GROUPS.md

The 23 vocal-burst class groups

23 groups over 118 vocal-burst label strings, each group holding names that denote the same or a very similar sound. Seeded semantically, then checked against the measured annotator confusion using directed lift, P(top1=B | requested=A) / P(top1=B), which divides out the frequency of the two attractor labels (deep_breath and exasperated_sigh are top-1 for 29 % of all events regardless of what was requested; merging on raw confusion would book a generation failure as a hit). Drift into an attractor measures lift 1.6-2.8, a genuine one-sound-two-names pair 5-24.

Why this exists

Many of our burst class names denote the same sound. snicker and chuckle are the same laugh; shriek and scream are the same cry under two names. Scoring only exact name agreement throws away successful generations because the annotator chose a different word.

The trap this scheme avoids

deep_breath and exasperated_sigh are the annotator's top-1 label for 29 % of all events regardless of what was requested — when the model cannot produce the sound asked for, it falls back on a breathy exhale. Grouping on raw confusion would book that generation failure as a hit. Groups here are checked with directed lift, P(top1=B | requested=A) / P(top1=B), which divides the attractor's own frequency out. The two populations separate cleanly: drift into an attractor measures lift 1.6–2.8, a genuine one-sound-two-names pair measures 5–24.

Measured effect on generation hit rate

	mean hit rate
exact class name	0.302
the 23 groups	0.537
15 coarse families	0.609
random grouping, identical group sizes (300 draws)	0.355
net gain over the random control	+0.182

Any grouping raises a hit rate, because a group is a bigger target. Only the last row is a result. Classes at or above a 0.15 hit rate: 28 exact, 44 grouped, 34.1 under the random control.

The group sob is semantically correct and empirically a failure: its members score 0.001-0.008 and the group scores below its own random control. Grouping does not substitute for the model being able to produce the sound.

Effect on usable training material

At a floor of 100 annotated segments in both corpus halves, 16 individual classes clear the bar (90,924 of 126,512 segments usable); grouped, 10 groups clear it (104,694 segments). Grouping does not create more trainable units -- it creates more material per unit. The bottleneck is the real half: hiccup has 2 real segments, swallow 0, hiss 3.

Are the groups right? A test each member had to pass

Groups were seeded semantically, so each member label was then checked against the data. The test is own-group lift — P(label | a member of the label's group was requested) / P(label) — and not the raw own-group share, because the share is depressed for attractors too: deep_breath concentrates in its own group only 14 % of the time and still plainly belongs in breath_calm.

58 of 58 testable member labels are enriched in their own group. Two were not, and both were removed on 2026-09-04:

• sigh — own-group lift 0.68 over 116 emissions; 2.6 % came from a sigh_pos request and none from sigh_neg. A generic fallback, spread evenly. Left unmapped. Removing it also disarms a substring collision: a substring matcher reads Exasperated Sigh as matching sigh, which would silently merge sigh_neg into sigh_pos -- the one pair these groups exist to keep apart.
• groan — own-group lift 1.46 over 37 emissions; 8.1 % came from a sigh_neg request, most of the rest from hum, sob and swallow. Left unmapped.

Labels whose group contains no requested class cannot be tested this way and are excluded rather than scored zero.

