Participants 2022

Lyrics Generation: OpenAI GPT-3 To generate the story, we inserted a starting prompt containing example sentences, and let the model proceed from there. Getting GPT-3 to continue a given story was fairly straightforward, and with a few iterations ended up with a story above. The model also generated dialogue for the story, which we decided to keep in the final song. Next, to generate the lyrics, we used the above story as context and tried to generate lyrics with GPT-3. This was much more challenging, as we had to carefully tweak the prompt  instructions. We did not want to edit or curate the generated lyrics too much, and instead used the insert and edit funtionalites of GPT-3 to generate alternate verses. Vocal Synthesis: Uberduck AI We used the Tacotron-2 and FastPitch models from Uberduck. We found that the models trained on rappers had a flow to them and could sit on a beat, and used this as the main voice for the song. We picked the voice of our favorite rapper, Kendrick Lamar. For the dialogues, we experimented with various voice models, and settled on two - one that had the tone of an AI voice (PotatOS from Portal), and the other a clearer distinct voice (Kratos from God of War). The challenges here were to get certain sounds pronounced correctly. We had to experiment with phonetics variations of words to get appropriate pronunciations from the voice synthesis models e.g. universe vs yuniverse Along with this, we also use simple apps from Magenta studio such as Generate, Continue to create musical ideas and Drumify and Groove to add dynamics to generated MIDI sequences. We always use AI tools as collaborators, as systems with which we bounce off ideas when we run out of them. Integrating AI tools as collaborators during music making vastly improved the fluidity of our process. We were never blocked due to running out of lyrical, vocal or musical ideas since the AI tools could fill in the gaps of our informal musical education. Specifically as a team, we found lyric writing and singing to be the most difficult elements to create by ourselves, since none of us are trained artists or vocalists. Hence, with the help of GPT-3 and vocal synthesis, we were able to give sound to ideas that we could not have otherwise expressed.