Improving Performance of Telephone- Based Mandarin Speech Recognition

Since telephone is the only ubiquitous communications device in current world, it is the largest potential application field for speech techniques. Telephony speech recognition is a core technique for such telephone-based speech applications. It is well known that the bandwidth of telephone line is limited to 300~3400Hz and there are many inherent variations within the telephone network. All these make speech recognition over telephone a more difficult task compared to its desktop pairs. Additionally, due to the freely speaking style required by real applications and the diverse background environment, a perfect laboratory system may become very vulnerable in real world. So the robustness is the life-and-death issue for such commercial systems. In this paper, we will introduce our recent progresses on improving the performance for a Mandarin telephony speech recognition system. Our improvements include a more robust and straightforward feature extraction block for telephony speech and a novel dynamic channel compensation algorithm. And then we will focus our discussion on the strategy of dealing with outof-vocabulary (OOV) utterances. Through all these amendments, the system’s performance obviously improves in real applications.