Getting the best results

This isn't a full manual — every control in the app is labeled, and most have a tooltip. This page is about the decisions that matter more than any one slider: when to reach for what, and why.

Start with the Analyzer

When you load a track, Latent's Analyzer looks at it and surfaces specific things worth your attention — harshness in a particular frequency range, a hard cutoff up top, phase issues between channels. These are suggestions, not a checklist to click through blindly. Read what each one says before applying it; they're tuned to what's actually measurably present in your file, not generic advice.

Why AI-generated tracks need different mastering

Generic mastering tools apply a chain that works for normal source material. AI-generated audio has known failure modes that normal material doesn't:

HF "swirl" / metallic sheen — a 14–18 kHz neural-codec shimmer that sits on top of the track. The de-harsh module has a "Sheen mode" tuned for exactly this.
Brickwall cutoffs — generators like Suno and Udio output at 32 kHz internally and upsample, leaving a hard cliff up top. Latent detects it, badges it, and synthesizes air above the cutoff.
Stereo anomalies — near-mono width, decorrelated bass, unstable side-channel phase. M/S correlation detection catches these; an elliptical bass-mono crossover fixes the low end.
Clipping / flat-topped waveforms — inherited from compressed training data. De-clip rebuilds the peaks with spline interpolation.

The "AI Fix" preset is the flagship one-click action: it turns on repair, de-harsh, glue comp, and stereo transient-expand in one tap. A user who touches nothing still gets a clean, loud, streaming-compliant master.

When to reach for Repair

De-click and de-clip catch real, specific damage — clicks/pops and clipped peaks, the kind of thing that shows up in lossy or poorly-exported AI-model output. De-noise is a heavier tool, meant for genuinely noisy source material (background hiss, artifacts from a rough export), not general polish. It's off by default for a reason — leave it off unless your track actually has a noise problem.

Reference A/B + EQ Match — and when not to use it

Load a second track as a reference, and Latent loudness-matches it so you can A/B fairly against your own mix. EQ Match then nudges your track's tone toward the reference's spectral shape. This is great for "make my track sit next to this other track I like" — it is not a fix for a fundamentally different-sounding arrangement or mix underneath. EQ Match reshapes tone; it doesn't rebalance instruments or fix structural problems.

Codec Preview

Almost everywhere your track ends up — Spotify, YouTube, SoundCloud — re-encodes it with lossy compression before anyone hears it. Codec Preview lets you actually hear that result before you upload anywhere, by running your master through a real Opus or AAC encode/decode round trip right in your browser. If something sounds off after compression that didn't before, this is how you catch it while you can still fix it.

The signal chain, in order

Latent's chain is opinionated and fixed — fewer footguns, predictable results. Every module is a toggle first, a full control panel second. The order:

Input gain → DC removal → Repair → 5-band EQ → De-harsh → Glue comp → Stereo imaging → Exciter → Soft clip → True-peak limiter
Loudness normalization runs at export time only
Fades, trim, and TPDF dither are export-stage only

Default on: DC removal, low-cut 24 Hz, EQ (flat), gentle glue comp, true-peak limiter (−1 dBTP), loudness norm (−14 LUFS). Default off: Repair expander/denoise, de-harsh, stereo imaging, exciter, soft clip — but the Analyzer may suggest enabling them.

Export formats and options

WAV — 16-bit, 24-bit (default), or 32-bit float; native rate, 44.1 kHz, or 48 kHz
TPDF dither — applied automatically when reducing bit depth
Fades — linear, log, or S-curve, with adjustable in/out times
Head/tail silence trim — with padding so you don't clip the fade
Loudness targets — off, Spotify (−14), Apple Music (−16), YouTube (−14), Club (−9)
Loudness report — the export dialog shows achieved vs. target LUFS with a check or warning

FLAC and MP3 320 export are planned for a later phase.

Privacy: nothing leaves your browser

Latent runs entirely in your browser via WebAssembly. Your audio is decoded locally, processed locally, and only leaves your machine if you choose to export it. No silent uploads, no server-side processing, no per-minute metering. This is a real trust signal for a browser-based audio tool — and it's not a marketing claim, it's how the architecture works.

A short glossary

LUFS: Loudness, measured the way streaming platforms actually measure it — not the same as peak volume.
True Peak: The real peak level after accounting for inter-sample peaks that a simple sample-peak reading can miss — this is what causes distortion after lossy encoding if it's too high.
Correlation: How similar your left and right channels are. Near +1 means close to mono; near 0 or negative can mean phase problems that cause thin or hollow sound on mono playback systems.
LRA: Loudness Range — how much the loudness varies across the track. Low LRA is consistent; high LRA is dynamic.
PLR: Peak-to-Loudness Ratio — the gap between your loudest peaks and your average loudness. Higher PLR means more punch and headroom.
M/S: Mid/Side — a way of working with stereo as the center (Mid) and the stereo difference (Side). Useful for keeping bass mono or widening highs.