Two-microphone speech enhancement using a learned binary mask

Ideal binary mask speech enhancement is shown to increase the speech quality as well as speech intelligibility. But, this property depends highly on the accurate separation of speech and masker time-frequency units of the input spectrum, which is a difficult task in real situations. Ordinary binary mask methods are single-microphone methods and so, can obtain little information from the environment. In this paper, we devise a two-microphone method that uses a classifier to distinguish speech-dominated and masker-dominated time-frequency units. The classifier uses simply computable two-microphone features which enable it to be used in real-time scenarios. These proposed features empower the classifier to reach to classification accuracies near 80%. This high accuracy in turn, empowers the Ideal binary mask method to obtain higher SNRI and NPLR values in comparison to state-of-the-art noise reduction methods. These results indicate that the proposed two-microphone features have high information content for speech/masker separation.