There has been a wealth of research effort focused on algorithm development, computational architectures, and interface design with a steady decrease in word error rates of about 10% per year (Deng 2004). Most speech recognition algorithms, however, are evaluated in controlled environments with the desired speaker in close proximity to a microphone. It is well known that the performance of most recognition algorithms is severely degraded with even modest amounts of background noise. At present, the most widely used speech recognition algorithms are based on the familiar hidden Markov models (HMM), in which words are constructed from a sequence of states (Young 1989, Young 1990, Furui 2002). Although hidden Markov models have been successful in performing speech recognition, there is still much work to be done. The goal of distributed listening research is to enhance speech recognition and natural language understanding. Distributed listening takes a different approach to addressing speech recognition and understanding using traditional HMM recognition systems. Researchers have found that humans perform some sort of distributed listening. In psychology, this is called Dichotic Listening (Bruder 2004). In dichotic listening, subjects listen to two voices, one in each ear, at the same time. Why don't existing speech recognition systems perform dichotic listening? Distributed listening research aims to enable systems to hear in a similar manner to humans.
展开▼
