Measured locator and decoding recommendations
Recommended starting settings
MOSS decoding: temperature 0.0 (greedy), one beam, KV cache enabled. Sampling at 0.3 or 0.7 with top_p=0.9, top_k=0 did not improve caption fidelity consistently in this pilot. Keep a seed when sampling. These recommendations concern MOSS generation, not the temperature of the reviewer, which stayed at zero.
Locator: start at onset threshold 0.6, minimum interval duration 0.08 s, and overlap suppression at 0.5 of the shorter interval. This overlap rule is not IoU 0.5. Keep all selected events rather than a fixed top-three cap for dataset annotation. Preserve the raw candidate list alongside the selected intervals.
- 0.5: recall-oriented; it was selected on the first ten clips.
- 0.6: a provisional practical compromise; highest composite on the ten confirmation clips.
- 0.7: more conservative; fewer false proposals, with lower recall in this pilot.
- 0.8–0.9: too many omissions in the selection recordings to recommend as a general default.
The selected threshold was not the confirmation winner. This small study therefore does not establish a universal optimal threshold or a statistically reliable temperature advantage.
python inference.py recording.mp3 --temperature 0.0
# In the burst model repository: full audio plus both event endpoints
python inference.py recording.mp3 --start 4.2 --end 5.1 --temperature 0.0
# Optional sampled comparison, matching the tested sampling policy
python inference.py recording.mp3 --temperature 0.3 --top-p 0.9 --seed 20261011
Use select_events.py on the separately produced Humaneness Ears event proposals; MOSS is the captioner and does not itself return these detector probabilities. Humaneness Ears Medium and MOSS require separate Transformers environments as described in the model README.
Experimental protocol
Twenty complete recordings: ten synthetic acting performances and ten Emilia speech clips. Ten (five per source) selected settings; ten different clips confirmed the locked shortlist and greedy threshold controls. These are the same twenty recordings used in the earlier procedural-caption study, so the confirmation split is held out only from this MOSS decoding search, not an independent human benchmark. The grid contains five thresholds × three temperatures = fifteen combinations. Generation was performed on four GH200 GPUs; repeated interval/temperature requests were cached across threshold settings. Nonzero temperatures have one seeded generation each; there is no multiple-seed study.
The voice model gets the whole recording. The burst model also gets the whole recording, with start AND end seconds; it receives no intended emotion, script, type label or reference caption. The Gemini Flash audio reviewer listened to the actual audio at fixed temperature 0. Candidates were anonymized and shuffled. The reviewer rated whole-caption fidelity, interval-caption fidelity, timing, event precision and recall on a 0–5 scale.
The composite uses 30% whole-caption fidelity, 25% burst-caption-set fidelity, 20% timing, 15% event precision and 10% event recall. Empty predicted sets receive burst-caption-set fidelity zero when the reviewer heard events, and five when it heard no events. † The table's burst-caption column is the conditional mean over proposed event captions and excludes empty sets; the composite instead uses a clip mean including that empty-set policy. ‡ Precision and recall columns are reviewer ratings, not measured precision/recall fractions or percentages.
Selection: ten clips
| Threshold | Temperature | Clips | Events | Voice caption | Burst caption† | Timing | Event precision‡ | Event recall‡ | Composite |
|---|---|---|---|---|---|---|---|---|---|
| 0.5 | 0.0 | 10 | 36 | 3.70 | 3.36 | 3.40 | 3.50 | 3.80 | 3.57 |
| 0.5 | 0.3 | 10 | 36 | 3.40 | 3.33 | 3.40 | 3.50 | 3.80 | 3.45 |
| 0.5 | 0.7 | 10 | 36 | 3.70 | 3.33 | 3.40 | 3.50 | 3.80 | 3.56 |
| 0.6 | 0.0 | 10 | 29 | 3.70 | 3.34 | 3.00 | 3.50 | 3.30 | 3.24 |
| 0.6 | 0.3 | 10 | 29 | 3.40 | 3.34 | 3.00 | 3.50 | 3.30 | 3.15 |
| 0.6 | 0.7 | 10 | 29 | 3.70 | 3.34 | 3.00 | 3.50 | 3.30 | 3.23 |
| 0.7 | 0.0 | 10 | 17 | 3.70 | 3.71 | 3.10 | 3.70 | 3.10 | 3.30 |
| 0.7 | 0.3 | 10 | 17 | 3.40 | 3.71 | 3.10 | 3.70 | 3.10 | 3.19 |
| 0.7 | 0.7 | 10 | 17 | 3.70 | 3.65 | 3.10 | 3.70 | 3.10 | 3.31 |
| 0.8 | 0.0 | 10 | 7 | 3.70 | 3.71 | 3.20 | 3.90 | 2.60 | 3.31 |
