Thomas Wagensommerer

A com­pos­i­tion envir­on­ment that treats a folder of record­ings as a corpus. Every file is ana­lysed into frame-level descriptors (loud­ness, onset dens­ity, pitch and pitched­ness, spec­tral shape, chroma and key, stereo width, an MFCC-derived timbre space); pieces of arbit­rary length are then re-ordered accord­ing to expli­cit order­ing prin­ciples — sort­ing, glid­ing, con­trast, morph­ing, formal clus­ter­ing, Markov gram­mars learned from the sources, serial rows, modal and har­monic retun­ing, aud­ible sort­ing algorithms — and rendered as high-res­ol­u­tion audio with a full event log. A pic­ture add-on cuts frame-exact video to the res­ult­ing sound. The tool makes the order­ing itself the com­pos­i­tional mater­ial: struc­ture becomes aud­ible as pro­cess.

The per­form­ance branch of sed­i­ment. Up to eight lanes hold com­posed mater­ial that can be looped, stuttered, re-sorted and trans­formed in real time; the timbre map of each lane becomes a play­able sur­face on which ges­tures are drawn, recor­ded and recalled. Sets store audio and state, MIDI learn binds every con­trol. Designed to run along­side a DAW in con­cert.

A real­time effect unit apply­ing sed­i­ment's order­ing modes to live input. Incom­ing audio is writ­ten to memory and ana­lysed as it arrives; up to eight lanes re-order pieces of the imme­di­ate past in time and route them to arbit­rary dis­crete out­puts. Archi­tec­tur­ally it sep­ar­ates a com­piled, alloc­a­tion-free audio body from an ana­lys­ing ‘brain’ that sched­ules events ahead, which makes mul­tichan­nel oper­a­tion at low latency pos­sible.

Firm­ware that ports sed­i­ment · fx to the Elec­tro­s­mith Daisy Pod. Stereo line input is held in 87 s of memory; six­teen ‘moves’ (scan, loop, sorts by loud­ness, bright­ness, pitch, noise and flux, mirror, shadow, glide, con­trast, drift, onset, grid) re-order pieces of it. One knob spans the grain con­tinuum from 2 ms buzz to 2.5 s wash, the other con­trols the move. Plate: spec­tro­grams rendered off­line from the firm­ware's own engine on a radio record­ing.

Reads a video frame by frame and builds a soundtrack from a sample folder. Image descriptors — bright­ness, hue axes, motion and its dir­ec­tion, edge dens­ity, cut impulses — are routed through a mod­u­la­tion matrix to gran­u­lar para­met­ers: which sample sounds, where in it grains are read, pitch, grain size, scan speed, fil­ter­ing, space and stereo pos­i­tion. Pic­ture and sound are bound sample-accur­ately; regions can be re-rendered loc­ally.

A gen­er­at­ive mont­age system for sur­veil­lance foot­age. Every frame of every camera is described once — colour, light, motion and optical flow, pres­ence, com­pos­i­tion, CLIP semantics, detec­tions, faces, pose, depth and real clock time — and the mater­ial is then cut by sort­ing rather than by chance: 74 pre­sets in fam­il­ies such as per­cep­tual, con­tinu­ity, semantic, sur­veil­lance, people, faces, image, depth, time, struc­ture and audio. Con­tinu­ity con­straints, curve read­ings of every rank­ing and a use-once rule turn sort­ing into edit­ing; each render is repro­du­cible from its side­car (CSV, OTIO, JSON) and can be streamed as OSC for son­i­fic­a­tion.

Com­pan­ion to sort­cut (T06): sort­cut makes the cut, score­cut makes it sound. The cut video is decom­posed into scenes and cam­eras; each camera becomes a ‘sta­tion’ of an audio corpus, switched at the exact frame of the cut, while image fea­tures drive a spec­tral and spa­tial treat­ment of the sound. Thirty pre­sets in eight fam­il­ies; cut-aware smooth­ing guar­an­tees that every change lands on its frame. Output: high-res­ol­u­tion audio, the stream-copied pic­ture with new sound, and a cue list.

A stan­dalone tool for con­trolled image decay. Two chain­able engines: a flow engine in which meas­ured motion is reshaped through 45 para­met­ers and applied to a canvas that is never refreshed — repeat­able at any res­ol­u­tion — and a bit­stream engine that deletes I-frames and duplic­ates P-frames in real MPEG-4 data, leav­ing the damage to the decoder. Thirty-two pre­sets in five fam­il­ies (smear, glitch, fluid, dream, brutal); identical res­ults on CUDA and CPU back-ends.

