Generated with a machine learning model I've been working on; a few bass notes were removed and some vocals bleed through in the song but as the model trains that'll get better. New model is back to a dense net setup that utilizes the transformer architecture - calling it a frame transformer. Basically, just a u-net where each decoder is preceded by a bottleneck to 1 channel with a 3x1 kernel and an evolved transformer block with relative positional encoding which is then concatenated to its input and then upsampled in the following decoder; the frequency bins are treated as the embedding dimensions and attention is computed between each frame. This is only the 4th epoch, 1st at the highest learning rate yet has surpassed the validation loss of all previous versions; will post more videos after further training.
Will probably unlist this video when First Fragment releases the official instrumental version of this album.
11