Mask estimation in non-stationary noise environments for missing feature based robust speech recognition