Abstract: In this paper, an experiment of percussion detection was designed and feature extraction methods of percussion signals were studied, and then the possibility of using speech recognition technology to recognize internal defects in metal materials was verified. The results show that defects appeared to cause the frequency spectrum of the percussion signals to move to low frequency band or to be split in more peaks.GMM-HMM (Gaussian mixture-hidden Markov model) built on MFCC (Mel-scale frequency cepstral coefficients) parameters of percussion signals could identify different types of specimens effectively but were liable to be affected by noise. The model improved with “binary information fusion+noise generalization” has a higher recognition rate in the extreme noise circumstance. Under the noise (10 dB SNR) circumstance, the recognition rate of the improved GMM-HMM reached the optimal value 99.3% with the sound information fusion rate at 0.6.