AUDIO_STACK.md

pipelines/AUDIO_STACK.md

Audio Stack — Music & SFX Generation Landscape (2026-07-15)

Status: RESEARCH BRIEF — informs the audio/cinematic Translation Docs still pending per

docs/PIPELINE_REVIEW_MISSION.md Phase D, and is the evidence base for correcting

_source/01_Tier_1_Foundation/T1_Build_Pipeline_Contracts [ACTIVE v1.0].md §3 (AIVA), §10 (SFX), and

the audio-consumption half of §5 (Cinematic Scene Build) — which were locked May 3, 2026 on

vendor names (AIVA, ElevenLabs, generic "SFX") without live verification. Not canon, not a build

order — a decision input, same register as the seven sibling briefs in this folder (see

LOCAL_3D_ASSET_GEN.md for the pattern this follows).

VERIFIED vs INFERRED. Every load-bearing claim below was checked against the live web on

2026-07-15. VERIFIED = fetched directly from a primary source (vendor pricing page, ToS/EULA/LICENSE

file, official docs, GitHub repo/license file, Epic's own documentation). INFERRED = secondary/

aggregator source, community report, cross-referenced search synthesis, or my own reasoning applying

verified facts to this project — not independently confirmed against a primary document. Read the tag

before trusting a number. A prior same-day pass, Q3_2026_MODELS_REFRESH.md §4, already did a "light

pass" on voice/AIVA/Stable-Audio-Open/licensed-libraries and confirmed no vendor swap was needed; this

brief re-verifies those findings, resolves two items it left open (the Stable Audio 3.0 SFX-model

weights license, the AIVA API's actual existence/shape), and is the first pass to cover open-weights

local music generation, the MetaSounds/Quartz adaptive-authoring requirement, licensed orchestral/SFX

libraries, and creature-vocalization workflow at all — none of that was previously researched anywhere

in this repo.

The ruled constraints this research operates inside (do not re-litigate):

1. Shipped-asset path defaults to UNRESTRICTED (MIT-class) or ROYALTY-FREE-OWNED components;

territory-restricted, revenue-capped, or non-commercial tools are never load-bearing for shipped

content (previz/reference lanes may use anything) — Josh's licensing ruling.

2. Hardware is one RTX 5090 (32GB VRAM), 9950X3D, 128GB RAM, local-first where quality permits; paid

cloud only where it demonstrably wins.

3. Voice is already ruled and out of this brief's lane (Kokoro-82M bulk NPC + ElevenLabs hero behind a

clone-consent gate) — see the "Voice — not re-litigated" note closing §1 for the one update found.

---

0. VERDICT (top-line — see §1/§2 for full reasoning)

Music. The incumbent contract's AIVA pick is confirmed, not overturned — but with two real

corrections. First, AIVA's usable commercial license lives only on the Pro tier (`Copyright owned

by YOU`, €33/month billed annually) — the Free and Standard tiers both explicitly say "Copyright owned

by AIVA" on AIVA's own pricing page, so anything generated below Pro cannot ship inside the game at

all, only (per AIVA's ToS) monetize on YouTube/Twitch/TikTok/Instagram. Second, **AIVA has no

self-serve public API** — its own End User License Agreement states API access "is ruled by a separate

Licensing Agreement, to be negotiated and signed between the parties," with no published pricing or

endpoint docs found anywhere live. The existing contract's §3.4 orchestration-agent design ("submits to

the AIVA API") assumes a REST integration that may not exist as a self-serve product — this is a real

operational gap, not a licensing one (§4 Unknowns #1). For bulk regional/chapter theming at 79-node

scale, this research adds a genuinely new, previously-unresearched option: **ACE-Step 1.5 (MIT) and YuE

(Apache 2.0)** are real, fully unrestricted, locally-runnable-on-the-5090 open-weights music models with

instrumental modes, unlike Meta's MusicGen/AudioCraft (CC-BY-NC-4.0 weights — non-commercial,

disqualified) or Stability's Stable Audio Open/3.0 (Community License, free only under $1M annual

revenue — a standing cap-risk, same pattern already flagged for Stable Fast 3D in

LOCAL_3D_ASSET_GEN.md §1.5). Suno and Udio are not viable for the shipped path today: Suno's

paid tiers assign output rights but explicitly provide no indemnification against the labels'

still-active infringement suits (Suno settled with Warner only; UMG and Sony remain in litigation), and

Udio disabled all downloads in its walled-garden UMG deal — there is no way to extract a file to ship

even though its Pro tier nominally grants commercial rights.

SFX. The incumbent two-track design (procedural API + licensed library, §10.4) is confirmed sound

— this research fills in which vendors actually work. ElevenLabs' SFX endpoint is real, API-driven, and

improved (30s/generation, seamless looping, 48kHz WAV) — good for the standard track. A genuinely new

finding: MOSS-SoundEffect v2.0 (Apache 2.0, May 2026) is a real, fully unrestricted, 48kHz,

30-second-capable open text-to-SFX model that did not exist in any prior research pass here — a strong

local candidate to sit alongside or replace Stable Audio Open (which is capped, and by its own model

card is "better at generating sound effects... than music," i.e., it was always more of an SFX tool than

a music one). Licensed libraries (Sonniss's free GDC bundle, BOOM Library's perpetual buyouts, Soundly's

subscription) are all confirmed genuinely royalty-free for shipped commercial games, with one universal

catch: **every major SFX library license (Sonniss verified directly, EastWest by extension) explicitly

prohibits using the licensed audio to train AI/ML models** — a clean line for direct use, a hard stop if

this project's own generative-entity runtime layer ever wants to fine-tune on owned/licensed audio.

Creature vocalization (132-creature bestiary) is confirmed NOT a generative-AI task at all — none of

the text-to-SFX tools evidence creature-specific work; the real, industry-standard workflow is layered

animal/human source recordings run through dedicated pitch/formant creature-design plugins (Krotos

Dehumaniser 2 + Reformer Pro, $399 each, perpetual license), which maps directly onto the contract's

already-specified sound-designer-in-loop canonical track — this research names the actual tool, it

doesn't change the workflow design.

Adaptive authoring. UE 5.8 MetaSounds + Quartz require generated music to arrive as **separate,

tempo/bar-metadata-tagged stems** (bass/drums/melody/percussion/etc., each a loopable asset referenced by

a data asset carrying tempo and per-layer bar-occupancy, triggered at Quantization Boundaries by a Quartz

clock) — not a single mixed-down file. Of every tool evaluated, only AIVA natively exports stems (its

Pro-tier "Stems" download format, sourced from its own compositional instrument layers, not blind audio

source-separation). Every other generator (Suno, Udio, ACE-Step, YuE, HeartMuLa, Stable Audio) outputs a

single mixdown; Demucs (MIT, Meta) is the open, unrestricted bridge tool to retrofit stems from any

of them, though it is trained for pop-song separation (vocals/drums/bass/other) and is a poor fit for

orchestral section-splitting (strings/brass/woodwinds/percussion/choir) — a real quality gap, not just an

integration one, detailed in §1.6.

