Support vector machines for automatic data cleanup

Accurate training data plays a very important role in training effective acoustic models for speech recognition. In conversational speech, in several cases, the transcribed data has a significant word error rate which leads to bad acoustic models. In this paper we explore a method to automatically identify such mislabelled data in the context of a hybrid Support Vector Machine/hidden Markov model (HMM) system, thereby building accurate acoustic models. The effectiveness of this method is proven on both synthetic and real speech data. A hybrid system for OGI alphadigits using this methodology gives a significant improvement in performance over a comparable baseline HMM system.