论文部分内容阅读
在语音识别实际应用中,由于噪声的多样性,会造成训练和测试的失配,导致系统性能下降.特征补偿作为鲁棒性语音识别的一种重要方法,通过对训练和测试环境之间差异的研究,在特征空间中修正语音特征,使得修正后的测试语音特征能够更加接近训练语音特征.本文介绍一种实用的基于环境模型矢量泰勒级数(VTS)近似的特征补偿算法.首先验证传统的VTS离线算法在实际车载环境下的有效性;其次由于离线算法本身运算量很大,为了使其实用化,本文对算法进行改进,使其在提高效率的同时又能够保证与离线时相当的性能.通过实验结果验证,本文提出的实用化VTS算法在识别性能上相当接近离线时最好的性能.
In the practical application of speech recognition, due to the diversity of noise, the mismatch between training and testing will result in the degradation of system performance. As an important method of robust speech recognition, feature compensation, through the difference between training and testing environment , The modified speech features are corrected in the feature space so that the modified test speech features can get closer to the training speech features.This paper presents a practical feature compensation algorithm based on the VTS approximation of the environment model.First, VTS off-line algorithm in the actual vehicle environment; secondly, because the off-line algorithm itself is very computationally intensive, in order to make it practical, this paper improves the algorithm to make it more efficient and at the same time off-line Performance.According to the experimental results, the practical VTS algorithm proposed in this paper has the best performance in recognition performance when it is offline.