A Front-End Technique for Automatic Noisy Speech Recognition

The sounds in a real environment not often take place in isolation because sounds are building complex and usually happen concurrently. Auditory masking relates to the perceptual interaction between sound components. This paper proposes modeling the effect of simultaneous masking into the Mel frequency cepstral coefficient (MFCC) and effectively improve the performance of the resulting system. Moreover, the Gammatone frequency integration is presented to warp the energy spectrum which can provide gradually decaying the weights and compensate for the loss of spectral correlation. Experiments are carried out on the Aurora-2 database, and frame-level cross entropy-based deep neural network (DNN-HMM) training is used to build an acoustic model. While given models trained on multi-condition speech data, the accuracy of our proposed feature extraction method achieves up to 98.14% in case of 10dB, 94.40% in 5dB, 81.67% in 0dB and 51.5% in -5dB, respectively.