---

1. Music lane

Quick-reference table

ToolLicense classCommercial-ship viable?Stems?API/batch?VRAM (local)
AIVA (Pro tier)SUBSCRIPTION-OWNEDYes, Pro tier onlyYes, nativeEnterprise-negotiated only, no self-serveN/A (cloud)
Suno (Pro/Premier)SUBSCRIPTION-OWNEDLegally usable, high litigation risk, no indemnificationNoNo official API (early-access intake only as of July 2026)N/A (cloud)
UdioTERRITORY-OR-USE-CAPPEDNo — downloads disabled, walled gardenNoNoN/A (cloud)
ElevenLabs Music (Eleven Music v2)SUBSCRIPTION-OWNEDYesUnconfirmedVia ElevenLabs API familyN/A (cloud)
Meta MusicGen / AudioCraftNON-COMMERCIALNo (weights CC-BY-NC-4.0)NoLocal, but disqualified by license~8-16GB (INFERRED)
Stable Audio Open 1.0TERRITORY-OR-USE-CAPPEDYes, under $1M annual revenueNoLocal only~6GB (DiT step) / ~14.5GB (full pipeline incl. decode)
Stable Audio 3.0 (Small/Medium, open)TERRITORY-OR-USE-CAPPEDYes, under $1M annual revenue (same license family, precise cap not independently re-confirmed on this specific model page)UnconfirmedLocal onlyNot found (Small variants target on-device)
YuEUNRESTRICTEDYes (Apache 2.0, weights + output)No (single mixdown)Local only~24GB reported for full pipeline (INFERRED)
ACE-Step / ACE-Step 1.5UNRESTRICTEDYes (MIT)No (single mixdown)Local only<4GB (2B turbo) up to 24GB+ (XL, best quality)
HeartMuLa (oss-3B)UNRESTRICTEDYes (Apache 2.0)No (single mixdown)Local onlyNot found; smallest released tier (3B)
Spitfire SSO/BBC SO DiscoverROYALTY-FREE-OWNEDYesYes (real recorded stems, per-instrument)N/A (sample library)N/A
EastWest Composer Cloud+SUBSCRIPTION-OWNED (compositions persist after cancel; re-editing does not)YesYes (real recorded stems)N/A (sample library)N/A
Demucs (stem-separation bridge)UNRESTRICTEDYes (MIT)N/A — this IS the stem toolLocal, scriptableModest (INFERRED, not independently benchmarked here)

1.1 AIVA — VERIFIED unless noted

3 minutes, MP3+MIDI only, "Copyright owned by AIVA." Standard Annually — €11/month billed

annually (€132/year, "26% discount"), 15 downloads/month, tracks up to 5 minutes, MP3+MIDI only,

"Copyright owned by AIVA." Pro Annually — €33/month billed annually (€396/year, "33% discount"),

300 downloads/month, tracks up to 5:30, all formats including high-quality WAV, **"Copyright owned by

YOU."**

copyright on the pricing page itself. Cross-referencing AIVA's End User License Agreement (VERIFIED,

fetched aiva.ai/legal/1): three license types exist — Non-Commercial (Free tier), Limited Commercial

restricted to "a limited set of third-party websites: Youtube, Twitch, Tik Tok and Instagram" (reads

as the Standard tier — AIVA retains copyright but grants a narrow monetization license on those four

platforms specifically, which does not cover shipping music inside a purchased video game), and

Full Copyright ("all copyrights of the MIDI and/or Audio Composition to Licensee" — Pro tier). At

€396/year this is cheap relative to a composer budget, but it is a hard gate: nothing generated below

Pro can ship.

aiva.crisp.help/en/article/what-formats-can-i-download-the-music-in-1ppr6no/): downloadable formats

are "MP3 Audio," "Orchestrated MIDI," "Reduced MIDI," "16bit & 48kHz uncompressed WAV file," and

"Stems" — plus chord data as text. This is the one vendor across the entire music lane that

natively exports separated instrument-layer audio, because AIVA composes with actual instrument

tracks internally (confirmed via a second search-corroborated source: "The AI generates a full-fledged

track... separated into different instrument tracks, which you can edit or even export into a MIDI

file") rather than a single end-to-end diffusion mixdown. The helpdesk page does not state which tier

gates Stems specifically (INFERRED: almost certainly Pro-only, given Pro is the only tier offering

"all file formats").

assumption.** AIVA's own ToS (aiva.ai/legal/1, VERIFIED): "If the Licensee wishes to get access to

AIVA's APIs, this use case is ruled by a separate Licensing Agreement, to be negotiated and signed

between the parties." AIVA's own site (VERIFIED, fetched aiva.ai/) confirms "For Enterprises" exists

only as a contact-us section header with no published pricing, docs, or endpoint reference found

anywhere live. Separately, "Enterprise" is explicitly defined in the ToS as "a business with 3 or more

employees AND that generated more than $300k of revenues in the past year" — a solo developer under

that bar may not even qualify for the negotiated tier without a direct conversation. **T1_Build_

Pipeline_Contracts §3.4 currently describes an orchestration agent that "submits to the AIVA API" as

a routine, self-serve REST call — this assumption needs re-checking against what AIVA will actually

sell a solo dev before that agent is built** (§4 Unknowns #1). One important naming-collision flag,

same pattern as the Hunyuan3D-vs-Hy3 collision noted in LOCAL_3D_ASSET_GEN.md: docs.aivaconnect.ai

is an unrelated call-center product that happens to share the "AIVA" name — not the music

generator. Do not confuse the two when searching.

style models" / "Upload an audio or MIDI influence" feature exists (confirmed via pricing-page

mention), but the ToS carries a two-way asymmetry worth flagging: **the Licensee is explicitly

forbidden from using AIVA's own output "as part of a training dataset for any Machine Learning, Deep

Learning or statistical algorithm," while simultaneously "Licensee provides Licensor with a

worldwide, non-exclusive, commercial and transferrable license to train AIVA's AI systems on the

Uploaded Influence, in perpetuity."** If this project ever uploads its own composed reference themes

to seed AIVA's style model, AIVA retains a perpetual right to train on that upload — worth knowing

before uploading anything proprietary as a style seed.

1.2 Suno — VERIFIED unless noted

its right, title and interest in and to any Output... generated... during the term of your paid-tier

subscription" — but immediately hedged: "due to the nature of machine learning, Suno makes no

representation or warranty to you that any copyright will vest in any Output." No indemnification

— the obligation runs the other way: users must "defend, indemnify, and hold harmless" Suno against

claims arising from their use, including third-party infringement claims on the output itself. Free

tier is non-commercial only, attribution required.

scope details): Suno settled with Warner Music Group (November 2025). **UMG and Sony remain in

active litigation** — reported as a stalled impasse over licensing fees/equity as of April 2026. A

parallel GEMA (German collecting society) verdict in Munich was scheduled for July 31, 2026 — after

this research date, worth a calendar re-check. The core US fair-use question is now expected to slip

into 2027.

reported by Music Business Worldwide): Suno does not currently offer an official public API —

it is "exploring" one, taking early-access applications from "a curated group of partners," no

timeline disclosed. Every "Suno API" found in search results is an unofficial third-party wrapper

scraping/reselling access to Suno's own consumer product — a ToS-risk layer this project should not

build a pipeline around.

