Voiceprint Identification for Limited Dataset Using the Deep Migration Hybrid Model Based on Transfer Learning.

Area of Science:

• Machine Learning
• Biometrics
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) excel in voiceprint recognition but require extensive data.
• Limited datasets hinder deep neural network convergence and performance in practical voice biometrics.
• Existing methods struggle with effective voiceprint recognition using small sample sizes.

Purpose of the Study:

• To develop an effective voiceprint recognition method for small sample datasets.
• To improve the convergence and accuracy of deep learning models in voice biometrics.
• To propose a novel deep migration hybrid model for enhanced voiceprint recognition.

Main Methods:

• Utilized Transfer Learning to adapt pre-trained models to limited voiceprint datasets.
• Replaced fully-connected layers with Restricted Boltzmann Machine (RBM) layers in the hybrid model.
• Implemented Data Augmentation to expand the voiceprint dataset size.
• Incorporated fast batch normalization for accelerated network convergence.

Main Results:

• The proposed TLCNN-RBM model achieved an average accuracy exceeding 97%.
• This hybrid model outperformed traditional CNN and Transfer Learning-CNN (TL-CNN) approaches.
• The method demonstrated effective voiceprint recognition with limited training data.

Conclusions:

• The deep migration hybrid model (TLCNN-RBM) offers an effective solution for small-sample voiceprint recognition.
• Transfer learning combined with RBM and data augmentation significantly improves model performance.
• The approach accelerates training time and enhances convergence speed for practical applications.