A Study on Robust Speech Recognition in Real Environments and Its Applications

Speech is the most natural way of communication for human being, thus, techniques to communicate with machines such as speech oriented information input and retrieval are strongly required to an advanced information-oriented society. Speech recognition is one of the techniques to realize this kind of communication, and has achieved nearly 95 % recognition accuracy at present in restricted conditions. However, in spite of such great technical progress of speech recognition, the convenient and innovative systems equipped with speech recognizer have not yet been developed and distributed. This circumstance of speech recognition tech-niques is exactly caused by serious problems that present speech recognition techniques are not robust for differences of speakers, utterance styles, adverse noises, and so on. Hence, it requires extremely restricted conditions for accurate performance. For realistic system devel-opment, a speech recognizer should suppose that a user speaks to a microphone located in a distant place in noisy environment, spontaneously. However, in this situation, speech recog-nition accuracy considerably deteriorates, and the reliability of the system equipped with the speech recognizer also deteriorates. Consequently, it is impossible to develop and distribut