(March 26, 2026, adds "Custom Models" — fine-tuning on user-uploaded tracks, paralleling AIVA's style

models). One direct game-audio review tested Suno for a cinematic game-intro cue and found it "polished

and clear... but the overall vibe felt a bit off... coming across more like lighthearted cartoon

background music than an epic cinematic intro" — a real qualitative gap against AIVA, which the same

source describes as specializing "in cinematic, orchestral, and game-score composition."

the indemnification gap and the still-active UMG/Sony suits — reasonable for scratch/reference/temp-

track work only, consistent with Q3_2026_MODELS_REFRESH.md's prior finding.

1.3 Udio — VERIFIED

October 29, 2025 announcement and follow-up Reddit statement): following the UMG partnership, **Udio

disabled downloading of audio, video, and stems entirely.** A 48-hour exception window (Nov 3, 2025)

let users retrieve pre-deal songs under the old ToS; that window has long closed. Udio is now

operating as a "walled garden" for 2026 — users can create and stream inside Udio's own app but

cannot export, download, or distribute to any external platform. Sony and Warner remain in active

litigation against Udio (UMG alone settled).

license terms.** Even though Udio's Pro plan reportedly grants commercial rights on paper, there is

currently no mechanism to extract a file to import into UE5. A "fully licensed launch" with downloads

restored is reportedly expected later in 2026 (INFERRED, unconfirmed timeline) — worth a re-check at

benchmark-gate time, not a reason to plan around it today.

1.4 ElevenLabs Music (Eleven Music v2 / "ElevenMusic") — VERIFIED unless noted

already ruled in for hero voice**, which makes it worth flagging as a consolidation option even

though it doesn't replace AIVA's composer-in-loop design.

released April 1, 2026 as a consumer app; ElevenLabs states the model is "trained only on licensed

data and cleared for commercial use" (secured licensing deals with independent music-rights

organizations ahead of an August 2025 launch of the underlying model). Free tier: 7 songs/day. Pro:

$9.99/month, 500 tracks/month. Sits alongside the existing ElevenLabs API pricing tiers already

verified in Q3_2026_MODELS_REFRESH.md §4.1 (Free $0/10k credits through Business $990/6M credits).

vendor already integrated for voice, but nothing found here suggests it beats AIVA's purpose-built

orchestral/cinematic composer-in-loop workflow or its native stems export for the canonical theme

track. Worth a bench comparison against AIVA at benchmark-gate time, not a recommended swap today.

1.5 Meta MusicGen / AudioCraft — VERIFIED

Creative Commons Attribution-NonCommercial 4.0 International covers the model weights for both

MusicGen and AudioGen (the filename covers both). Code is MIT, but "this does not extend to the

pre-trained model weights themselves." Trained on ~20,000 hours "owned by Meta or licensed

specifically for this purpose."

exception. Previz/reference lane only, and even there, ACE-Step/YuE below are strictly better choices

(better license, comparable or newer architecture) with no reason to reach for MusicGen instead.

1.6 Open-weights local music generation — the new research this brief adds

Stable Audio Open 1.0 (VERIFIED unless noted) — Repo/model card:

huggingface.co/stabilityai/stable-audio-open-1.0, github.com/Stability-AI/stable-audio-tools.

License: Stability AI Community License (VERIFIED, fetched the LICENSE.md directly) — commercial use

permitted, free only under USD $1,000,000 annual revenue ("any licenses granted to You under this

Agreement shall terminate" above that threshold; an enterprise license is required beyond it), output

ownership "to the extent permitted by applicable law," registration required for commercial

distribution, "Powered by Stability AI" attribution required if distributing. 1B parameters,

transformer-based diffusion (DiT) in latent space, 44.1kHz stereo, maximum 47 seconds of output

(VERIFIED, HF model card). VRAM: the diffusion step itself uses ~5.9-6GB, but the full pipeline

including waveform decoding rises to ~14.5GB (VERIFIED via a HF discussion thread doing direct

measurement); community benchmarks suggest 24GB is the practical comfortable minimum, with 16GB

workable under tighter optimization (INFERRED). **This clears the 32GB 5090 easily even before

optimization.** Explicit quality limitation straight from Stability's own model card (VERIFIED): "The

model is not able to generate realistic vocals" and, tellingly, **"The model is better at generating

sound effects and field recordings than music"** — Stable Audio Open was always more of an SFX tool

wearing a music-model label; see §2.2 for its SFX-lane treatment.

Stable Audio 3.0 (VERIFIED unless noted) — released May 20, 2026 (VERIFIED via Stability's own

announcement + TechCrunch coverage). A family: Small, Small SFX (459M params, targets on-device/mobile-

class hardware), and Medium, all released as open weights on HuggingFace, trained specifically on

"fully licensed data" (a deliberate answer to the Suno/Udio infringement-litigation problem) — six-

minute track generation, a genuine capability jump over 1.0's 47-second cap. License: same-family

"Stability AI Community License" confirmed by name on the Small SFX model card (VERIFIED it is this

license family; the exact revenue-threshold clause text for this specific model page could not be

re-fetched verbatim in this pass the way it was for 1.0 — treat the $1M cap as INFERRED-but-likely for

3.0 until directly re-confirmed, tracked as §4 Unknowns #2). **Same standing risk as Stable Fast 3D in

the 3D-asset brief: fine today for a solo dev, a real constraint if the game crosses $1M revenue.**

YuE (VERIFIED) — Repo: github.com/multimodal-art-projection/YuE (M-A-P / HKUST). **License:

Apache 2.0, covering both the model weights and generated outputs** (VERIFIED, fetched the LICENSE file

directly: "The YuE model (including its weights) is now released under the Apache License, Version

2.0"). This is the cleanest license in the entire music lane — genuinely unrestricted, no revenue cap,

no attribution requirement beyond the license's own boilerplate (a "credit YuE by HKUST/M-A-P" is

described as encouraged, not mandatory, per secondary corroboration). A lyrics-to-full-song model

(vocals + instrumental accompaniment together) — explicitly marketed by its own authors as usable for

"soundtracks for films, video games, and virtual content" (VERIFIED, project materials). VRAM:

community reports cluster around 24GB for the full pipeline (INFERRED, not independently benchmarked

here) — fits the 5090 with headroom.

ACE-Step / ACE-Step 1.5 (VERIFIED unless noted) — Repo: github.com/ace-step/ACE-Step-1.5.

License: MIT (VERIFIED, fetched directly: "This project is licensed under MIT"). Scales cleanly

across the entire VRAM range this project cares about: 2B "turbo" variant runs under 6GB (INT8 +

CPU offload) up through 8-16GB with a vLLM backend; the XL (4B) variant needs ≥12GB with offload/

quantization or ≥20GB without, fitting the 5090's 32GB comfortably at full quality with room to spare.

