Abstract
This paper explores the application of deep learning (DL) techniques, specifically Convolutional Neural Networks (CNN), for speaker recognition using Mel-Frequency Cepstral Coefficients (MFCC) as the feature extraction method. The speaker recognition, which relies on unique vocal characteristics, faces challenges such as speech variability, background noise, and microphone differences. These factors complicate the creation of stable speech templates, making accurate recognition systems difficult to implement. To address these challenges, the study employs MFCC for feature extraction along with a novel CNN1D model for recognition tasks. This suggested method is evaluated using various datasets, including 50-speaker audio data, the DARPA TIMIT dataset with 454 speakers across 8 dialect regions, and other sample sets. Experimental results demonstrate the effectiveness of the MFCC + CNN approach, achieving an accuracy of 93% for 2462 speech samples and up to 98.01% for 750 samples. The study highlights the significant impact of dataset size on model performance, with larger datasets yielding higher accuracy. Despite the promising results, challenges related to speech variability and environmental factors remain. Future work will explore the use of CNN2D with spectrogram images as input, potentially enhancing performance further. This research contributes to the ongoing development of robust speaker recognition systems built on DL(deep learning) techniques.

This publication has 11 references indexed in Scilit: