论文部分内容阅读
口语对话系统中,集外词的存在会引起很多识别错误,为了有效地发现并拒绝集外词,提高系统性能,研究利用置信度打分进行语音确认的方法,发现并拒绝识别错误。提出上下文相关的置信度特征,充分考虑当前待确认词与其前序词和后序词之间的相关性。实验结果表明:上下文相关的置信度特征能够很好地提高拒识性能,对符合识别文法的句子,错误拒绝率为2.5%或5%时,对比没有使用上下文相关的置信度特征时,错误接受率分别下降了29%和36%;基于置信度打分的语音确认策略在拒识性能上优于系统已有的在线垃圾模型。
In colloquial dialogue system, the existence of extra words will cause a lot of recognition errors. In order to effectively find and refuse the extra words and improve the system performance, this paper studies the method of confirming the voice by the confidence score, and discovers and rejects the recognition errors. This paper proposes the context-dependent confidence features and fully considers the correlation between the current to-be-confirmed words and their predecessors and consequent words. The experimental results show that the context-dependent confidence features can improve the rejection performance well. When the false rejection rate is 2.5% or 5%, the error acceptance Rate dropped by 29% and 36% respectively. The voice recognition strategy based on confidence scoring outperformed the existing online spam model in rejection performance.