Audio goes in. Measurements come out, in language rather than meters. Nothing about your music leaves the machine, and nothing is presented as more certain than the analysis allows.
| A finished track | Drop a file. You get whole-track measurements, structure, chords, melody and lyrics. |
|---|---|
| Stems you exported | Drop a folder from your DAW. Every part is measured separately — the most detailed read WavRead can produce. |
| Straight from your DAW | The Bridge plugin records what passes through it. No exporting at all. |
| A mix, separated | WavRead can estimate parts from a finished mix. Those stems are guesses and the document says so — real parts always beat estimated ones. |
These are three different kinds of claim and WavRead labels every number with which one it is. That labelling is the product, as much as the numbers are.
Arithmetic on the samples: loudness, true peak, crest factor, clipping, spectrum, stereo width. Two runs agree exactly.
Found by an algorithm with a defined method: tempo, section boundaries, onsets. Reliable, occasionally wrong at half or double the tempo.
Inferred by a model: key, chords, melody, lyrics, separated stems. A good first draft, carrying a confidence where one exists.
And a fourth state that matters just as much: unavailable. A dash never means zero. If a window was too short, a take had no quiet gap, or an analysis predates a measurement, the document says which — a number that isn't there beats one that's wrong.
| Integrated loudness | ITU-R BS.1770, in LUFS. |
|---|---|
| Loudness range | EBU Tech 3342, in LU. |
| True peak | BS.1770 Annex 2 — 4× oversampled, so inter-sample peaks a sample meter misses are caught. Reported in dBTP and marked as the estimate the standard says it is. |
| Sample peak and clipping | Measured across the real channels. A mono fold-down understates a peak living in one channel by up to 6 dB — a bug WavRead shipped once, and now has a regression test against. |
| Spectrum | 1/3-octave bands, grouped into named regions and reported as a share of measured energy. |
One number for a whole song hides the song. WavRead measures loudness, stereo width and mono compatibility across the track, so it can tell you the second chorus is quieter than the first, or that the intro collapses in mono while the rest survives it. Section comparison answers the smaller question: how this part differs from that one, in numbers.
WavRead has no processors, and that is a deliberate boundary rather than a missing feature. It will show you a suggested EQ curve and let you hear it level-matched against the original — so the louder one doesn't simply sound better — but the file it writes is a document, never a new mix. Your DAW does the editing.
Assistants can't hear audio, but they read Markdown well. The Track Document holds everything one would need to talk about your mix specifically rather than generically. That is optional — the built-in Feedback Report is written by rules, not by a model, and works with the network cable out.
WavRead is a commercial product standing on open research. Each component
keeps its own licence, reproduced in full in the
THIRD_PARTY_LICENSES file that ships with the app.
| Project | Used for |
|---|---|
| Demucs (Meta AI, MIT) | Neural stem separation |
| Whisper (OpenAI, MIT) via faster-whisper | Lyric transcription from the isolated vocal stem |
| Basic Pitch (Spotify, Apache-2.0) | Melody transcription to notes |
| librosa (ISC) | Core audio analysis |
| pyloudnorm (MIT) | Broadcast-standard LUFS measurement |
| JUCE (AGPLv3 / commercial) | The AU and VST3 capture plugin |
Every analysis records which of these produced it, and at which version, in the document's provenance line. A measurement is only reproducible if you know what made it.