J-Master turns a raw track into a release-ready master. It was built for the output of AI music generators (SUNO, Udio, Riffusion, or local models like MusicGen and Stable Audio) but it masters any WAV from any source, DAW bounces included. Load it, and the console tells you the tempo, the song's sections, where it runs hot, and what's wrong with it. Shape it with eight macro processors, or press one button and watch the analysis do it, with every decision shown. Export at an exactly-solved loudness for any streaming platform, tagged, in four formats, or assemble a whole album into a replication-ready CD image. Everything runs locally. Nothing leaves your machine.

Built by JMW Software for Jamware Records. Music, manufactured.

---

Eight macro processors, tuned so that every knob does something musical and none of them can ruin your track:

Macro	What it does
TONE	Warm ↔ bright spectral tilt, ±4.5 dB around 700 Hz
SHAPE	Scooped ↔ forward mid contour (presence vs. wall-of-sound)
AIR	13 kHz shelf, up to +6 dB of sheen
SMOOTH	Dynamic high-band tamer that cuts harsh AI-generation zing only when it appears; transparent otherwise
CHARACTER	4× oversampled tanh saturation with even-harmonic bias for tape/tube colour at unity small-signal gain
DENSITY	Slow glue compression with auto-makeup: thickness, not pumping
IMPACT	Transient contour: soften ↔ punch, ±6 dB on attacks
WIDTH	Stem-aware tilted M/S: bass anchored below 140 Hz, mids widened gently, highs opened most

Plus BALANCE with one-click AUTO-CENTER, four STEM LANES (bass / drums / vocal / air component trims at ±3 dB of phase-coherent DSP extraction, with the lane interface ready for a neural separator), a 6-band advanced parametric EQ drawer with draggable handles over the live spectrum, loudness-matched REF bypass (tap R; hold it for a momentary compare), LIM Δ (hear only what the limiter removes), a MON monitor matrix (mono fold-down, side solo, or a single channel, metered exactly as heard and never exported), full undo/redo, and A/B snapshot slots. Click any knob's readout to type an exact value, press ? for the complete keyboard map, and the whole console re-flows cleanly down to a 720 px window instead of clipping.

---

Every load runs a full analysis pass:

• Tempo detection: spectral-flux onsets → autocorrelation, phase-locked to kick and bass. Drives a bar/beat GRID overlay and a CLICK metronome (mixed in after metering; never exported).
• Section detection: checkerboard novelty over band energies finds INTRO / PEAK / OUTRO boundaries, snapped to bars, which also anchors bar 1 to the music.
• Track diagnosis: a check sheet tuned to AI-music pathologies: bass smeared into the sides, phasey width (windowed correlation), image lean, harsh highs. Found issues open with measured values and pre-checked one-click fixes. Nothing is ever applied silently unless you arm AUTO-FIX ON LOAD.
• Dynamics report: PLR, EBU R128 loudness range, and a delivery table showing exactly what Spotify / Apple / YouTube / Tidal normalization will do at your current target.
• AUTO →: the one-button master. Picks a genre preset from tempo and spectral signature, applies the diagnosed fixes, sets a target, and shows its complete reasoning on a report card, with one button to undo the lot.
• MATCH: load a reference track you trust; its tonal balance, loudness and width are measured against your source and a capped ±6 dB correction curve is applied in one click.

The waveform is a workspace: wheel-zoom to transient level, drag fade handles with four curve shapes, and flip to a log-frequency spectrogram that shares the same zoom, grid, playhead and loudness lane. SPLIT stacks the source above the processed master on one timeline, and LOOP (the L key, or a double-click) cycles the detected section under the cursor while you work the console:

---

The export pipeline solves loudness exactly: measure → gain → true-peak limit → verify, iterating until the master lands within 0.15 LU of target with the ceiling held in dBTP.

Format	Engine	Notes
WAV	in-house	48 kHz, 24/16-bit, TPDF dither, RIFF INFO tags
FLAC	in-house (RFC 9639)	per-frame stereo decorrelation, LPC to order 12, partitioned Rice; verified bit-exact lossless
MP3	lamejs (LGPL)	320/256/192 CBR, ID3v2.3 tags
Ogg Opus	WebCodecs + in-house RFC 7845 muxer	256/192/128 kbps, OpusTags metadata, zero dependencies

• Release metadata on every export: title, artist, album, year, genre, catalog number, written in the native tag format per container.
• Cover art: click or drop an image (auto-scaled to a 1000 px JPEG) and it embeds natively everywhere: an ID3v2.3 APIC frame, a FLAC PICTURE block, Ogg Opus METADATA_BLOCK_PICTURE (the muxer spans pages when the art demands it), and a WAV id3 chunk. Covers travel with .jmaster projects and batch exports.
• Codec audition: loop the loudest section and A/B the actual encoded result against the lossless master before you commit.
• Batch albums: queue tracks, reorder, per-track preset overrides and per-track diagnosed fixes, auto-numbered tags.
• Album / CD assembly: a replication-ready 44.1 kHz/16-bit image WAV (frame-aligned tracks, configurable gaps, streamed to disk) with a CUE sheet carrying CD-TEXT, per-track ISRCs and album UPC, plus a manifest.
• Projects: save the whole session as a .jmaster file; double-click to reopen exactly where you left off.