The groups

group	coarse family	members	annotated segments (real / DramaBox)	clears training floor
laugh_soft	laugh	chuckle, breathy_giggle, childlike_giggle, snicker, nervous_giggle, snorting_giggle, giggle, snickering, snickering_giggle, chuckling, titter	1,037 / 14,528	yes
laugh_loud	laugh	cackle, guffaw, laughter, laugh	42 / 3,983	no
scream	scream	scream, shriek, screech, screaming, yell, shout	199 / 4,169	yes
wail	scream	mournful_wail, crying	25 / 364	no
sigh_neg	sigh	exasperated_sigh, frustrated_groan, exhausted_groan, displeased_grunt, effort_groan	928 / 22,684	yes
sigh_pos	sigh	relief_sigh, contented_sigh, wistful_sigh	360 / 9,023	yes
gasp	breath	surprised_gasp, fearful_gasp, gasp, sharp_inhale, inhale	538 / 7,877	yes
breath_calm	breath	deep_breath, deep_breathing, normal_breathing, slow_breathing	557 / 19,260	yes
breath_fast	breath	panting, heavy_breathing, fast_breathing, pant	496 / 2,449	yes
hum	hum	humming, soft_hum, resonant_hum, low_mumble, whispered_mumble, purr, murmur	355 / 12,999	yes
grunt	grunt	affirmative_grunt, effort_grunt	182 / 2,838	yes
throat	throat	cough, coughing, clears_throat, ahem, throat_clearing, clearing_throat	67 / 3,870	no
hiccup	throat	hiccup, hiccups	2 / 2,530	no
swallow	mouth	gulps, swallows, nervous_gulp, drinking_noises, slurping_noises, gulp, swallow	0 / 1,080	no
sob	sob	sobs, quiet_sob, convulsive_sob, sob, trembling_whimper, whimper, weep	9 / 931	no
moan	moan	pain_moan, pleasure_moan, moan	35 / 1,332	no
sniff	breath	sniff, snort, sniffle	38 / 2,350	no
growl	growl	growl, snarl	23 / 1,614	no
hiss	hiss	hiss	3 / 2,164	no
yawn	yawn	yawn	190 / 4,025	yes
mouth	mouth	lip_smack, smacks_lips, smack_one_s_lips, tongue_click, clicks_tongue, click_one_s_tongue, tsk, kissing_sounds, kissing_noises, blowing_a_kiss, licking_sound, chewing_noises, sucking_noise, spitting, smack	16 / 995	no
whistle	whistle	soft_whistle, sharp_whistle, person_whistling_playfully, person_whistling_to_get_attention, wolf_whistle, whistle, whistling	54 / 44	no
misc_body	misc_body	burp, sneeze, snore, snoring, gurgling, shivering, hand_slaps, slap_face, hand_scratching_head, finger_snaps	5 / 89	no

Rationale, group by group

laugh_soft — Leises bis mittleres Lachen. Gemessen: snicker->chuckle 52 %, nervous_giggle->chuckle 50 %, snorting_giggle->chuckle 33 %, childlike_giggle->chuckle 21 %.

laugh_loud — Lautes, offenes Lachen. Stärkstes wechselseitiges Paar im ganzen Satz: cackle<->guffaw, min-Lift 8,3 (guffaw->cackle 31 %, cackle->guffaw 11 %).

scream — Schrei. shriek->scream 44 % bei Lift 14,9 -- der Annotator hat schlicht einen Lieblingsnamen für denselben Laut.

wail — Klagelaut. Bewusst NICHT im Schrei: mournful_wail geht zu 29 % auf scream, aber zu ebenfalls 29 % auf yawn -- das ist keine stabile Zuordnung.

sigh_neg — Genervter/erschöpfter Seufzer und Stöhnen. Einziges Paar mit echter Rückrichtung in die Attraktor-Familie: frustrated_groan->exasperated_sigh 37 %, Rücklift 4,0.

sigh_pos — Erleichterter/zufriedener/wehmütiger Seufzer. wistful_sigh->contented_sigh 9 % und ->relief_sigh 9 %; deep_breath->relief_sigh 20 %. Das nackte sigh wurde am 4.9. wieder entfernt, siehe unten.

gasp — Scharfes Einatmen und Luftschnappen. fearful_gasp->surprised_gasp 35 % bei Lift 24,2 -- der zweitstärkste Beleg überhaupt.

breath_calm — Ruhiges, tiefes Atmen. ACHTUNG: deep_breath ist der größte Attraktor (14,7 % aller Events). Die Lifts hier liegen bei 1,8-1,9, also im Drift-Bereich -- die Gruppe ist semantisch begründet, nicht messtechnisch.

breath_fast — Hecheln und schweres Atmen. heavy_breathing<->panting wechselseitig, min-Lift 5,8.

hum — Summen und Gemurmel. soft_hum->humming 57 % bei Rücklift 1,6. whispered_mumble geht zu 67 % auf humming -- Lift 7,9, aber Rücklift 0,0: das ist Vokabel-Asymmetrie, kein Beleg, dass beide gleich klingen.

grunt — Kurzer Kehllaut (zustimmend oder bei Anstrengung). Beide halten ihren eigenen Namen überdurchschnittlich (45 % bzw. 30 % Selbsttreffer).

throat — Husten und Räuspern. cough<->coughing wechselseitig, min-Lift 10,8.

hiccup — Schluckauf. Reiner Singular/Plural-Split derselben Sache: hiccups->hiccup 24 %, Lift 17,6.

swallow — Schlucken. Stärkstes Paar der gesamten Matrix: gulps<->swallows, min-Lift 18,4; dazu gulps<->nervous_gulp, min-Lift 7,9.

sob — Schluchzen und Wimmern. Semantisch eine Gruppe, messtechnisch ein Totalausfall: alle Mitglieder liegen bei 0-8 % Selbsttreffer und streuen in die Attraktoren.

