Run-time information fusion in large vocabulary continuous speech recognition.

摘要

Continuous speech recognition systems are environmentally sensitive and suffer from the great variability of speech. In order to achieve recognition robustness, there's a strong interest among researchers on how to fuse different information sources for speech recognition. A common problem of those approaches is that complementary information is lost either before or after recognition.; To avoid this unrecoverable information loss, and to better utilize this complementary information, we proposed a run time information fusion scheme. The hypothesis of this thesis is that by performing fusion at different levels and stages of a Large Vocabulary Continuous Speech Recognition (LVCSR) system, especially inside the decoder, more reliable and efficient fusion is possible.; The hypothesis is first tested in a speech segmentation task, which is essential to the performance of an LVCSR system. Furthermore, three different approaches of run time fusion are proposed and implemented inside an LVCSR decoder. The experiments demonstrate the effectiveness and potential of these approaches.