Speaker Identification in Noisy Environments for Forensic Purposes

Abstract

The speech is a biological or physical feature unique to each person, and this is widely used in speaker identification tasks like access control, transaction authentication, home automation applications, among others. The aim of this research is to propose a connected-words speaker recognition scheme based on a closed-set speaker-independent voice corpus in noisy environments that can be applied in contexts such as forensic purposes. Using a KDD analysis, MFCCs were used as filtering technique to extract speech features from 158 speakers, to later carry out the speaker identification process. Paper presents a performance comparison of ANN, KNN and logistic regression models, which obtained a F1 score of 98%, 98.32% and 97.75%, respectively. The results show that schemes such as KNN and ANN can achieve a similar performance in full voice files when applying the proposed KDD framework, generating robust models applied in forensic environments.