Maximum output length: 600 seconds (10 minutes) — far beyond any competitor here, useful for long

ambient/exploration cues. Has an explicit "Instrumental Mode" ([instrumental]/[inst] token) for

vocal-free generation — directly relevant since most of a 79-chapter game score is instrumental, and

this sidesteps ACE-Step's weakest area. Independent (non-vendor) quality read (INFERRED, aggregator

synthesis): vocal output carries "noticeable audio artifacts... metallic timbre, sibilance distortion,

occasional pitch glitches," described as sitting "closer to or below Suno v4" in vocal expressiveness —

but this is a vocal-quality critique that should not be read onto instrumental orchestral output,

which is a materially different generation path and was not itself independently benchmarked in the

sources found. No independent (non-vendor) instrumental-specific benchmark was located — a genuine gap,

tracked in §4 Unknowns and matching the sibling brief's own benchmark-gate discipline.

HeartMuLa (oss-3B) (VERIFIED unless noted) — Repo: github.com/HeartMuLa/heartlib, paper

arxiv.org/abs/2601.10547. License: Apache 2.0 (VERIFIED, confirmed as updated in the repo). A

January 14, 2026 initial open release (HeartTranscriptor-oss, HeartMuLa-oss-3B, HeartCodec-oss) from an

academic-scale reproduction effort claiming "comparable performance with Suno in terms of musicality,

fidelity and controllability" at an internal 7B scale — but **only the 3B checkpoint appears actually

released**; whether a larger open checkpoint has since shipped was not confirmed in this pass (§4

Unknowns #4). Genuinely new and unrestricted, but the least independently proven of the three open

MIT/Apache candidates — treat as a benchmark-gate entrant alongside ACE-Step and YuE, not a default pick

yet.

The stem problem — the core adaptive-music gap. None of YuE/ACE-Step/HeartMuLa/Stable Audio/Suno/

Udio natively output separated stems; each returns one mixed-down file. Demucs (VERIFIED, repo

github.com/facebookresearch/demucs, now community-maintained at github.com/adefossez/demucs):

MIT license, "free for personal and commercial use... permits commercial use without restrictions"

— the correct unrestricted bridge tool to pull stems back out of any generator's mixdown. The real

caveat (ANALYSIS, applying a verified fact to this project's specific need): Demucs is trained for

popular-music source separation — its output classes are vocals/drums/bass/other. An orchestral

game score wants section-based stems (strings/brass/woodwinds/percussion/choir) to build the kind

of vertical remixing MetaSounds+Quartz is designed for (§3). Demucs was not evaluated against orchestral

section-separation quality in this pass and there is real reason to doubt a pop-tuned 4-stem separator

generalizes well to that — this reinforces, rather than undercuts, **AIVA's native compositional-layer

stems as the right mechanism specifically for anything that needs true adaptive layering**, with

Demucs as a fallback for lower-stakes bulk cues where a rough vocals/percussion/other split is good

enough.

1.7 Licensed orchestral libraries — the quality-benchmark alternative

Spitfire Symphony Orchestra Discover / BBC Symphony Orchestra Discover (VERIFIED unless noted) —

free (email-subscriber-gated), 44 instruments recorded at AIR Studios Lyndhurst Hall — "home to

countless film, TV and game scores." Commercial use, including in commercial games, is permitted under

the developer EULA (VERIFIED via Spitfire's own FAQ page). License class: ROYALTY-FREE-OWNED — no

revenue cap found, no subscription, install-and-own. Spitfire's wider paid catalog operates on the same

perpetual-license model. Real recorded per-instrument stems (not AI-separated) — this is the actual

quality and stem-cleanliness ceiling any generative option should be benchmarked against.

EastWest Composer Cloud+ (VERIFIED unless noted) — subscription access to "70 full collections."

EastWest's own EULA (VERIFIED, fetched soundsonline.com terms/EULA pages directly): compositions are

"100% royalty free for commercial use" including explicitly "scoring A-list game titles," and — the

important nuance — **"nothing will happen to your compositions if you pause or cancel your

membership"; already-created work stays usable and distributable forever. The catch: you lose the

ability to re-edit the MIDI performance or re-render/re-mix after cancelling** — a genuine risk on a

project with a multi-year build cycle, where a canonical cue might need revision long after any given

subscription period lapses. License class: SUBSCRIPTION-OWNED, but functionally closer to

ROYALTY-FREE-OWNED for anything already locked/rendered before a lapse. EastWest's EULA also explicitly

prohibits using its content to train any machine-learning/AI music-generation system — same pattern

as the SFX libraries in §2.

1.8 Music lane — slot recommendations

**Bulk regional/chapter theming (79 nodes × cultural instrumentation, budget-constrained, high

volume). Recommend ACE-Step 1.5 and YuE, benchmarked head-to-head on the 5090 in instrumental mode**

against representative per-region prompts, both fully unrestricted (MIT / Apache 2.0) with zero

per-generation cost once running locally. Strongest objection: neither model was purpose-built for

cinematic orchestral game scoring the way AIVA was (both are general song/lyrics-to-music models whose

public benchmarks skew toward vocal pop), and no independent instrumental-orchestral-specific quality

evidence was found for either in this pass — this is a real, not hypothetical, quality-ceiling risk that

only a direct benchmark-gate run resolves, exactly the posture LOCAL_3D_ASSET_GEN.md §5 already

recommends for the mesh-gen decision. Budget a composer-touch-up pass regardless of which model wins the

gate, the same way the existing contract already assumes for the canonical theme track.

**Hero set-piece score (Grand Sage reveal, Cassius three-phase boss, three endings, Vimana activation —

the contract's composer-in-loop canonical theme track). AIVA Pro remains the right pick, confirmed

not overturned.** It is purpose-built for exactly this content class, its Full Copyright tier resolves

the ownership question cleanly, and it is the only vendor evaluated with a genuine per-instrument stems

export. Strongest objection: the orchestration-agent design in the existing contract (§3.4) assumes

a self-serve API that this research could not confirm exists as a product — before that agent gets

built, someone needs to email AIVA directly and get real API terms/pricing, or the canonical-theme-track

automation may need to fall back to a manual browser-based workflow (still compatible with the

composer-in-loop discipline the contract already specifies, just not the "orchestration agent submits

via API" automation layer it currently assumes).

Adaptive layering mechanism (the MetaSounds/Quartz stem requirement, §3). **AIVA's native Stems

export is the only clean path found**; treat Demucs-based retrofit stem-splitting on Suno/ACE-Step/YuE/

Stable-Audio output as a fallback for lower-stakes bulk ambient cues only, given Demucs' pop-tuned

4-stem separation is a poor structural match for orchestral section layering. Strongest objection:

this concentrates the adaptive-music-critical content path on a single vendor (AIVA) with an unclear,

possibly-negotiated-only API — a real single-point-of-failure risk for the one music category (canonical

set pieces) the game's hard-line moments depend on most.

Voice — not re-litigated, per scope. Q3_2026_MODELS_REFRESH.md §4.1-4.2 (same-day, live-verified)

already confirmed ElevenLabs (hero) + the Kokoro-82M addition (bulk NPC, Apache 2.0, no cloning

capability by design) as sound, with Chatterbox (Resemble AI, MIT) flagged as a higher-quality but

consent-risk-shifted alternative. Nothing found in this pass changes that. The one addition: ElevenLabs'

own new "Eleven Music" app (§1.4) and Eleven v3's Text-to-Dialogue API show the vendor consolidating

voice+music+SFX — worth knowing as a single-vendor option at benchmark-gate time, not a reason to change

the ruling today.

