基于邻接空间的鲁棒语音识别方法

Robust Speech Recognition Based on Neighborhood Space

提出了一种基于邻接空间模型的鲁棒语音识别方法,解决测试集和训练集差别导致的识别正确率过低的问题.在以声学模型为中心的邻接空间中计算贝叶斯预测概率密度值,作为观察概率输出分值进行识别.实验表明,相对于传统语音识别方法,鲁棒识别方法在保证干净测试集的识别率没有很大下降的前提下,对含噪测试集的识别率获得了较大的提高.

This paper presents an approach to robust speech recognition based on neighborhood space, which can achieve performance robustness under mismatch between training and testing conditions. This approach uses neighborhood space of each underlying model to produce Bayesian predictive density as observation probability density. Experimental results show that the proposed method improves the performance robustness.