Optimizing Automatic Speaker Recognition with Feature Fusion and Machine Learning Models: Insights and Performance Metrics

In the contemporary landscape where security is paramount, biometric systems, particularly speaker recognition, have gained significant prominence. This paper explores the advancements in automatic speaker recognition systems, focusing on text-independent approaches using a variety of utterance datasets. Speaker recognition utilizes unique vocal characteristics such as pronunciation patterns, rhythm, and accent to identify individuals. The paper details methodologies for feature extraction by utilizing MFCC (Mel-frequency Cepstral Coefficients) and LPC (Linear Predictive Coding), alongside machine learning models such as GMM (Gaussian Mixture Models) and SVM (Support Vender Machines). The research demonstrates the effectiveness of these techniques in clean environments and addresses challenges such as voice variability and environmental noise. Notably, the fusion of MFCC and LPC features achieves superior results, with SVM achieving 98.0% accuracy and GMM achieving 99.4% for single-word utterances. For sentence utterances, SVM and GMM reach 93.2% and 96.8% accuracy, respectively. The study also highlights the limitations of the dataset size and the potential for future work involving larger, noisier datasets and multi-user scenarios. The findings underscore the benefits of feature fusion and the promise of advanced machine learning techniques in enhancing speaker recognition accuracy.