Low-Key Shallow Learning Voice Spoofing Detection System

Abstract

This paper creates a Gaussian shallow learning Mixture Model (GMM) voice-replay detector using the MATLAB low-key machine learning and statistics libraries. Our model extracts the Mel frequency cepstrum coefficients (MFCC) and constant Q cepstrum coefficients (CQCC) from the input voice signal in the front-end feature extraction stage. The collected characteristics are fed to the constructed GMM classifier to categorize the input voice as either authentic from a live source or replayed from a prerecorded source. The GMM is trained using large datasets of voice feature samples representing both classes. The classifier’s performance is measured using the Equal Error Rate (%EER) metric. To optimize performance, we subject the trained GMM to substantial development and assessment datasets in diverse scenarios and settings of reduction, normalization, and filtration. The best %EER results for the GMM classifier are 11.2237% for the development set and 22.5429% for the evaluation set.