Goats to Sheep: Can Recognition Rate Be Improved for Poor Tangora Speakers?

GOATS TO SHEEP: CAN RECOGNITION RATE BE IMPROVED FOR POOR TANGORA SPEAKERS? Catalina M. Dan.is User Interface Institute IBM Thomas J. Watson Research Center Yorktown Heights, NY 10598 ABSTRACT This paper reports on a study of recognition performance for a group of new users during their first month of experience with the TANGORA system. TANGORA is a 20,000 word, speaker dependent, isolated-word system which transcribes speech input into text in real-time. Twelve users, six males and six females, participated in 21 sessions each, during which they read aloud unrelated sentences selected from a corpus of office correspondence. Their goal was to develop a speaking style which minimized TANGORA's recognition error. Hypotheses were generated about users' speech habits which may have lead to increased recognition error and suggestions were made to them on how to modify their speaking style accordingly. On average, recognition errors decreased by 33% from the first to the fourth week. Some characteristics of successful speakers have been identified. INTRODUCTION There is a great deal of variability in the accuracy with which users of large vocabulary automatic speech recognition (hereafter ASR) systems are recognized. In a typical finding, Brown, Vosburgh & Canetti (in preparation) reported that recognition error for a group of first time users of a 20,000 word ASR system varied from 2.0% to 14%. Two conclusions may be drawn from such results. First, the technology is good enough to produce a high degree of recognition accuracy for some speakers. Second, there are some speakers who encounter severe problems and, for them, the technology is probably not usable. This research was motivated by the latter group. It is concerned with whether recognition accuracy can be improved through behavioral means for speakers who are initially poorly recogniT.ed by an ASR system. Can a user modify his or her speaking style in ways which will be acceptable to the user and will result in a significant improvement in recognition, thereby making ASR systems more widely usable? IBM's experimental TANGORA system, implemented on the Personal Computer AT, was used in this investigation. This system functions in real-time and has the capacity to recognize 20,000 words. TANGORA is an isolated word system; this requires that users pause briefly between words. Further, it is a speaker-dependent system and must be "trained" to the user's voice. Such a system is most accurate when it has a description or model of the acoustic characteristics of a user's voice. This speaker model is generated by TANGORA from a sample (1200-2400 words) of the user's speech, collected during a "training session." A description of the TANGORA system can be found in Averbuch et al., (1986). This investigation had four general goals. The first was to investigate recognition performance for a group of new users during their first month of experience with the TANGORA system. The focus was on determining the rate and amount of improvement, if any, in recognition accuracy. It is an important, but unanswered question, whether poor ASR speakers can improve substantially with experience. …

Content truncated to 500 words