论文部分内容阅读
鲁棒性问题是决定语音识别技术能否在实际中得以应用和推广的关键问题之一。概括起来说，导致语音识别系统性能变坏的原因大体上来自三个方面，即噪声（加性噪声、卷积噪声）、信道变化和不同的讲话者（不同的声道形状、不同的发育方式等）。本文对三种高鲁律性自适应语音识别方法进行了研究和改进，并对它们的性能进行了比较，这三种方法分别是VQ码本自适应法、HMM参数自适应法和基于正则相关分析的谱变换补偿方法。实验结果表明，这三种方法都能提高非特定人语音识别系统对信道以及说话人的鲁棒性，而且基于正则相关分析的稻变换补偿方法具有最好的性能，它能够补偿由三种失真源同时引起的训练条件与测试条件之间的不匹配，因此适合作为一种通用的自适应方法。
Robustness is one of the key issues that determine whether speech recognition technology can be applied and promoted in practice. To summarize, the reasons leading to the deterioration of speech recognition systems generally come from three aspects: noise (additive noise, convolution noise), channel variation and different speakers (different channel shapes, different developmental modes Wait). In this paper, three kinds of GA-adaptive speech recognition methods are studied and improved, and their performances are compared. The three methods are VQ codebook adaptive method, HMM parameter adaptive method and based on regular correlation Analysis of spectral transform compensation method. The experimental results show that these three methods can improve the robustness of the speaker-independent speech recognition system, and the rice transform compensation method based on canonical correlation analysis has the best performance, which can compensate for the three kinds of distortion The sources also cause a mismatch between the training conditions and the test conditions and are therefore suitable as a general adaptive method.