Unsupervised versus Supervised Training of Acoustic Models

摘要

In this paper we reports unsupervised training experimentswe have conducted on large amounts of the English Fisherconversational telephone speech. A great amount of workhas been reported on unsupervised training, but the majordifference of this work is that we compared behaviors ofunsupervised training with supervised training on exactlythe same data. This comparison reveals surprising results.First, as the amount of training data increases, unsupervisedtraining, even bootstrapped with a very limited amount (1hour) of manual data, improves recognition performancefaster than supervised training does, and it converges tosupervised training. Second, bootstrapping unsupervisedtraining with more manual data is not of significance if alarge amount of un-transcribed data is available.