A voice-activated telephone exchange system and its field trial

Abstract

Speaker-independent speech recognition systems that can accept telephone quality speech may open opportunities for introducing new user-friendly services over the public switched telephone network (PSTN). The authors are currently engaged in a project to introduce an automatic speech recognizer over PSTN. They have developed a voice-activated telephone exchange system by combining a continuous speech recognizer and a private branch exchange system (PBX), and conducted field trials. The system has been installed in the R&D laboratories for daily use since June 1993, in order to investigate its performance in a real environment and collect man-machine dialogues. More than 5,000 man-machine dialogues have been collected, and incorrect recognitions have been analyzed and categorized into three categories such as (1) incorrect detection of speech, (2) out-of-vocabulary responses, (3) incorrect recognition with inadequate hidden Markov models of speech and noise. The authors have improved system performance by mainly attacking the issues (1) and (3). They have just developed a new version of the system, using the improved scheme obtained by analyzing the collected speech data. In order to collect more man-machine dialogues, they are planning to carry out the second phase field trial in which the new system will be installed in branch offices.<>