LIVE GENERATIVE MEDIA SYSTEM / 2020
Biden Unleashed
Biden Unleashed was an automated Twitch stream featuring a synthetic Joe Biden avatar that responded to chat. I built it with local text-generation models, generated speech, 3D animation, and a two-computer system for getting everything on the air.
One machine handled Twitch input, generation, and synthetic output while a second machine handled playback, character media, and the live broadcast.
The recovered training workflow used the 355M GPT-2 model, with later experiments moving into aitextgen and custom-tokenizer workflows.
The project ran publicly on Twitch long enough to reach Partner status before I retired the stream.
I used remote SSH access and a network-connected power strip to recover the unattended streaming machines when they froze while I was away from home.
This started with speech transcripts and segmented audio, then grew into a live synthetic-media system. I trained and tested local language-model workflows, generated character responses, produced synthetic speech, and tied the output to an animated 3D performance instead of stopping at a command-line text demo.
The live system was split across two machines so generation did not have to compete with playback and broadcast work for the same GPU and memory. One computer listened to Twitch, generated the response, and published the resulting text/audio across the home network. A second machine watched for new output, updated the on-screen message/character presentation, played the generated media, and fed the scene into the Twitch stream.
Reliability became its own engineering problem. I was often away from the apartment during the pandemic, so I configured remote access through my home network and used a network-connected power strip as a last-resort recovery path when one of the PCs froze. Both machines were configured to boot back into the scripts they needed. It worked well enough to operate remotely, but intermittent lockups and network failures could still turn into days of downtime when I could not safely get back to the hardware.
The recovered files now cover transcript preparation, GPT-2 training/generation, later aitextgen experiments, and part of the iClone lip-sync workflow. The model checkpoints and the exact Twitch listener, network-transfer, second-machine watcher/player, startup, and remote-recovery scripts have not yet been found. I am documenting those missing pieces explicitly rather than rebuilding them from memory and presenting reconstruction as original source.
The character, generated text, and generated voice were a fabricated performance. None of the generated output should be presented as an authentic quotation or statement by Joe Biden. I keep the project as an early experiment in live generative media, distributed systems, and automated performance: not as political messaging.
Original footage and project documentation.
Preserved media from the original project archive. These are not newly generated mockups.
Model, corpus, and iClone source recovered
The recovered source covers transcript/audio preparation, GPT-2 and aitextgen experiments, synthetic text generation, and part of the iClone lip-sync integration. Large model checkpoints and media corpora are intentionally not bundled with the portfolio.
Two-computer Twitch deployment
The live deployment separated generation from media playback and broadcasting. The exact Twitch/network/player source from the second machine is still missing from the recovered archive.
Original synthetic-media capture
The archived livestream recording shows the satirical avatar and on-screen audience interaction.