论文部分内容阅读
网络电话语音的识别具有广阔的应用前景 ,而相对较低的话音质量突出了关键词捕捉的重要意义。关键词捕捉的核心问题是关键词可信度的估计。在零虚警假设下提出了关键词的后验可信度 ,在测度中结合了声学层分数和基于拼音的统计语言模型分数 ,利用动态规划推导了计算测度的前向后向算法。在 2 4 0个关键词的捕捉实验中 ,后验可信度下的关键词识别率高于 88%。基于拼音格 HMM(隐 Markov模型 ) ,对可信度估计、关键词捕捉、最优部分路径搜索及拼音多候选重排进行了统一的解释。
Internet phone voice recognition has broad application prospects, while the relatively low voice quality highlights the importance of keyword capture. The key issue of keyword capture is the estimation of the credibility of keywords. Under the null false alarm hypothesis, the posterior credibility of the key words is proposed. In the measurement, the acoustics layer score and the statistical language model score based on pinyin are combined, and the forward-backward algorithm for calculating the measure is derived by using dynamic programming. In 240 captures experiments, the keyword recognition rate of posterior credible was higher than 88%. Based on pinyin HMM (Hidden Markov Model), a unified interpretation of credibility estimation, keyword capture, optimal partial path search and pinyin multiple candidate rearrangement is made.