Two themes, both from the Jamware Records design system: PLATE (the mastering floor at night) and PAPER (the spec sheet):

---

The core guarantee is WYSIWYG audio: the real-time preview worklet and the offline export renderer run the same TypeScript DSP code, and preview gain is auto-calibrated against the loudest section of the track, so what you hear is what renders.

flowchart LR
    subgraph UI["React UI · zustand"]
        C[Console] --> S[Store]
    end
    subgraph RT["AudioWorklet (real-time)"]
        W["MasterChain\n+ BS.1770 meters"]
    end
    subgraph OFF["Workers (offline)"]
        R["MasterChain\n+ loudness solve\n+ encoders"]
        A["Analysis:\ntempo · sections\ndiagnosis · profile"]
    end
    S -->|params| W
    S -->|params| R
    A -->|"BPM · fixes · LUFS"| S
    W -->|meters| S
    R -->|"WAV / FLAC / MP3 / Opus"| D[(Disk)]

Loading

src/audio/dsp/ is the shared engine: RBJ biquads, a 4× oversampled saturator, glue compressor, transient shaper, stem lanes, tilted M/S width, dynamic de-harsh, lookahead true-peak limiter, and a full BS.1770-4 implementation (K-weighting, gated integration, LRA, true peak). The FLAC encoder and the Ogg Opus muxer are written from scratch in this repo: no WASM blobs, no native modules.

Deeper details live in docs/ARCHITECTURE.md.

---

Every DSP claim in this project was verified by measurement during development, using a scriptable hook (window.__jmaster) that drives the real app:

Claim	Measured
Export hits loudness target	−11.58 LUFS on a −11.5 target
FLAC is lossless	0 errors across full decode roundtrips (incl. LPC + M/S frames)
True-peak ceiling held	−0.98 dBTP against a −1.0 ceiling
BPM detection	100.24 detected on a 100.00 ground-truth track
Opus loudness through codec	−11.63 LUFS decoded vs −11.5 target
CD image frame alignment	track 2 INDEX at exactly 00:34:00 for 32 s + 2 s gap
Balance correction	+2.03 dB measured on a +2.02 dB expected shift
MON matrix is honest	correlation 1.00 in MONO; side solo −4.7 dB on a centred source
Cover art survives every container	APIC / PICTURE / METADATA_BLOCK_PICTURE parsed back from real renders; 657 KB art spans 13 Ogg pages and still decodes
Idle repaint cost	3 s paused = 1 waveform draw + 0 spectrum draws (was ~180 / ~540)

---

npm install
npm run dev        # renderer in a browser (http://localhost:5183)
npm run dev:app    # full Electron app against the Vite dev server
npm run dist       # Windows x64 installer + portable exe → release/
npm run dist:arm64 # Windows ARM64 builds (no native modules, same JS)

Requires Node 20+. The renderer also runs standalone in a Chromium browser (file input + download fallback), which is how the automated verification drives it. Screenshot and banner assets regenerate with node scripts/make-banners.mjs and node scripts/capture-screens.mjs.

Windows packaging note: if electron-builder fails with EPERM renaming win-unpacked.tmp, something (usually an editor's file watcher) is holding directory handles in the project tree. Build to an outside folder: npx electron-builder --win --x64 "-c.directories.output=%TEMP%\jm-release"

---

GPL-3.0: see LICENSE.

Copyright © 2026 JMW Software and Jamware Records. J-Master is genuinely open source; copyleft keeps derivatives open, while all copyright in the software, its design language, and the J-Master / JMW Software / Jamware Records names and marks remains with the companies. The GPL covers the code; it does not grant rights to the brand.

Third-party notice: MP3 encoding uses @breezystack/lamejs (LGPL-3.0). All other DSP, the FLAC encoder, and the Ogg Opus muxer are original to this repository. UI set in the Jamware Records design language (Archivo + IBM Plex Mono, bundled under the Open Font License).

_{JMW SOFTWARE · JAMWARE RECORDS · MUSIC, MANUFACTURED.}