Identity Vector Extraction by Perceptual Wavelet Packet Entropy and Convolutional Neural Network for Voice...

Area of Science:

• Speech Processing
• Biometrics
• Machine Learning

Background:

• The identity vector (i-vector) model has significantly improved voice authentication accuracy.
• Existing i-vector extraction methods face challenges in noisy conditions and optimal feature representation.

Purpose of the Study:

• To propose a novel i-vector extraction method for enhanced voice authentication.
• To improve the accuracy and robustness of speaker recognition systems.

Main Methods:

• A perceptual wavelet packet transform (PWPT) converts speech into wavelet entropy feature vectors.
• A Convolutional Neural Network (CNN) estimates frame posteriors from these features.
• i-vectors are extracted using the estimated frame posteriors.

Main Results:

• The proposed method demonstrated effective i-vector extraction on TIMIT and VoxCeleb datasets.
• Experimental results showed a reduction in the equal error rate (EER).
• Improved voice authentication accuracy was observed in both clean and noisy environments.

Conclusions:

• The PWPT and CNN-based i-vector extraction method offers superior performance.
• This technique enhances the accuracy and reliability of voice authentication systems.
• The method provides a promising direction for future research in speaker recognition.