PIANO MUSIC ANNOTATION SYSTEM
Annotation for piano performance analysis and generation
Analyzing a performance based on a teacher's evaluation can greatly enhance performer's skills. By categorizing evaluation criteria according to musical feature and implementing a system where performances can be listened to and evaluated, performers could potentially refine their playing.
Utilizing evaluation data, we can analyze music performance and the teacher's music style. Moreover, by training the correlation between performance and annotations, we can develop models for automated annotations. With these goals, we have developed an annotation system and models using annotation data.
Our System
Demo
Please follow the link to enter the website.
Please follow the link to see the user manual.
Motivation
To improve music performance, consistent practice with appropriate guide is essential.
However, the number of instructors who can provide one-to-one musical guidance is limited, and the costs in terms of time and money are very high.
Also in the case of music performance, there are subjective evaluation standards that are difficult to evaluate with objective standards, so it is important to learn the instructor's evaluation standards well rather than generating objective evaluation standards.
We can solve this problem by automatically generating annotations through a model that has trained with instructors’ annotation data about performance.
In the case of evaluating music performances, there may be subjective interpretations by the evaluator, so it will be necessary to train a model by capturing which points the evaluator focuses using the evaluation method.
In the future work, considering those factors, we could generate customized annotations to help students learn in desired style by training model according to the instructor's play style.
GOALS
Our main goals are:
Generate a model that predicts intervals and values for each feature of label annotation for student's performance
Generate personalized evaluation by learning the evaluation style for each teacher, student's feedback and community informationation.
Using LLM, analyze non-standardized data like text annotation and audio annotation.
Model
Performance features are extracted through embedding of performance and music scores, and generate an annotation by predicting the annotation value after searching for the interval in which the annotation will exist using a model learned from existing annotations and performance data.
