基于支持向量机的非特定人孤立数字语音识别

为了识别一组非特定人、不连续的数字语音信号，本文提出了一种基于支持向量机理论的语音信号识别算法．具体过程主要包括训练过程和识别过程．其中训练过程为：先使用预先建立起来的语音库对选定的支持向量机进行训练，得到一组与该语音信号相关的支持向量；在识别过程中，首先获取被测语音信号，并根据MFFC理论提取特征向量，然后使用训练后的支持向量机进行识别．此外，还提出使用短时区域能量谱的方法对语音信号进行端点检测．结果表明，与目前流行的隐马尔可夫算法比较，本文算法具有识别速度快、准确率高等优点．

Recognition of Discontinuous Numeric Speech by Nonspecifical Person Using Support Vector Machines

To distinguish a group of nonspecial discontinuous numeric speech signals （nDNSS） ,a algorithm using Support Vector Machines （SVM）was proposed. The whole process included two main steps of training and recognition. SVM was selected in the training by using a preestablished nDNSS database. During recognition process,an eigenvector of a nDNSS was drawn out primarily using MFFC theory and then taken into the selected SVM to be recognized as input parameters. Furthermore,an algorithm for end-point detection of speech signal based on short-term-region-energy spectrum （STRES）was presented for the first time. Experiments show that the proposed algorithm can achieve better performance than the existing popular method （HMMD） on rapidity,accuracy and robustness.