Hugging Face's logo

Datasets:
ACloudCenter
/
fast-vibevoice-notes

Modalities:
Text
Formats:
text
Size:
100K - 1M
Libraries:
Datasets
Dataset card Data Studio Files
xet
 Community
1
fast-vibevoice-notes / MASTER_HANDOFF.md
ACloudCenter's picture
ACloudCenter
Master handoff: complete project record — goals, phases, results, techniques, artifacts, failures, open threads — self-contained for a fresh reviewer
4444a2e verified
preview code
|
Raw History Blame Contribute Delete
12.7 kB

MASTER HANDOFF — Emotion-Steerable Finetuned VibeVoice

Project: Fast, emotionally-steerable, multi-party dialogue TTS on VibeVoice-1.5B Dates: 2026-09-26 → 2026-10-03 | Budget: $250 granted; ~$110 spent, ~$110 unspent Read this first. Every claim below is backed by a banked artifact. Depth lives in the per-phase files listed at the end.

1. ORIGINAL GOALS (user's words, distilled)

  1. VibeVoice-class model fine-tuned so it can be emotionally steered per speaker per turn (documentary: narrator + witnesses; multiparty characters talking back and forth)
  2. Much faster than stock VibeVoice (~1 min audio per 1 min generation on beefy GPU — user wanted a fraction of that)
  3. Keep VibeVoice's strengths: multi-speaker long-form, prosody that carries across turns, voice consistency

2. FINAL DELIVERABLE (what exists NOW, all on the Hub)

Artifact Where Status
Joint-finetuned model (dialogue) ACloudCenter/vibevoice-lora-joint → fullv3/lora/ ✅ ear-certified: intelligible multi-speaker dialogue, WER 0.356 vs stock 0.029, SIM 0.931 vs 0.937
Emotion-tuned model (tags v1) same repo → emo3/lora/ ✅ trained (warm-start from fullv3), eval_loss 0.918→0.841
Emotion+quality model (ship) same repo → emo4/lora/ ✅ trained on clean data (warm-start from emo3), eval_loss 0.67 stable
Merged experimental model ACloudCenter/vibevoice-1.5b-joint-merged ⚠ quality-chaos — DO NOT DEPLOY (kept as evidence)
Curated final samples (27 wavs, 9 story-ordered folders + README) ACloudCenter/vibevoice-final-samples ✅ START HERE for listening
All intermediate samples ACloudCenter/vibevoice-eval-jointv1, ACloudCenter/vibevoice-emotion-probe ✅ ~60 wavs
Research/ops notes (29 files) ACloudCenter/fast-vibevoice-notes (this repo) ✅

Ship recipe (verified): emo4 adapter (load via vibevoice-community lora_loading.py) · 20 DDPM steps (4× high-band vs 10, user-confirmed cleaner) · <emotion> tags in script text · per-speaker clean emotional reference clips · CFG 1.3 · greedy text.

3. THE CENTRAL DISCOVERY (the paper's spine)

VibeVoice's LLM and diffusion head are a tightly co-adapted pair. Modifying the head alone fails catastrophically in closed loop — no matter how well it trains offline.

Evidence chain (all controlled, all banked):

  1. 5 head-only attempts failed (4 ours: v-pred family escalating recipes; 1 user's: CFM/MeanFlow LongFlow) — all trained to better-than-stock offline fit, all produced babble/runaway in real generation
  2. Mechanism found (C/E mismatch 2×2): stock head is condition-robust (teacher-forced 0.96 / closed-loop 0.96); finetuned head collapses offline-on/online-off (0.50 / 1.36) — teacher-forced "leak" training destroys the condition-robustness the loop needs
  3. Joint adaptation works (the control): LLM-LoRA + head trained TOGETHER on DailyTalk dialogue → first finetuned model that speaks (WER 0.356, ear-verified). n=5 failures vs n=1 success, one variable changed.
  4. Stability-boundary corollary: any weight perturbation (3 independent merges, extra epochs) reshuffles WHICH samples fail (WER band 0.32–0.41 adapter / 0.53–0.79 merged), mean barely moves. Only stock is fully stable.
  5. Reference conditioning is TOTAL: reference clips transmit emotion AND accent AND noise floor into generation (F0 300→184 Hz across emotion arms; accents transferred; noise floor transferred 60×).

