Each channel,
its own sonic profile.

Psychoacoustics is the study of how the human auditory system perceives sound — not just physically, but cognitively and emotionally. The ear and brain together form a remarkably sensitive analysis system, tuned by evolution to detect threat, presence, and meaning in acoustic environments. That sensitivity works for and against recorded audio.

Degraded audio triggers threat response. When background noise, distortion, or inconsistent levels force the brain to work harder to extract the signal, cognitive load increases and emotional access narrows. The listener is no longer inside the story — they’re fighting to hear it. This is not a subjective response. It is a measurable physiological effect.

The brain rates credibility through sound. Film cognition research consistently shows that viewers judge overall production quality — professionalism, trustworthiness, emotional authenticity — primarily through audio, not image. A pristine visual with compromised audio reads as amateur. The inverse rarely applies.

A cloud or local job is not “run the UNet.” It is an ordered pipeline. Each stage is logged. If repair moves a channel too far from the original, QC falls back to the untouched ISO. Originals stay in the packet.

RestoreUNet is one step. It is a 4-stage spectrogram U-Net (12.83 million parameters, 49 MB checkpoint). That is one denoise step, not the product. The shipped engine is ffmpeg + DeepFilterNet and/or RestoreUNet + loudness measurement (BS.1770). A custom edit is Lemonwater in-house, quoted from the results page — not a heavier software tier, and not a third-party plugin the customer has to own.

Nothing runs in the browser. lemonwater.io is checkout, upload, and status (Next.js on Vercel). Vercel does not do GPU inference. Cloud jobs are queued to a long-lived worker with a working directory — hours of multichannel BWF, not a serverless function. The privacy path is the downloadable app on the operator’s machine. Uploading for the cloud intro means files leave the set.

LUFS (Loudness Units relative to Full Scale) is the measurement standard for perceived loudness in broadcast, streaming, and film delivery. Unlike peak meters, which measure instantaneous amplitude, LUFS integrates loudness over time — reflecting how the human auditory system actually experiences volume.

Every delivery platform and submission specification uses a loudness target for a reason: when all content is normalized to the same perceived loudness, the dynamic range within your mix — the quiet moments and the loud ones — carries meaning instead of just reflecting how hard your recorder was driven.

True peak is a separate measurement — the highest instantaneous amplitude in a digital signal after sample-rate conversion. True peak limits prevent intersample clipping that would otherwise distort on playback systems that reconstruct audio between samples. Festival and broadcast specs typically require a true peak ceiling of –1dBTP or –2dBTP.

Premix can normalize the optional reference stereo to a loudness target you select. Festival names on the specs form are presets for LUFS and true peak only. They are not a representation that a festival, network, or union will accept the film. For custom delivery, you type the numbers.

The live path is boom, lav, and multichannel BWF: labels, a written report, optional assistive cleanup, mixer on the job. Music finishing, meditation and frequency work, and audiobook enhancement can run through the same repair stages — they are not coequal SKUs at checkout.

Therapeutic-specific finishing (binaural separation, Solfeggio tuning, somatic/brainspotting delivery) is on the roadmap, not in the automated offer. If that finish matters, say so in project notes and it is handled as a human review request.