A Mandarin Voice Organizer Based on a Template-Matching Speech Recognizer

Language, Information and Computation(PACLIC 11),1996, 289-296 A Mandarin Voice Organizer Based On a Template-Matching Speech Recognizer Jhing-Fa Wang, Jyh-Shing Shyuu, and Chung-Hsien Wu Institute of Information Engineering National Cheng Kung University wangjf@server2.iie.ncku.edu.tw Abstract On the observation of current available voice organizers, all of them accept only voice commands or word-based commands. Using natural spoken language to operate organizer is still a difficult problem. In this paper, a template-based speech recognizer which accepts near(constrained) spoken language is proposed. Since the template-based recognizer is a domain-dependent speech recognition system, representing and matching of sentence templates become the main tasks of the recognizer. We use finate state networks(FSNs) to represent the sentence templates and propose a vowel-based, syllable-scoring method to match a correct template. By replacing the template sets, this method can be easily applied to other domains. Besides, two main functions, voice recording and voice message query, are implemented on our organizer using a fast CELP encoder/decoder to compress/decompress the voice data in realtime. Experimental results shows that the collected 31 sentence templates can greatly improve the voice interface between the user and the voice organizer. 1. Introduction A voice organizer uses the voice input and output to process personal data. For example, for recording voice data or querying a voice message from the voice database, a voice organizer must first recognize the input voice and then respond the user with voice output. Hence, a voice organizer includes at least the following voice processing modules. First, a voice recognizer is used to decode the input speech into operating commands. Secondly, the voice output module responds the user with the information that he wants. Thirdly, a voice storage module is used to save the large amount of voice data. On the observation of current available voice organizers, all of them accept only a few voice commands or word-based commands(Huang 94). Using natural spoken language to operate voice organizer is still a difficult problem. In this paper, we proposed a template-matching speech recognizer that allows the user to operate the voice organizer using near natural spoken language. The templates that we define here are those sentence patterns which are frequently used to operate the voice organizer. Among so many sentence patterns, they are classified into several so called "templates". The voice recognizer then match the input voice based on these templates. Hence, the input styles are more flexiable than those of voice-command based voice organizer. 289 In the following Sections, Section 2 describes the system block diagram of our proposed voice organizer. Section 3 go through details of the templated-matching speech recognizer. Section 4 describes the voice response module and voice compression module. The experimental results are given in Section 5. Finally, a conclusion remark is given in Section 6. 2. The System Block Diagram The overall system block diagram is shown in Fig. 1. First, the input voice is recognized into syllable-lattice by the acoustic signal recognizer. The syllable-lattice is then decoded into operating commands by the template-matching speech recognizer based on pre-created …

Content truncated to 500 words