Abstract: To improve robustness of automatic speech recognition system, a post-processing module called Confidence Measure（CM） is usually employed in state-of-the-art system to serve as error detection, Out-Of-Vocabulary（OOV） rejection, etc. In a command word speech recognition system, a traditional filler-based CM is adopted, the performance of which is poor especially for OOV detection. This paper presents an improved syllable loop based confidence measure method which simplifies the decoding network to meet the real application requirements. Experimental results on a Mandarin command word recognition show that the proposed method achieves significant improvement in both performance and efficiency when comparing to the baseline.

Key words: speech recognition, confidence measure, syllable loop

摘要： 为获得较为鲁棒的识别性能，一般的语音识别系统中都会在后端加入一个置信度判决模块，以实现识别错误检测和集外词拒识等功能。针对命令词语音识别系统，传统的基于Filler模型的置信度方法由于自身模型结构的限制，性能相对有限，尤其是对集外词的检测效果不好。为此，使用了一种基于音节循环的置信度判决方法，并对该方法的解码网络进行精简，以满足实用化的效率要求。在中文命令词测试集上的实验结果表明，该方法相对于基于Filler模型的置信度方法对识别效果与识别效率都有了较大的提升。

关键词: 语音识别, 置信度判决, 音节循环