| 0.8 | 0.3 | 10 | 7 | 3.40 | 3.57 | 3.20 | 3.90 | 2.60 | 3.19 |
| 0.8 | 0.7 | 10 | 7 | 3.70 | 4.00 | 3.20 | 3.90 | 2.60 | 3.34 |
| 0.9 | 0.0 | 10 | 1 | 3.70 | 4.00 | 2.60 | 3.70 | 2.40 | 2.77 |
| 0.9 | 0.3 | 10 | 1 | 3.40 | 4.00 | 2.60 | 3.70 | 2.40 | 2.69 |
| 0.9 | 0.7 | 10 | 1 | 3.70 | 4.00 | 2.60 | 3.70 | 2.40 | 2.77 |
Locked confirmation: ten other clips
| Threshold | Temperature | Clips | Events | Voice caption | Burst caption† | Timing | Event precision‡ | Event recall‡ | Composite |
|---|---|---|---|---|---|---|---|---|---|
| 0.5 | 0.0 | 10 | 74 | 3.90 | 3.38 | 3.30 | 3.20 | 3.80 | 3.55 |
| 0.5 | 0.3 | 10 | 74 | 3.80 | 3.31 | 3.30 | 3.20 | 3.80 | 3.51 |
| 0.5 | 0.7 | 10 | 74 | 3.80 | 3.27 | 3.30 | 3.20 | 3.80 | 3.50 |
| 0.6 | 0.0 | 10 | 63 | 3.90 | 3.52 | 3.30 | 3.40 | 3.60 | 3.60 |
| 0.7 | 0.0 | 10 | 50 | 3.90 | 3.54 | 3.20 | 3.70 | 3.10 | 3.41 |
Selection winner: 0.5 / greedy, confirmed composite 3.55/5. The 0.6 greedy control scored 3.60/5 on confirmation. The descriptive paired score difference between 0.5 and 0.6 over all twenty clips is +0.138, with a clip-bootstrap 95% interval [-0.075, +0.409]; this is not evidence for a universal threshold advantage.
Separate original listening review
The earlier review covered 50 clips, 141 proposed events, using a different listening selection (top three intervals, onset threshold 0.6). Whole captions averaged 3.94/5, event captions 3.68/5, and intervals 3.71/5. That set contained 45 English and five German clips. These scores are not a paired before/after comparison with the twenty-clip search.
Historical procedural caption search
The deterministic procedural pipeline separately selected 0.7 + overlap suppression, a speech-only reference of 80,265 Medium-scored clips (41,771 S10 + 38,494 multilingual Emilia), three distinctive VoiceNet phrases, up to two emotions with z ≥ +0.5, and a two-sentence template. Tempo is always included; demographic and authenticity wording is omitted while raw predictions remain saved. Content Enjoyment supplies pleasantness wording. Blend is a continuous regression target and never marks a clip burst-free.
| Paired procedural setting | Selection, 10 | Confirmation, 10 | Descriptive combined, 20 |
|---|---|---|---|
| Original captions | 2.675 | 2.485 | 2.580 |
| Reviewed default c12 | 3.735 | 3.405 | 3.570 |
The procedural composite has a different rubric (caption 30%, timing 25%, precision 15%, recall 15%, consistency 10%, readability 5%). Its scores cannot be compared directly to the MOSS composite above.
Artifacts and limits
Machine-readable configuration and results are in recommended_settings.json and evaluation/locator_decoding_pilot.json; procedural results are in evaluation/procedural_caption_pilot.json. Raw numeric predictions and historical normalization are retained. The original-50 review plus all procedural and decoding review calls in the shared budget ledger have an estimated total list-price cost of $0.5033, within the $10 ceiling. This is an estimate from token usage, not a billing invoice.
These scores come from an automated audio reviewer. Neither captions nor event boundaries were manually adjudicated. Small sample size, prior tuning on these sources, one sampling seed, and imperfect reviewer judgments limit generalization. Use the settings as a documented starting point and inspect events from your own domain.
Current procedural format and paired 50-audio review
Current procedural settings are threshold 0.7, two VoiceNet properties, up to one emotion at z ≥ +0.5, compact tags and the same 80,265-clip speech reference. Every spoken segment and selected event has a numeric duration; pauses use square brackets. Exact grammar and overlap policy.
On the same 50 recordings (45 EN, five DE), the shared audio-judge composite is 3.863/5 for uncapped greedy MOSS and 2.704/5 for compact procedural captions. Twenty-five recordings selected the procedural configuration, and twenty-five different recordings confirmed it. Compact tags tied previous prose overall; there is no demonstrated general wording-quality gain. These are automated reviewer ratings, not human benchmark metrics. This shared rubric differs from the earlier twenty-clip search.
Listen to all 50 paired annotations and inspect six Humaneness Voice training formats · Complete paired metrics · Procedural code and defaults.