Questions about using magenta for other instruments(Voilin, Flute)

[Yoon Kang]

Hi everyone,

I am thinking about trying a mini-project, using magenta or fixed magenta 

to train auto-transcription for other basic instruments like violin or flute.

However, I am totally new to this audio domains so I reached here to get some advice, if allowed.

I have few questions before I dive into this project. Thanks for your attention in advance.

1) Do you think it is a good idea to apply magenta project to other instruments?

2) Is it simpler to transcript violin or flute? since those instruments have no issues

like piano solo regarding  polyphonic issues, note decaying issues and etc.

I am afraid that if I apply magenta to other items without considering easier and 

more efficient models to transcript or train violin score, it would be useless attempt.

3) Is there any guide to train magenta for other instruments?

(ex. transfer learning, audio data augmentation for target inst.)

Thank you for take your time to read this.

It would be much appreciated if you can give a tip about any of above questions.

[Evan Templeton]

Hey Yoon! I’m working on this right now using Apple’s MLSoundClassifier. I’d love some help, if you’re interested.

Evan

On Jun 24, 2021, at 2:13 AM, Yoon Kang <epon...@gmail.com> wrote:

Hi everyone,

--
Magenta project: magenta.tensorflow.org
To post to this group, send email to magenta...@tensorflow.org
To unsubscribe from this group, send email to magenta-discu...@tensorflow.org
---
To unsubscribe from this group and stop receiving emails from it, send an email to magenta-discu...@tensorflow.org.

[Jamie Hlusko]

I think the problem will be finding ground truth midi-audio pairs like the maestro dataset. If you find/create a similar dataset for another instrument please let us know!

[John Thickstun]

My MusicNet dataset might be helpful here. It has only a little bit of solo violin and solo flute, but the overall diversity of the dataset means that it might generalize better to solo violin & flute recordings than a model trained on piano music. If you're really determined, you could use a dynamic time warping protocol like the one I used to construct MusicNet to create new aligned midi-audio pairs for training a model on instruments of your choosing.

Best,

John

[Jamie Hlusko]

I converted the musicnet labels to midis, hope this helps!