4. PHASE-BY-PHASE (chronological, what/learned/cost)

Phase What happened Key result Cost
P0 Grounding Trainer CLI/data contract verified; manifest prep built; preflight harness born prep_dailytalk.py, preflight.py ~$5
P1 Head-only attempts (v1-v4 + LongFlow) 4 escalating head recipes — all offline-good, closed-loop babble Negative result + diagnosis target ~$15
P2 Diagnosis 2×2 teacher-forced vs closed-loop matrix; capture pipeline (own rollouts) C/E mismatch mechanism ~$10
P3 User's LongFlow head CFM/MeanFlow head trained on captured states Offline-good, loop-babble; KEEPING stopping proved stopping/content separability ~$5
P4 Joint finetune v4 LLM-LoRA r8 + full head FT + connectors, 950 steps DailyTalk, non-reentrant ckpt BREAKTHROUGH: speaks. WER 0.356/SIM 0.931, ear-verified ~$5
P5 Eval + polish 9-dialogue closed-loop harness (6-layer fixes); 2-epoch polish (eval_loss 0.968→0.888) joint2 = statistical wash vs fullv3; chaos confirmed 3rd way ~$12
P6 Merge campaign 3 merges (bf16/fp32/re-run), all mechanically clean RTF fixed (1.56→~1.0) but WER chaos 0.53–0.79 → MERGE CLOSED ~$8
P7 Speed campaign cProfile + ladder: LLM 47%, head 8%, python glue 44%; compile 1.40 vs 1.37 (no gain); vLLM absent (0 hits) Speed verdict: launch-bound; honest ceiling ~1.3–1.5×; documented for the Claude session ~$6
P8 Emotion research Sub-agent lit scan: R1 reference-clip emotion (community+paper verified), R3 tags, IndexTTS2 mechanism, datasets EmoV-DB/CREMA-D/RAVDESS/ESD mapped ~$0
P9 Emotion probes v1 CREMA-D (F0 300→184 Hz); v2 palette/drama/switch (user: "I hear the emotions"); v3 RAVDESS US-accent 8-emotion palette R1 ADOPTED; accent transfer discovered (fixed via RAVDESS) ~$8
P10 R3 emotion-LoRA Warm-start from fullv3, 800 steps CREMA-D tags (+DailyTalk anti-forgetting); 3 tag-eval runs (2 with named bugs, fixed) Tags shift F0 −87 Hz with refs held constant; R3 ADOPTED as tag+matching-refs; tags-alone → voice-drift artifact (documented negative) ~$12
P11 Quality diagnosis + emo4 Noise-floor A/B: static is REFERENCE-carried (CREMA-D floor 0.003→gen 0.0059 vs RAVDESS 0.0001); 20 DDPM = 4× high-band; emo4 = 1600 steps on clean RAVDESS+filtered data Quality fixed: emo4 banked, eval_loss 0.67 stable; ship = 20 DDPM ~$10

5. RESULTS TABLE (the numbers)

Configuration WER (closed loop) SIM Duration behavior RTF (a10g-small)
stock 0.029 0.937 sane 0.89–1.14
fullv3 joint (ship lineage) 0.356 (band 0.32–0.41) 0.931 sane 1.36–1.79
joint2 polish 0.317/0.407 (2 runs) 0.929–0.931 sane ~1.43–1.59
emo3 (tags v1) not WER-eval'd — — —
emo4 (clean scale-up, 20 DDPM) ear-check pending — full-length in eval draws ~1.53 @ 20 steps
merged ×3 0.53–0.79 0.90–0.92 chaotic ~1.0
All 5 head-only ~1.0 (babble) 0.45–0.70 runaway 137–172 s —
Run-to-run WER variance ±0.1 (DDPM sampling noise); WER is single-draw noisy — multi-seed eval recommended for future claims.

6. TECHNIQUES / RECIPES (copy-paste ready)

Training recipe (the one that works): LLM-LoRA r=8 α=32 on q/k/v/o/gate/up/down + FULL diffusion-head FT + connector adapters · CE(text) 0.04 + diffusion MSE 1.4 · bs1×accum16 · lr 2.5e-5 fresh / 1e-5 continue · cosine 3% warmup · max_grad_norm 0.8 · gradient_checkpointing use_reentrant:false (REQUIRED for LoRA grads) · voice_prompt_drop 0.2 · DailyTalk-format JSONL {text, audio, voice_prompts:[...]} (voice_prompts MUST be lists). Warm-start recipe: load previous stage via peft's own _load_language_model (raw load_state_dict is a key-name no-op!) + diffusion_head_full.bin + connector .bins — chain: base → fullv3(dialogue) → emo3(tags) → emo4(quality). Emotion recipes: R1 = emotional reference clips per speaker slot (zero training; verified). R3 = <emotion> plain-text tags + emotion-matched DIFFERENT-clip same-actor prompts (800 steps minimum; tags-alone → voice drift; ship = tags + matching refs). R4 (unbuilt): IndexTTS2-style emotion-embedding prefix (arXiv 2506.21619 mechanism). Eval recipe: 9 fixed 2-speaker dialogues (3 tiers), Carter/Maya refs, 10 DDPM (now 20), CFG 1.3, greedy; Whisper-large-v3 return_timestamps=True; resemblyzer SIM; per-arm F0/RMS garnish; user ears = certified instrument (Whisper hallucinates fluent text on babble; F0 ≠ perceived emotion). Inference quality knobs: 20 DDPM steps (+13% time, 4× high-band) · clean references (noise floor < 0.001 filter) · noise is ref-carried (60× measured).

