【摘要】 语音识别系统的确信度判决用于对未登录词(Out-of-Vocabulary,OOV)的拒识.本文提出了一种有效的确信度判决的方法.在本方法中包含了两种模型:填充模型和噪声模型.填充模型能对无关语音进行拒识;噪声模型则用于强化对噪声的拒识,联合使用两种模型起到了较好的拒识效果.这两种模型中使用的声学模型单元均利用基本识别器已有的模型单元,无须额外的训练数据和单独训练.本文设计了一遍识别过程,识别和确信度判决在一次过程中完成,使系统的运算量增加极小.实验结果表明该方法能有效地处理OOV问题.更多还原

【Abstract】 An effective method of confidence measure in speech recognition is proposed, which is used to reject Out-of-Vocabulary words (OOV). In this method two models are included: filler model and noise model. Filler model can reject irrelative input speech and noise model emphasizes on the rejection of noise. The rejection performance is beetter when both are used jointly. The acoustic model units of both models are all from those of basic system and don’t need additional training data or procedure. A one-pass recognizing procedure is also designed, which performs recognition and confidence measure in one pass. So the increase of calculation is very little. Experiment results show that the method is effective for OOV rejection.更多还原