---

2. SFX lane

Quick-reference table

ToolLicense classCommercial-ship viable?Best fit
ElevenLabs SFX endpointSUBSCRIPTION-OWNEDYesBulk ambient/foley/impacts, API-driven
Stable Audio Open 1.0TERRITORY-OR-USE-CAPPEDYes, under $1M revenueBulk ambient/field-recording-style SFX (its actual strength)
Stable Audio 3.0 Small SFXTERRITORY-OR-USE-CAPPED (same family; exact terms not re-confirmed on this page)Likely, pending confirmationOn-device/lightweight SFX generation
MOSS-SoundEffect v2.0UNRESTRICTEDYes (Apache 2.0)Bulk ambient/Foley, fully local, zero legal friction
Meta AudioGenNON-COMMERCIALNo (weights CC-BY-NC-4.0)Disqualified
Sonniss GDC bundlesROYALTY-FREE-OWNEDYes, freeFoundational library backbone
BOOM LibraryROYALTY-FREE-OWNEDYes, $99-199/library one-timePremium foundational library
SoundlySUBSCRIPTION-OWNED (shipped-project rights persist after cancel)YesSearchable aggregator library
freesound.orgUNRESTRICTED (CC0 subset only — mixed licenses per-sound)Yes, if curated to CC0/CC-BYFree bulk curation, labor-intensive
UE 5.8 MetaSounds procedural synthesisUNRESTRICTED (engine-native)YesUI, footsteps, simple runtime-variable SFX
Krotos Dehumaniser 2 / Reformer ProROYALTY-FREE-OWNED ($399 each, perpetual)YesCreature vocalization — the actual professional tool

2.1 ElevenLabs SFX — VERIFIED unless noted

30 seconds maximum per generation; output as MP3 for all effects, **WAV at 48kHz for non-looping

effects**; a dedicated looping mode is designed for effects that "repeat without perceptible

start/end points" for use beyond a single 30-second clip. Pricing: 40 credits per second when

duration is explicitly specified (an auto-duration mode exists at a different, undocumented-in-this-

fetch rate). This sits within the same overall ElevenLabs credit-pricing tiers already verified in

Q3_2026_MODELS_REFRESH.md §4.1 (Free through Business/Enterprise).

over what the original contract assumed — this pass independently re-confirms the same 30s/48kHz/

looping facts directly from Eleven's own docs page rather than relying on the prior pass's citation.

terms** — the general paid-tier "full output rights, no volume cap" language already verified in the

sibling brief is assumed to extend to SFX outputs (INFERRED by extension, not independently re-quoted

for the SFX endpoint specifically).

2.2 Stable Audio Open 1.0 / Stable Audio 3.0 for SFX — VERIFIED (cross-referenced from §1.6)

The load-bearing fact for this lane, already surfaced in §1.6: Stability's own model card for Stable

Audio Open 1.0 states **"The model is better at generating sound effects and field recordings than

music"** — meaning this tool's real strength was always SFX, not music, despite the branding. Same

license (Stability AI Community License, free under $1M revenue), same VRAM profile (~6-14.5GB,

comfortable on the 5090). Stable Audio 3.0 Small SFX (459M params, open weights, same license

family) is explicitly designed for on-device/lightweight SFX generation and is the natural successor to

evaluate first at benchmark-gate time, once its exact weights-license terms are pinned down (§4

Unknowns #2).

2.3 MOSS-SoundEffect v2.0 — VERIFIED, genuinely new finding not present in any prior research pass here

MOSS-TTS and the HuggingFace model page huggingface.co/OpenMOSS-Team/MOSS-SoundEffect-v2.0`

independently surfacing the same facts): **1.3B parameters, Diffusion Transformer + Flow Matching

architecture, DAC VAE + Qwen3 text encoder, 48kHz output, up to 30 seconds per generation.** License:

Apache 2.0 (VERIFIED, confirmed on the HF model card directly).

environmental audio from text descriptions... ambience, urban scenes, creatures, human actions, and

short music-like clips," explicitly framed as a no-library-required Foley/ambience generator.

license than Stable Audio (no revenue cap at all), higher native sample rate (48kHz vs Stable Audio

Open's 44.1kHz), purpose-built for SFX rather than a music model repurposed for it. No independent

(non-vendor) quality benchmark was located (§4 Unknowns), so it earns a benchmark-gate slot alongside

Stable Audio 3.0 Small SFX rather than an outright replacement recommendation, but it is a genuine

new option the existing contract's "Stable Audio Open or comparable as fallback" language did not

and could not have known about in May 2026.

2.4 Meta AudioGen — VERIFIED (cross-referenced from §1.5)

Same LICENSE_weights file covers AudioGen and MusicGen jointly: CC-BY-NC-4.0, non-commercial only.

Disqualified for the shipped path for the same reason as MusicGen. No reason to reach for it over

MOSS-SoundEffect or Stable Audio Open, both of which are equal-or-better on quality claims and

materially better on license.

2.5 Licensed SFX libraries — the professional default, VERIFIED unless noted

Sonniss #GameAudioGDC bundles (VERIFIED, fetched sonniss.com/gdc-bundle-license/ directly, quoted

verbatim): "Licensee may use the licensed sound effects for the purposes of synchronization with audio

and visual projects the Licensee is involved with, which includes but is not limited to: games, films,

television & interactive projects" — commercial use permitted, no attribution required ("Licensee

may use and modify the licensed sound effects for personal and commercial projects without attribution

to the original creator"). Cannot resell standalone, cannot claim original authorship, governed by

English law. The 2026 bundle (released ~March 2026 for GDC) is ~7.47GB / 347+ files, free; the

community-maintained historical archive across nine prior years totals roughly 200GB, also free.

License class: ROYALTY-FREE-OWNED. One hard, explicit, and universal restriction worth flagging for

this project specifically: **"The Licensee is expressly prohibited from using any sound effects licensed

under this Agreement for the purpose of training artificial intelligence technologies"** — a clean line

for direct-use-as-shipped-asset (this project's actual need), but a real wall if the generative-entity

runtime-layer vision (memory: generative-entity-system-vision) ever wants to fine-tune a model on owned

SFX content that includes Sonniss-sourced material.

BOOM Library (VERIFIED unless noted) — one-time purchase per library, **$99-199, perpetual

royalty-free license for commercial and non-commercial projects**, "pay once and own the sounds

forever" (buyout tier excludes future product updates). Flagship "BOOM ONE" bundle: 60,000+ files,

199,000+ total effects. License class: ROYALTY-FREE-OWNED.

