High-Level CNN and Machine Learning Methods for Speaker Recognition.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Speech Processing

Background:

• Speaker Recognition (SR) is crucial in AI sound analysis, utilizing diverse methodologies like Deep Learning (DL) and traditional Machine Learning (ML).
• Evaluating these methods on diverse datasets with emotional variations is essential for real-world applicability.

Purpose of the Study:

• To compare the performance of a custom Convolutional Neural Network (CNN) against pre-trained networks and a Naïve Bayes model for speaker recognition.
• To investigate the effectiveness of different input representations (spectrograms, MFCCs) and feature selection techniques.

Main Methods:

• A custom CNN and pre-trained networks (e.g., AlexNet) were trained using spectrograms and Mel-frequency cepstral coefficients (MFCCs).
• A traditional ML approach employed acoustic feature extraction, selection, and a Naïve Bayes classifier.
• Experiments were conducted on the DEMoS dataset, comprising 8869 audio files from 58 speakers with varying emotional states.

Main Results:

• The custom CNN achieved the highest accuracy (90.15% on grayscale spectrograms).
• AlexNet demonstrated comparable performance (89.28% on spectrograms).
• The Naïve Bayes classifier offered 87.09% accuracy and high AUC (0.985) with faster training and better interpretability. Key features identified include F0, MFCC, and voicing-related parameters.

Conclusions:

• Custom CNNs, particularly when trained on grayscale spectrograms, show superior accuracy for speaker recognition in emotionally diverse scenarios.
• Traditional ML methods like Naïve Bayes provide a viable, interpretable, and computationally efficient alternative.
• The DEMoS dataset's characteristics, including emotional content and sample size, facilitate robust model generalization for speaker recognition.