GFM-based methods for speaker identification

Area of Science:

• Speech Processing
• Artificial Intelligence
• Pattern Recognition

Background:

• Speaker identification is crucial for security and personalized systems.
• Existing models face challenges with noisy environments and feature extraction.
• Hybrid models offer potential for improved accuracy and robustness.

Purpose of the Study:

• To propose three novel speaker identification methods.
• To leverage the strengths of hidden Markov models (HMM) and generalized fuzzy models (GFM).
• To evaluate model performance in real-world and noisy conditions.

Main Methods:

• Method 1: HMM for feature extraction, GFM for speaker identification.
• Method 2: Gaussian mixture model (GMM) and GFM for speaker identification.
• Method 3: A novel approach inspired by human social network analysis for speaker identification.

Main Results:

• All proposed models (HMM-GFM, GMM-GFM, HMM-GFM fusion) showed superior performance.
• Models were validated on VoxForge and NIST 2003 datasets.
• Performance remained strong even with corrupted speech data across various noise levels.

Conclusions:

• The proposed hybrid HMM-GFM and GMM-GFM models offer significant improvements in speaker identification.
• The fusion and social-network-inspired methods show promise for enhanced accuracy.
• These models provide a robust solution for speaker identification in challenging acoustic environments.