Research On Automatic Chord Arrangement Of Piano Based On Deep Learning

Chord is a combination of multiple notes.It plays an important role in enriching the melody expression.Arranging chords for melody requires a lot of knowledge of music,including the basic music theory,chord rules and other professional knowledge.With the development of deep learning and neural network technology,artificial intelligence is widely used in music retrieval,music creation and music teaching.This thesis aims to replace artificial arrangement with machine arrangement by means of deep learning,which provides a powerful auxiliary tool for piano music creation.The research in the thesis is aimed at the piano music in WAV format.The process of chord arrangement is divided into three main subtasks: note detection,multiple pitch estimation and model training.Music Signal is divided into several segments after note detection and the main note and chord components of each segment are extracted by multiple pitch estimation,which will be input into neural network as the feature and the label to train the model,so that the model will obtain the ability to arrange chords.The main research work and innovations of this thesis are:(1)The thesis studies on the algorithms of note onset detection,most of which do not take into account the prior knowledge of musical instrument pitch.Considering that the piano is a twelve-tone instrument,the thesis designs a group of filter bank based on the twelve-tone equal temperament,which aims to obtain the distribution of the frequence of each frame in the music signal and thus,detect the notes.This method takes the characteristics of the tones into account and combines with the specific pitch information to improve the effect of note detection.(2)The thesis studies on the algorithms of multiple pitch estimation and finds that the most common idea is to decompose the spectrum of the music into the spectrum of each note,but the spectrum contains a lot of redundant frequency information.The thesis designs a group of filter bank based on harmonic structure of piano keys.The filter banks aim to extracte the features of fundamental frequency and harmonics from music signal,and next we can estimate multiple pitches by decomposing the features of the music into the features of each note,which means the features of music are expressed as the weighted sum of the features of each note.(3)The thesis studies on the deep learning algorithm.Since the chord arrangement relies on the distribution of nearby notes,the thesis selects the bidirectional recurrent neural network based on LSTM and combines the encoder-decoder model with attention model.After training the model,we can get the chord sequence by inputting mono sequence into the model.Now we can generate music with chord sequence and the information of note onset,so as to make comparison with the mono music.(4)At last we build a network scoring platform and upload original mono music and arranged music to the server,so that music lovers can listen to the music online and rank the chord arrangement.The result shows that most of the songs scored high,indicating that the auditors are quite satisfied with the chord arrangement,which verifies the effectiveness of the chord arrangement system in the thesis.