一种基于分而治之的语音识别错误纠正方案

Speech recognition error correction scheme based on divide-and-conquer

介绍了一种基于分而治之的语音识别错误纠正方案,首先利用混淆网络把连续语音识别问题转换为顺序的、独立的分类子任务。每个分类子任务可看做是孤立词识别问题,通过训练专门的支持向量机来区分混淆网络的识别候选。提出了一种快速的基于码本转换的语音向量对齐方法,解决了变长语音向量无法直接作为支持向量机输入的问题。通过一个普通话音节识别任务的实验结果表明,该方案能有效提高系统的正确率。

This paper introduced a divide-and-conquer speech recognition error correction scheme. Firstly transformed continuous speech recognition problem into sequential,independent,classification tasks using confusion network（ CN） . Each of these sub-tasks could be taken as an isolated word recognition problem and specialized support vector machines （ SVMs） were trained and applied to each problem to discriminate the recognized candidates from CN. Proposed a fast codebook transformation based speech vector alignment method to address the problem that variable length speech vector could not be directly acted as the input vector for SVM. Experiment on a mandarin syllable recognition task shows the proposed scheme can improve the recognition accuracy effectively.