论文部分内容阅读
传统的语音端点检测方法在低信噪比环境下可靠性会急剧下降。本文提出了两种特征处理方法:特征的似然度加权和基于散度的维数缩减,来提高噪声下端点检测的性能。通过加权增加动态特征在似然度计算中的比重,可以提高端点检测的噪声Robustness。缩减散度值较小的特征维,对检测精度只有很小的影响,但可以提高检测效率。似然度加权对维数缩减之后的特征同样有效。在Aurora2数据库上的实验结果显示,在干净数据训练的检测模型下,似然度加权可以显著提高噪声下的端点检测性能。对维数缩减后的特征进行似然度加权,获得了与原始特征似然度加权相当的检测性能。这说明本文提出的方法是有效的。
The traditional voice endpoint detection method in a low signal to noise ratio environment will be a sharp decline in the reliability. In this paper, two feature processing methods are proposed: feature likelihood weighting and divergence-based dimension reduction to improve the performance of endpoint detection under noise. By weighted increasing the proportion of dynamic features in the likelihood calculation, Robustness of endpoint detection can be improved. Reducing the feature dimension with a smaller divergence value has only a small effect on the detection accuracy but can improve the detection efficiency. Likelihood weighting is also valid for features after dimension reduction. Experimental results on the Aurora2 database show that in the clean data training detection model, likelihood weighting can significantly improve endpoint detection performance under noise. By weighting the dimensionality-reduced features by likelihood, the detection performance weighted by the original feature likelihood is obtained. This shows that the method proposed in this paper is effective.