Bgm

Run the Preflight first — no credential is not a green light to silently generate locally. Before generating, complete the sign-in Preflight (see ../SKILL.md → Preflight): run npx hyperframes auth status, recommend signing in, and STOP for the user’s choice (sign in for HeyGen’s music library, or continue offline with local generation). This applies to a one-off “generate a BGM” request just as much as inside a full workflow.

Driving it from the request

audio_request.json → bgm: { mode?, query?, prompt? }:

• mode — retrieve | generate | none. Omit for auto (retrieve when credentialed, else generate). An explicit retrieve is strict: no credential ⇒ skip, never a detached generate (so a caller with no wait-bgm step, e.g. product-launch, can’t get a pending job it won’t await).
• query — the mood, used for retrieval and as a fallback prompt seed (e.g. a storyboard’s music: field, falling back to message → arc → "calm cinematic underscore").
• prompt — an explicit full prompt for generation; omit and the engine infers one (see Mood inference). Optional blob / archetype / arc feed that inference.

HeyGen retrieval (default)

searchSounds(query, "music", { limit: 5 }) → GET /audio/sounds?query=<mood>&type=music&limit=5. Take the top result (ranked by score), download its presigned audio_url → assets/bgm/track.mp3. Synchronous. No match → skip (BGM is optional; never fail the render over it). Cue written to audio_meta.json:

volume comes from the engine’s bgmDefaultVolume(): BGM_BED_VOLUME (currently 0.12 ≈ -18 dB — a bed under the voice) under narration, BGM_SILENT_VOLUME (currently 0.9) for a silent film (no voice). Tune those constants in scripts/lib/bgm.mjs, not call sites. An explicit volume in audio_meta.json always overrides this default. bgm_pending is false — the file is on disk when the engine returns.

For short launch videos, do not assume the beginning of the retrieved file is the best edit point. Check the opening against later five-second sections. If the track starts with a quiet build but a later section has a stronger, clean musical entrance, trim from that section and apply a short fade-in and longer fade-out. Repeat this check whenever the composition duration changes; the final music file must cover the full cut without a silent tail.

Local generation (fallback) — Lyria → MusicGen

Spawned detached so voice work isn’t blocked; audio_meta.bgm_pending: true and bgm_pid / bgm_log are set until it finishes. Run scripts/wait-bgm.mjs before assembling — it polls the output file / process / log, detects crashes, and writes bgm_status.json (status: ready | failed | timeout | disabled). A failed/absent track is simply omitted; it never blocks voice/SFX.

Output → assets/bgm/track.wav, target = total voice duration. MusicGen generates one seed clip (≤28–30s, under the decoder’s positional limit) then crossfade-loops it up to the target (or trims down if shorter), avoiding per-segment seams. Backend selection is by what can actually run: Lyria only when import google.genai succeeds, else MusicGen; if neither can be made to run, BGM is skipped (voice + SFX still render).

Mood inference (the generate prompt)

inferBgmPrompt() in scripts/lib/bgm.mjs: an explicit prompt wins; otherwise industry-keyword base → narrative-archetype shape → emotional-arc tiebreaker.

Archetype then reshapes the arc — PAS → “MINOR to MAJOR” build; BAB / future-pacing → aspirational rising; feature-cascade → +10 BPM driving; demo-loop → −8 BPM minimal. The emotional arc breaks remaining ties (tension→relief, excitement, trust/reassurance).

Lyria knobs (direct recipe use)

The engine bakes BPM / scale into the prompt text (via the inference above) and passes only --output / --duration / --prompt to the recipe. If you invoke scripts/lyria-recipe.py directly you can also set: --bpm (90–110 calm, 110–130 energetic), --brightness (0–1, ≥0.7 promotional), --density (0–1, higher = fuller), --scale (MAJOR / MINOR / PENTATONIC / …), --negative-prompt (styles to exclude). MusicGen ignores all of these — put the mood in the prompt.

Failure modes

BGM failure never blocks a render.