The Karasawa model gives 1.4% for the mean error rate, 0.9964 correlation, and 12.7% for RMS.
The P4 feature set was evaluated as the best one with the mean error rate of 13.45%.
The results showed that multiple correlated coefficient R2 is greater than 0.96, and mean error rate δ is less than 15.
We can see that the proposed method obtains the smallest mean error rate, which demonstrates the effectiveness of the proposed method.
The YIN algorithm with median aggregation performs even better with a mean error rate of only 3% for polyphonic music, whereas the "aperiodic power" variant performs extremely poor.
This estimation approximately corresponds to the result of the classification method with DFT features which yields a mean error rate of 2% in our study.
When it comes to the mean error rate, d-Confidence does not perform has well has when relying on a SVM base classifier.
The results showed especially a significant effect of temporal strains on performance and a strong increase in mean error rate at the very beginning of the processing of the second task.
For onset detection, the target variable is the F measure; for instrument recognition, it is the classification error rate; and for pitch estimation, it is the mean error rate using a tolerance of a half semitone.
The mean error rate over all tones is applied to measure the performances of the approaches, i.e., it is assumed that all tones are equally important, regardless of their length.
