Adjustment of initial transcription offsets, also affects pitches

[00sapo]

MMSP2021-Audio2ScoreAlignment/alignment/evaluate_audio2score.py

Line 185 in f37fafc

self.perfm -= np.min(self.perfm[:, 1:3])

MMSP2021-Audio2ScoreAlignment/alignment/evaluate_audio2score.py

Line 124 in f37fafc

self._ground_truth -= np.min(self._ground_truth[:, 1:3])

MMSP2021-Audio2ScoreAlignment/alignment/evaluate_audio2score.py

Line 118 in f37fafc

self.score -= np.min(self.score[:, 1:3])

should be like:

 self.perfm[:, 1:3] -= np.min(self.perfm[:, 1:3]) 

Such an error could affect TAFE, EIFE, SEBA, and transcription evaluation. Specifically:

• Since the transcription first onsets are likely the same as the ground-truth, the transcription evaluation is only partially affected, likely in a negligible way
• Since in ASMD the first onset of the score is 0, the SEBA method is not affected at all, because it only uses the pitches of the score
• TAFE and EIFE could may be affected by such error in that they compare the score pitches to the transcription pitches, and this error makes them not comparable; we can expect an improvement in both the two methods

As consequence, the main conclusions of the paper, should not change

• EIFE is far better than TAFE on piano music
• SEBA is better than TAFE and EIFE on non-piano music
• Bytedance model works better than Omnizart model

[00sapo]

Fixed in 09041c9

New results are being computed.