论文部分内容阅读
Tandem技术是目前主流语音识别系统中提升性能的重要手段之一,它基于训练数据和其所对应的正确标注用有监督的方式训练神经网络的参数。该文提出一种基于解码中竞争信息对传统Tandem起到补充作用的方法,能够增强Tandem技术的区分性。为了获取竞争信息,首先在词图中生成识别解码中的竞争片段,再分别使用基于时长重叠信息和后验概率信息的挑选策略来选取对神经网络训练最为有效的竞争信息,以提高竞争网络对识别性能的补充作用。实验结果表明:加入竞争信息的改进Tandem系统获得了超过传统系统的性能。
Tandem technology is one of the most important means to improve performance in the mainstream voice recognition system. It trains the neural network parameters in a supervised manner based on the training data and its corresponding correct annotation. This paper proposes a method based on the competitive information in decoding that plays a complementary role in traditional Tandem, which can enhance the distinction of Tandem technology. In order to obtain the competitive information, the competitive segments in the recognition decoding are first generated in the word map, and then the selection strategies based on the overlapping information of duration and the posterior probability information are respectively used to select the most effective competition information for neural network training to improve the competition network pair Identify the performance of the complementary role. The experimental results show that the improved Tandem system with competitive information gains more performance than the traditional system.