VS Code is quickly becoming my DAW in the composition phase.
I write each piece as a JSON song doc in VS Code; my VS extension plays the synth engine as I edit. Every save is heard within a few seconds, without stopping the music.
A bit about the newest part below.
@boysbasbleu That’s basically the motivation for what I’ve been building (see my pinned post). A JSON5 format that defines the music and its instrument params, tools that enable an AI to “hear” the result, and a set of skills that enable the AI to use those to collaborate with me.
@Princesah009@X I’m finding that even if I can successfully vibe code something really cool and useful, I still need to be comfortable with devops and marketing to successfully share it with the world. Let’s connect!
@figarotokii@jakemclain_ That’s basically what I did. I created a synth/composition sequencer that uses plain JSON5. Then trained a set of Claude skills that know music theory and how to work with that JSON5 format. check my pinned posts if you’re interested.
Lance, you might be interested in the synth/composer/sequencer-as-code toolset I’m so close to releasing open source. It basically turns VS Code into a DAW where you can live-edit a composition as plain JSON5. But you can also send the tracks straight into Ableton as well - all in realtime.
As someone who's published a couple hours-long generative pieces, I can verify (no pun intended) that verifying whether one is actually good does take real time and effort..
ffmpeg and my own Rust-based analyzers go a long way. Statistical and frequency analysis catch a lot of issues fast, without listening hours. On a recent piece I'm working on, my tooling flagged a part that turned too dissonant about 20 minutes in. The culprit: the synth's fm_ratio was allowed to drift too far.
So the bigger win is upstream: a generative engine that's rules-based, music-theory-aware, and lets me tightly constrain how the composition and synth parameters drift. That's been a major focus of the synth/composer/sequencer-as-code toolset I've been building.
By the way, I plan to open-source the core Rust crates and publish my VS Code extension very soon. Stay tuned 👀
Anyone else doing automated QA on generative audio? Curious what you measure..
"How do you know a 10-minute song is good? You kinda have to spend 10 minutes." @kane of @rootvc on why generative music lags text and code: verification itself takes real time.
@kris_craw Yep, just making a change to the JSON5 doc and then hearing the change right away. I also have a set of Claude skills that can work with the same file.
@AIAllar@suno This is just kind of the nature of an AI that generates the wav files directly versus AI tools the collaborate with you on the pre-render composition. The later gives you the control to quickly make those kinds of very specific changes. Good luck to you!
@kris_craw I’ve been thinking similarly. In my case, my synth/composition sequencer runs as a VSX extension and LSP so I can compose and live edit music in an IDE (such as VS Code) with the song and sounds fully represented in plain JSON5. (See my pinned posts)
@johnrush The greatest performance factor in “coding with AI” is now the product/project management skill. I believe that is the core of your observation.
I create long-form (1 to 5 hours) unique, non-looping tracks and found both lacking for that.. and neither provide the fine control I wanted.
So I built my own engine that leverages AI to collaborate with me in the composition phase but leaves me in full control because the composition, instrument params, generative sequencing rules, etc are all in a JSON doc I and an AI collaborator can work on in live-edit mode.
Once I release it to the world, I’m going to hit you up for a review ;-)