Handwritten Music Recognition Improvement through Language Model Re-interpretation for Mensural Notation

Abstract

Handwritten Music Recognition studies techniques for computers to transcribe handwritten musical notation that is registered in document images into electronic format, and to make this music available to the public. This task has been of great interest lately, as the technologies improve and can get better and better results on this problem. Recent machine intelligent approaches based on Deep and Recurrent Neural Networks have already shown how they work significantly better in the problem than traditional HMM-based approaches, especially when we are talking about Mensural Notation. These Neural Network-based researches have investigated the task of recognizing Mensural Notation as another written text recognition task, but have not explored the characteristics of musical elements in depth. Other papers have tried to dig deeper into analyzing musical elements and the extraction of their characteristics from segmented symbols, without reflecting this in holistic way. In this paper, we will try to make a complete recognition system directly from the scores, using techniques that enhance information obtained from symbols. We explore other language model interpretations and test our proposal on a publicly available dataset. In our experiments, we have made a 31% relative improvement in regards to error at the symbol level. With this, we have gone from a 3.91% absolute error rate, using Neural Network-based technology, to a 2.70% absolute error rate, by using language model re-interpretations.