Abstract
Speech is the simplest modality to be considered for Unimodal Biometric System. The accuracy however does depend on the quality of the signal which tends to be compromised due to the conditions under which it is spoken or recorded. To implement a robust system the challenge is in enhancing the quality and intelligibility of the noisy speech signal. Various speech enhancement techniques can be applied to the speech signal on understanding the type of noise present in the signal. This paper compares the performance of the Speaker Recognition System in terms of the percentage accuracy of recognition achieved for a raw noisy signal and the enhanced signal. Spectral Subtraction method has been used for Speech enhancement while the Text Independent Automatic Speaker Identification has been implemented using Vector Quantization (VQ) and Gaussian Mixture Model (GMM). The VidTimit data base has been used for the same. An accuracy of 96.11% is achieved using enhanced speech information and GMM as compared to 74% using the raw signal. GMM has also proven to give better accuracy over the VQ method.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call