Dilated RNN for Music Transcription

[faraazn]

I was reading through the dual-objective piano transcription paper (really cool and *super* thorough paper btw!) and I was wondering if you ever attempted transcription using Dilated Convolutions as described in Wavenet? I would guess that given the model architecture (predicting onsets/frames) and structure of the dataset (only piano being played, nothing crazy), the CNN and LSTM combination makes more sense. The task of finding onset frames seems especially better suited towards CNNs. But I could also imagine a model using the Dilated RNN that simply predicted the frames given some kind of Q-transformed audio / mel scaled spectrograms. Just curious to see if you ever tried that and noticed anything (lacking) about the results. Thanks!

Faraaz

[faraazn]

Oops title of this post should be dilated convolutions for music transcription :P

[Adam Roberts]

We did try this when we were developing Onsets and Frames but did not have luck, most likely due to the small size of the training sets. Once there are larger datasets available, I think this would be worth revisiting!

--
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.

[Qiuqiang Kong]

I did tried dilated CNN but was not able to outperform RNN. Dilated CNN introduced many "noise" in transcriped scores but RNN is more smooth. 

[faraazn]

Got it, thanks!