Projects 01
A professional, explainable audio analysis and intelligent mastering engine, written in modern Python.
What WPAudio Engine is
Every decision WPAudio Engine makes answers three questions — what, why, and how. Not as a comment in the source, and not as a paragraph in a manual. Each answer carries the metric IDs it used, the thresholds it compared against, the confidence it had, and what it chose to protect at the cost of something else.
Mastering tools that work well tend to be either a cloud service you upload to and hope, or a chain of processors that will not tell you what they did. WPAudio Engine is an argument that an engine can make good decisions and account for them.
01
Five stages, with strict boundaries between them. Three are built, one is partly built, and one is an empty package with a name.
01Built
Immutable buffer, loader
02Built
Five analyzers, 39 metrics
03Not yet built
Milestone M2 — empty package
04Partially built
Reference mastering only
05Built
HTML, JSON, CSV
One line through five stages. A built stage is a filled node joined by a solid segment; the partly built stage is a ring joined by a long dash; the planned stage is a quiet ring joined by a short dash. Each stage says its state in words above its name.
Those are three of twelve rules the project holds itself to. The one that matters most for explainability is blunter: the “why” is not a comment — it is a test. A rule that cannot be asserted against a table of inputs and expected decisions does not ship.
02
Ruff and strict MyPy clean. Python 3.12+, on numpy, soundfile and pyloudnorm — three runtime dependencies, deliberately.
A note on names: WPAudio Engine is the product. aurora is the code name and the Python namespace it was born in — the package is aurora-dsp and the command is aurora. WPAudio Engine is what it is called; import aurora is what the code says.
03
Not a mockup — the engine running locally, mid-session, on a track from Wings & Prayers. It analysed the file, mastered it to a stated target, then re-opened the written file from disk and checked its own work with a second, independent meter.
04
Album-level analysis exists because another Studio project needed it. The twelve-track Wings & Prayers album was WPAudio Engine’s first real production workload — thirty-three minutes of finished masters, analysed end to end in seventy-nine seconds. Those reference masters have since shipped: the album was released on streaming platforms on 11 August 2026, byte for byte from the files this engine verified.
Running a real record through it surfaced things a synthetic test signal never would. The clearest: silence was reporting an integrated loudness of 0.0 LUFS — maximal loudness, the worst possible sentinel a loudness metric could pick. It now reports negative infinity, and the JSON exporter serialises non-finite values as null rather than emitting invalid JSON.
One project’s engine, mastering another project’s music. That is the whole studio thesis inside a single working directory.
The full story — every pre-master measured, every reference master verified, and the actual audio to A/B for yourself — is now a case study of its own.
Case studyMastering Wings & Prayers — measure, decide, verify, listenRead it, then hear itA month later the same engine mastered the same songs again — the Second Flight record — to a different brief: one platform-reference level for every track, and every figure cross-checked by a meter from outside the project before it was published.
Case studyMastering Second Flight — one target, every number measured twiceRead the measurements05
Reserved, with URLs held. Each stays empty until there is something real behind it.
The signal chain and module contracts, drawn out.
Eight exist in the repository. They belong here.
A harness exists; no results are published yet.
Read from the repository, not maintained twice.
Getting started, then reference, as the API settles.
A later chapter. Not a product, and not a waitlist.
06
DropTrack's Music Analyzer scored Wings & Prayers: Second Flight between 94 and 97 out of 100. The full results, what the tool says the number is, and the line between a proprietary score and a measurement someone else can repeat.
The mastering case study forced a decision about evidence — quote the private engineering reports, or measure the published files. Choosing the second changed what the page could honestly say, and produced ADR 0007.