一种新的基于DBN的声学特征提取方法

A New Feature Extraction Method Based on Bottleneck Deep Belief Network

大词汇量连续语音识别系统中,为了进一步增强网络的鲁棒性、提升深度置信网络的识别准确率,提出一种基于区分性和ODLR自适应瓶颈深度置信网络的特征提取方法。该方法首先使用鲁棒性较强的瓶颈深度置信网络进行初步特征提取,进而进行区分性训练,使网络的区分性更强、识别准确率更高,在此基础上引入说话人自适应技术对网络进行调整,提高模型的鲁棒性。利用提出的声学特征在多个噪声较强、主题风格较为随意的多个公共连续语音数据库上进行了测试,识别结果取得了22.2%的提升。实验结果表明所提出的特征提取方法有效性。

In order to further improve the robustness and recognition rate of deep belief network in Large Vocabulary Continuous Speech Recognition system,this paper presented a novel bottleneck deep belief network to extract new features, which was based on speaker adaptation and discriminative training.Firstly, a bottleneck deep belief network was adopted to get the feature.And discriminative training performed on this basis gave a more distinguished network to improve the recognition accuracy.Simultaneously,a more robust speaker adaptation method was introduced to adjust the network.The proposed method was tested on several public continuous speech databases with strong noise and casual themes and a relative 6.9% promotion of the recognition accuracy was obtained.The result proves the superiority of the proposed method compared to the conventional one.