Psychoacoustic model compensation with robust feature set for speaker verification in additive noise
Ashish Kumar Panda · 2014
This paper addresses the problem of speaker verification in the presence of additive noise for resource deficient languages. Psychoacoustic model compensation (Psy-Comp) has been shown to impart noise robustness to Gaussian Mixture Model (GMM) based speaker verification systems using Mel Frequency Cepstral Coefficients (MFCCs). This work extends the idea of Psy-Comp to incorporate a more robust feature set, which includes Cepstral Mean Subtraction (CMS) and Δ coefficients along with the MFCCs. We propose a model domain CMS operation following the psychoacoustic compensation for improved performance in additive noise. An advantage of this approach is that it does not require specialized developmental data and hence it may be suitable for resource deficient languages. Experiments conducted with the NIST-2000 database corrupted with real-life street noise show improved performance with the proposed method.