Soft Machinery: making room inside a programmed record
Soft Machinery is an eight-piece instrumental album by MATT BARGAR. It lasts 32 minutes and 52 seconds before any player overlap. The starting point was simple: a machine does not have to sound forceful. Small, rounded sounds can make a moving record without making everything feel urgent.
This is a separate October 12 release. Matt explicitly asked for another album while the female-vocal project waits for a singer. Nothing here substitutes for that performance. There is no hidden vocalist, cloned voice, session band or recorded acoustic ensemble. The companion lyrics are a separate written work.
The shape of the record
A Small Permission establishes sparse electric-key-like attacks, sustained tones and a slow bass pulse. Pocket Motor brings in a more syncopated figure and a small shuffle in the high percussion. Peach Glass opens the register: the upper voice becomes a bell-like sound above quicker broken chords. Sideways Sun displaces the wood-like percussion against the bar rather than simply making the previous track louder.
The Quiet Hinge is the centre. It removes the drum kit entirely, slows to 70 BPM and lets the bell and sustained voices carry the space. Everything Can Bend returns motion through a seven-step selection pattern laid over eight subdivisions per bar. Those patterns do not line up identically at every bar boundary. Open Assembly is the rhythmic crest, at 116 BPM and with more continuous upper phrases. Nothing Left Running returns to the opening tonal area and leaves a final major triad rather than the unresolved major seventh used elsewhere.
This is not eight unrelated random seeds. The pieces share three harmonic families and an interval collection, but change transposition, pulse, density, register and which instrument carries the foreground. Each has a sparse opening, a middle reduction and an ending. That shared architecture is also a limitation: the formal vocabulary is intentionally narrower than a band improvising through eight separate sessions.
What actually made the sound
The source is an original Python program using NumPy arrays. It writes a deterministic score for every piece: instrument, beat, pitch, duration and velocity. Track-specific seeds make the noise percussion reproducible. No external sample pack or commercial synthesizer was used. GarageBand, Ableton and LANDR were not used for this record.
The key voice is a sine oscillator with a decaying phase-modulation component and a quiet third harmonic. The modulation falls away faster than the body, so the attack is brighter than the sustained note. The bass is a fundamental plus a restrained second harmonic. The pad combines a fundamental, a slightly detuned partner and a second harmonic, with a slow onset.
The bell combines three partials at ratios 1, 2.76 and 5.4, each with its own decay. These are deliberately not a conventional harmonic series. The small wood-like hit is two quickly decaying sine components; it is not a recording of wood. The kick uses a descending-frequency sine. The high percussion is seeded noise with a short envelope and a simple first-difference emphasis.
Every event has a short attack and release envelope to avoid hard-edged note boundaries. Equal-power panning distributes the voices across stereo. Four quiet cross-channel delay taps, at 137, 293, 431 and 719 milliseconds, give the parts a common space. This is a small synthetic delay field, not a measured room response, a binaural capture or a physical studio recording.
Finishing and verification
The floating-point mixes were peak-normalized to approximately 0.82, about -1.72 dBFS sample peak, with gentle entrance and exit fades. This is a simple finishing pass, not a claim of expert attended mastering. Different densities still produce different perceived loudness. No loudness-normalization target or true-peak certification is claimed.
FFmpeg encoded original 24-bit, 44.1 kHz stereo PCM WAV masters and tagged FLAC release copies. For every track, the WAV and FLAC were decoded to the same signed 24-bit PCM representation and their SHA-256 hashes compared. All eight matched. The render also checked finite values and sample peaks below full scale. AAC listening files were encoded separately from the WAVs. The lossless files were never made from those AAC copies.
Numerical checks are not listening tests. No human listening session, speaker comparison, or artistic acceptance is asserted here. The source scores and verification report are retained so revisions can be traced to actual changes.
The cover and the missing seed
Matt's designated artwork library was searched recursively for image files, excluding videos and cloud-only placeholders. No locally accessible image remained. The cover therefore uses the authorized original-art fallback, not a purported transformation of Matt's work. No seed image was downloaded, overwritten or modified, and no signature was invented.
The built-in image tool was asked for interlocking translucent peach and smoky blue glass forms, amber light and midnight teal space. The rounded sculpture echoes the record's overlapping attacks and decays. The exact title and artist name were included in the prompt and visually checked in the returned image. The prompt and seed limitation are preserved in the provenance record. Matt subsequently chose the cover billing MATT BARGAR + ChatGPT. The final cover is a text-edit variant of the initial image; earlier versions are preserved.
Cost and credit
No new purchases, commissions, subscriptions or paid mastering were initiated. Direct new spending for this run is $0; existing subscription costs have not been allocated per album. Creative direction is credited to Matt Bargar. Created with ChatGPT credits the actual composition, arrangement, synthesis programming, text, artwork development and production-support contributions.
What transfers to a procedural game
The event score is more useful than a stereo loop when a game needs variation. Tempo, harmonic family, density, instrument role and seed are separate controls. An AI director could lower density before changing harmony, withhold percussion for a quiet encounter, or bring back a recognisable interval at a later location. The seven-step pattern shows how a small rule can avoid bar-for-bar repetition without abandoning a stable pulse.
The reusable material here is the method and saved event data. This release does not yet include a tested adaptive game-audio engine, runtime stems or seamless loop exports. Those would need dedicated timing and transition tests. The final stereo tracks have fades and finite endings; treating them as seamless loops without editing would be incorrect.