Aiming at the problem that the common speech features such as the Mel frequency cepstrum coefficient（MFCC）feature and the perceptual linear prediction coefficient feature cannot effectively extract the frame correlation information before and after the speech

and the redundant information is too much

resulting in poor recognition effect

a sparse bottleneck feature and The composite feature construction method combining the MFCC features to improve the characterization and modeling ability of speech

and perform Gaussian mixture model hidden Markov model（GMM-HMM）acoustic modeling.The experimental results on the TIMIT database show that the acoustic model constructed based on acoustic composite features has excellent recognition performance

which proves the effectiveness of the method.