Speech polarity determination: A comparative evaluation

The performance of various speech processing applications may be dramatically affected by an inversion of the speech polarity, which depends upon the recording setup. As a consequence, automatically detecting the speech polarity is a necessary preliminary step to guarantee a correct behaviour of such methods. The goal of this paper is two-fold. First a new approach for polarity determination based on the calculation of higher-order statistical moments is introduced. These moments oscillate at the local fundamental frequency with a phase shift which is dependent on the speech polarity. Secondly, a thorough comparative evaluation between the proposed method and three other state-of-the-art techniques is carried out. Experiments are led on a large amount of data with 10 speech corpora. In addition to an analysis in clean conditions, the robustness of these methods to both an additive noise and to reverberation is also investigated.