weavesignal/spike/talkshow/README.md

The Blether: a synthetic talk show for WeaveSignal tests

A test fixture, not product code (ADR-0078 in the project folder). A fictional show on the fictional Kettle FM that exercises the decided features: a browser guest, cross-talk, carts, the screener talking to callers off air, a profanity dump, the news on a hard post, playout down (OPEN), and the overlap handover to a remote DJ.

• show.json: cast, music, say_as respellings for the text-to-speech, per-cue seeds, and the cues (each with the feature it tests). In the live run the operator takes the host's lines; the others wait for the host.
• synth.sh: every speech cue through VoiceLab's tts-serve on 127.0.0.1:7500 into out/cues/.
• build.py: the timeline (out/timeline.json), one stem per participant (out/stems/), and two references: out/room.wav (what the room hears) and out/listeners.wav (dump cut, fallback while playout is down).

Ignored by git: voices/ (consented personal-use clones and VCTK references, see voices/ATTRIBUTION.txt), music/ (copyrighted, trimmed from the operator's files) and out/. Never stream any of it publicly.

Voices: host, guest and caller copy the VoiceLab persona references (N, Trevor, Drukhi). Screener, newsreader, farmer and DJ are VCTK speakers p249 (Aberdeen), p262 (Edinburgh), p292 (Belfast) and p286 (Newcastle): their first two available utterances, silence trimmed, joined with 0.35 s gaps, 24 kHz mono, about 12 s, with the corpus transcripts as ref.txt. OmniVoice's voice design has no Scottish, Northern Irish or Geordie accent; a clone carries its reference's accent. Whisper (VoiceLab STT) checked every line; three were re-seeded.