MusicNet (2016)

Re:MusicNet (2016) by dpflan on Thursday October 15, 2020 @08:48AM (#24788023)

Somewhat silly question: To what extent does analysis of music/sound by looking at spectrogram images provide "enough" information for usage in deep learning systems (like ResNet) compared to something like MusicNet?

[ Parent | on HN ]

Re:MusicNet (2016) by dijksterhuis on Thursday October 15, 2020 @11:24AM (#24789551)

They're pretty fundamental. speech to text networks like DeepSpeech convert audio to an MFCC power spectrogram. Others use an Stft magnitude spectrogram.

It's a 2-d (amplitude by Freq and time) representation of something that's usually 1-d (amplitude over time).

I think some networks have tried using the discrete wavelet transform too,but that's outside my knowledge area.

[ Parent | on HN ]

Re:MusicNet (2016) by dpflan on Thursday October 15, 2020 @11:46AM (#24789822)

Thanks for the explanation. I've checked out your HN profile and then SoundCloud: are many of the tracks you've posted generating as part of your research into "adversarial audio examples"?

[ Parent | on HN ]

Re:MusicNet (2016) by dijksterhuis on Tuesday October 20, 2020 @03:47PM (#24841630)

Sorry, I just saw this.

Ha, noooo. I own a bunch of synths and samplers and noodle about for fun.

Although I haven't played with them much recently. I've gotten all super serious @_@

People do have to host their adversarial examples somewhere when publishing their research, not sure why no one has used SoundCloud yet... Odd.

[ Parent | on HN ]