Design, Compilation and Processing of CUCall: A Set of Cantonese Spoken Language Corpora Collected Over Telephone Networks

Design, Compilation and Processing of CUCall: A Set of Cantonese Spoken Language Corpora Collected Over Telephone Networks W.K. LO, P.C. CHING, Tan LEE and Helen MENG The Chinese University of Hong Kong wklo@ee.cuhk.edu.hk, pcching@ee.cuhk.edu.hk, tanlee@ee.cuhk.edu.hk, hmmeng@se.cuhk.edu.hk Abstract The design and compilation of the CUCall telephone speech corpora is described in this paper. Speech database is an indispensable resource for research and development of state-of-the-art spoken language technology. These speech recognition systems rely greatly on a huge amount of well-designed and appropriately processed speech data for parameters training. On the other hand, as telephony applications are becoming more demanding and complicated, natural language interface is gaining more popularity than the traditional touch tone operation. Therefore, large telephone speech databases are required for such system building. Separate speech corpora are needed for telephone systems since there exist significant differences due to the channel difference. In this paper, we will describe the design and processing of a set of spoken language corpora for Cantonese that are collected over fixed line as well as mobile telephone networks. The corpora are intended as a versatile set of training data for general purpose application systems that adopt a statistical approach to spoken language processing. The designed set of corpora will be made up of over 1000 speaker calls. 1 1 Introduction Speech data collected over telephone network is an essential resource for telephone based spoken language systems. The increasing penetration of remote system or service access over telephone networks has created a great driving force for collecting a huge amount of telephone speech data from a large speaker population and for different languages. Since the current state-of-the-art speech recognition techniques are statistically based, the availability of annotated data is particularly important. In general, the greater the amount and coverage of the data, the better the speech applications developed. In order to build a spoken language system over telephone network, the speech data has to be collected over telephone network and properly transcribed. The goal of this work1 is to collect and compile a set of general purpose Cantonese telephone speech data from a large group of people of both genders. With the availability of this set of corpora, the rapid growth in the spoken language applications over telephone networks for the Cantonese speaking community is made possible. Over the past decades, many telephone based spoken language systems have been developed with great success. They all take advantage of the existence of several spoken language corpora compiled in recent years. Examples include the Jupiter from MIT [25], HMIHY2 from AT&T [17] and the European Union projects such as ACCeSS [27] and ARISE [28] etc. Nowadays, there are quite a large number of companies that make use of simple automatic telephone service systems to reduce the cost of employing human operators. Many of them have upgraded or wish to upgrade their touch-tone based system to speech enabled versions. It is obvious that continuous efforts are needed to enhance these services via speech technologies as much as possible. For …

Content truncated to 500 words