A large corpus of music written in lute tablature, spanning some three-and-a-half centuries, has survived. This music has so far escaped systematic musicological research because of its notational format. Being a practical instruction for the player, tablature reveals very little of the polyphonic structure of the music it encodes—and is therefore relatively inaccessible to non-specialists. Automatic polyphonic transcription into modern music notation can help unlock the corpus to a larger audience, and thus facilitate musicological research.ududIn this study we present four variants of a machine-learning model for voice separation and duration reconstruction in 16th-century lute tablature. These models are intended to form the heart of an interactive system for automatic polyphonic transcription that can assist users in making editions tailored to their own preferences. Additionally, such models can provide new methods for analysing different aspects of polyphonic structure.ududWe have experimented with modelling only voice and modelling voice and duration simultaneously, applying each in a forward- and in a backward-processing approach. The models are evaluated on a dataset containing 15 three- and four-voice intabulations. Each processing approach has its advantages, and the results vary between the models. With accuracy rates between approximately 80 and 90 per cent, both for voice prediction and for duration prediction, the best models’ performance is promising. Even in this early stage of the research, such models yield a useful initial transcription system.
展开▼
