Every variant produces the voice-acting caption (GENERAL + per-sentence SCRIPT with cues,
inline bursts and [pause]) for the same 100 clips. Each caption is scored 1–10 by
Gemini-3.5-Flash, which listens to the audio and rates General, How-to-say (delivery
cues), Inline-text (transcription), Vocal-burst accuracy, and Overall. We compare model size (4B vs 8B),
reasoning (Thinking on/off), and the upstream ASR (Parakeet-v3 vs Qwen3-ASR vs VibeVoice).
🧬 Evolutionary prompt search for vocal-burst captioning → — top MOSS-8B prompting structures (find bursts, few hallucinations) validated on 100 clips.
| Variant | General | How-to-say | Inline text | Vocal-burst acc. | Overall | gen/clip |
|---|---|---|---|---|---|---|
| 4B · Thinking | 8.09 | 7.98 | 8.26 | 9.07 | 7.57 | 8.1s |
| 4B · No-Think | 8.11 | 7.95 | 8.32 | 9.07 | 7.56 | 4.3s |
| 8B · Thinking | 8.63 | 8.45 | 8.3 | 9.15 | 7.84 | 20.6s |
| 8B · No-Think | 8.22 | 8.26 | 8.54 | 9.11 | 7.79 | 4.1s |
| 8B·No-Think + BurstLocator | 7.68 | 7.64 | 8.54 | 6.68 | 6.98 | 5.5s |
| 8B·Think + Qwen3-ASR | 8.45 | 8.39 | 8.43 | 9.06 | 7.91 | 18.9s |
| 8B·Think + VibeVoice-HF | 8.64 | 8.59 | 9.03 | 9.16 | 8.23 | 18.6s |
| 8B·Think + VibeVoice-events | 8.24 | 7.9 | 6.27 | 8.77 | 6.48 | 19.4s |
Green = best in column. gen/clip = mean MOSS generation time (compute cost).
Thinking on vs. off:
The MOSS config is held fixed (8B · Thinking); only the ASR transcript/timestamps fed in changes. Because these demo clips have no ground-truth transcript, the fairest quality signal is Gemini's inline-text and overall scores on the resulting caption (Gemini hears the audio). An alternative ASR beat Parakeet-v3.
| ASR into 8B·Think | General | How-to-say | Inline text | Vocal-burst acc. | Overall | gen |
|---|---|---|---|---|---|---|
| 8B·Think + Parakeet-v3 (baseline) | 8.63 | 8.45 | 8.3 | 9.15 | 7.84 | 20.6s |
| 8B·Think + Qwen3-ASR | 8.45 | 8.39 | 8.43 | 9.06 | 7.91 | 18.9s |
| 8B·Think + VibeVoice-HF | 8.64 | 8.59 | 9.03 | 9.16 | 8.23 | 18.6s |
| 8B·Think + VibeVoice-events | 8.24 | 7.9 | 6.27 | 8.77 | 6.48 | 19.4s |
Idea (as requested): run laion/vocalburst-locator at a looser threshold 0.7 (0.9 has many false negatives), pass the flagged time regions to MOSS-8B (no-think) as hints, give it the complete VocalBurst taxonomy, and prompt it to prioritise identifying/labelling bursts (inline + in GENERAL). Result on 95 judged clips:
Finding: the aggressive locator-hint + burst-emphasis approach HURT quality. Pushed to look for bursts, MOSS hallucinated them (e.g. a "Guffaw"/"Chuckle" where the audio has none) or mislabeled the locator's false-positive regions (a breath/gasp called a chuckle). The baseline scores high on burst accuracy precisely by staying conservative (correctly adding none on burst-free clips). Recall went up but precision collapsed. A gentler operating point (higher threshold, "only if clearly audible" prompt) would be the next thing to try. Example clips the judge penalised for a hallucinated/mislabeled burst:
Each card: one clip's audio + the 4B/8B × Thinking/No-Think captions, with that clip's Gemini scores (O=overall) and the judge's one-line comment.