一种基于互补声学模型的多系统融合语音关键词检测方法

采用一种基于互补声学模型的多系统融合方法来获得高性能的语音关键词检测系统: 1)在基线系统的基础上, 使用不同的音素集进行声学建模, 并引入基于神经网络的声学建模方法, 获得另外两套具有建模差异性的声学系统; 2)在多套关键词检测系统的基础上, 通过选择有效的系统融合准则, 将多个系统的输出进行整合, 获得更好的语音关键词检测结果. 该方法充分利用了差异性声学建模系统之间的互补性, 在不增加训练数据的情况下, 显著地提升了最终系统的性能. 和基线系统相比, 该方法在2005年国家863电话语音关键词检测技术评测集上, 在等错误率(Equal error rate, EER)指标下, 获得相对21.6%的显著性能提升.

A System Combination Based Keyword-spotting Method Using Complementary Acoustic Models

In this work we explored a system combination based keyword spotting (KWS) method by using complementary acoustic models. The main steps included: 1) constructed two complementary acoustic models by using different modeling units (phone set) and different modeling methods (neural networks), respectively, as well as the baseline system; 2) combined the outputs of the above three systems with appropriate method and obtained a better result. The proposed approach exploited the complementary features of different systems to improve the system performance without using any additional training data. With this method, a significant relative reduction of 21.6% in EER (equal error eate) was obtained over the baseline system for the Mandarin CTS (conversational telephone speech) KWS task in 2005 China National ``863' Evaluation.