Guitar Tab Mining, Analysis and Ranking

With over 4.5 million tablatures and chord sequences (collectively known as tabs), the web holds vast quantities of hand annotated scores in non-standardised text files. These scores are typically error-prone and incomplete, and tab collections contain many duplicates, making retrieval of high quality tabs difficult. Despite this, tabs are by far the most popular means of sharing musical instructions on the internet. We have developed tools that use text analysis and alignment for the automatic retrieval, interpretation and analysis of such tabs in order to filter and estimate the most accurate tabs from the multitude available. We show that the standard means of ranking tabs, such as search engine ranks or user ratings, have little correlation with the accuracy of a tab and that a better ranking method is to use features such as the concurrency between tabs of the same song. We also compare the quality of top-ranked tabs with state-of-the-art chord transcription output and find that the latter provides a more reliable source of chord symbols with an accuracy rate 10% higher than the ranked hand annotations.