Upload showcase-log.txt with huggingface_hub

v3: strip_tags removes ONLY the <tag> token (keeps 'Speaker N:' prefixes — the v1 untagged arm crashed on speaker-less lines)

Upload showcase-log.txt with huggingface_hub

v3: strip_tags removes ONLY the <tag> token (keeps 'Speaker N:' prefixes — v1/v2 untagged arm crashed on speaker-less lines)

Upload showcase-log.txt with huggingface_hub

v3: strip_tags removes ONLY the <tag> token (keeps 'Speaker N:' prefixes — untagged arm crashed on speaker-less lines)

Upload showcase-log.txt with huggingface_hub

Upload showcase-log.txt with huggingface_hub

v3: strip_tags keeps 'Speaker N:' prefixes (removes only <tag> token) — fixes untagged-arm parse crash

v2: extended to ~1,500 words/40 turns (true 10-min target), strip_tags keeps Speaker prefixes via re.sub

v2: regex tag-strip (keeps Speaker prefixes — v1's split killed speaker lines), extended script ~1,400 words (~10 min), guard N_WORDS>=1200

10-min showcase v3: extended ~40-turn script, single-pass + chunked (3 acts) strategies, emo4 @ 20 DDPM

Upload showcase-log.txt with huggingface_hub

10-min showcase wrapper (literal quoting): fetch showcase_10min.py (no patch), preflight incl. emo4, run, log push

10-min 4-speaker showcase: 'The Last Keeper of Cape Ross' documentary, emo4 @ 20 DDPM, tagged + untagged arms, F0/RMS garnish

Master handoff: complete project record — goals, phases, results, techniques, artifacts, failures, open threads — self-contained for a fresh reviewer

Copy manifest: curated final samples from eval-jointv1 + emotion-probe → vibevoice-final-samples with story-ordered folders

Upload tagged-eval-emo4-log.txt with huggingface_hub

emo4 tagged eval wrapper (literal-quoting patch: ADAPTER_SUB→emo4/lora, DDPM steps 10→20)

Upload emo4-train-log.txt with huggingface_hub

emo4 scale-up wrapper: DailyTalk 500 + emo4 prep (clean-filtered), warm-start from EMO3 via peft loader (_ilu2/_sys2 fixed), 1600 steps, push emo4/

emo4 prep: FULL RAVDESS (noise-floor filtered, actor-paired) + clean-filtered CREMA-D + DailyTalk sample → mixed_train/val

Upload quality-probe-log.txt with huggingface_hub

Quality probe: 4 arms (CREMA-D refs / RAVDESS refs / stock studio refs / RAVDESS@20DDPM) with noise-floor + high-band metrics on refs AND generations

User ear verdicts: R1 confirmed; probe v2 emotions audible; R3 tags+matching-refs WINS (clip 3), tags-alone insufficient + voice-drift artifact (clip 2)

Upload tagged-eval-v3-log.txt with huggingface_hub

Fix: direct full-word emotion compare (ANGRY/SAD/NEUTRAL as measured), no 3-letter remap

Upload tagged-eval-v2-log.txt with huggingface_hub

v2: matching-refs arm actually uses ANGRY/SAD refs (v1 bug overrode all arms to neutral)

Upload tag-eval-log.txt with huggingface_hub

Upload tagged-eval-log.txt with huggingface_hub

Tagged eval: emo3 adapter, 4 arms — no-tags control, tag+neutral-refs (stretch), tag+matching-refs (trained pairing); F0/RMS garnish

Tagged eval: emo3 model, same dialogue × 5 tag arms (untagged/angry/sad/mixed/happy), neutral refs held constant, F0/RMS garnish

Upload emo-train-log-v4.txt with huggingface_hub

Fix: import sys as _sys2 in inserted block (sys unbound in finetune main)

Upload emo-train-log-v3.txt with huggingface_hub

Fix: _ilu2.module_from_spec (importlib name unbound in patch context)

Upload emo-train-log-v2.txt with huggingface_hub

v2: warm-start LM via peft's own loader (_load_language_model — fixes 730-missing-keys no-op); data fix retained

Fix: voice_prompts as LIST (matches DailyTalk schema — pyarrow mixed-type crash)

Upload emo-train-log.txt with huggingface_hub

R3 emotion-LoRA warm-start wrapper: prep mixed data, fetch fullv3 as init, warm-start patch, 800-step train, push emo3/

R3 data prep: CREMA-D emotion-tagged turns (6 emotions, 250/emo cap, self-paired voice prompts) + DailyTalk sample → mixed_train/val.jsonl

Upload emotion-probe-v3-log.txt with huggingface_hub

v2: dynamic cast selection (first 3 actors complete in all 8 emotions, prefer even/odd/even mix)

Upload emotion-probe-v3-log.txt with huggingface_hub

Probe v3: RAVDESS US-accent refs (8 emotions, mixed-gender cast 02F/03M/04F), palette + drama + switch dialogue

Upload emotion-probe-v2-log.txt with huggingface_hub

Probe v2 wrapper: setup → preflight → plain fetch (v2 script standalone) → run → log push

Probe v2: 6-emotion palette (same sentence) + 3-speaker drama + per-turn switch dialogue; 3 actors, all 6 CREMA-D emotions