A sweep­ing bar crosses the moving image and reveals, behind it, one of 35 ana­lyt­ical read­ings — ter­rain (slope, curvature, rug­ged­ness), pho­to­met­ric (glint, water index, haze), motion (flow, diver­gence, vor­ti­city, time-to-con­tact), detec­tion (land­mass, hori­zon, rec­ti­lin­ear anom­alies) and depth. The frame is thus never fully in one epi­stem­o­logy. Output: an image sequence, a pre­view film and a per-frame scan log.

Cap­tures every frame an applic­a­tion actu­ally presents — includ­ing irreg­u­lar, frame-gen­er­ated or lag­ging sources — with high-res­ol­u­tion timestamps, and con­forms them to an even timeline (hold, blend, optical flow or RIFE inter­pol­a­tion). A render mode lets Unity drive cap­ture frame by frame with syn­chron­ised audio over UDP, turn­ing real-time engines into determ­in­istic off­line ren­der­ers.

Lossless con­cat­en­a­tion of numbered video parts (name_partNN) by stream copy: groups by base name, sorts numer­ic­ally, veri­fies identical stream para­met­ers across parts and checks the merged dur­a­tion against the sum of the parts. Coun­ter­part to AVs­plit­ter.

Fully local trans­la­tion dub­bing for long record­ings. Speech is tran­scribed (Whis­per, lan­guage locked after detec­tion), trans­lated by a local lan­guage model that is told each line's dur­a­tion, syn­thes­ised, fitted into the ori­ginal timing by con­dens­a­tion and time-stretch­ing, con­ver­ted through a trained RVC voice model and mixed over the ori­ginal or a sep­ar­ated back­ground. Output: video with Eng­lish dub and ori­ginal track plus sub­titles.

Local sub­titling pipeline for record­ings of three to seven hours: chunked tran­scrip­tion with voice-activ­ity detec­tion and word timestamps, trans­la­tion through a local lan­guage model, typo­graphic styl­ing as ASS, and either soft muxing or hard­ware-encoded burn-in. Batch-cap­able; shares its listen­ing and trans­la­tion core with dub­trans.

A local text-to-speech envir­on­ment whose voice can be a model trained on arbit­rary, includ­ing non-vocal, audio. Typed text or any input file is syn­thes­ised and passed through a trained RVC model; ref­er­ence voices, express­ive tags, extreme slow-down and stereo-pre­serving con­ver­sion are avail­able, and train­ing jobs can be queued from the same inter­face.

A Blender add-on that fetches elev­a­tion and satel­lite imagery for any region on earth, builds a met­ric­ally cor­rect (un-warped Mer­cator) ter­rain mesh and derives normal, 16-bit height and ambi­ent-occlu­sion maps, expor­ted with dated tex­tures as FBX/OBJ for game engines. Plate: Groß­g­lock­ner, 20 × 20 km.

Fuses sev­eral image sets or videos through Depth Any­thing 3 (depth and camera pose) into an aver­aged land­scape height­field or a free-stand­ing sculp­ture, expor­ted as tex­tured FBX for Unity, and as per-frame point clouds for 4D sequences.

A pipeline for neural-gen­er­ated (GNM) heads. The rig add-on clas­si­fies regions from mater­ial slots, solves the ana­tom­ical frame from eye and teeth centroids and builds a repro­du­cible Root/Neck/Head/Jaw/Tongue/Eye rig with baked viseme and blink shape keys for Unity. The tex­ture kit bakes pos­i­tion and normal maps per UV region and syn­thes­ises seed-driven skin, eyes and mouth; the pro­jector con­forms any frontal (e.g. AI-gen­er­ated) face photo onto the head via land­mark align­ment and thin-plate-spline warp­ing.

A rig-agnostic rag­doll gen­er­ator for Unity: col­liders and joints are fitted from skin weights through the bind pose, so bipeds, quad­ru­peds, insects, tentacles and mech­an­ical rigs work without humanoid map­ping; unrigged meshes are chunked into bodies. A driver adds eight motion layers, 26 pre­sets and audio input.

An OSC con­trol sur­face for Res­ol­ume built for slow per­form­ance: long fades, few ges­tures, noth­ing cut­ting that should glide. The fade engine runs in the bridge rather than the browser, so a twelve-minute cross­fade keeps its exact curve even if the inter­face is closed; para­met­ers, soft trig­gers, clip grids and slow sequences are arranged as nodes, with cues and remote con­trol.