CNN Trained Speaker Recognition System in Electric Vehicles

Abstract

Speaker recognition is the technique of determining a person's identity based on their voice features. Speaker recognition modules are now included in several commercial products because of the speaker recognition revolution. One such application is in electric vehicles, where a speaker recognition system is used for voice authentication in unlocking the vehicle. The performance was affected due to the background noise in the existing model which was improved using the proposed Least Mean Square (LMS) filter and Kalman filter. For reducing background noise, the LMS filter performed much better, while the Kalman filter performed better for Additive White Gaussian Noise (AWGN). In this work, Features of a speech are extracted using Mel Frequency Cepstral Coefficient (MFCC) which is trained on Convolutional Neural Network (CNN) classifier algorithm employing 16000 PCM speech samples dataset. Recognizing speakers from different recording conditions creates numerous challenges for the system. The recognition accuracy increased to 92.8%. Superior results were obtained using the presented MFCC-CNN model with filtering approaches. Hence the experimental results conveys that the implemented model for external noises in speaker recognition system is better.