Speech presence detection in the time-frequency domain using minimum statistics
Karsten Sørensen, S.V. Andersen · Lund University Publications (Lund University) · 2004
The contribution of this paper is a time-frequency domain speech presence detection method that classifies power bins in the time-frequency domain as containing speech or not. An initial decision rule is based on ratios between optimally time-smoothed signal-plus-noise periodograms and weighted noise periodogram estimates, obtained from minimum statistics as proposed by Martin [1]. The initial decision rule is generalized into a weighted decomposition where the weights are obtained from off-line training by, means of an artificial neural network. Experiments show that the method can be configured to be very sensitive to speech presence even in very high levels of noise and without classifying much of the noise as speech. It is shown that a fixed set of weights gives good performance at different signal-to-noise ratios indicating that the terms in the decision rule have been adequately chosen.