Development of High Accuracy Classifier for the Speaker Recognition System

Area of Science:

• Speech processing
• Biometric recognition
• Machine learning

Background:

• Speech signals contain rich features for biometrics, gender, and emotion recognition.
• Background noise and reverberation challenge speech feature consistency and recognition accuracy.
• Existing models struggle with feature shifts caused by adverse channel conditions.

Purpose of the Study:

• To develop a hybrid speaker identification model for consistent speech features.
• To achieve high recognition accuracy despite challenging channel conditions.
• To enhance noise immunity in speaker recognition systems.

Main Methods:

• Incorporated pitch frequency coefficient into Mel frequency spectrum coefficients (MFCC).
• Proposed a single hidden layer feed-forward neural network (FFNN).
• Tuned the FFNN using an optimized particle swarm optimization (OPSO) algorithm.

Main Results:

• Achieved 97.83% recognition accuracy in clean environments.
• Demonstrated significantly less impact from noisy channels compared to baseline classifiers (FFNN, RF, KNN, SVM).
• The hybrid model showed enhanced noise immunity across various Adaptive White Gaussian Noise (AWGN) levels (0-50 dB).

Conclusions:

• The proposed hybrid model effectively maintains speech feature consistency and recognition accuracy.
• The FFNN optimized by OPSO offers superior noise immunity for speaker identification.
• This approach presents a robust solution for speaker recognition in real-world noisy environments.