RESPITE: Annual Report 1999: Background

(RESPITE logo)

Background and Objectives

Recent improvements in Automatic Speech Recognition (ASR) have resulted in products for dictation and speech interaction. However, use of these systems is restricted to predictable, quiet situations. For ASR to become useful in noisy, real-world environments, it is necessary to develop new techniques to achieve the kind of robustness shown by human listeners in the face of signal corruption. The RESPITE partners have introduced two new and complementary approaches to this crucial problem: missing data and multi-stream. RESPITE opens up an opportunity to develop and combine these approaches. Scientifically, the approaches are principled both in terms of their theoretical exposition and in their links to speech perception. Commercially, they point the way to making ASR viable across a range of applications in which current techniques fail. The objectives of RESPITE are: