Retrieving speech correctly despite the recognition errors
Mikko Kurimo and Ville Turunen
In: 2nd Joint Workshop on Multimodal Interaction and Related Machine Learning Algorithms(2005).
This work studies ways to recover from speech recognition errors in
retrieving spoken documents.
The methods are evaluated by Finnish news reading data using
an unlimited vocabulary recognizer with
language models for unsupervised morpheme-like units.
Recognition errors can naturally be reduced by improving the recognizer,
but the focus here is on the attempts to improve
the search index more directly, namely by adding new index
terms to replace those lost due to recognition errors.