论文部分内容阅读
在针对电话语音的语种识别系统中,训练语音和测试语音之间存在不同说话人的个性差异带来的干扰,是影响系统识别性能的一个重要因素.基于此,本文首先对当前语种识别系统中消除此影响的方法进行研究,对比分析它们各自的优缺点,选择将锚模型方法引入语种识别系统中,该方法将语料映射至说话人无关的锚超矩阵进而消除说话人相关信息.针对锚超矩阵的选择存在语种混淆和信息冗余等问题,本文并提出一种结合支持向量机的锚模型训练算法,该方法下得到的锚超矩阵更具语种区分性,并去除了混淆信息的影响,增强了矩阵的紧致性.实验结果表明,新方法下的锚模型映射方法能有效提高基线系统的识别性能,并降低了语种识别系统训练和识别时的计算量.
In the language recognition system for telephone voice, the interference caused by the differences in personality of different speakers between the training speech and test speech is an important factor that affects the performance of system recognition.Therefore, in this paper, This paper studies the methods to eliminate this influence, compares and analyzes their respective advantages and disadvantages, and chooses to introduce the anchor model into the language recognition system, which maps the corpus to the speaker-independent anchor hyper-matrix and eliminates the speaker-related information. There are problems such as linguistic confusion and information redundancy in matrix selection. In this paper, an anchor model training algorithm based on SVM is proposed. The anchor super-matrix obtained by this method is more language-specific and removes the influence of confusion information. Which enhances the compactness of the matrix.The experimental results show that the anchor model mapping method under the new method can effectively improve the recognition performance of the baseline system and reduce the computational load when the language recognition system is trained and identified.