|RESPITE: Annual Report 1999: Approach|
Our ASR framework involves two principal stages, which are depicted below together with solutions based on recent innovations developed largely by the partners.
This approach is motivated by perceptual findings which suggest that listeners make sense of speech in everyday conditions by selecting parts of the spectro-temporal evidence which belong to the attended source, then exploiting signal redundancy to base recognition solely on these selections. The RESPITE consortium brings enabling techniques to both identifying reliable evidence and dealing with missing data. In contrast to other approaches, our methodology makes no assumptions about the number of background acoustic sources and requires no models for them. In addition, the multi-stream formalism allows for evidence combination across whatever sources of information are reliable. RESPITE will develop and compare these techniques, with application to ASR in cars.