Constant Q Cepstral Coefficients and Long Short-Term Memory Model-Based Automatic Speaker Verification System

Automatic speaker verification is an advancement in authentication property of a security system like biometric at organizations, IVR system at call centers, an unlocking system of smartphones, etc. Significant components of these sensitive systems are speech feature extraction unit as frontend and a classification model for backend. This paper provides the development details of an ASV system that is trained by ASVspoof 2019 dataset. Knowledge of speech signal processing is essential to design the frontend of any speech-based system. Process of constant Q cepstral coefficient (CQCC) feature extraction is discussed with the spectrographic views of outcomes of signal processing operations. A long short-term memory (LSTM) model with the time distributed wrapper layers is used for the backend of the proposed system. Proposed system is achieving 0.42% EER for logical access (LA) set and 0.51% for physical access (PA) set of the ASVspoof 2019 dataset.