Topic segmentation of speech data based on keyword detection

提出了一种基于关键词发现技术的电视新闻语音主题识别方法。基于互信息和词的长度准则，选取了3000个名词作为主题词。通过根据所发现的单词的声学得分和单词的主题概率计算主题的可能性来识别新闻的主题。如果正确的主题出现在主题识别结果的前三位，则主题识别正确率为66.5%。为了在保持较高的主题识别率的前提下减少误报，提出了一种结合韵律信息和音素信息的关键词识别方法。韵律之间的差异性的关键字和输入语音的F0轮廓的DP匹配测量。基于这两个度量的总得分用于检测关键字。在基于F0模型进行平滑之后，为每个关键字存储F0模板。F0信息的引入使电视新闻语音在相同检测率下的虚警率降低了30%~ 50%，提出了一种基于词长度的词识别准确率估计方法，并与基于词长度的估计方法进行了比较。虚警计数估计的一个新的措施，通过模拟语音识别的音素序列，语言模型产生的计算。基于仿真的措施显示出更好的性能估计虚警。

A method of topic identification is proposed for TV news speech based on the keyword spotting technique. Three thousands of nouns are selected as keywords which contribute to topic identification, based on criterion of mutual information and a length of the word. The topic of news is identified by calculating possibilities of the topics in terms of an acoustic score of the spotted word and a topic probability of the word. Topic identification rate is 66.5 percent assuming that identification is correct if the correct topic is included in the first three places of the result of topic identification.A new method of keyword spotting using prosodic information as well as phonemic information is discussed in order to reduce false alarms keeping high detection rate. Prosodic dissimilarity between a keyword and input speech is measured by DP matching of F0 contours. A total score based on these two measures is used for detecting keywords. The F0 template are stored for each keyword after smoothing based on the F0 model. The introduction of F0 information reduced false alarms by 30% to 50% for the same detection rate for TV news speech.A method of estimating accuracy of word spotting is proposed and it is compared with the estimation based on the length of the word. False alarm counts are estimated by a new measure calculated by simulation of speech recognition for phoneme sequences that the language model generates. The simulation-based measure shows better performance for the estimation of false alarms.