Soundly (VERIFIED, fetched Soundly's own FAQ/EULA pages): cloud-searchable aggregator, subscriptions

from $14.99/month, "sounds in Soundly Pro and Soundly Free libraries are cleared for commercial use,

including podcasts, films, and video games." Critically: **"Any projects created while subscribing to

Soundly Pro stay cleared even if you end your subscription... podcasts, films, video games"** — the one

restriction is that "the sounds must be used as part of a project with other media," i.e., you cannot

resell the raw sound files themselves after cancelling. **License class: SUBSCRIPTION-OWNED, but

functionally ROYALTY-FREE-OWNED for anything already shipped** — same pattern as EastWest in §1.7.

freesound.org (VERIFIED unless noted) — community CC-licensed repository, no single license:

individual sounds carry CC0 (public-domain-equivalent, no attribution), CC-BY (attribution required), or

CC-BY-NC (non-commercial — must be excluded from the shipped path per the ruled posture). Free, huge

volume (56,000+ new uploads in 2024 alone per one corroborating source), but **requires per-sound license

curation at ingestion time** — a real labor cost at scale, not a "download and use" library the way

Sonniss/BOOM are. License class: UNRESTRICTED only for the CC0 subset; treat as a supplementary

long-tail source, not a primary bulk backbone, unless a curation filter step is built into the SFX

orchestration agent's licensed-library lookup path (§10.4 of the existing contract already frames this

as a "licensed library lookup" surface where such filtering would live).

2.6 UE 5.8 MetaSounds — native procedural SFX synthesis, VERIFIED

Confirmed directly from Epic's own UE 5.8 documentation (VERIFIED, `dev.epicgames.com/documentation/

unreal-engine/metasounds-the-next-generation-sound-sources-in-unreal-engine` and the companion

"Creating Procedural Music with MetaSounds" page): MetaSounds exposes a full DSP graph — oscillators,

filters, envelope generators, wave players, random selectors, math nodes — wired together in a

node-based editor, generating audio synthetically at runtime with sample-accurate timing, "no

programming experience" required for basic use, and each graph compiles to "an optimized static,

non-virtual C++ object" for performance. License class: UNRESTRICTED — it is the engine itself, zero

external licensing question. This is genuinely the right tool for **UI sounds, footstep variation

(procedural pitch/velocity randomization layered onto a small sample set), and simple runtime-reactive

SFX** where paying a generation or licensing cost per variant is wasteful — but it is not a content-

generation tool for complex, information-dense sounds like creature vocalizations or detailed foley; it

shapes and combines existing/simple synthesized material rather than inventing rich new source audio

from nothing.

2.7 Creature vocalization (132-creature bestiary) — confirmed NOT a generative-AI task

Directly answering the brief's question: the real, professional workflow is **layered source recordings

(animal + human + non-literal sources like metal scrapes or balloon squeaks) run through dedicated

pitch/formant creature-design processing** — not text-to-SFX generation. This is confirmed by two

independent lines of evidence: (1) industry sound-design sources (Abbey Road Institute, A Sound Effect,

Boom Box Post, Krotos' own tutorials — INFERRED synthesis across several corroborating articles)

consistently describe a three-layer model (low-end animal-growl/synthetic-rumble foundation,

mid-frequency human-plus-animal texture layer, high-frequency detail layer from bird/snake/metallic

sources), processed with pitch-shifting and formant filters to "transpose recordings cleanly, allowing

voices to reach necessary low frequencies for monster voices"; and (2) **none of the generative text-

to-SFX tools evaluated in this brief (ElevenLabs SFX, Stable Audio Open/3.0, MOSS-SoundEffect) claim or

evidence creature-vocalization-specific capability** — their demonstrated strengths are ambience, Foley,

and impacts, not character-driven monster vocal performance.

The actual tool: Krotos Dehumaniser 2 and Reformer Pro (VERIFIED, cross-corroborated pricing across

multiple retailers) — $399 each, perpetual/one-time license ("you own them rather than subscribing"),

purpose-built creature/monster vocal-design plugins, industry-standard per multiple sound-design

tutorial sources referencing them directly for this exact task. License class: ROYALTY-FREE-OWNED.

This maps cleanly onto the existing contract's already-specified **sound-designer-in-loop canonical SFX

track** (§10.4.1) — this research doesn't change that workflow design, it names the actual tool that

belongs in the sound designer's hands, and confirms the category correctly deserves canonical-track

(human-in-loop) treatment rather than the agent-only standard track, given no automated alternative

exists.

2.8 SFX lane — slot recommendations

**Bulk ambient + foley (biome ambience across 69 regions, footsteps per terrain, weapon foley, item

handling). Recommend the existing contract's hybrid design, now vendor-confirmed: licensed libraries

first** (Sonniss's free archive as the zero-cost backbone, a BOOM ONE buyout for depth/consistency) to

cover generic, high-volume categories cheaply and immediately; **MOSS-SoundEffect v2.0 or Stable Audio

Open** (both now confirmed viable, MOSS-SoundEffect fully unrestricted) filling gaps the libraries don't

cover, run locally on the 5090 at zero marginal cost; MetaSounds procedural synthesis layered on top

for runtime variation (pitch/velocity jitter on footsteps, simple impact layering) to cut asset count.

Strongest objection: licensed libraries are generic by construction — none were recorded for this

specific 90M-word bespoke fantasy world's cultures and regions, so bulk-licensed SFX still needs the same

per-region cultural-authenticity ear-check the contract's §10.8 already mandates for generated content;

cheap/free licensing is not a substitute for that audit pass, just a cost reduction on the raw material

feeding it.

Creature vocalization (132 creatures). Recommend the contract's existing sound-designer-in-loop

canonical track, now with Krotos Dehumaniser 2 + Reformer Pro ($798 total, one-time, perpetual)

named as the actual tool, fed by licensed or recorded animal-source material. Strongest objection:

this is the single most human-labor-intensive SFX category in the whole pipeline — 132 creatures, each

needing multiple vocalization states (idle, alert, attack, hurt, death, at minimum), each requiring

actual sound-design craft, not prompt iteration — and the existing contract does not currently budget

sound-designer seat-hours against this explicitly. This is a scheduling/scope risk, not a tool-choice

risk; the tool pick is solid.

UI. Recommend MetaSounds procedural synthesis as the default (zero asset weight, zero licensing

question, sample-accurate, engine-native) for standard interaction feedback, supplemented by a small

curated set from a licensed library or a short ElevenLabs SFX/MOSS-SoundEffect generation batch for

bespoke polish tones (magical vril-tier UI stings, ability-cast confirmations) that are hard to

synthesize convincingly from raw oscillators alone. Strongest objection: procedural synthesis trades

one skill cost for another — it requires actual DSP-graph sound-design competence inside the MetaSound

editor, which is real engineering/craft time, not a "type a prompt" workflow, and that time is not

currently budgeted separately from the generative-SFX time in the existing contract.

---

3. The UE 5.8 integration seam

3.1 Import automation — Python asset import for audio

Confirmed directly from Epic's own UE 5.8 documentation (VERIFIED, `dev.epicgames.com/documentation/

en-us/unreal-engine/importing-audio-files): supported source formats are **.wav, .ogg, .flac`,

.aif, .opus, .mp3**, any sample rate, 16- or 24-bit (16-bit recommended — "24-bit audio lacks

dithering during the conversion process"; **all audio is internally converted to 16-bit WAV regardless

of source format**), mono/stereo/4.0/5.1/7.1 channel configurations all supported. On import, every

Sound Wave is compressed per the project's configured Default Audio Compression Type — Bink Audio

(lossy, up to 10:1, the default perceptual codec), ADPCM (lossy, up to 4:1, fixed-quality

time-domain), PCM (lossless/uncompressed, higher memory but fastest decode), or **Platform

Specific**, all overridable per-platform and per-asset.

