Higher order statistics based Gaussianity test applied to on-line speech processing

Detection of speech in noisy recordings becomes a challenging problem when the noise does not follow the usual whiteness, stationarity and high signal-to-noise ratio assumptions. A robust speech detector can affect significantly the performance of several speech processing tasks, such as endpoint detection, segmentation, and finally recognition, if we deal with real life data, as opposed to laboratory or controlled environment recordings. The detector proposed is based on a Gaussianity test that employs third-order cumulants of the data to decide on the binary hypotheses of noise only versus speech plus noise. Speech intervals are detected by exploiting the third-order information present in the speech signal. The detector can handle a large family of additive noises, thanks to its third-order statistics basis. The sample-adaptive and decision feedback variations proposed, provide the detector with a tracking ability both with respect to the time variations of speech and the possible nonstationarity of noise. Experiments carried out using real data, recorded in a moving car interior, show satisfactory performance of the proposed algorithms down to -6 dB signal-to-noise ratio