Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

A Study of Applying Adaptive Learning to a Multi-Module System

Tung-Hui Chiang, Yi-Chung Lin, Keh-Yih Su

Department of Electrical Engineering, National Tsing Hua University, Hsinchu, Taiwan, R.O.C.

Adaptive learning procedures have been widely used in speech recognition tasks for minimizing system error rate. However, they were usually applied to a single module. For many applications, systems are composed of multiple modules. A variety of learning strategies, including isolated, incremental and joint learning, therefore, are proposed in this paper to train the parameters of a multi-module system. We have employed these learning procedures to Chinese phonetic typewriter systems for recognizing 1,000 sentences in the speaker-independent, isolated character input mode, with a very large vocabulary of 90,495 words. According to the simulation, the isolated learning strategy is basically unable to achieve the best system performance. It even deteriorates the performance in some cases. In general, the joint learning procedure is recommended for its potential superiority. However, the joint learning strategy cannot show its benefit if the later modules are modeled in terms of a large number of parameters. In this circumstance, the incremental learning procedure is suggested.

Full Paper

Bibliographic reference.  Chiang, Tung-Hui / Lin, Yi-Chung / Su, Keh-Yih (1994): "A study of applying adaptive learning to a multi-module system", In ICSLP-1994, 463-466.