钢琴音乐自动记谱可应用于钢琴辅助教学、音乐辅助创作、智能乐谱翻译等诸多领域，但由于复音音符间存在复杂的时频重叠，其记谱准确率距离实际需求还有相当大的差距。本项目以钢琴复音音乐为研究对象，借助信号处理及机器学习中的理论、模型与技术，结合音乐理论知识，研究层次化的自动记谱方法，实现音频到五线谱的完整转录。具体内容：研究基于高阶统计量的时频特征表示，解决复音音乐信号的频域谐波重叠问题；研究机器学习与人工规则相结合的音符识别方法，解决小样本下系统泛化能力差的问题；研究音符间音乐学约束的建模，使音符级别的记谱结果更加符合音乐学规范；针对缺乏可读性的计算机记谱形式，借助音乐结构分析实现五线谱形式的转录及评价。相关研究可推广至其它类型的复音音乐，并为智能音乐交互技术的实现提供有力支持。

Automatic transcription of piano music can be widely used in many applications, such as computer-aided teaching of piano, music creation and musical score translation. Since the polyphonic notes overlap in the time and frequency domain, the performance of automatic piano music transcription cannot satisfy the requirements of practical applications. Based on the theories, models and technologies in signal processing and machine learning, combined with music theory knowledge, this project will study the hierarchical transcription of piano polyphony music, which will convert the audio to musical score completely. The major research content includes four parts. Firstly, the time-frequency representation based on higher-order statistics will be extracted to solve the problem of harmonic overlaps in spectrum. Secondly, the note recognition method combining machine learning and hand-craft rules will be studied to overcome the small data challenge in deep learning. Thirdly, a music language model will be built to describe the structural regularity in polyphonic note sequence, which can improve the performance of note-level transcription. Finally, the piano-roll representation will be transcribed due to lack of readability. The musical score will be obtained and evaluated through music structure analysis. The proposed methods can be applied in automatic transcription of other polyphonic music. Furthermore, the research can provide support for the realization of intelligent music interaction.