基于动态垃圾评价的语音确认方法

语音关键词识别和确认方法在语音对话系统中得到了广泛应用。评价此类系统性能的一个重要指标就是处理非关键词（垃圾）的能力。处理垃圾的传统方法是状态下进行垃圾建模。但该方法并不能很好的描述大量的系统词库以外的词，且训练规模较大。该文提出了动态垃圾评价方法，不对垃圾本身建模，而是在识别过程中对输入语音进行可信度评估，从而对识别结果进行确认，解决了传统垃圾模型灵活性差及在线垃圾建模方法的确认能力不足等问题。同时由于在垃圾评价中增加了反关键词信息，减少了关键词之间的识别错误。

Utterance Verification Based on Dynamic Garbage Evaluation Approach

Spotting and utterance verification is widely used in special area dialog system with small vocabulary. The capability to deal with the words out of vocabulary is an important factor to evaluate a dialog system. Keyword verification is an effective means to reduce false alarm rate in which computing likelihood ratio is the essential approach. The key problem is how to choose the appropriate alternative model when calculating likelihood ratio. The traditional explicitly-trained garbage models, such as using HMM, can not describe the words out of vocabulary very well and it is difficult to train. On-line garbage modeling does not consider the recognition error between keywords. This paper presents a dynamic garbage evaluation approach based on speaker verification algorithm. This method does not attempt to explicitly define a garbage model, instead, it evaluates the confidence of the utterance in speech recognition with strong flexibility. The dynamic garbage evaluation score represents the common information of the keywords in vocabulary. Based on the score, the system can more easily discriminate nonkeywords between keywords. Furthermore, the anitikeyword model is used to enhance the ability of distinguish keywords. The new approach integrates the dynamic garbage evaluation approach with statistic hypotheses test frame of utterance verification. A two-pass strategy is adopted, consisting of recognition followed by verification. The dynamic garbage evaluation is more flexible and with lower computation than the explicit garbage modeling and reduced recognition error rate between keywords. Recognition and verification on isolated and continuous words are tested. The dynamic garbage evaluation performes well. Based on dynamic garbage evaluation, the keyword detection rate is 96%, when the false alarm rate is 6.2%.