I Used My Live Guitar as a Sentiment Signal Against an AI Script

Sam Castillo — Fri, 16 Oct 2026 10:38:38 +0000

Here is the setup. AI voices read a podcast script. I sit with my guitar and play live underneath, with no plan and no rehearsal. I am just reacting. Afterward, we ran the clean guitar track against the script, line by line, to see whether the playing was saying something. It was.

What the playing said

Agreement sounds full. When a line lands, I answer with a big, settled chord that resolves home.
Disagreement sounds tense. Loud playing in a minor, diminished, or off-key color shows up where I push back on a claim.
Pushing back sounds like a chug. Short, repeated staccato notes show up as resistance to the script.
Avoidance sounds like stepping back. On topics I would rather not engage, the guitar drops out, comes back, and drifts sharp. It settles once the subject moves on.

Ground truth first

We anchored the analysis on one moment I confirmed myself (a five-second passage where I know I disagreed). The method has to call that moment correctly before we trust it anywhere else.

A personal fingerprint

We also scored how much each recording sounds like me: usual chord families, go-to ii–V moves, tuning habits, tone. On a 0–100 signature score, the podcast sessions average 58; every earlier era of my channel averages 46–48. Reacting to ideas in real time brought out more of my own voice, not less.

Why it is hopeful

Most talk about AI and music is about replacement. This points the other way. The AI supplied the script and the voices; what came back was a clearer picture of the human part: how I listen, where I agree, where I hold my ground. Used well, AI is a lever for human skill, not a stand-in for it.

Caveats: these are inferences from the music, not mind reading. Chords and keys are machine estimates, the sample is one player, and the method is still being tested.

How we approached it (dev view)

Isolate the guitar stem from the mix.
Extract per-window features: loudness, onset density (staccato "chug"), estimated chord and key, pitch deviation from tuning, and silence/dropouts.
Align windows to script lines by timestamp.
Map feature patterns to four labels (agree / disagree / push back / avoid) with context rules, then validate on the one human-confirmed moment before reading anything else.

Biggest lesson: one confirmed ground-truth point is worth more than a dozen clever features. Without it you are just telling stories with spectrograms.

I would love feedback on better ways to validate a single-subject signal like this.

The full write-up lives on my site: https://www.predictiveinsightsai.com