Neural Network Front-ends Based Speech Recognition in Reverberant Environments
Zhen Zhang, Peng Li · 2016
This paper presents an investigation of reverberant speech recognition using frond-ends based methods.A 2-channel dereverberation method is adopted to achieve robust dereverberation under different reverberant conditions.Also a 2-channel spectral enhancement method is used where the gain of each frequency bin is controlled by acoustic scene, which is detected based on the analysis of full-band coherent property.Deep Neural Network (DNN) is also presented as a feature extractor.The DNN based front-end allows a very flexible integration of meta-information.Bottle neck features is extracted in place of MFCC features used in HMM-GMM system.We evaluated our methods on the data provided by REVERB challenge.On simulated data, the DNN front-end yields more than 33% relative reduction in Word Error Rate (WER).