Influence of speech recognition errors on topic detection

Abstract

The effect of speech recognition errors on a system of unsupervised, synchronous clustering of broadcast news stories, is investigated using topic detection and tracking (TDT) algorithms. The TDT algorithm is used to impose an organization on a collection of documents, such that the underlying topical structure is exposed. It was found that the performance of the system was significantly degraded by automatic speech recognition errors, particularly for small clusters.