Single-microphone blind channel identification in speech using spectrum classification

We propose an algorithm for blind estimation of the magnitude response of a channel using the observations of a single microphone. The algorithm employs channel robust RASTA filtered Mel-frequency cepstral coefficients as features and a Gaussian mixture model based classifier to generate a dictionary of average speech spectra. These are then used to infer the channel response from speech that has undergone spectral modification in the capturing process. Simulation results using babble noise, car noise and white Gaussian noise are presented, which demonstrate that the proposed method is able to estimate a variety of channel responses to within 3-4 dB in terms of weighted spectral distance; and it is more accurate than a previously published method.