A selection of spectrogram & audio samples, produced by applying transformations to neural network architecture, using the Brave synthesiser. Here, I present a spectrogram/raw audio file of the samples, along with a description of the transformations that have been applied.
The darbouka RAVE model without any transformations applied to the underlying network. This should act as a baseline when inspecting the other samples.
Applying Δy = 0.47 to encoder.net.1.bias.
Applying Δy = 0.3 at encoder.net.2.bias.
Applying a scale factor of -0.79 to encoder.net.7.weight.
Applying Δy = -0.5 at encoder.net.10.weight.
Applying Δx = 0.195 at encoder.net.14.weight.