BridgeNets: Student-Teacher Transfer Learning Based on Recursive Neural Networks and its Application to Distant Speech Recognition

摘要

Despite the remarkable progress achieved on automatic speech recognition,recognizing far-field speeches mixed with various noise sources is still achallenging task. In this paper, we introduce novel student-teacher transferlearning, BridgeNet which can provide a solution to improve distant speechrecognition. There are two key features in BridgeNet. First, BridgeNet extendstraditional student-teacher frameworks by providing multiple hints from ateacher network. Hints are not limited to the soft labels from a teachernetwork. Teacher's intermediate feature representations can better guide astudent network to learn how to denoise or dereverberate noisy input. Second,the proposed recursive architecture in the BridgeNet can iteratively improvedenoising and recognition performance. The experimental results of BridgeNetshowed significant improvements in tackling the distant speech recognitionproblem, where it achieved up to 13.24% relative WER reductions on AMI corpuscompared to a baseline neural network without teacher's hints.