Similar to issue #36, including experimental evidence that addressing this would have significant inference benefits, but requiring a different solution. Possible solution: Implement pitch-center detection and "correction" during training data prep and inference data prep so that if the pitch center is not a half-step multiple away from A=440 Hz, transpose the performance so that it is.
Having this problem is a weakness of the CQT algorithm, but the success of CQT for CSI probably justifies the extra work and processing to detect and adjust performance pitch.
Similar to issue #36, including experimental evidence that addressing this would have significant inference benefits, but requiring a different solution. Possible solution: Implement pitch-center detection and "correction" during training data prep and inference data prep so that if the pitch center is not a half-step multiple away from A=440 Hz, transpose the performance so that it is.
Having this problem is a weakness of the CQT algorithm, but the success of CQT for CSI probably justifies the extra work and processing to detect and adjust performance pitch.