Deep Learning Techniques for Automated Speaker Recognition System

This project introduces an advanced automated speaker recognition system designed for precise multi-speaker identification using cutting-edge deep learning techniques. The system is built on a curated subset of the CMU ARCTIC dataset, incorporating high-quality audio recordings from 18 diverse speakers (13 male and 5 female) encompassing a range of accents: US English (9 speakers), Indian English (6), Canadian English (2), and Scottish English (1). This diversity ensures robustness across linguistic variations.

Several models were developed and rigorously evaluated: a Random Forest classifier serving as the baseline, a hybrid CNN-GRU architecture for capturing spatial and temporal patterns, and a lightweight MobileNet combined with LSTM for efficient deployment. Training utilized a comprehensive set of acoustic features, including Mel-Frequency Cepstral Coefficients (MFCC), delta and delta-delta coefficients, chroma features, and spectral contrast, to extract discriminative speaker-specific traits.

To facilitate practical usability, a secure web application was implemented with user registration/login mechanisms and real-time speaker prediction functionality upon audio file upload. The system demonstrates high accuracy and reliability, positioning it as a promising solution for biometric authentication, voice-based security systems, and forensic speaker verification applications.

Keywords

Speaker Recognition, Deep Learning, CMU ARCTIC Dataset, MFCC, Delta Coefficients, CNN-GRU, MobileNet-LSTM, Random Forest, Acoustic Features, Chroma, Spectral Contrast, Web Application, Biometric Authentication, Forensic Analysis