As the semester comes to an end, so too does the Orchestrator project. The blog hasn’t been updated in a while, but we’ve been hard at work on some cool final projects!

• James and Alicia wrote a music classifier based on the M5 convolutional network architecture and trained it on both raw audio samples and latent vectors produced by the pretrained UMTN encoder.
• Jim and Alicia used PCA and Gaussian mixture models to analyze latent vectors produced by the pretrained UMTN encoder and generate new latent vectors as well.
• Jasmine and James wrote a Seq2Seq network for music translation and trained it using a dataset of single notes rendered with different MIDI instruments.
• Praveen and Andrew modified the UMTN source code so that training in ‘single-node’ mode (on a single machine) still trained all of the different decoders separately.

To learn more about these projects, check out our final presentation slides, which we presented at the Launchpad Showcase on May 6.