project ct voice profile caption audio quality gate 20260715T193334
name: Voice-profile ingest needs to filter overlaid audio (music/memes), not just ASR hallucination description: CreatorTrack's Gus Outdoor voice-profile pipeline: auto-captions transcribe the entire audio track (music, meme soundbites) not just the creator's own speech; the planned ingest quality gate must detect 'this isn't the creator speaking', not just hallucination type: project
2026-07-15: Justin caught a nonsense line in the Gus Outdoor Co voice card ("holy crap Donald Trump hello Peter welcome to fortnite"). Initial read was ASR hallucination; Justin corrected it, the line was an accurate transcription of a meme soundbite Gus had overlaid on the clip, not model hallucination. Same root cause for a separate flagged line that turned out to be real song lyrics from background music.
Why this matters: auto-captions capture whatever is on the audio track, music, sound effects, overlaid meme audio, not just the creator talking. For a voice profile (whose whole point is capturing how the creator talks), an overlaid meme soundbite or song lyric is exactly as useless as a hallucination, arguably worse since it reads as confidently real. 18 of 515 ingested videos (17 music-montage, 1 short) were quarantined as "no usable speech" via a speech-density + music-marker sweep; corpus is now 497 verified spoken-word transcripts.
How to apply: the still-queued ingest quality gate (mentioned as item 2 of 3 follow-ups after Task 16) should be framed as "detect that the captioned audio isn't the creator speaking" (music-marker dominance, speech density, short-clip-matches-meme/lyric-pattern), not "detect ASR hallucination". Perfect detection isn't achievable from captions alone, so the human-readable voice card plus a human eyeballing it stays part of the loop by design, don't try to fully automate this away. Relevant to any future CreatorTrack workspace running the same Scripts/voice-profile ingest pipeline, not just Gus Outdoor.
[auto-memory session 60d132de-4659-485c-8044-427af0ccce7f, confidence 0.68, mode staged]