Microphone Distance Adaptation Using Cluster Adaptive Training for Robust Far Field Speech Recognition

Microphone distance adaptation is an important and challenging problem for far field speech recognition using a single distant microphone. This paper investigates the use of Cluster Adaptive Training (CAT) to learn a structured Deep Neural Network (DNN) that can be quickly adapted to cope with changes in the distance between the microphone and speaker at test time. A speech corpus was created by re-recording the Wall Street Journal (WSJ0) audio using far-field microphones with 8 different distances from the source. Experimental results show that unsupervised adaptation of the CAT-DNN model achieved up to 0.9% absolute word error rate reduction compared to the canonical model trained on multi-style data.