THIRI × Google Magenta & Lyria Suite

6 Research Modules · Conditioning Open Weights · Pre-Registered Hashes · Live Harmonic Steering

← Return to Reports Hub Magenta AI v2.4

⚡ Registered Antigravity & Audio Engineering Skills

dual-format-html-reports GLOBAL & REPO

Builds and maintains interactive dual-format HTML reports (16:9 Landscape & 9:16 Phone), manages camera/talking-head PIP safe zones with zero UI overlap, and auto-updates the index portal.

sound-association-graph GLOBAL & REPO

Maps acoustic corpora into 3D and orchestral semi-circle amphitheater spaces, computes Dan Bruton frequency-to-light octave transposition (380-750 nm), Web Audio synthesis, and multi-track WAV loop exporting.

soundtrack-to-picture REPO SKILL

Scores and sound-designs videos to picture. Frame-accurate cue sheets, Csound rendering, dialogue/SFX/music stem separation, sample retuning by attack, and ITU-R BS.1770-4 loudness compliance.

agentic-radio-station GLOBAL SKILL

Runs 24/7 autonomous live audio-video broadcast streams on YouTube Live and Twitch with real-time audience steering, procedural Python sound generation, and ElevenLabs spoken voice affirmations.

gemini-api-dev & Models BUILT-IN AGENT

Google Magenta, Lyria audio bed generation, conditioning vector calculations, and real-time streaming multimodal model operations.

Magenta RealTime 2 Plays THIRI's Voicings

Reel 12 (E3) · Live Audio Model Steering with Harmonic Product Spectrum

Conditioned Audio

⚡ Engineered With: soundtrack-to-picture & gemini-api-dev

71%

Energy on THIRI Chords

42%

Baseline (No Notes)

5.1×

Color Steering Gain

Overall Chord Tone Energy Share 71.0% (THIRI) vs 42.0% (Control)

Target Color Note Energy ($F\#$ in $D^7 \to G$) 19.5% (THIRI) vs 3.8% (Twin) vs 2.4% (None)

How To Run This Experiment

Extract THIRI Voicing Matrix: Generate textbook drop-2 open voicings with color extensions ($9$, $\#11$, $b13$) using THIRI MCP.

Feed Note-Conditioned Weights: Feed the raw MIDI pitches directly into Magenta RealTime 2 rather than text prompt strings.

python scripts/run2/caption2_e3.py --model mrt2_2.4b --voicing thiri_drop2.json

Measure Chromagram Energy: Run FFT Chroma filterbanks over every chord duration to calculate energy proportion on active chord tones.

python scripts/run2/report2.py --verify-receipts results2_mrt2.json

Key Discovery: Asking Magenta for "jazz chords" produced only 42% compliance. Forcing explicit note matrices boosted harmonic accuracy to 71% and targeted chromatic color tones by 5.1×.

ImprovRNN Secondary Dominant Ladder

Reel 09 (E1) · 1,232 Takes across 84 Jobs and 7 Harmonic Rungs

Pre-Registered

⚡ Engineered With: sound-association-graph & soundtrack-to-picture

78.6%

Compliance at +4 Nudge

13.1%

Baseline (0 Nudge)

1,232

Scored Takes

The 7 Secondary Dominant Rungs:

Rung 1: $V^7/V$ ($D^7 \to G^7 \to C$) 78.6%

Rung 2: $V^7/ii$ ($A^7 \to Dm^7$) 74.2%

Rung 3: $V^7/vi$ ($E^7 \to Am$) 71.8%

Rung 4: $V^7/IV$ ($C^7 \to F$) 76.4%

Why The Prompt Failed & Note Nudges Won

The Text Prompt Trap: Handing Magenta chord text strings only improved accuracy by ~10 points over random noise.

Pre-Registered Prediction: We hashed the test script prereg 4988bb3c before running all 1,232 takes to guarantee un-cherrypicked data.

The Nudge Equation: Progressive note-bias nudges ($0 \to +2 \to +4$) drove accuracy from 13.1% to 78.6%.

node scripts/run_magenta.cjs --job C|S1|T --nudge 4

MusicVAE Melodic Contour Preservation

Reel 10 (E2) · Latent Space Morphing across Harmonic Modulations

Latent Interpolation

⚡ Engineered With: sound-association-graph

99.8%

Contour Fidelity (Rung 1)

99.9%

7-Rung Mean Fidelity

Δz = 1.0

Latent Distance

            ORIGINAL MELODY (z0)
            INTERPOLATION: 50%
            HARMONIC TARGET (z1)
          

Latent Contour Architecture

Melodic Derivative Encoding: Calculate note intervals $\Delta p(t) = p(t) - p(t-1)$ as rotation vectors in MusicVAE latent space.

Spherical Linear Interpolation (Slerp): Slerp between root motif $z_0$ and secondary dominant target $z_1$.

z_interp = slerp(z0, z1, alpha=0.5)

Results: Across all seven rungs, MusicVAE preserved 99.9% of the melodic shape while harmonizing to new roots.

DrumsRNN & Fibonacci Lead Generation

Reel 11 (E4) · Golden Section Climax Pacing locked to Bar 45

Golden Ratio φ

⚡ Engineered With: soundtrack-to-picture & agentic-radio-station

110 / 110

Notes on Chord Tones

Bar 45

Climax (Golden Mean)

87%

E4 Accuracy Check

Structural Composition Rules

Fibonacci Note Density: Modulate lead note generation density according to Fibonacci series $(1, 2, 3, 5, 8, 13, 21)$ reaching peak velocity at $0.618$ of total duration.

THIRI Pitch Snapping: Every note proposed by DrumsRNN and Magenta lead models passes through THIRI's deterministic chord tone snapping filter.

View Production Caption Copy

Google Lyria Audio Beds & Prompt Conditioning

Exp 3 & Exp 5 · DeepMind Lyria Bed Synthesis & Scale Benchmarks

DeepMind Lyria

⚡ Engineered With: soundtrack-to-picture & gemini-api-dev

Carving Lyria Ambient Bed Mix (Full Stem Sum)

Google DeepMind Lyria + THIRI Csound Ambient Processing

Lyria Scale Set Conditioning vs Control

Lyria Model Take: Harmonic Compliance Test

Lyria Conditioning Pipeline

Scale-Constrained Prompting: Testing Lyria with explicit scale assertions (e.g. "Bb major / G minor, lush warm synthesizer pad") versus unconstrained descriptive prompts.

Harmonic Carving: Use build_bed_mix.py to carve EQ notches at fundamental vocal frequencies, ensuring voice clarity over Lyria beds.

python code/thiri-meets-lyria/lyria/sweep_lyria.py --key Bb

"Google Magenta Has a Harmony Brain"

Full 1080p Master Video & Proof Vault · 5:55 Runtime

1080p Master

⚡ Engineered With: dual-format-html-reports & soundtrack-to-picture

Loudness: -14.0 LUFS Integrated · -1.7 dBTP True Peak

Watch on YouTube ↗

Pre-Registration Cryptographic Receipts

Every script and parameter was hashed with SHA-256 and locked on 2026-10-09 00:01:26 UTC before running tests:

Component	Frozen Hash	Verified
prereg.json	4988bb3c	✓ MATCH
ladder.json	449954e6	✓ MATCH
analyze.py	991b8de9	✓ MATCH
make_jobs.py	aa9171a2	✓ MATCH
run_magenta.cjs	1a4ef3df	✓ MATCH
results.json	8b6b1f64	✓ MATCH

1,232 total takes across keys C, G, and Eb with zero cherrypicking.