All reviewers recommend acceptance. Their main concerns were about empirical evaluation, and the overall motivation and potentially limited scope of the work. I share the latter concern, but I do not believe that it should be grounds for rejection, given the quality of the work. A remaining concern is the quality of the training data, since labels are produced by an existing algorithm (onsets-and-frames) which is not 100% accurate. Please make sure to adequately address the implications of this in the updated manuscript.