Methodology

The journey for this project so far has been experimental to say the least. Fortunately, the methodology for creating and producing the work has been less experimental. 

The practice-as-research methodology employed in the production of the artefacts engages several phases of development:

• Composition of musical material (Melody and Harmony) as input.
• Generation of the output with the (AI) collaborator and reflection on it.
• Development of the artefacts with a DAW.

While there is nothing unusual about the development phases compared to a more run-of-the-mill collaboration, the composition relies on a non-human participant, evoking Actor-Network Theory (ANT). This runs counter to Technological Determinism, which would posit that the technology itself is having an impact on the output, rather than the complex interactions of the inputs to the model (a technology), the model (another technology, the output, my ability to have some control over the output via weights, and any other human and non-human factors that create a complex network of relationships that can affect the final product. 

To make this a lot easier to digest, I’ve created a 5-step methodology (Figure 1) that captures the process of working with my AI collaborative partner though to producing a track.

The breakdown of the phases goes as follow:

1. Generate a melody with ST4RT+ based on the input from my written melody and harmony.
2. Check the output from the model, if the output sounds wrong, modify the output weights and regenerate if necessary.
3. Document the weights used to create the output melody. These weights may be different for each melody to create a good output.
4. Generate a second melody (melody only) that is based on the first input using the same weights as last time. 
5. Develop the materials in a DAW, reflect on the output and regenerate if necessary.

The regeneration parts of this methodology aren’t deleterious, in this case meaning that I keep all the previous outputs as both reference and as opportunities for remixing with other outputs. I’d do this with a human collaborator, so I may as well do it with the AI model. I guess the best way to look at this is if a human collaborator came up with an amazing guitar riff for a song, but the verses were terrible, you wouldn’t throw out the amazing riff because the verse was bad. You’d simply work on getting the verses better. 

In this way, the AI has felt like a collaboration partner. Sometimes it gets it right. Other times I feel it misses the mark, but by working with it (rather than against it) it can sometimes surprise and delight when we find a solution that is better than the sum of its parts.