7. ENVIRONMENT PINS + GOTCHAS (proven; full detail in preflight-discipline.md)

torch==2.6.0 · transformers==4.51.3 (4.57+ rejects checkpoint) · peft==0.15.2 (≥0.17 conflicts diffusers 0.29) · diffusers==0.29.0 · datasets==3.6.0 · setuptools<81 (resemblyzer→webrtcvad pkg_resources) · librosa 0.11 · code: vibevoice-community/VibeVoice @ d5895e7 (inference+lora_loading), voicepowered-ai/VibeVoice-finetuning @ f743686 (trainer). GOTCHAS (each cost a job): HF cache = relative symlinks (copyfile, never os.replace) · trainer -e . upgrades torch (re-pin after) · sys/importlib unbound in patched main() contexts · nested-quote string surgery fails (patch in committed .sh files) · dataclass exec needs sys.modules registration · eval OOM → expandable_segments + small val · a10g-large sometimes = 22.3GB · aggregate 'loss' logging spikes are accumulation-window artifacts · log tail only ~20 lines (grep pushed logs via uv jobs) · wrapper order: install deps BEFORE preflight · --label_names input_ids required for eval_loss with this collator · early_stopping_patience is a callback arg not a TrainingArguments field.

8. OPERATIONS DISCIPLINE (institutionalized after failures)

preflight.py (binaries/exact-pins/repos/paths) runs before every job · kill-fast on failed asserts · every submission named ml-intern-* · verify pushes within minutes · one variable per run · user ears = certification instrument · foreign repo writes (tucson/az-civic/pima) = the user's OTHER project, no conflict, shared budget only.

9. OPEN THREADS (handed off / optional)

  1. Speed research — IN PROGRESS by Claude (separate session): brief = speed-verdict-for-research.md. Closed levers: merge/compile/vLLM/DDPM-reduction. Open: python-glue surgery (44% of gen time; ceiling ~1.3–1.5×), CFG batching, production batching, hardware scaling, distillation. Chaos constraint: any weight/precision change must be re-evaluated closed-loop.
  2. R3 scale-up (optional, ~$10): 1600 steps on clean US emotional data (RAVDESS done; add EmoV-DB US actors w/ noise filter; Expresso = premium if licensed) — stronger tag channel. emo4 is the warm-start.
  3. SOAR on-trajectory training (optional, ~$15–20): the only method targeting the stability-boundary chaos; train on the model's own rollouts. Scripts sketched: train_soar.py/train_soar_v2.py in this repo (earlier head-only era — needs rework for joint model).
  4. Paper (optional, ~$0–2): complete arc banked; spine = §3 evidence chain; systems findings (peft tax, merge chaos, compile dead-end, ref-quality transfer) + emotion-steering wedge as the positive result.
  5. Multi-speaker >3: untested post-finetune; stock demo caps at 4 slots; probe ran 3 successfully.

10. FILE INDEX (this repo, fast-vibevoice-notes)

Handoff/state: MASTER_HANDOFF.md (this) · project-state.md · session-resume.md · master-summary.md Research: emotion-research.md · speed-verdict-for-research.md · paper-feasibility.md · eval-design.md · longflow-audit.md Results: eval-results-joint.md · eval-results-joint2.md · ear-check-joint.md · emotion-probe-verdicts.md · baseline-results.md · v3-results.md · diagnosis-results.md · eval-results.md · eval-results-v4.md Ops: preflight-discipline.md · preflight.py · session logs (emo-train-log*.txt, merge-log*.txt, eval-log-.txt, quality-probe-log.txt, emotion-probe-.txt) Code: prep_dailytalk.py (manifest builder) · emo_data_prep.py · emo4_data_prep.py · eval_joint.py (canonical closed-loop eval) · eval_2x2.py · emotion_probe.py/_v2/_v3 · emotion_tagged_eval.py · quality_probe.py · speed_profile.py · speed_cprofile.py · speed_ladder_adapter.py · collect_final_samples.py · train_soar.py/_v2 (legacy) · wrappers (*.sh — proven job chains) Data/capsules: v4-caps.tar.gz (capture-era), epoch2-losses/summaries, mf640_gen.py (LongFlow-era)

11. HOW TO VERIFY (30 minutes)

  1. Listen: vibevoice-final-samples in folder order (README has the guide)
  2. Metrics: this file §5; per-phase files for raw numbers
  3. Weights: vibevoice-lora-joint → emo4/lora (ship), fullv3/lora (dialogue-only)
  4. Reproduce an eval: eval_joint.py + emotion_probe_v3.py + emotion_tagged_eval.py with the pinned env (§7)
  5. Resume anything: session-resume.md + preflight.py + the wrappers