Research On Automatic Generation Method Of Piano Fingering Based On Q-Learning

Due to a wide range of application scenarios and flexible multi-part music performance,piano has become the first choice for the public to learn musical instruments.With regard to massive piano learning needs,the current piano education teachers are very limited and unbalanced.Piano online education will benefit piano learners more widely and more targetedly.Among them,piano fingering teaching is the most important and difficult part.Using computer technology to study the automatic generation method of piano fingering will promote the online teaching of piano fingering.In recent years,the automatic generation method of piano fingering has also made some achievements,but there are methods for feature extraction that do not consider piano key positions and supervised learning are subject to the quality of fingering data.Aiming at these problems,this thesis proposes an automatic generation method of piano fingering based on QLearning.The main research contents and innovations are as follows:(1)In terms of feature extraction,this thesis proposes a piano key feature extraction method based on track separation.This method is based on the structure principle of MIDI music files,analyzes the note feature matrix,screens and separates the piano track blocks,and extracts the key sequence of the notes,respectively.The key position feature of the track separation takes into account the specific key position information of the finger in the actual process of piano playing,and the situation that different hands correspond to different parts and tracks,which is conducive to the establishment of the fingering generation model.(2)In terms of piano fingering evaluation,this thesis proposed a rule-based piano fingering evaluation method.Summarizing the fingering habits and general rules of modern piano performance,taking into account the physiological characteristics of the fingers and the key position information of the front and back notes,under two consecutive notes under different circumstances,the evaluation is performed with the grade and score as fingering,forming a quantitative evaluation system for piano fingering.The rule-based piano fingering evaluation method considers the influence of the key positions of the front and back notes and the conversion of the fingers,can objectively describe the quality of the piano fingering,and plays a role in the feedback and improvement of the fingering generation model.(3)In terms of model establishment,this thesis proposes an automatic generation method of piano fingering based on Temporal-Difference reinforcement learning model.This method uses Q-Learning,a Temporal-Difference reinforcement learning algorithm,as a learning model for special scenes generated by piano fingering.It is concise and clear,and the learning speed is fast.It also takes into account the influence of the fingers before and after actual playing.Reinforcement learning does not require fingering data,but automatically generates fingerings based on piano key features and rule evaluation.It scientifically regulates and overcomes the limitations of fingering data quality and data size,and is applicable to a wider range of scenarios.In the experimental results,the score rate of various fingerings generated by machines reaches about 0.9,and the difference with professional fingerings does not exceed 3%,indicating that the method proposed in this paper achieves the automatic generation of piano fingerings.