Spoken Document Classification of Broadcast News

Abstract

Two systems for spoken document classification are implemented by combining an automatic speech recognizer with the two classification algorithms naive Bayes and logistic regression. The focus is on how to handle the inherent uncertainty in the output of the speech recognizer. Feature extraction is performed by computing expected word counts from speech recognition lattices, and subsequently removing words that are found to carry little or noisy information about the topic label, as determined by the information gain metric. The systems are evaluated by performing cross-validation on broadcast news stories, and the classification accuracy is measured with different configurations and on recognition output with different word error rates. The results show that a relatively high classification accuracy can be obtained with word error rates around 50%, and that the benefit of extracting features from lattices instead of 1-best transcripts increases with increasing word error rates.