AI media · MusicAI music

One world, 76 tracks

An AI-generated soundtrack for the life-sim Heart Apartment — directed by ear, checked by numbers, shipped in two days.

Style anchor · home, night

夜裡的小爵士 Little Night Jazz

Finish on par with published music.

Challenge

A soundtrack has to belong to one world across many scenes. Technical checks alone cannot tell whether a track fits a quiet room or an argument.

My contribution

I wrote scene briefs, listened to every render, and chose each track's place in the game. AI agents built the production scripts; the audio model generated the music.

Evidence

Listen to the selected tracks, inspect the filter that failed its own test, and follow the checks used for game delivery.

Related capabilities Creative direction & experience design Verification & quality judgment Project delivery & cross-functional collaboration

76tracks shipped · 54 themes, 15 story cues, 7 ambience beds

244renders, each listened to, with the verdict stored next to the file

~1.5days from first planning note to game integration

0renders reached a listener without passing automatic checks

Overview

Fifty tracks that have to agree

A cozy life-sim needs hours of music that sounds like one world and stays out of the way. A model can make any single track. The hard part is fifty that agree — when the only quality test is one person saying “this feels wrong”.

Me: one-sentence brief per scene, every render heard, every verdict stored word for word
AI agents: batch runner, analysis and ranking scripts, loudness normaliser, listening page
Model: Stable Audio 3 Medium, local, in ComfyUI

Scene brief→Batch job→Auto-checks→Listen & comment→Ship

Rejected takes loop back as a rewritten description. Highlighted steps are human decisions.

Three calls that changed the product

No loops — full three-minute pieces; long sessions need variety, not seamless repeats
Three ambience beds cut — too noisy, and of little use in the game
Every use has a place — a bright Shibuya-kei take moved from title theme to “on the way out”

Highlights

Taste, numbers, and knowing which to trust

1 · “It sounds wrong” became a rule you can write down

My assessment	The number	The rule
Monotonous, and noisy overall	Harmonic change 0.10–0.11 vs 0.21–0.56 for liked tracks	Spell out the harmony: maj7 / m9 chords, slow progression
Arrangement too grand; feels oppressive	50–80 % of energy below 60 Hz vs 0–12 %	No drones, no sub-bass, no string pads
Too many instruments; sounds cluttered	Spectral centroid 492 Hz vs 120–250 Hz	Three to five instruments; space between phrases

Two heatmaps of note energy over 30 seconds. The rejected take shows long flat bands on one or two notes; the style anchor shifts between notes every few seconds. — Left: rejected as monotonous on listening (0.11) — the same notes for seconds at a time. Right: the style anchor (0.22) keeps moving.

What “harmonic change” measures

The track is split into short windows; each window's spectrum is folded into the 12 musical notes; the measure is how much that note profile shifts from one window to the next. A track stuck on one chord scores low; one that moves through a progression scores high.

2 · I tested my own filter, and it failed

Script said pass79 renders

48%

Script said cut162 renders

46%

Cut for extreme brightness11 renders

Share I liked · dashed line: all 241 renders, 47 %. Of the 244 verdicts, 3 were too mixed to count as liked or not, so 241 are shown. The 11 bright renders are part of the 162 the script would cut.

As a gate, the script was a coin toss — tracks it would cut were liked as often as tracks it passed
Most rejections were about mood and fit: a tone close to a phone ringing, or a mood suited to an argument scene, neither right for the setting. No spectrum threshold can see that
What changed: numbers diagnose a complaint; a person screens. The effort went into making listening cheap

3 · The process changed when the evidence said so

AI prompt expansion kept failing — it added vinyl crackle and hip-hop drums and turned “slightly uneasy” into horror. Prompts now name concrete instruments and a tempo in one sentence
30-second auditions hid problems — a quiet theme passed at 30 s; all four 3-minute versions failed when a double bass entered at 1:00. Approval moved to full length
Percussion wouldn't behave — generate as a drumless duo, or lower the drums afterwards with source separation

Read the detail

The built-in expander added vinyl crackle, shakers and hip-hop drums, and turned “slightly uneasy” into a horror-film atmosphere that did not fit the game. From round 7, prompts named concrete instruments and a tempo in one sentence and avoided mood words like ambient or cinematic; nearly every 30-second test in that round passed.

A quiet reading theme was rated well at 30 seconds. All four three-minute versions were then rejected because a double bass came in at the one-minute mark. Quiet scenes became solo piano, and approval moved to full length.

Clap rolls and over-dense snaps could not be controlled. Two fixes: generate as a melodic duo with no drums, or remove or lower the drums afterwards with source separation, leaving the other instruments untouched.

4 · Built to ship, not to demo

Every render verified before anyone listens — seed, prompt, settings, real length, format; 95 jobs logged
One loudness standard — −18 LUFS music, −27 ambience, true peak ≤ −1 dBTP, fixed gain only
Provenance recorded — each shipped file maps to its render, seed and job; Stability AI Community License
In the game — all 76 tracks as 192 kbps MP3 with a playlist file

Read the detail

A batch runner starts one ComfyUI job per process with a 600-second limit, then decodes each file and checks seed, prompt, settings, real length and format against the job file. All 95 jobs are logged with the prompt hash and peak GPU memory.

All 76 tracks were measured to ITU-R BS.1770-4 and brought to −18 LUFS for music and −27 LUFS for ambience, true peak at or below −1 dBTP. The fix is a fixed gain only — no compression or limiting — so the sound is unchanged; six quiet tracks were left slightly under target rather than pushed through a limiter. The meter was self-tested against the standard's reference signal first.

Listen

The anchor, a favourite, and one description three ways

New batches didn't start from a blank prompt. Each job reused the full generation settings of a track I had already approved — first the night-jazz take below, later a morning piece I made myself in ComfyUI. Only the text description changed.

Style anchor · home, night

夜裡的小爵士 Little Night Jazz

Finish on par with published music.

Home, daytime

晴天的木琴 Sunny Marimba

A very high level of finish.

收拾好的午後 The Afternoon, Tidied — one description, three kept takes

Four seeds, three kept, each heard differently. This is the process itself.

Take A

Usable.

Take B

Strong sense of story.

Take C

Well-built atmosphere.

Excerpts, 60–75 s, from the loudness-matched masters.

Where the 76 tracks play

Home, day	Morning · weekend brunch on the sofa · chores · afternoon reading · cooking and dinner together
Home, night	Night rotation · livelier chat time · quiet evening · winding down after work · reading side by side
Work	Two characters' workdays · wrapping up and overtime
Out and about	Neighbourhood · park bench and walks · commute home at dusk · on the way out
Shops	Supermarket · plaza and department store · lifestyle store · cinema lobby
Theme	Title · weekly recap
Story cues	Being remembered · light teasing · waiting for the boss's reply · exhausted pause · loss and parting · apology and making up · good news · end of the week
Ambience	Home · office · park pond · supermarket · rain

Status & limits

What exists, what's proven, what isn't

Built

Batch runner with per-render checks
Ranking script, drum removal, loudness normaliser
Listening page with feedback capture

Verified

76 tracks approved and integrated into the game (2026-09-15)
Loudness targets met — six quiet tracks left slightly under rather than limited
Ranking script tested against 241 verdicts: it doesn't work as a gate (2026-09-22)

Limits & next

The numbers explain complaints; they don't predict taste
The model can't follow bar-level structure — tracks were chosen, not composed
Separate music and ambience sliders: specified, not built