An improved DNN-based spectral feature mapping that removes noise and reverberation for robust automatic speech recognition.