Onsets and Frames Piano Transcription Updates

[Curtis "Fjord" Hawthorne]

Hi all, we've made a few improvements to the automatic piano music transcription model and wanted to share them with you!

Onsets and Frames now supports velocity estimation. This results in much more natural-sounding transcriptions. We've updated our paper on arXiv with details:

https://goo.gl/magenta/onsets-frames-paper

We also have a site with several examples of transcriptions with velocity:

https://goo.gl/magenta/onsets-frames-examples

The code on GitHub is also updated:

https://goo.gl/magenta/onsets-frames-code

We've also added an easier to use command line utility for creating transcriptions. Once you install the pip package and download the checkpoint, you can just run this command to transcribe piano audio to midi files:

onsets_frames_transcription_transcribe \

  --acoustic_run_dir="${CHECKPOINT_DIR}" \

  <piano_recording1.wav, piano_recording2.wav, ...>

There's also a colab you can use without installing any software:

https://goo.gl/magenta/onsets-frames-colab

Finally, we'll be presenting our research at ISMIR 2018 in Paris this fall. We'd love to talk with you more about it, so please come to our talk and stop by our poster afterward!

Happy transcribing!

-Fjord

[Circu Virtu]

Great news, tnx Fjord for imrpoving the model and adding velocity. That makes it sound alot better.

I'm still getting additional very low and high notes. Is there any advicer best practice  to maybe preprocess the audio file

for best results? Are there any parameters in the model I might tweak for improvements?

-cv

[Curtis "Fjord" Hawthorne]

One thing that seems to lead to better results is to use the original .wav file, not a file that has been compressed with MP3 or something like that. I'm guessing this works better just because we trained the model on uncompressed .wav files. It would be interesting to try expanding the training dataset to include different versions of the files created with different lossy compression techniques.

--
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
---
You received this message because you are subscribed to the Google Groups "Magenta Discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to magenta-discu...@tensorflow.org.