Research On Piano Music Generation And Similarity Calculation Under The Framework Of Autoregressive Language Mode

In recent years,with the development of music artificial intelligence technology,language model-based technology has been widely concerned and has become the focus of this field.This paper focuses on two typical applications of the auto-regressive language model,the task of symbolic piano music generation and similarity calculation.First,the diversity sampling problem in the generation task of the auto-regressive language model is studied.The probabilistic trajectories of degenerate samples are studied,and the technique of enhancing diversity based on the perturbation of high-frequency words is proposed.An interquartile range inverse probability weighting sampling algorithm is proposed,which determines the set of high-frequency words by quartile and suppresses the probability of high-frequency words by inverse probability weighting.Experiments on English text generation show that the proposed algorithm can enhance the diversity of the generated samples without affecting the quality.Furthermore,a random perturbation sampling algorithm is proposed for the problem of controllable diversity generation.The range of random perturbation is explicitly controlled,and the magnitude of diversity enhancement is further controlled.The performance of the proposed algorithm is tested on the Chinese lyric generation problem,and the scheme of constructing vocabulary based on Chinese word segmentation and the pre-trained language model for Chinese lyrics are proposed.An end-to-end Chinese lyric generation system is built and deployed.Second,the generation of symbolic piano music with performance information is studied.Aiming at the problem of music vocabulary construction,a method of piano music vocabulary construction based on vocabulary aggregation is proposed.The difference between the piano music language model and the natural language model is analyzed,and a method of piano music generation with controllable performance information is proposed.By using vector parameters to control different types of performance information,the controllable generation of performance information for piano music is realized.Experimental results show that the proposed method can make up for the lack of diversity issue of piano music generation by traditional algorithms,and achieve the control of performance information.A case study reveals the influence of the proposed method on piano music generation.Third,the similarity calculation of symbolic piano music is studied.Using the music sequence representation and the embedding calibration of the language model as the technical approach,the music sequence representation is computed from the music symbol representation of the language model by using the embedding calibration method,and the automatic evaluation scheme is constructed by using the objective information of the dataset to find the optimal combination of the embedding calibration methods.The embedding calibration technique is analyzed by the structure index of the embedding space and the clustering experiment.Experimental results show that the method can greatly improve the similarity of symbolic piano music without manual annotation and model modification.The influence of the embedding calibration technique on the embedding space is revealed by the structure analysis,the clustering index,and the visualization results.