Research And Implementation Of A CNN-based Piano Music Transcription Algorithm

With the improvement of living standards,people’s pursuit of spiritual entertainment has been continuously improve.The enthusiasm for learning to play piano has been rising,along with an upsurge of piano education.Piano music transcription technology can detect the pitch and onset of each note in the piano music acoustic signal,and can be applied to evaluate piano performance automatically and objectively,to help piano learners discover performance errors in time,and improve the learning efficiency.However,it is still a challenging problem to implement an accurate piano music transcription algorithm.To address the above issues,a piano music transcription algorithm based on convolutional neural network(CNN)has been studied and implemented.The time-frequency analysis is first conducted on the input piano music acoustic signal.Then note onsets are detected.Finally,multi-pitch estimation is performed to detect pitches of the newly played notes at each onset.The main work of this thesis can be summarized as follows:(1)An extensive research on the research progress and related work is conducted,and automatic music transcription techniques are reviewed;(2)The time-frequency analysis is implemented using the short-time Fourier transform and the constant-Q transform respectively.And based on the CNN model,the note onset detection and multi-pitch estimation are implemented.The architecture and the training method of the CNN and post-processing methods are optimized;(3)Comparative experiments are implemented to evaluate the performance of CNN networks,when using different time-frequency representation as input,and when using different training methods.At last,the detection errors are analyzed.The proposed CNN-based piano music transcription algorithm achieves 86.80%on note-wise F-measure when it is evaluated using the first 30s of the piano pieces in the MAPS ENSTDkCl(allowable deviation range of the onset is ±100ms).The evaluation results show that the algorithm proposed in this paper can achieve the highest F-measure among existing CNN-based piano music transcription algorithms,which can provide efficient technical support for computer-assisted piano tutoring.