Demo: audio synthesis and transcription results (accompaining IEEE Transactions on Audio Speech and Language Processing paper)

Performance rates have been averaged with respect to all test musical pieces, and the respective standard deviation is also represented.


Onset-only metric

In the Onset-only metric a correct note implies a correct onset with a deviation up to 50 ms.

Individual results

Onset-only benchmark

Onset-Offset metric

This metric also presents the results as Recall, Precision, F-measure and Mean Overlap Ratio (MOR). In the Onset-Offset metric a correct note implies a correct onset with a deviation up to 50ms and a correct offset with a deviation of up to 20% of the note length or 50ms.

Individual results

Onset-only benchmark

Decay/Sustain score

This metric is the one that best correlates with the human hearing perception. Results are presented as Decay Score, Sustain Score and Final Score: Decay Score is used for percursive pitched instruments and employs a note oriented approach considering only pitches and onsets, generating a score ([0-100]%) for each note; Sustain Score is used for sustain musical instruments (eg.: woodwind) and employs a time oriented approach, measuring the overlap between the original and transcribed notes; the Final Score is the average value between Sustain Score and Decay Score.

Individual results

Onset-only benchmark