The programmatic import path (VERIFIED via the class references + cross-confirmed against the same

generic pattern UE_BUILD_AUTOMATION.md already documented for mesh import): unreal.SoundFactory /

unreal.ReimportSoundFactory expose an automated_import_data property of type

unreal.AutomatedAssetImportData, "data for how to import files via the automated command line

importing interface" — the same AssetToolsHelpers.get_asset_tools().import_asset_tasks([task]) pattern

already confirmed for mesh assets applies directly to audio, just with a Sound-specific factory. This

means the SFX/AIVA orchestration agents' "write generated file → import to UE5" step is a solved,

scriptable problem — the open question is entirely upstream, at getting the generated/licensed audio

files in the first place (§1-§2), not at the UE5-side import mechanics.

3.2 MetaSounds wiring — Python-scriptable, with real limits

Confirmed via Epic's own Python API docs (VERIFIED, cross-referenced across the MetaSoundBuilderSubsystem,

MetaSoundSourceBuilder, and MetaSoundBuilderBase class references): a MetaSound Builder API

exists specifically to "author MetaSounds procedurally from gameplay code" and "create MetaSounds in

memory without using the MetaSound Editor" — MetaSoundBuilderSubsystem.create_source_builder()

constructs a builder; MetaSoundSourceBuilder exposes add_graph_input_node(), add_node_by_class_name(),

audition(), and build() to assemble and test a graph in script. This means an agent pipeline can, in

principle, programmatically wire a generated audio stem into a MetaSound playback graph.

The real limit (VERIFIED, directly from the same Epic docs): "The Builder API does not currently

support variables, and paged inputs and graphs have limited support and are not authorable by Blueprint

nor outside of edit time from the frontend Builder API." Cross-referenced against an independent

middleware comparison (INFERRED, aggregator synthesis but internally consistent and specific):

**MetaSounds "is workable for simple adaptive music but lacks dedicated music authoring tools... you are

essentially programming music logic from scratch using audio nodes. This works for simple systems

(two-layer adaptive music) but becomes unwieldy for complex interactive scores"** — contrasted against

Wwise ("the most sophisticated interactive music system in the industry... hierarchical music

organization") and FMOD ("visual and intuitive workflow for musicians... audition music transitions...

without running the game"). **This project's engine target (UE 5.8) and the task framing both assume

MetaSounds+Quartz as the delivery mechanism, so this is not a recommendation to adopt third-party

middleware** — it is a flag that fully agent-automated construction of anything beyond simple two-to-

three-layer adaptive graphs likely still needs a human audio designer building complex canonical-set-

piece graphs by hand inside the MetaSound editor, which is actually consistent with (not a gap against)

the existing contract's composer-in-loop/sound-designer-in-loop/director-in-loop discipline already

scoping human intervention to exactly the canonical/hero content tier.

3.3 Quartz-driven adaptive playback — the concrete technical shape

Confirmed via Epic's own UE 5.8 documentation and a detailed third-party implementation writeup

(VERIFIED for the Epic-documented mechanism, INFERRED for some implementation specifics from the

third-party source): music cues are built from multiple layers/stems (bass, drums, melody,

percussion, chordal, etc.) referenced by Data Assets carrying **tempo, song duration in bars, and

which bars/beats each layer occupies** (e.g., "bass lines are eight-bar phrases"). Blueprint logic

listens to a Quartz clock subsystem and triggers layer playback at defined **Quantization

Boundaries**, enabling vertical remixing (layers fade in/out in sync) as gameplay state changes. The

MetaSound Wave Player node supports sample-accurate concatenation so a queued next-loop plays with

no audible seam.

**This is the concrete answer to "what does authoring adaptive music actually require of the generated

stems": each stem needs (1) to exist as a separate audio asset, (2) known tempo, (3) known bar-length/

phrase-length, (4) clean loop points with no baked-in fades that would break seamless looping.** Cross-

referencing this against the existing contract's schema (§3.5-§3.6 of T1_Build_Pipeline_Contracts,

read directly for this brief): T0_Theme_Registry's theme_substrate object carries mood_tags,

instrumentation_substrate, cultural_substrate, era_substrate, recurrence_anchors — **but no

tempo, bar-count, or stem-role field**. This is a genuine, concrete gap between what the ratified

contract schema captures and what the UE-side adaptive-music mechanism actually needs to consume a

generated cue as a true adaptive layer rather than a single static music_cue_id_ref playback. Per this

project's do-not-invent discipline this is flagged, not silently patched — it belongs with the

Phase 5M T0_Schema_Dictionary work T1_Build_Pipeline_Contracts §3.11 already lists as open ("music_

mood_tags object expansion covering... dynamic mix layering parameters, stem-track granular reference

defer to Phase 6 region page build feedback") — this brief's finding narrows exactly what that expansion

needs to add: tempo, bar-length, and stem-role/layer-name fields, grounded in the real MetaSounds/Quartz

Data Asset shape rather than left abstract.

3.4 What the Cinematic pipeline consumes — grounded directly in the ratified contract

Read directly from T1_Build_Pipeline_Contracts §5 (not web research — this is the project's own ratified

schema, restated here for the audio-stack seam): the Cinematic Scene Build agent reads

T0_Scene_Spec_Registry rows carrying mesh_refs, voice_refs, music_cue_id_ref, `sfx_category_

array, and sfx_id_ref_array` — **and will not proceed until every referenced asset resolves to an

accepted/locked generation status** (the blocked_on_dependencies state, §5.7). Audio-family mixing

intent is captured per-scene by two fields worth highlighting for this brief's purposes: `audio_

priority_layer (an enum: voice_primary, music_primary, ambient_primary, silence, layered_

balanced — which layer of the three-pipeline audio family "drives" a given scene) and scene_audio_

intent` (free text capturing authorial mix intent, e.g. "voice carries dramatic weight, music swells at

reveal, ambient quiets to silence"). Scene assembly itself executes via the NeoStack Cinematic Sequencer

Python surface — confirmed elsewhere in this repo's own research (UE_BUILD_AUTOMATION.md, VERIFIED

class existence) to use unreal.MovieSceneAudioTrack/unreal.MovieSceneAudioSection as the native

Sequencer audio-track classes, added via the same add_track pattern used for camera/animation tracks.

**Net for this brief: the Cinematic pipeline's audio contract is schema-complete and grounded (it

already knows exactly which three audio outputs it needs and in what priority), and the only real gap

this research surfaces is one level upstream — whether the music/SFX generation agents that are supposed

to fill music_cue_id_ref/sfx_id_ref_array in the first place can actually reach their vendor APIs

the way the contract assumes (§4 Unknowns #1, #7).**

---

4. Unknowns register

1. AIVA's actual API terms, pricing, and self-serve availability. Confirmed to exist only as a

negotiated enterprise agreement (§1.1); no public pricing, documentation, or endpoint reference found

live anywhere. How to settle: email contact@aiva.ai directly and request API terms/pricing

before building the orchestration agent T1_Build_Pipeline_Contracts §3.4 currently assumes; budget

for the possibility that the canonical-theme-track generation stays a manual/browser-driven,

composer-in-loop-only workflow rather than a scripted API call.

