An enquiring system of unknown words in TV news by spontaneous repetition (application of speaker normalization by speaker subspace projection)

摘要

We have constructed a system in which we can enquire about unknown words included in TV news speech by repeating them spontaneously. For example, we might hear "Japan would join the PKO" from the TV news, and if "PKO" is an unknown word then we can enquire about it by saying "What is the PKO?" The system recognizes the word "PKO" and explains its meaning. In this system, it estimates a common section between the announcer's speech and the user's speech, and recognizes the word corresponding to the common section. We solved the problem of speaker difference in extracting the common sections by speaker subspace projection.