Chiron · source-audio audit · 2026-07-13

Are the 8 "source-audio" 10-card renders really the creator's voice?

Or was any one silently rendered with the Lester XTTS clone, the way two earlier _sourceaudio files (wpUCkqhE8v4, mihAR69tVtc) turned out to be?

Verdict: all 8 are the creator's voice. Zero Lester. No tracker row needed flipping.
Two independent speaker methods agree. The deciding one is a resemblyzer d-vector run, whose two references separate cleanly at 0.596, so it can actually tell the two voices apart. Every render sits near the creator and far from Lester, by a wide margin.

Method 1 — resemblyzer d-vector cosine (deciding evidence)

Speaker-identity embedding, the gold standard for this yes/no question. Reference separation creator vs Lester = 0.596 (well apart). Green bar = similarity to creator, rust bar = similarity to Lester.

RendercreatorLestermarginowner

Method 2 — MFCC mean+std cosine (corroboration)

Timbre and channel, not pure speaker identity. Same ranking, creator wins every row, but the margins are thin (0.010 to 0.017), which is exactly why Method 1 is the one the verdict rests on.

RendercreatorLestermarginowner

What this proves, and what it does not

Proven: each of the 8 renders' voice matches a confirmed creator reference (90s pulled from the creator's own YouTube channel, id TrluemyPA2c) far more than it matches a confirmed Lester reference (from the known-Lester chalkboard render wpUCkqhE8v4_chalkboard_LESTER_clean.mp4). A 60s window at t=60s was scored for each render.
Not claimed: this does not audit voices beyond these 8, and it does not check the two earlier flagged files (wpUCkqhE8v4, mihAR69tVtc) — those were separately confirmed Lester and are out of scope here. Confidence is high but not a legal identity proof; it is an embedding-distance decision on a 60s sample per file.

Tracker outcome

The chiron-video-tracker tencard rows for all 8 carry the truthful chiron-sourceaudio-… route and no Lester label. The task said "fix Voice Owner + badge if wrong." Nothing was wrong, so nothing was flipped. Reference: feedback_chiron_tracker_strict_voice_and_render_rules.

Evidence files beside this page: verify_v3_resemblyzer.py + verify_v3_report.json (deciding run), verify_v2.py + verify_v2_report.json (MFCC), the two reference wavs, and eight <slug>_60s.wav clips. Local audit page, not deployed.