2. Stable Audio 3.0 Small SFX's exact weights-license terms. Confirmed to be a "Stability AI

Community License" by name on the model card, but the specific $1M-revenue-threshold clause text for

this model page specifically (as opposed to Stable Audio Open 1.0's, which was quoted verbatim) was

not re-obtained in this pass. How to settle: direct-fetch and quote the LICENSE file linked from

the stable-audio-3-small-sfx HuggingFace repo before relying on it at build time — this was also

flagged as open by the prior Q3_2026_MODELS_REFRESH.md pass and remains open here.

3. **Sonniss/EastWest-style "no AI/ML training" clauses vs. this project's own generative-entity

runtime-layer vision.** Both major licensed-library vendors checked (Sonniss directly, EastWest

directly) explicitly forbid using their licensed content to train AI/ML systems. Irrelevant to this

project's immediate need (shipping the audio directly as licensed assets), but a real wall if the

memory-tracked generative-entity-system-vision runtime layer ever wants to fine-tune a model on

owned/licensed audio content that includes library-sourced material. How to settle: a legal skim

at the point any such fine-tuning is actually planned, not before.

4. Whether a larger open HeartMuLa checkpoint exists beyond the released "oss-3B." The project's own

paper references an internal 7B version claiming Suno-comparable quality; only the 3B tier appears

actually released as of this research date. How to settle: re-check `github.com/HeartMuLa/

heartlib/releases` at benchmark-gate time.

5. **How much MetaSound Builder API's "no variables support" limitation actually constrains

agent-automated adaptive-music graph construction** at the scale this game needs (potentially

per-region × per-integrity-level interactive layering, not just simple two-layer cases). **How to

settle:** a scripted spike — attempt to build a representative 3-4-layer adaptive ambient graph via

the Python Builder API against a live 5.8 editor session and see exactly where the automation breaks

down and hand-authoring becomes necessary.

6. The T0_Theme_Registry / T0_SFX_Registry schema gap for tempo/bar-length/stem-role metadata

(detailed in §3.3) — the ratified contract schema doesn't yet carry the fields MetaSounds/Quartz

actually needs to consume a cue as a true adaptive layer. This is a canon/contract-schema decision,

not something this brief should resolve unilaterally. How to settle: fold into the Phase 5M

T0_Schema_Dictionary work the contract's own §3.11 already tracks as open, using this brief's

concrete field list (tempo, bar-count, stem-role/layer-name) as the starting proposal for Josh's

review under the standing decision-analysis protocol.

7. Whether NeoStack's Cinematic Sequencer Python surface has higher-level audio-track helpers (e.g.

an "add scored cue with fade envelope" convenience call) beyond the generic MovieSceneAudioTrack/

MovieSceneAudioSection classes, or whether the agent must hand-key volume automation itself.

Already flagged as INFERRED/unconfirmed by UE_BUILD_AUTOMATION.md; not independently re-verified in

this pass since it sits outside this brief's audio-content-generation scope. How to settle:

confirm against the actual shipped SequencerScripting Python examples once NeoStack access is live,

per that sibling doc's own residual-unknowns tracking.

8. **Whether ACE-Step/YuE/HeartMuLa's instrumental-orchestral output quality holds up against AIVA's

purpose-built cinematic scoring** for this project's specific need — every quality signal found for

these three models is either vendor-claimed or focused on vocal/pop performance, not independently

benchmarked on instrumental orchestral cues. How to settle: exactly the benchmark-gate protocol

LOCAL_3D_ASSET_GEN.md §5.2 already models for the 3D-asset decision — fixed test set (2-3

representative regional themes, one hero-adjacent cue), run through AIVA + all three open candidates,

score on instrumentation fidelity, cultural-substrate accuracy against the Care Doctrine, and stem/

loop usability, not generation speed.

---

Appendix: primary sources fetched directly (VERIFIED tier)

project's own ratified contract, read directly, not web research)

"light pass" on voice/AIVA/Stable-Audio-Open/licensed-libraries — reused findings re-verified where

re-fetched, cited rather than re-derived where not)

Appendix: secondary/aggregator sources used (INFERRED tier — cross-referenced, not primary)

Suno/Udio litigation and commercial-terms tracking: chartlex.com, digitalmusicnews.com, aivortex.io,

happycapyguide.com, hollywoodreporter.com, dynamoi.com, ailawsuittracker.com, routenote.com,

musically.com, billboard.com, ageofllms.com, techjacksolutions.com, undetectr.com, once.app,

medium.com (@J.S.Matkowski), terms.law, blog.dubspot.com, sunowatermark.com, licenseorg.com ·

AIVA API/enterprise: apitracker.io, forcefriction.com, aitools-directory.com, videosdk.live,

solomonsignal.com, max-productive.ai · Suno API status: musicbusinessworldwide.com (primary-adjacent —

reports Suno's own CPO announcement directly), docs.sunoapi.org and other third-party wrapper sites

(flagged as unofficial, not relied on for terms) · Open music models: marktechpost.com, heart-mula.com,

communeify.com, andrew.ooo, studio.aifilms.ai, spheron.network, arxiv.org (YuE and HeartMuLa papers),

heartmula.github.io · Stable Audio 3.0 coverage: stability.ai (own announcement), techcrunch.com,

musicbusinessworldwide.com, technology.org, blockchain.news, winbuzzer.com, aiinsiders.net, toknow.ai ·

Orchestral libraries: spitfireaudio.com and support.spitfireaudio.com (vendor-primary-adjacent),

production-expert.com, vi-control.net, bedroomproducersblog.com, soundsonline.com (EastWest,

vendor-primary-adjacent), pluginfox.com and other retailer listings (price cross-checks only) ·

SFX libraries: sonniss.com (gdc.sonniss.com, primary-adjacent beyond the directly-fetched license page),

boomlibrary.com, getsoundly.com (vendor-primary-adjacent), freesound.org and its own forum threads,

opengameart.org, sendshort.ai, gamineai.com, hackingtons.com · Creature vocalization: shapingwaves.com,

abbeyroadinstitute.com.au, asoundeffect.com, krotosaudio.com (vendor-primary-adjacent for pricing/

licensing claims), boomboxpost.com, blog.prosoundeffects.com, mysoftwaredeals.com, jrrshop.com,

equipboard.com, gearspace.com, musicplayers.com, plugin-alliance.com · UE MetaSounds/Quartz: abovenoise

studios.com (detailed third-party implementation writeup, cross-referenced against Epic's own docs

fetched directly), forums.unrealengine.com, unrealcommunity.wiki, cdm.link, malachigregory.com,

strayspark.studio, aircada.com, sfxengine.com, generalistprogrammer.com, toolsinfo.com.

Generated by harness/site/structure_site.py — the URL path is the repo path. review root