Abstract
In this paper, we propose a robust voice activity detection (VAD) method that uses a density ratio model. For VAD under highly noisy environments, the likelihood ratio test (LRT) is effective. Conventional LRT constructs speech and noise models, calculates the likelihood of each model, and takes the ratio of those likelihoods to detect speech. Although some improved LRT have been proposed, in conventional LRT, it has not been taken into account that the likelihood ratio of speech and noise model is required, not the likelihood of each model. The proposed method directly estimates the likelihood ratio without calculating each likelihood using an density ratio model obtained in advance by density ratio estimation procedure. Moreover, there is the problem of determining thresholds, which are used for VAD and significantly affect its performance. We propose a method that automatically determines thresholds using discriminant analysis. The experiments show that the proposed method is more effective than conventional methods especially under non-stationary noisy environments.
