Unsupervised Domain Adaptation for Speech Recognition via Uncertainty Driven Self-Training

The performance of automatic speech recognition (ASR) systems typically\ndegrades significantly when the training and test data domains are mismatched.\nIn this paper, we show that self-training (ST) combined with an\nuncertainty-based pseudo-label filtering approach can be effectively used for\ndomain adaptation. We propose DUST, a dropout-based uncertainty-driven\nself-training technique which uses agreement between multiple predictions of an\nASR system obtained for different dropout settings to measure the model's\nuncertainty about its prediction. DUST excludes pseudo-labeled data with high\nuncertainties from the training, which leads to substantially improved ASR\nresults compared to ST without filtering, and accelerates the training time due\nto a reduced training data set. Domain adaptation experiments using WSJ as a\nsource domain and TED-LIUM 3 as well as SWITCHBOARD as the target domains show\nthat up to 80% of the performance of a system trained on ground-truth data can\nbe recovered.\n