On combining acoustic and modulation spectrograms in an attention LSTM-based system for speech intelligibility level classification