Research And Implementation Of A Neural Network-based Automatic Drum Transcription Algorithm

Music can cultivate sentiment,improve life taste and personal temperament.With the improvement of living standards,people’s enthusiasm for learning drums has been rising,along with a boom of drum education.Automatic Drum Transcription(ADT)technology aims to generate drum notes from music recordings,which can be used for automatic score generation or automatic performance recognition to help drum kit beginners improve their learning efficiency.A robust ADT system plays an important role in drum automation education and music intelligent entertainment software.However,it is still a challenging problem to implement a high-performance ADT algorithm.Aiming at the problem of automatic drum transcription,this paper researches and implements a convolutional neural network-based automatic drum transcription algorithm.Moreover,different algorithm optimization strategies are designed.The main contributions of this paper are as follows:(1)A convolutional neural network(CNN)based transcription algorithm is designed and implemented.The time-frequency information of the audio signal is extracted using short-time Fourier transform(STFT)with Mel-filter and then processed by CNN.Finally,the peak picking algorithm is applied to detect the onset events of each drum instrument respectively.(2)Specific to the problem of the insufficient generalization ability of the neural network,a training method using label augmentation is proposed.Using labels that combine music information and drum category information with the self-distillation method can effectively improve the recognition ability of the neural network.(3)Specific to the problem of the limited coding ability of the convolutional neural network,an involutional neural network suitable for drum transcription is designed.The transcription ability is compared with other neural network models experimentally.(4)Specific to open vocabulary transcription,a fewshot learning-based model using a relation network is implemented,and the transcription capabilities of the model under fixed vocabulary and open vocabulary are compared and analyzed.(5)Based on the proposed drum transcription algorithm,the application of drum score recognition is designed,which verifies the effectiveness of the transcription algorithm.The evaluation result shows that the proposed algorithm has the highest average FMeasure in the general dataset compared with the existing neural network-based drum transcription algorithms.The proposed algorithm provides core algorithm support for the realization of intelligent drum education software.