基于混合高斯模型和倒谱变换的歌手识别算法研究

发布时间：2018-11-07 16:13

【摘要】：如今互联网上的多媒体资源不断增多，于此同时伴随着网络技术的飞速发展。对于众多的音视频资源，如何对这些数据进行分类处理就成为一个当前迫切需要解决的问题。特别是大量的音乐资源，手工分类已无法应对，人们需要对它们有高效快速的识别方法。音乐流派分类和音乐歌手识别这两方面在音乐分类算法领域的研究中开展的比较广泛。其中音乐歌手识别通常采用机器来识别和分类。本算法的设计思路是利用两组倒谱参数之间的变换来减弱背景伴奏的干扰实现歌手识别。本算法首先利用说话人识别算法，对语音数据进行预处理，提取特征向量。但和其他算法的不同点在于，我们提取两组特征向量，一组是歌手的清唱声音，一组是歌唱声音。然后通过研究清唱和歌声之间的MFCC倒谱系数的变化，并用高斯混合模型描述这种变化，再在得到待识别语音特征向量时应用该变换，实现对未知音频资源中背景音乐伴奏的减弱，得到歌手的语音信息，最后就是利用模板库匹配，，识别出不同的歌手。通过实验的设计和实现，我们验证了该方法应用于歌手识别领域的可行性，实验结果证明了该算法对歌手识别效果有一定的改进，同时还与其他论文的方法和结果进行了对比。
[Abstract]:Nowadays, the multimedia resources on the Internet are increasing, which is accompanied by the rapid development of network technology. For many audio and video resources, how to classify these data has become an urgent problem. In particular, a large number of music resources, manual classification has been unable to cope with, people need to have efficient and fast recognition methods. Music genre classification and music singer recognition are widely studied in the field of music classification algorithms. Among them, musical singer recognition usually uses machine to recognize and classify. The design idea of this algorithm is to use the transformation between two sets of cepstrum parameters to attenuate the interference of background accompaniment to realize singer recognition. Firstly, the speaker recognition algorithm is used to preprocess the speech data and extract the feature vectors. But different from other algorithms, we extract two groups of feature vectors, one is the singing voice, the other is the singing sound. Then by studying the change of MFCC cepstrum coefficient between singing and singing, we describe the change with Gao Si's mixed model, and then apply the transform when we get the speech feature vector to be recognized. The background music accompaniment of unknown audio resources is weakened, and the voice information of the singer is obtained. Finally, different singers are identified by using template library matching. Through the design and implementation of the experiment, we verify the feasibility of applying this method to the singer recognition field. The experimental results show that the algorithm has a certain improvement on the singer recognition effect. At the same time, it is compared with the methods and results of other papers.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN912.34

【相似文献】