一种按节拍动态分帧的歌曲有歌唱部分检测新方法

Automatic Singing Voice Detection for Songs Based on a Dynamic Slicing Method with Beat

歌曲中的有歌唱部分是音乐的精华所在,也是音乐检索中用户检索的主要部分,传统的歌曲有歌唱部分检测采用的短时固定长度分帧方法没有考虑到音乐信号与一般音频信号的不同,针对于此,文章中提出一种结合音乐信号的节拍特征进行动态分帧的方法,该方法结合乐理知识,首先将音乐信号按照节拍进行分割,在此基础上提取特征,训练分类器,从而检测歌曲中的有歌唱部分,经实验表明,该方法能够对不同风格音乐信号有歌唱部分进行有效的检测,相对传统方法,能够提高检测的精确度.

Singing voice is the soul. of a song, and it is also the main target for music information retrieval. Traditional methods for singing voice detection use a short and fix-sized time slicing algorithm without concerning the differences between music signal and ordinary audio signal. According to this fact, this paper introduces a novel framework for singing voice detection using beat information of the music signal to segment the song dynamically. As the music theory, first, the song is segmented by the beat information. Then the CQT features are extracted from the beat segments and transformed to a feature matrix for training and classification. Experimental results show that this method has a better performance of singing voice detection, compared to the traditional method, and it works well for all kinds of music genres.