Channel Normalization for Unsupervised Spectral Subtraction
Guillaume Lathoud · 2006
Abstract. Application domains such as in-car human-machine interaction require noise-robustfront-ends in order to cope with the noisy situations encountered in practice. Moreover, whenspeech is captured through a cellphone, the phone channel characteristics are often unknown. Itis thus desirable to estimate and remove both phone channel characteristics and ambient noise,in an online manner. The main contributions of this paper are twofold. First, a novel channelnormalization method is proposed, that is used before noise reduction, at the magnitude spectro-gram level. It removes the convolutive channel, and reduces the stationary part of the ambientnoise. Second, an alternative to classical spectral subtraction is proposed, called “UnsupervisedSpectral Subtraction” (USS), which does not require any parameter tuning. Channel normaliza-tion followed by USS (two steps) permit to reach an ASR performance very similar to that ofthe ETSI Advanced Front-End (Wiener filtering, with many steps and parameters). The compu-tational cost of the proposed approach is very low, which makes it fit for real-time applications.Furthermore, channel normalization followed by the ETSI Advanced Front-End leads to a majorimprovement in noisy conditions, and best overall results.