May 10, 2019

Attendees: Laurent, Yaolong, Silvain, Tim, Nestor Regrets: Ich

Yaolong: People in music theory have not paid a lot of attention into the fact that the way we annotate is ambiguous, and this is problematic in a space like machine learning. We can’t expect good accuracy with ambiguous annotations.

Silvain: Some musics may not represent the object of study of the task (tonal music). For example, maybe Bach chorals are not the best corpus for studying harmony and string quartets would be a better corpus. Should we consider that?

Tim: The inspiration about the task of melodic pattern finding is that if you can find something that is significant, that pattern will probably be repeated in the future. Knowing this information could help us to “fill the gap”, “predict the future”, “extend the music”.

Nestor: What is the single most relevant feature in finding cadences?

Laurent: Voice-leading!

Yaolong: Voice leading is also important in harmony

Laurent to Silvain: Why are computers and all the stuff we do important for a music theorist?

Silvain: The use of computers in music theory is critical. A career of a theorist that analyzes 100 pieces is already productive, however, it doesn’t even scratch the surface of what is a musical style. We have a biological limitation to analysis. We can overcome that barrier only with the speed of computers

Nestor: Is there something we can do together to publish a paper that each of us cannot complete individually?

After discussion at the table, the creation of a dataset with annotations for several of our taks emerged as the way to put us all in the same track.

Maybe, for this, we could have a shared google docs document