音乐库中检索特征音调歌曲智能识别仿真

The Simulation of the Intelligent Recognition of the Retrieval of Characteristic Tone Songs in the Music Database

对于音乐库中特征音调歌曲的智能识别,可使用户通过哼唱音乐片段的方式,在较大规模乐曲库中准确、快速搜索到目标歌曲。检索特征音调歌曲需要对歌曲的起始点音符智能识别,提取旋律基音短时谱结构特征来完成歌曲检索。传统方法利用复数帧组成音符起始点特征参数向量,通过计算分布函数来识别音符,但无法获得时谱的结构特征,过程复杂且准确性差。提出基于削波的检索特征音调的哼唱音符起始点智能识别方法。上述方法先对音符进行预处理,滤除高频部分噪声,利用历史采样值与激励信号的线性组合表述音符短时语音信号当前的取样值,采用中心削波法给出特定哼唱音符起始点范围阈值,并利用阈值消除哼唱音符信号的低幅值部分,提取基音短时谱结构特征和包络特征,完成检索特征音调歌曲。仿真结果表明,所提方法识别精确度高,可以满足哼唱检索实时性要求。

An intelligent recognition method for retrieving feature tone humming note initial point is proposed based on the clipping retrieval. Firstly, the note is preprocessed and the noise of high frequency part is filtered. Then, the history sampling value and linear combination of drive signal are used to express the current sampling value of short time voice signal, and the specific range threshold value of humming note initial point is given out by using the center slicing method. Moreover, the threshold value is used to eliminate the low amplitude value part of humming note signal, and the structure feature of keynote short - time spectra and envelop feature is extracted. Finally, the re- trieval of feature tone song is completed. The simulation results show that the method has high recognition precision. It can satisfy the instantaneity requirement of humming retrieval.