摘要： 针对短训练语音的说话人识别系统，提出一种基于决策层融合的识别算法。识别时运用经验模式分解法对语音信号进行处理，对获取的固有模态函数分量提取语音特征序列，分别进行匹配，通过决策层融合算法，将所得的匹配结果与传统独立识别结果相结合，最终输出识别结果。利用信号分解的方法，实现待测语音信号的重复识别，同时采用决策层融合算法优化识别结果，从而在短训练语音情况下，使系统的识别率得到保障。实验结果表明，该算法在短训练语音识别系统中的识别效果优于传统方法。

关键词: 短语音, 说话人识别, 美尔频率倒谱系数, 经验模式分解, 决策层融合

Abstract: An algorithm of decision-fusion for the speaker recognition systems is presented that uses little speech data for training models. In the phase of recognition, it decomposes the speech signal by using empirical mode decomposition processing method, extracting the speech features and repeating speech recognition based on some gained intrinsic mode function components. Meanwhile, by using this algorithm, it can fuse the results of repeat recognition and original system, and get the final output result. With the method of signal processing, repeating recognition can be implemented the original results and the accuracy rate of the recognition system based on little speech data is guaranteed. It proves the algorithm is advanced in recognition systems based on little speech data than the traditional ones by simulation experiments.

Key words: little speech data, speaker recognition, Mel Frequency Cepstrum Coefficient(MFCC), Empirical Mode Decomposition(EMD), decision level fusion

中图分类号: 

• TN912.3