This paper proposes robust methods for processing EMG (electromyography) signals in the framework of audio-EMG-based speech recognition. The EMG signals are captured when uttered and used as auxiliary information for recognizing speech. Two robust methods (Cepstral Mean Normalization and Spectral Subtraction) for EMG signal processing are investigated to improve the recognition performance. We also investigate the importance of stream weighting in audio-EMG-based multi-modal speech recognition. Experiments are carried out at various noise conditions and the results show the effectiveness of the proposed methods. A significant improvement in word accuracy over the audio-only recognition scheme is achieved by combining the methods.