Hi,
I have a dataset with many audio files and their beat timestamp. How can I process the data into a dataset and build up a beat tracking model? I think it will be a combination of CNN and LSTM as per the state of the art.
With Regards,
Suti
Welcome @sutirtha38!
Someone else will have to answer about the data setup, but you will find LSTM as a Recurrent component in the Deep learning components section - it looks like this, with settings
I like your plan! The beat timing apps I have tried previously (for setting my running pace) weren’t that good… it would be very cool if this works well!
Good luck!
1 Like
Hi, welcome to the forum @sutirtha38!
Very cool usecase!
Best way right now is if you can turn your timeseries data into an image (similar to this or newer examples: https://datascience.aero/time-space-time-series-train-cnn/). This is because we don’t yet have support for Arrays or Timeseries datatypes in the tool, although they are on their way.
Hope that will help!
Best,
Robert
1 Like
Thank you so much. I will try to preprocess it as suggested by @robertl. Thanks @JulianSMoore for your prompt reply.
HI @sutirtha38
I also took a look at the link suggested by @robertl and just in case it is useful to you, I started to investigate plotting recurrence diagrams.
I found this StackOverflow post which also addresses the issue of large datasets for recurrence plots. (but see also pyts)
If you just turn time-data into an image (1 pixel per unit time, for example)… 10,000 samples per second gives you (Nyquist) 5kHz bandwidth, which means in a 1k x1k image you could get 100s of data.
Obviously you can reduce the effective sample rate… it depends on the frequencies you expect to find/use in the “beats”.
Hope that helps.
1 Like