Integrating tonal information into Mandarin name recognition with different strategies

Name recognition is a practical application of speech recognition technology. As Chinese is well known to be a tonal language, tonal information has an important influence on this task. In this paper we integrate tonal information into a speaker-independent Mandarin name recognizer, and two combination strategies: feature combination and posterior combination are investigated first. The recognizer is evaluated on an extremely challenging Mandarin name corpus, which includes 100 tonally confusing pairs. Although a significant improvement in the recognition accuracy can be achieved with either strategy, the system has a poor flexibility. Based on the analysis of the experimental results we propose a two-step process to improve the system performance further. It is shown that a maximal improvement of 29.96% in word accuracy can be achieved. At the same time the system has good flexibility with tonal information being integrated dynamically.