on emotional speech.

Main Results:

• The proposed constrained CNN demonstrates potential for speaker recognition.
• Quantization methods are explored to reduce model complexity for edge deployment.
• Robustness in emotional speech conditions is a key focus of the evaluation.

Conclusions:

• A simple, constrained CNN offers a viable approach for speaker recognition, particularly for resource-limited environments.
• Quantization techniques are effective in creating smaller, efficient models for real-time edge applications.
• Further research is needed to fully assess robustness across diverse emotional speech datasets.