【作者】 张超； 刘轶； 郑方；

【Author】 ZHANG Chao~(1,2),LIU Yi~1,ZHENG Thomas Fang~1 (1.Center for Speech and Language Technologies, Division of Technology Innovation and Development, Tsinghua National Laboratory for Information Science and Technology,Beijing 100084,China; 2.Department of Computer Science and Technology, Tsinghua University,Beijing 100084,China)

【机构】 清华信息科学技术国家实验室技术创新与开发部语音和语言技术中心； 清华大学计算机科学与技术系；

【摘要】 该文提出了应用声学似然分作为置信度来生成可靠口音相关单元的方法。基于可靠口音相关单元构造声学模型,并通过声学模型重构的方法将它们融合到标准普通话模型中,以改善普通话语音识别器对带多方言口音语音的识别效果。另外,还提出了使用增量式决策树融合及根据支配度选择Gauss混合2种方法来减少冗余的Gauss混合,从而提高了重构后的声学模型的效率。实验表明:该方法在不降低对标准普通话的识别率的前提下,对粤、吴口音的绝对音节错误率分别下降了9.25%和9.21%。更多还原

【Abstract】 The acoustic likelihood score is used as a confidence measure to generate reliable accent-specific units and to merge such reliable accent-specific units through acoustic model reconstruction.The decision tree merge and acoustic model reconstruction efficiencies are improved by reducing redundant Gaussian components through an incremental decision tree merge procedure and selection of Gaussian components according to their dominance.Tests on Cantonese and Wu accents show that this approach yields significant 9.25%and 9.21% absolute syllable error rate(SER) reductions without degrading the performance on standard Putonghua.更多还原