Comparison of Statistical Model-Based Voice Activity Detectors for Mobile Robot

This paper deals with the problem of voice activity detection in adverse acoustic conditions, namely high and varying noise scenarios. For robotic applications, we need the voice activity detector to be computationally light, robust to varying levels of background noise, and have a low latency, especially if we are tracking moving speakers. We analyze three voice activity detectors—two model the discrete Fourier transform coefficients by Gaussian and generalized Gaussian distribution, while the third models the spectral envelope as having either Rayleigh or Rice distribution—and we present them in a unifying and consistent manner, with respect to a statistical hypotheses ratio measure and a joint noise spectrum estimation algorithm. Moreover, we compare the performance under various noise conditions; three types of noises, three different signal-to-noise ratios and six different speakers, by means of receiver operating characteristic curves and area under a curve score. The results showed that the Rayleigh-Rice model had on average better results and medium computational demand.