Same audio, same recipe (rank 16, α 32, 5 epochs, lr 1e-4, base
sft3/export), same prompts, same seeds. Only the burst labels and the burst spans
differ.
Shriek zero times, used
8 of its 83 labels and gave 70 % of its detections to two labels; asked blind, Gemini 3.8
Flash put the requested burst in its top-3 on 53/60 against the detector's 3/60. An adapter trained
on Gemini's labels may therefore produce a burst the detector cannot name. A null on strict
hit rate is ambiguous, not negative — which is why this page exists and why the ear decides.
Rows are re-bucketed by the label the row's own rendered script names, not by the bucket
the detector originally filed them under: a row whose only Gemini event is an
Exhausted Groan would otherwise train an adapter named affirmative_grunt
on a cue that contradicts its own name. The floor is 100 rows; below that a rank-16 LoRA
on 5 epochs memorises.
| class | Gemini rows | detector-control rows | shipped adapter rows | Gemini adapter | control adapter |
|---|---|---|---|---|---|
| affirmative_grunt | 132 | 17 | 1600 | yes | no |
| breathy_giggle | 131 | 213 | 1600 | yes | yes |
| chuckle | 675 | 227 | 1600 | yes | yes |
| deep_breath | 409 | 201 | 895 | yes | yes |
| exasperated_sigh | 473 | 101 | 1600 | yes | yes |
| exhausted_groan | 150 | 215 | 1600 | yes | yes |
| frustrated_groan | 144 | 32 | 568 | yes | no |
| heavy_breathing | 110 | 13 | 1357 | yes | no |
| humming | 194 | 37 | 348 | yes | no |
| panting | 194 | 17 | 765 | yes | no |
| relief_sigh | 123 | 37 | 1582 | yes | no |
| scream | 154 | 200 | 891 | yes | yes |
| sharp_inhale | 392 | 207 | 1600 | yes | yes |
| wistful_sigh | 121 | 223 | 1600 | yes | yes |
| yawn | 164 | 199 | 1600 | yes | yes |
Excluded (below the 100-row floor, Gemini count in brackets): soft_hum (93), snicker (85), contented_sigh (82), displeased_grunt (57), surprised_gasp (40), effort_grunt (26), ahem (24), cackle (24), growl (22), mournful_wail (21), sniff (20), cough (19), clears_throat (18), deep_breathing (18), pain_moan (18), childlike_giggle (16), guffaw (16), shriek (14), resonant_hum (14), snort (13), fearful_gasp (13), low_mumble (13), pleasure_moan (11), soft_whistle (11), trembling_whimper (8), sharp_whistle (7), tsk (6), fast_breathing (6), lip_smack (4), coughing (3).
Every difference is paired at the prompt level: the runner seeds on the prompt
index alone, so the arms drew identical sampling noise. d hit (strict) is the change
in the fraction of prompted cues the detector named as the target class;
d hit (family) counts a near miss inside the same burst family as a hit.
d wrong and d WER are the guardrails: a hit-rate gain bought with bursts
of the wrong kind, or with speech nobody can transcribe, is not a gain.