WaveNet by Deepmind and/or Alphabet, Inc. and/or Google has created a very impressive audio signal modeling system. Although Adobe’s Voco is cool, my feeling is WaveNet is much further along in development, it’s just missing the flashy live video demonstration that went viral from Adobe.
A quick summary of what WaveNet is capable of.
They trained WaveNet on a dataset of classical piano music. This dataset was generated by modeling the raw waveforms of the classical music pieces, one sample at a time with approximately 16,000 samples per second. They did not provide a musical score, structured sequence, or order, they simply allowed WaveNet to generate it’s own musical compositions based on what it learned from the music dataset. Then WaveNet takes all those many thousands of tiny audio soundwave samples and reconstructs it own compositions. Really quite amazing.
Without further ado, here’s WaveNet’s original piano compositions.