关键词: 大词表连续语音识别, 识别错误, 分类

Abstract:

Large vocabulary continuous speech recognition system consists of several components, and recognition errors are caused by different factors. Developers need to know how the errors occur. In this paper we deduced the principle to classify the recognition errors from the recognition theory and put each error according to its cause into one of four classes: the decoding, the acoustic model, the language model, and the acoustic and language model. We then performed statistical analysis of classified errors. Experiments show that classification analysis of recognition errors provides guidance for improving the system.

Key words: large vocabulary continuous speech recognition, recognition, classification