Abstract: In language recognition,due to the insufficiency of information in each frame,traditional MFCC feature extraction is easily suffered from noise pollution．Meanwhile,the general method of SDC feature extraction depends on artificially setting in parameter selection which increases the uncertainty of recognition performance．In order to overcome these drawbacks,the deep learning method was introduced and a novel feature extraction approach named BN-DBN which is based on deep learning was proposed．Finally, the relevant comparative experiments for the bottleneck layer size,the number of hidden layers and the position of the bottleneck layer were carried out in NIST2007database．Experimental results show that extraction method of the bottleneck features based on deep belief networks are more effective in language recognition,compared with traditional methods.

Key words: Language recognition,Bottleneck features,Deep belief networks