Hello everyone,
Small (but hopefully interesting) update!
A while back I showed how I was able to memorize music and play it back using HTSL. Now with NeoRL I can not only remember but also generate more music based on sample data.
As is usually done with these predictive-generative scenarios, I added some noise to the input as it runs off of its own predictions. This causes it to diverge from the original data somewhat, resulting in semi-original audio.
Here is some audio data from a song called “Glorious Morning” by Waterflame:
Here is a sample of some audio I was able to generate, after training off of raw audio data, without preprocessing:
Training time: about 1 minute.
Now a problem with this is that it is just being trained off of one song right now, so the result is basically just a reorganized form of the original plus noise. I am going to try to train it on multiple songs, extract end-of-sequence SDRs, and use these to generate songs with a particular desired style based on the input data styles. Longer training times should help clear up the noise a bit too (hopefully).
Full source code is available in the NeoRL repository. It is the Audio_Generate.cpp example. Link to repository here.
Until next time!
Impressive work indeed, considering the quality/time ratio.
I wonder if it’s possible to make a language learner. I know a NN model called ANNABELL (http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0140866) can do this. It uses no backprop as well.