Speaking-style dependent lexicalized filler model for key-phrase detection and verification

A task-independent ller modeling for robust keyphrase detection and veri cation is proposed. Instead of assuming task-speci c lexical knowledge, our model is designed to characterize phrases depending on the speaking-style, thus can be trained with large corpora of di erent but similar tasks. We present two implementations of the portable and general model. The dialogue-style dependent model trained with the ATIS corpus is used as a ller and shown to be e ective in detection-based speech understanding on di erent dialogue applications. The lecture-style dependent ller model trained with transcriptions of various oral presentations also improves the veri cation of key-phrases uttered during lectures.