Composing and editing piano performances using a Digital Audio Workstation (DAW) can often become a lengthy task as the proper timing, duration and velocity of each note must be set precisely in order to convey the intended emotional effect. If one does not have access to a MIDI keyboard, this task might even become impossible to handle.
In this project, we propose to rely on an Artificial Intelligence to interactively compose piano performances through a constant dialogue with a human user. To this end, we chose to train our A.I. on a single specific task that we called piano inpainting task. Similar to its counterpart in image edition, the piano inpainting task consists in restoring a piano performance where some parts have been occluded by filling in the missing parts. After being trained on a large dataset of piano performances, this A.I. is able to generate convincing and always-renewed proposals in a wide variety of styles.
With access to such an A.I.-assistant, a user can ask the model to generate any region of a MIDI piano performance with little effort. This comprises connecting two parts together, creating a new piece from scratch or even continuing a first sketch recorded on a MIDI keyboard. The user now acts as an artistic director, validating or invalidating the proposals made by the A.I. We hope that this novel approach to music composition will be entertaining and that it will make piano composition accessible to non-professional musicians.