---

Transcription demonstrations

---

Demonstrations I: transcription of music from CDs

Note: More recent transcription demonstrations, using the system developed together with Matti Ryynänen, are here.

The transcription examples below are produced by an automatic transcription system, which gets only the acoustic musical signal as input (with no side-information) and produces a musical notation for the sounds that make up the piece (i.e., a MIDI-like representation). Please notice, that the overall transcription system has not yet been optimized at all. It is simply:

• onset detection (signal segmentation) + multipitch estimation at each onset

Typical problems in the transcription are illustrated by the last few examples in the table. Such problems are still very common in the automatic transcription of real-world music.

Demonstrations II: transcription of synthesized MIDI

Accurate and realistic evaluation of a transcription system is best achieved by transcribing synthesized MIDI songs. In this case, the exact reference score is readily available. Also, high-quality MIDI songs are complex enough to simulate the real case.

Transcription of synthesized MIDI gives somewhat optimistic transcription results. It is more difficult to transcribe music from CDs. However, MIDI songs allow much more accurate comparison between the reference and the transcription and this is more important. Pure listening of transcription results is not sufficiently informative.

Below, 20 randomly selected MIDI pieces and their transcription results are shown (I excluded 3-4 pieces due to bad transcription. Considering it now it is very pity because the selection is no more completely random. On the other hand, preparing a new and completely random set of transcriptions is too tedious, no time for such now.)

Again, the transcriber gets only the acoustic musical signal as input. Musical meter estimation was used for temporal segmentation, which improves the system accuracy compared to using one-by-one onset detection. No musical knowledge or the context was utilized. Each temporal segment was transcribed independently using a multiple-F0 estimator.

More? Check the DAFx2000 demonstrations page.