fast-vibevoice-notes / collect_final_samples.py
ACloudCenter's picture
Copy manifest: curated final samples from eval-jointv1 + emotion-probe β†’ vibevoice-final-samples with story-ordered folders
917463d verified
Raw History Blame Contribute Delete
6.71 kB
"""collect_final_samples.py β€” gather the curated final audio set into one repo.
Manifest: (source_repo, source_filename, dest_folder/dest_name). Fails loudly on any miss.
Prints COLLECT DONE. Also writes the README.
"""
import os
from huggingface_hub import HfApi, hf_hub_download
EVAL = "ACloudCenter/vibevoice-eval-jointv1"
PROBE = "ACloudCenter/vibevoice-emotion-probe"
DEST = "ACloudCenter/vibevoice-final-samples"
MANIFEST = [
# 01 β€” the stock baseline (what we started from)
(EVAL, "stock_medium_0.wav", "01_dialogue_baseline_stock/stock_medium_0.wav"),
(EVAL, "stock_long_0.wav", "01_dialogue_baseline_stock/stock_long_0.wav"),
# 02 β€” the ear-verified joint model (fullv3): 5 prior head attempts failed; this speaks
(EVAL, "joint_short_2.wav", "02_dialogue_our_model_fullv3/joint_short_2.wav"),
(EVAL, "joint_medium_2.wav", "02_dialogue_our_model_fullv3/joint_medium_2.wav"),
(EVAL, "joint_long_0.wav", "02_dialogue_our_model_fullv3/joint_long_0.wav"),
(EVAL, "joint_long_1.wav", "02_dialogue_our_model_fullv3/joint_long_1.wav"),
# 03 β€” the polished twin (emo-lineage predecessor, joint2)
(EVAL, "joint2_short_1.wav", "03_dialogue_polished_joint2/joint2_short_1.wav"),
(EVAL, "joint2_medium_0.wav", "03_dialogue_polished_joint2/joint2_medium_0.wav"),
(EVAL, "joint2_long_2.wav", "03_dialogue_polished_joint2/joint2_long_2.wav"),
# 04 β€” emotion steering R1: reference-clip emotion (CREMA-D actors, mixed arm = the money clip)
(PROBE, "emo_ANGRY_ANGRY.wav", "04_emotion_reference_steering_cremad/emo_ANGRY_ANGRY.wav"),
(PROBE, "emo_SAD_SAD.wav", "04_emotion_reference_steering_cremad/emo_SAD_SAD.wav"),
(PROBE, "emo_ANGRY_SAD.wav", "04_emotion_reference_steering_cremad/emo_ANGRY_SAD_mixed.wav"),
# 05 β€” emotion steering R1, US-accent: the 8-emotion palette (same sentence, same voice)
(PROBE, "v3_palette_NEU.wav", "05_emotion_us_accent_palette/v3_palette_NEUTRAL.wav"),
(PROBE, "v3_palette_CAL.wav", "05_emotion_us_accent_palette/v3_palette_CALM.wav"),
(PROBE, "v3_palette_HAP.wav", "05_emotion_us_accent_palette/v3_palette_HAPPY.wav"),
(PROBE, "v3_palette_SAD.wav", "05_emotion_us_accent_palette/v3_palette_SAD.wav"),
(PROBE, "v3_palette_ANG.wav", "05_emotion_us_accent_palette/v3_palette_ANGRY.wav"),
(PROBE, "v3_palette_FEA.wav", "05_emotion_us_accent_palette/v3_palette_FEARFUL.wav"),
(PROBE, "v3_palette_DIS.wav", "05_emotion_us_accent_palette/v3_palette_DISGUST.wav"),
(PROBE, "v3_palette_SUR.wav", "05_emotion_us_accent_palette/v3_palette_SURPRISED.wav"),
# 06 β€” 3-speaker drama scene (angry / fearful / surprised)
(PROBE, "v3_drama_04-ANG_01-FEA_03-SUR.wav", "06_emotion_3speaker_drama/v3_drama.wav"),
# 07 β€” per-turn emotion switch (same actor angry -> relieved mid-dialogue)
(PROBE, "v3_switch_04-ANG_to_HAP.wav", "07_emotion_per_turn_switch/v3_switch.wav"),
# 08 β€” SHIP CONFIG: emo4 @ 20 DDPM steps, tags + matching refs
(PROBE, "tag2_no_tags_neu_refs.wav", "08_ship_config_emo4_20ddpm/tag2_control_no_tags.wav"),
(PROBE, "tag2_tags_neu_refs.wav", "08_ship_config_emo4_20ddpm/tag2_tags_neutral_refs.wav"),
(PROBE, "tag2_tags_match_refs.wav", "08_ship_config_emo4_20ddpm/tag2_tags_matching_refs_BEST.wav"),
# 09 β€” the quality A/B (webcam-mic noise was reference-carried)
(PROBE, "q_A_crefs_10ddpm.wav", "09_quality_ab/q_A_cremad_refs_noisy.wav"),
(PROBE, "q_B_rrefs_10ddpm.wav", "09_quality_ab/q_B_ravdess_refs_clean.wav"),
(PROBE, "q_D_rrefs_20ddpm.wav", "09_quality_ab/q_D_ravdess_refs_20ddpm_crisp.wav"),
]
README = """# VibeVoice Final Samples β€” the whole story in audio
Joint-finetuned VibeVoice-1.5B (LLM-LoRA + full diffusion head, warm-started chain
fullv3 -> emo3 -> emo4). All samples generated with our emotion-steering stack.
