Keep the
unexpected.
Borrow a sound. Find what falls away.
Make a little world out of it.
Your selected layer, shaped into a texture.
Start somewhere
New audio keeps Weather, input shaping, rhythm, repair and Texture settings by default. Reset clears input and repair choices but keeps your Texture effects. A chosen DSP passage follows the same relative region in the new clip.
Weather is optional. At zero it is a bit-exact bypass of the imported, resampled audio.
Shape the input without Weather
Two serial stages, in the chosen order. Chorus controls above and phaser controls here are independent. Level matching happens once, after both.
One gain for both channels; capped at +12 dB and available peak headroom. This changes the level sent to the learned repair, which can change its split.
Make a rhythm
16 repeats your first eight steps as a starting point. Returning to 8 keeps the first half. Step length below sets the note timing.
Tap a step: full → half → closed. A rhythmic volume envelope, not a noise detector.
Opening, closing & depth
Opening and closing are fractions of a step, with at least 1 ms smoothing. The pattern starts fresh at the beginning of each rendered loop.
Before-repair rhythm applies when you prepare the split.
Find the interesting part
The blend changes which details go into the change layer. “Difference” is learned change minus DSP change; it does not follow the blend.
Explore a layer pair Give the body and the detail different jobs
Choose two layers, then listen to A, the pair, or B through the same Texture effects. These edits leave the repair bank and the comparison channels above unchanged.
Layer A
Layer B
Guide Δ is silent until guidance makes a change. Mid/side uses the existing stereo image: mono B has no side signal. No layer is automatically normalized; signed layers can cancel. The level controls are creative, not a fair repair comparison.
Design the DSP removal Change the decision, then hear the disagreement
These settings change the DSP spectral mask itself. Rebuild DSP keeps the current Input and learned repair exactly as they are. Guide the split below acts afterward.
A passage may contain music: selecting it asks the DSP to suppress spectral material relative to that reference. This is not instrument separation. The reference comes from the currently prepared Input. Analysis windows extend around their centers, so passage boundaries are approximate.
Sensitivity changes which spectral cells qualify. Maximum removal sets a spectral gain floor, not a guarantee about overall loudness. It is not an output-volume control.
How the mask moves
Narrow detail can sound chattery; broader smoothing joins nearby frequency decisions. Each FFT bin is about 23 Hz. Editable modes apply one mask to both stereo channels; Classic uses its original independent channel masks.
Classic is the current DSP endpoint.
The plot shows the applied DSP mask before optional guidance, averaged into display cells. Changing the DSP endpoint can change shared peak headroom. Compare channels within the same prepared recipe.
Guide the split Explore what falls away
The neural model stays fixed. These controls use measured features to weight its existing change layer, then return the rest to Kept. Nothing is trained on your clip.
ALLOW INTO CHANGE
Soft, overlapping regions around 300 Hz and 3 kHz. Lower values return more of that material to Kept. These are filters, not instrument labels.
Attacks are fast rises in measured energy. Quietness is relative to this clip. Neither identifies music or noise. Both channels follow the same envelope. Attack protection follows rises within the clip; an attack already present at its start may not register as a rise.
Before guide is the raw change at your current DSP/learned blend. Guide Δ = guided Change − Before guide. Use Stable source reference or One linked gain for this comparison; peak-each mode changes their relative levels.
What information does the model actually see?
Patina’s denoising autoencoder receives 16 nearby log-scaled frequency magnitudes from each of three adjacent time frames: 48 values. It predicts the middle frame’s magnitudes. We limit that prediction to attenuation and retain the input phase.
The controls above act after that learned prediction. Atlas inspired the measurement approach; its Python analyzer, speech recognition and text–audio model do not run on this page. Adding new model inputs would require a separate training experiment.
Give it somewhere to go
Pitch applies to the grains when Cloud is above zero.
Shape the grains Keep the octave. Change the way it moves.
Starting points change only grains, pitch and Cloud mix. Your chosen layer, echo and repair stay put.
Size is each fragment’s duration. Density is how often fragments start. Long, frequent grains overlap; short, sparse ones leave gaps.
Near zero, grains follow the loop’s position. Higher values borrow more widely from across the clip.
Direction, width & flutter
A chance per grain, separate from reversing the entire feed above.
Per-grain left/right balance. At zero, the source’s existing stereo image remains.
Smooth pitch motion within the grains; 100 cents is a semitone. Speed locks to complete cycles in this clip, with the actual rate below. Both channels share the same motion.
The same settings and arrangement repeat on each loop and are kept in project backups. Raising pitch is a texture effect, not transparent pitch shifting; bright material can acquire extra grit.
Give it space Distinct echoes → diffuse clouds
Grains → Double → Space → after-Texture rhythm. These affect Texture; the repair layers stay fixed. With ambience enabled, release a slider to render the new clip; playback then restarts.
Space starting points preserve your feed, grains and Double settings. Select Texture to hear the result.
Wait before sound enters Space; the dry sound stays in place.
Echo time sets repeat spacing; identical left/right input remains centered in Echo mode. Cloud scale spreads several delay lengths around that time. Decay is a nominal target; tone and diffusion change the actual tail.
Tone, movement & ducking
Ducking turns down the wet output while material arrives, then lets it return. Both channels follow one detector. Damping darkens successive repeats; movement gently shifts the readout timing.
Cloud starts with separate delay lines and builds toward the selected diffusion, measured from the start of this clip. Zero starts fully diffused.
The complete tail is included in Keep, project backups and WAV. Reusing a result longer than 12 seconds opens a region chooser. This is a rendered clip with a tail, not a continuously running reverb.
A little stereo double Small detunes, separate from the octave
Mono is a listening check only. Saved sounds and WAV exports retain their stereo channels.
A little more control
Stable reference keeps comparison levels fixed while you edit Texture or blend. Changing guidance recalculates common peak headroom. Texture alone gets a downward guard if needed. Linked mode includes Texture in the common gain; individual peak mode is for preparing sounds, not comparing removal.
Your discoveries
0Keep saves on this browser. WAV and project backup save portable files.
Preparing a file is not proof it was downloaded. If nothing appears, try the link or sharing sheet. Opening another browser does not transfer this unsaved sound.
Keep stores the source, both repairs, settings and exact selected sound on this browser. A project backup lets you reopen elsewhere. Clearing website data removes local keeps.
History keeps recent ancestor recipes, not earlier audio. Back up each version you want to revisit.
What is happening here?
Weather is ordinary DSP. The learned endpoint is Patina’s actual trained spectral denoiser; it predicts attenuation, not new music. Learned repair runs at strength 0.65. Classic DSP uses its original 0.65 setting; the editable DSP methods have separate detection and depth controls. Their removed parts can contain music as well as noise.
The blend is an ordinary crossfade between repair outputs, not a latent map. Reverse, echo, grains, Double and Space are DSP. Input EQ, chorus, phaser and rhythmic gating are optional DSP; they can precede the repair. This demo uses 24 kHz audio and up to 12 seconds of source, plus up to 8 seconds of optional Texture tail; it is not a live microphone effect or a restoration-quality claim.
Pocket 01–06 projects still open with their frozen sound. Subsequent edits use the current controls. Projects preserve frozen audio as well as settings. Reopening does not need to run the model. The page itself needs a connection to open; this is not yet an offline installed app. Your sounds stay in this browser unless you export or share a file.
iPhone playback, downloads and interruptions still need your real-device check. If an embedded browser blocks a download, open this link in Safari and use Share / Save to Files.