moan — Stöhnen. Bewusst von sigh_neg getrennt: beide halten ihre eigenen Namen (12 % bzw. 10 %) und wandern nicht ineinander.

sniff — Schniefen und Schnauben. sniff hält 27 % Selbsttreffer, snort 11 %.

growl — Knurren. Die sauberste Einzelklasse im Satz: 56 % Selbsttreffer, keine nennenswerte Wanderung.

hiss — Zischen. 57 % Selbsttreffer, kein Nachbar -- bleibt bewusst allein.

yawn — Gähnen. 29 % Selbsttreffer, wandert aber zu 34 % in exasperated_sigh; bleibt allein, weil es sonst die halbe Seufzer-Gruppe schlucken würde.

mouth — Mundgeräusche. Zusammengelegt, weil jede einzelne Klasse zu selten ist, um getrennt trainiert oder gemessen zu werden.

whistle — Pfeifen.

misc_body — Restliche Körper- und Nebengeräusche. Keine Trainingsklasse, nur damit die Abdeckung vollständig ist.

If your scorer matches labels by substring, read this

Some scorers match labels by substring rather than equality -- reward._same_class in this codebase does. Under such a scorer any pair where one label contains the other counts as a match, in both directions, with no sign that it happened. Where both members sit in the same group that is simply an early partial application of this scheme. Where they sit in DIFFERENT groups it silently merges two groups.

Five bare stems were removed from the map for this reason (sigh, groan, breathing, grunt, screaming_yawn): sigh and groan failed the lift test outright. breathing, grunt and screaming_yawn are too rare to test but each is a bare stem straddling a group boundary, so a substring matcher would merge two groups through it.

The map names every one of the 82 burst labels the shipped 83-class detector can emit (no_burst is deliberately outside the scheme -- it names the absence of a burst). This matters beyond tidiness: an emittable label with NO group is incoherent under a substring-matching scorer, because the detection counts as a strict hit for a target it shares a stem with while contributing no group hit, so the relaxed rate falls BELOW the strict rate for that clip. Four labels -- Kissing Noises, Wolf Whistle, Person Whistling to Get Attention, Slap Face -- were unmapped until 2026-09-04 and are now in mouth, whistle, whistle and misc_body.

One cross-group pair remains and cannot be removed: snort x snorting_giggle. snort (sniff) x snorting_giggle (laugh_soft) cannot be removed: both are real, frequently emitted labels, not bare stems. A snort is not a snorting giggle. A substring scorer will mis-credit this pair in both directions; an equality scorer will not.

How to use it

import json
spec = json.load(open('vocal_burst_groups.json'))
g = spec['label_to_group']
def group_of(label):
    return g.get(label.strip().lower().replace(' ', '_').replace('-', '_'))
# a group-level hit: group_of(predicted) == group_of(requested)

group_of(label) = groups[g]['members'] membership after lowercasing, replacing spaces and hyphens with underscores. A prediction counts as a group-level hit when the predicted label and the requested label fall in the same group. Report the random-grouping control alongside: any grouping raises a hit rate because a group is a bigger target, and for this scheme 0.054 of the +0.235 raw gain is that arithmetic.

Classifier accuracy under this scheme

Same data, same split, same head — only the encoder differs. Chance moves with the number of groups, so the ratio to chance is the only column comparable across schemes; the random control answers the separate question of whether these groups are the right ones.

encoder	licence	params	test set	17 classes	23 groups	chance	× chance	random control	net
voiceclap-large-v2	CC BY 4.0	7 B + LoRA	real, balanced	0.466	0.605	0.108	5.6×	0.493	+0.112
voiceclap-large-v2	CC BY 4.0	7 B + LoRA	DramaBox, balanced	0.574	0.697	0.106	6.6×	0.595	+0.102
voiceclap-commercial	CC BY 4.0	110 M	real, balanced	0.393	0.515	0.103	5.0×	0.425	+0.090
voiceclap-commercial	CC BY 4.0	110 M	DramaBox, balanced	0.536	0.628	0.102	6.2×	0.560	+0.068
voiceclap-small-v2	CC BY-NC 4.0	110 M	real, balanced	0.402	0.513	0.108	4.7×	0.433	+0.080
voiceclap-small-v2	CC BY-NC 4.0	110 M	DramaBox, balanced	0.551	0.675	0.103	6.6×	0.574	+0.102

What this does not cover

No human has listened. The entire scheme rests on the annotator's labels.