Research On Robust Evaluation Method Of Music

With the improvement of living standards, people are no longer satisfied with the experience of music listening and appreciation. There are more requirements about singing and feedback of music. Music evaluation is both entertaining and practical, it satisfies spiritual demands and is used in education. But existing music evaluation methods have the problem of inaccurate results and feedback, making against improving the level of singing. Therefore, the research on music evaluation is of great significance through improving the accuracy and objectivity of evaluating results.This thesis mainly studies singing scoring. In the traditional melody evaluation method, it focuses on the melody extraction and melody matching problems, improving some existing problems. In addition, based on the lacking of targeted rhythm scoring,this thesis puts forward the evaluation method of rhythm. The main innovations are shown as following:1. The melody evaluation methodWe study two main topics in melody evaluation method: melody feature extraction and melody similarity comparison algorithm. 1) Based on the inaccuracy of melody feature extraction and low noise robustness in evaluation method, we use SHS algorithm to extract melody which has high accuracy of feature extraction and make an improvement on the algorithm to improve the noise robustness of feature extraction and then improve the robustness of the melody evaluation method.2) Based on the fact that the DTW algorithm(dynamic time warping) fails in subsequence match to make method strictly requires the integrity of singing, improve DTW algorithm to meet the need of subsequence match and make evaluation method robust in the case of time offset and partial singing.3) Aiming at DTW has a high time and memory complexity, we use sliding window to reduce DTW’s memory complexity from O(MN) to O(N) and use a flexible threshold decision mechanism to reduce the time complexity significantly.2. The rhythm evaluation methodWe introduce the pertinence evaluation method of rhythm. Based on the acoustic model, we use frequency strength curve as feature of model training and recognition. Meanwhile, from the point of acoustics principle, we select reasonable music data to construct one in-beat and two out-of-beat statistical models. We realize the goal of evaluating singing rhythm based on results of the recognition of three rhythm statistical models.