基于先验知识的三音子模型聚类结构自适应策略-电子与信息学报.pdfVIP

第 29 卷第 9 期 电 子 与 信 息 学 报 Vol.29No.9 2007 年 9 月 Journal of Electronics Information Technology Sept...2007 基于先验知识的三音子模型聚类结构自适应策略 董 明 刘润生 (清华大学电子工程系 北京 100084) 摘 要: 该文提出了一种基于先验知识的三音子模型聚类结构自适应策略，可以在规模很小的自适应语音库条件下 改善三音子声学模型的聚类结构使之更适合应用对象的协同发音特点。以基本声学模型训练过程中的三音子模型聚 类结果作为先验知识的聚类中心，依据基本声学模型对自适应语音库的分割，按照最大似然准则迭代地重估新的聚 类中心和模型聚类结构。实验表明：基于先验知识的三音子模型聚类结构自适应策略可以在不足两小时的自适应语 音库上实现三音子模型聚类结构重估，在针对汉语母语说话人的英语声学模型实验中，该文的模型聚类结构自适应 策略可以将系统识别率从 74.59%提高到 83.63%。 关键词：语音识别；三音子模型；模型聚类 中图分类号：TN912.34 文献标识码：A 文章编号：1009-5896(2007)09-2050-04 Transcendental Information Based Triphone Model Tying Structure Adaptation Strategy Dong Ming Liu Run-sheng (Department of Electronic Engineering, Tsinghua University, Beijing 100084, China) Abstract: A Transcendental Information Based (TIB) triphone model tying structure adaptation strategy is delivered, and this strategy can improve the triphone model tying structure to suit the target co-pronunciation features with small amount of adaptation data. The TIB triphone model tying structure adaptation strategy uses the baseline acoustic model’s triphone model tying result as the transcendental model clustering center, with the adaptation data alignment by the baseline acoustic model, re-estimate the TIB triphone model clustering center and model tying structure recursively under maximum likelihood principle. The experiments show that the TIB triphone model tying structure adaptation strategy can improve the triphone model tying structure with only 2 hours’ adaptation corpus, and in the experiment of English acoustic model for Chinese speakers, the TIB strategy will increase the recogni