reference creatortrack scripts autocaption audio quality 20260715T191523
name: CreatorTrack Scripts app: auto-captions capture overlaid audio, not just creator speech description: YouTube auto-captions transcribe the entire audio track (memes, music, sound effects) not just the creator talking; voice-profile ingest must filter for this, not just ASR-hallucination. type: reference
Justin caught a bad line in Gus's Scripts app voice card ("holy crap Donald Trump hello Peter welcome to fortnite") that had been quoted as a prime cold-open example despite being someone else's meme soundbite overlaid on the video, not Gus talking.
Initial framing (auto-captioner "hallucinated" the line) was wrong per Justin's correction: it was an accurate transcription of an overlaid meme sound (the Trump "hello Peter, welcome to Fortnite" soundbite). Same root cause suspected for a "Hogwarts is my home" line, likely real song lyrics from background music.
Why: Auto-captions transcribe the whole audio track, not just the creator's voice. For a voice-profile pipeline, an overlaid meme/song is exactly as useless as ASR hallucination, both teach the generator to imitate someone else's voice.
How to apply: When building or auditing ingest quality gates for CreatorTrack's Scripts/voice-profile pipeline, frame detection as "is the captioned audio actually the creator speaking" (music-marker dominance, speech density, meme/lyric pattern match on short clips) rather than "detect ASR hallucination." Perfect detection isn't achievable from captions alone, the human-readable voice card plus a human eyeballing it stays part of the loop. As of 2026-07-15, 18 of 515 videos were quarantined this way (497 verified-real transcripts remain); a code-level quality gate for future ingests was still queued, not yet built.
[auto-memory session 60d132de-4659-485c-8044-427af0ccce7f, confidence 0.65, mode staged]