面向视唱评价的声乐自动标注系统

歌唱声音的自动标注是基于内容的音乐分析和检索领域的基础问题。该文在统计分析汉语歌唱声音声韵母时长和音高特征的基础上,提出了一种声乐自动标注模型。该模型将信号处理、语音学和音乐知识结合,通过韵母-乐谱音高对齐和基于声母时长分布的边界优化算法实现了高精度的音符切分。音高提取算法在移调检测的基础上,通过限定基音周期搜索范围的策略克服了信号频谱中半频/倍频点的干扰。结合音乐教学实践,通过检测音符内部最稳定音高成分、估计演唱速率和归一化音符时长的方法提取歌唱声音的音高、节奏和时长信息,并基于这3个客观物理量提出了一个视唱评价方法。实验表明,该文提出的算法能够准确地标注歌唱声音中的音符边界(平均误差26ms)和音高(平均误差0.23半音)。

Singing annotation approach for solfege evaluation

This paper proposed a automatic annotation method for solfege-oriented singing evaluation.Speech and singing were compared to demonstrate the characteristics of consonant and pitch in singing voice.A Dynamic Time Warping based audio-score alignment approach combined with consonant boundary refinement was used to perform note segmentation in a high time resolution.A time domain pitch estimation algorithm is proposed,which is able to eliminate octave errors by employing standard pitch derived from a reference score to restrict the search space.As for solfege scoring,criteria were proposed based on statistics of objective acoustic measurement,which were also used to investigate the performance of the proposed automatic annotation model.Experimental results show that the proposed model achieves high precision for both tasks of note segmentation and pitch estimation.