## Listening order
1. **01_dialogue_baseline_stock** β€” stock VibeVoice (the read-only baseline)
2. **02_dialogue_our_model_fullv3** β€” OUR model, ear-certified intelligible dialogue
(WER 0.36 vs stock 0.03; 5 prior head-only attempts produced babble β€” this is the joint-training proof)
3. **03_dialogue_polished_joint2** β€” the polished twin (more training: a wash statistically, different samples)
4. **04_emotion_reference_steering_cremad** β€” R1: swap the reference clip's EMOTION (same actors,
same dialogue): ANGRY_ANGRY / SAD_SAD / **ANGRY_SAD_mixed** (the per-speaker money clip).
Note: CREMA-D actors carry non-US accents β€” the accent transfer is itself evidence of total reference conditioning.
5. **05_emotion_us_accent_palette** β€” R1 US-accent (RAVDESS): same sentence, same voice, 8 emotions
(measured F0 tracks emotion: ANGRY 328 Hz ... SAD lowest)
6. **06_emotion_3speaker_drama** β€” 3-speaker scene: angry / fearful / surprised in one conversation
7. **07_emotion_per_turn_switch** β€” same actor goes ANGRY -> relieved mid-dialogue (dual-slot authoring)
8. **08_ship_config_emo4_20ddpm** β€” THE SHIP CONFIG: emo4 (clean-data scale-up) at 20 DDPM steps,
<emotion> tags + matching references. tag2_tags_matching_refs_BEST.wav = the final product demo.
9. **09_quality_ab** β€” the quality diagnosis: the 'webcam-mic static' was REFERENCE-carried
(CREMA-D refs: noise floor 0.006 in generation vs 0.0001-0.0004 with clean refs; 20 DDPM steps = 4x high-band)
## Ship recipe
emo4 adapter (vibevoice-lora-joint/emo4) | 20 DDPM steps | script tags <angry>/<sad>/... |
per-speaker emotional reference clips (clean recordings) | CFG 1.3
## Metrics
- Intelligibility: our model WER 0.32-0.41 band vs stock 0.03 (single-run; run-to-run +-0.1)
- Speaker identity: SIM 0.92-0.93 vs stock 0.94
- Emotion: F0 shifts 300->184 Hz (reference steering); tag channel -87 Hz with refs held constant
- Speed: architecture-bound ~1.4-1.5x slower than stock (launch-bound; measured ceiling)
"""
def main():
api = HfApi()
ok, miss = 0, []
for src_repo, src_name, dest in MANIFEST:
try:
local = hf_hub_download(src_repo, src_name, repo_type="model")
api.upload_file(path_or_fileobj=local, path_in_repo=dest, repo_id=DEST)
ok += 1
print(f"OK {dest}", flush=True)
except Exception as e:
miss.append((src_repo, src_name, str(e)[:120]))
print(f"MISS {src_repo}/{src_name}: {e}", flush=True)
api.upload_file(path_or_fileobj=README.encode(), path_in_repo="README.md", repo_id=DEST)
print(f"README OK", flush=True)
print(f"COLLECTED {ok}/{len(MANIFEST)}; missing {len(miss)}", flush=True)
if miss:
for m in miss:
print("MISSING:", m, flush=True)
print("COLLECT DONE", flush=True)
if __name__ == "__main__":
main()