钢琴音乐转写中音乐相关特性研究

Study of Music-related Characteristics in Piano Music Transcription

自动音乐转写,即利用程序将音乐音频转写成音符的形式,是音乐信息检索的一项重要课题.目前基于深度学习的转写算法多源自语音识别、自然语言处理等领域,缺乏对音乐自身特性的深入研究.本文采用Transformer架构进行钢琴音乐转写,利用合成的音乐数据进行预训练,并设计了符合音符事件结构的截止点识别网络,旨在尝试对转写中的音乐特性进行针对性设计.本文的方法一方面利用了音乐合成技术成熟、便捷的方案,易于数据扩充;另一方面从目标检测的角度进行剖析,对音符本身的事件特性加以探讨.实验表明,本文采用合成音乐进行预训练有助于音符识别,截止点识别网络可以提高音符级的转写结果,实现了音符起点F1=97.40%,带截止点音符F1=88.81%的转写结果.

Automatic music transcription,i.e.,the use of programs to transcribe music audio into note form,is an important topic in music information retrieval.Most of the current deep learning-based transcription algorithms come from the fields of speech recognition,natural language processing,etc.,and lack in-depth research on the characteristics of music itself.In this paper,we use Transformer architecture for piano music transcription,pre-train using synthesized music data,and design a cut-off point recognition network conforming to the structure of note events,aiming to try to design a targeted approach to the music characteristics in transcription.The approach in this paper takes advantage of the mature and convenient scheme of music synthesis technology for easy data expansion on the one hand;on the other hand,it explores the event characteristics of the notes themselves by profiling them from the perspective of target detection.Experiments show that this paper uses synthetic music for pre-training is beneficial to note recognition,and the cut-off point recognition network can improve the transcription results at the note level,achieving the transcription results of note starting point F1=97.40%and note with cut-off point F1=88.81%.