/explore seed: jingle packs + voice packs — two catalog kinds (Kokoro voice-designer .pt voices; instrumentals and jingles) feeding #708 beds and #702 casts

[genwave-radio]

/explore seed (Dean, 2026-09-06): two more community-shelf kinds in the ad-pack mould — jingle packs (short instrumentals and station jingles for background music under spots and imaging) and voice packs (custom Kokoro voices designed with https://github.com/n33kos/kokoro-voice-designer).

Why one epic

Both are catalog kinds with the same install path (manifest + assets → the box), and both have waiting consumers: #708 (generated spots get ducked background music) picks from installed jingle packs; #702 (ad cast pool) and persona cards pick from installed voice packs.

What a voice pack is

The designer exports a plain .pt voice tensor built by PCA-blending Kokoro's 54 stock voices (no training, no model swap). A pack = one or more .pt files + a manifest (name, gender/age hints, a demo clip, provenance) + licence text.

Plumbing already there

• kokoro-fastapi rescans its voices directory on every request (demo log: "Scanning for voices in path" per call) — a dropped .pt is live without a restart.
• GET /api/voices proxies Kokoro's live listing (KokoroVoiceLister), so the persona editor and the ad cast picker see new voices for free.
• Catalog kind tooling from catalog#74 (schemas, KindSpec, index branch, diff-scope guard, selftest, docs) is the template.

To settle in /explore

• Shared voices volume. Kokoro's voices dir is not mounted today; the api needs write, Kokoro read — the tts: volume pattern in compose.
• Provenance and consent. The designer's "fit to a target audio sample" mode is soft cloning. Manifest carries a "synthetic blend only, no real-person target" attestation — consent-first, as ruled for the House Voice; trust tiers per Plugin trust tiers: official / supported / community provenance on the loader (for /design) #676.
• Engine coupling. A .pt is Kokoro-only; if Spike: TTS bake-off — build a scored bench harness, run the self-hosted field against Kokoro #614's bake-off moves the house engine, voice packs become an engine-scoped kind — the manifest wants an engine field from day one.
• Jingle-pack assets. Licensing (CC0/CC-BY only?), size caps, formats, cue/loudness enrichment on install so beds are cue-trimmed and level-matched like every other authored row.
• Naming in the UI. Plain names (background music, jingles, voices) per UI/docs: radio jargon ("bed", "liner", "imaging", "ident", "patter") needs a plain name or a one-line explanation wherever it appears #707.

Sequencing: this epic first, then #708 and #702's pack-backed parts.

[genwave-radio]

Voice portability across engines (Dean's question, 2026-09-06): there is no conversion facility, and none is likely. A voice is a different object per engine — Kokoro: a 256-dim style tensor for a StyleTTS 2 decoder (the encoder that made the stock voices was never released); Piper: a whole VITS model per voice, trained on hours of audio; cloning engines (XTTS v2, F5-TTS, Chatterbox, Fish Speech, CosyVoice): a reference audio clip encoded into that engine's own speaker embedding at request time.

Consequences for the voice-pack manifest, from day one:

1. engine field + the native artifact (.pt for Kokoro).
2. A canonical reference clip (10–15 s neutral passage rendered by the pack's engine) — the only portable part. A cloning engine takes it as its prompt; the result approximates the voice and drifts per engine.
3. Provenance attestation stays "synthetic blend, no real-person target" — reference clips are renders, so re-cloning them is consent-clean.

Two bridges we would build ourselves, both approximations: (a) reference-clip cloning on engines that support it; (b) a zero-shot voice-conversion stage (seed-vc / OpenVoice v2; RVC for a closer match at the cost of per-voice training) over whichever engine wins #614 — a second model, real latency, a GPU expectation. Piper is reached by neither; supporting it means Piper's own voices only (today's Pi 4 posture).

[genwave-radio]

Epic PR is open: #711 (T410–T420 + T428, one commit per task, plus a T418 rider). T421–T423 recipes and the flags for the ruling are the first two comments on the PR. T429 (v5.7.0 pins + changelog) waits on that merge and on the catalog side (T424–T427, seed pack slug gw-first-beds).

[genwave-radio]

✅ Jingle packs + voice packs shipped in v5.7.0 and have been on the demo since; closing at v5.9.0 (https://github.com/GenWave-Org/genwave/releases/tag/v5.9.0, gate run https://github.com/GenWave-Org/genwave/actions/runs/35219533645). Owed items live on their own issues (#708 ear, #712 riders design).