The LucidHarmony Tech Stack: Modeling, Plugin, and Website

Published: December 17, 2025 • 10 min read

LucidHarmony is built as a full pipeline: we train transformer models offline, run real-time inference inside a JUCE audio plugin via ONNX Runtime, and support it all with a static website + infrastructure for shipping updates and documentation.

This post is a living inventory of the technologies we use across the three layers.

Modeling & Data Pipeline (Offline)

Programming language & runtime

• Python — the primary language for dataset extraction and training.
  • https://www.python.org/

Symbolic music + analysis

• music21 — corpus access, score parsing, and harmonic / Roman numeral analysis.
  • https://www.music21.org/

Machine learning

• Transformer (attention-based sequence model) — the primary architecture for learning harmonic progressions as of v1.3.0. Replaced the earlier LSTM approach for richer, more expressive output.

  • https://en.wikipedia.org/wiki/Transformer_(deep_learning_architecture)

• ONNX — models are exported from PyTorch to ONNX format for cross-platform inference.

  • https://onnx.ai/

• Temperature sampling — controls randomness during generation.

  • https://en.wikipedia.org/wiki/Softmax_function#Temperature

• Top-K sampling — restricts sampling to the K most likely tokens.

  • https://huggingface.co/blog/how-to-generate

Music representation

• Roman numeral tokens — functional harmony representation (e.g. I, V6, ii°, etc.).

  • https://en.wikipedia.org/wiki/Roman_numeral_analysis

• Quantized harmonic rhythm — extracting chords on strong beats (e.g. quarter‑note grid) to suppress passing-tone “chatter”.

Plugin (Real‑Time)

Languages & standards

• C++17 — core implementation language.
  • https://isocpp.org/

Frameworks

• JUCE — plugin framework (UI, audio/MIDI plumbing, file export).
  • https://juce.com/

Plugin formats / DAW integration

• Audio Units (AU)

  • https://developer.apple.com/documentation/audiounit

• VST3

  • https://steinbergmedia.github.io/vst3_doc/

• CLAP

  • https://cleveraudio.org/

Inference runtime

• ONNX Runtime — the plugin loads trained models exported to ONNX format and runs inference via the ONNX Runtime C++ API. This replaced a custom LSTM implementation in v1.3.0.

  • https://onnxruntime.ai/

• Softmax + sampling — generation uses softmax probabilities, temperature scaling, and top-K style filtering.

  • https://en.wikipedia.org/wiki/Softmax_function

Voicing / musical constraints

• Beam search / Viterbi-style path search — used to select voiced 4‑part realizations over time.

  • https://en.wikipedia.org/wiki/Beam_search
  • https://en.wikipedia.org/wiki/Viterbi_algorithm

• Constraint-based voice leading heuristics — avoid parallels, encourage stepwise motion, reward common tones, etc.

MIDI

• MIDI file generation — exports single-track and multi-track MIDI, with time signature meta events.

  • https://www.midi.org/specifications

• Drag-and-drop MIDI UX — DAW-friendly workflow to get generated harmonies into your project quickly.

Website (lucidmusician.com)

Framework

• Astro — static site generator. All pages are pre-rendered HTML at build time with zero client-side JavaScript by default.
  • https://astro.build/

Styling

• Tailwind CSS 4

  • https://tailwindcss.com/

• Tailwind Typography (prose styling for markdown content)

  • https://github.com/tailwindlabs/tailwindcss-typography

Content

• Astro Content Collections — blog posts and docs pages are authored in Markdown with typed frontmatter schemas.

  • https://docs.astro.build/en/guides/content-collections/

• @astrojs/sitemap — auto-generated sitemap at build time.

  • https://docs.astro.build/en/guides/integrations-guide/sitemap/

Performance

• lite-youtube-embed — lightweight YouTube facade that loads the real player only on click.
  • https://github.com/nicoulaj/lite-youtube-embed

Infrastructure / deployment

IaC

• AWS CDK (TypeScript)
  • https://aws.amazon.com/cdk/

Hosting + CDN

• Amazon S3 (static assets)

  • https://aws.amazon.com/s3/

• Amazon CloudFront (CDN)

  • https://aws.amazon.com/cloudfront/

DNS / certificates

• Amazon Route 53

  • https://aws.amazon.com/route53/

• AWS Certificate Manager (ACM)

  • https://aws.amazon.com/certificate-manager/

Release Packaging (macOS AU)

For macOS distribution, we automate AU installer creation with a small Bash script:

• scripts/package_au_pkg.sh

At a high level it:

• Stages the built .component into a temporary directory (mktemp, ditto).
• Code-signs the component (hardened runtime optional) using codesign, then verifies the signature.
  • https://developer.apple.com/documentation/security/notarizing_macos_software_before_distribution
• Builds an installer package using Apple’s packaging tools:
  • pkgbuild (component package)
  • productbuild (final signed installer)
  • https://developer.apple.com/documentation/installerproducts
• Notarizes the resulting .pkg with xcrun notarytool submit --wait.
  • https://developer.apple.com/documentation/security/notarizing_macos_software_before_distribution/customizing_the_notarization_workflow
• Staples the notarization ticket to the .pkg with xcrun stapler staple.

Related reading

• Modeling Harmonies: From Scores of the Masters to Real-Time AI

• Harmonic Generators for DAWs: State of the Union