Speaker Recognition for Surveillance Application

Abstract

This paper describes the speaker recognition problem regarding to the complex surveillance system. The proposed system extension enables identified the precise identity or at least the gender of the suspect by the captured voice analysis. Our solution is based on the text-independent approach by using Mel-Frequency Cepstral coefficients and fundamental frequency for extracting the identity from a voice signal. Gaussian Mixture Models up to 1024 mixtures were used to classify more than 20 speakers. In this paper the comparison and evaluation of speech based parametrizations and noise elimination techniques are presented regarding to the noisy acoustic data. This system extension could help to eliminate the vandalism and to increase the elucidation of crimes.

Keywords

• Speaker recognition
• Feature extraction
• Noisy environment
• Surveillance system