为了提高语音识别中方言识别由于差异性小混淆度高造成识别率低的问题,针对汉语普通话、青海方言和藏语安多方言设计一个基于子空间映射和分数归一化技术的GSV-SVM方言识别系统。利用最大后验概率准则自适应生成KL核的GSV,对GSV进行低维子空间映射;再利用映射后的空间训练多SVM模型进行得分域规整与识别。实验结果表明,采用该系统可以有效对混淆度高的方言进行识别。
To develop the low otherness and high confusion problem in dialect recognition, design a GSV-SVM dialect recognition system based on subspace mapping and score processing, which aims at mandarin, Qinghai dialect and Tibetan dialect. Firstly, adapting KL kernel based GSV under MAP rule, in order to map GSV into low dimension subspace. Then the subspace is used to train multiple SVM models to process scores and test. Experiments show that the proposed system could effectively recognize highly confused dialects.