As the complex spectrum of the target signal is restored from signals of
mutiple microphones for each of frequency points, the mel-filter bank
outputs of the target signal (clean speech) are calculated by making
weighted sums of restored spectrum
of the target
signal according to the mel-scaling. They are Fourier transformed to
Mel-Frequency Cepstral Coefficients (MFCCs) which are widely used as the
feature vector for speech recognition.