Transfer learning for acoustic modeling of noise robust speech recognition