vocal-bursts-segments

128,165 vocal-burst segments from two acoustically unrelated sources, cut with one policy, plus 46,494 verified no-burst segments. Every segment is a single burst — a laugh, a sigh, a gasp — and nothing else.

subtree	what it is	segments
real/	real speech (Emilia, LAION voice profiles, vocal-bursts-clean, Kartoffelphon), the 5,161 segments of laion/vocal-bursts-gemini-segments	5,161
dramabox/	DramaBox TTS output, cut out of laion/dramabox-burst-audio	123,004

Loadable as real, dramabox or all, so a model can be trained on either half alone — which is the point: the headline number in this card is what happens when you train on one and test on the other.

Why a new repo instead of extending vocal-bursts-gemini-segments

That repo is published with a single default config pointing at data/vbg-seg-*.tar, it is cited by a live Space, and its card documents a single-source release. Moving those files under real/ would break every loader that resolves that path and would silently change what an existing user gets. So the old repo is untouched and this is a new name. real/ holds the same audio bytes — copied out of the published tar without being decoded, no segment added, removed, re-cut or re-encoded — so the two are trivially cross-checkable. The real half's .json gains half, speech_overlap_frac_asr, speech_overlap_frac_word, det_overlap_frac, clamped_span and clamped_pad, and loses nothing.

⚠️ Read this before training on it

• No human has listened to any of this. Every label is gemini-3.8-flash's opinion, asked blind. Agreement numbers here are between models, not against ground truth.
• The dramabox/ half is generated by DramaBox TTS, and the LTX-2 Community Licence has not been assessed. Downstream training on that half is at your own risk. The real/ half is unaffected by this.
• The segments were never annotated as segments. The annotator heard whole clips and drew spans inside them; these files are cuts of those spans. See Span width.
• Use it for the label, not to train a locator.

The cut policy — identical on both halves, and that is the whole point

If the two halves were cut differently, the difference would be learnable and every cross-source number below would be measuring the cutter. So dramabox/ is cut with gt_pack.py's policy byte-for-byte (md5 2ec167f75251ce92a8c5be5eb6eff72f):

• the span the annotator asserted, plus 50 ms on each side, clamped to the audio;
• a cut shorter than 50 ms is dropped;
• nothing is narrowed. An energy-tightened window is computed and shipped as metadata (nucleus_start_s, nucleus_end_s) so a consumer can narrow deliberately and reversibly;
• 48 kHz mono OGG/Vorbis, level preserved, only a clipping peak scaled back;
• <key>_eNN.ogg / .json / .txt, the .txt holding the top label — the convention of laion/vocal-bursts-clean, so an existing loader reads this unchanged.

Clamping is recorded, not hidden. clamped_span = the annotator's own span reached outside its audio (297 segments). clamped_pad = only the 50 ms padding was truncated at a clip edge (2674 segments). §61 reported the first as 0 for DramaBox and this release reproduces it.

MP3 was rejected for a measured reason. LAME's decoded output leads its input by exactly 1,105 samples = 23.0 ms at 48 kHz and runs 32–40 ms long at the tail; lameenc cannot write the Xing/LAME frame a player uses to cancel it. Every span here would sit 23 ms off its own audio, and spans are the entire content. OGG through libsndfile is sample-exact.

The DramaBox source MP3 carries no Xing/LAME frame either (checked on the shipped bytes: it opens with an ID3v2 tag and contains no Xing/Info/LAME marker). So no decoder can strip the encoder delay, the annotator and this cutter decoded the same stream, and the spans needed no offset. Decoded length runs ~38 ms past the dur_s the generator recorded; the clamp is against the decoded length, never the stated one.

Counts

	real	dramabox
burst segments	5,161	123,004
median segment duration	0.76 s	1.37 s
mean	0.97 s	1.66 s
p90	1.60 s	2.94 s
total audio	1.39 h	56.77 h
distinct top-1 labels	59	97
no-burst segments (speech)	6,400	30,000
no-burst segments (silence)	94	10,000

The DramaBox half's spans are about 1.8× longer at the median. That is a systematic difference between the halves, it is not corrected, and it is the single most likely confound in any within-source result. The cross-source table is what controls for it.

Labels present in both halves

Only a class that both generators produce can appear in a cross-source cell at all.

label	real	dramabox
Deep Breath	533	18,121
Exasperated Sigh	557	17,771
Chuckle	793	9,893
Humming	225	10,449
Sharp Inhale	485	5,937
Relief Sigh	138	6,043
Yawn	190	4,089
Scream	184	3,637
Cackle	26	2,434
Hiss	3	2,199
Displeased Grunt	60	1,991
Frustrated Groan	150	1,810
Affirmative Grunt	152	1,699
Surprised Gasp	40	1,760
Breathy Giggle	139	1,601
Contented Sigh	89	1,621
Growl	23	1,633
Soft Hum	102	1,527
Guffaw	16	1,602
Wistful Sigh	133	1,478
Exhausted Groan	161	1,440
Coughing	3	1,522
Sniff	24	1,416
Childlike Giggle	16	1,423
Snicker	87	1,344
Cough	20	1,355
Heavy Breathing	159	1,212
Panting	327	975
Effort Grunt	30	1,196
Snort	14	975
Pain Moan	22	905
Hiccups	2	866
Normal Breathing	2	767
Deep Breathing	20	581
Shriek	15	582
Ahem	24	531
Clears Throat	20	523
Lip Smack	4	492
Resonant Hum	14	479
Pleasure Moan	13	435

59 labels appear in both halves; 0 only in real/ () and 38 only in dramabox/ (Blowing a Kiss, Burp, Chuckling, Click One's Tongue, Convulsive Sob, Crying, Drinking Noises, Effort Groan, Finger Snaps, Gasp, Giggle, Groan…).

The negatives, and the trap they were built to avoid

§61 of this project's protocol measured something that makes the obvious approach unusable: on twenty excerpts gemini-3.8-flash had itself left unannotated, it returned Scream or Shriek eleven times. "The annotator said nothing here" is therefore not evidence of "no burst". Negatives drawn naively from the gaps between annotated spans are poisoned with real bursts, and a detector trained on them learns to call bursts silence.

So a negative window here must clear two independent instruments and an acoustic condition:

1. ≥ 0.5 s clear of every span gemini-3.8-flash asserted in that clip;
2. ≥ 0.5 s clear of every span laion/vocalburst-locator v2 + laion/vocal-burst-detector-v2 found — a genuinely different instrument, and the one this project measured at a 3.57× lift over chance for containing a clip's loudest moment against the annotator's 1.27× (§55): it is better at where;
3. one of two acoustic conditions, which define the two sub-types:

sub_type	rule	why it exists
speech	≥ 60 % of the window covered by aligned Parakeet-TDT words, ≥ 1 whole alphabetic word, RMS ≥ 0.15× the clip's	the negative that matters. In production every decision this detector makes is speech-vs-burst
silence	no recognised word extent within 0.2 s, RMS ≤ 0.10× the clip's	pauses are real and must be represented — but this is the easy negative

The two are never pooled in any number in this card. A detector that has learned only "silence is not a burst" scores well on a pool half made of silence and is useless. Where a mean over negatives and a speech figure disagree, the speech figure is the real one.

Mix: 75 % speech / 25 % silence in the balanced set. Not 50/50 just because there are two sub-types: the operating distribution is overwhelmingly speech, but the corpus scripts contain explicit [N seconds pause] regions, so silence must appear and must not dominate.

Silence is not automatically burst-free, and it is checked as its own arm. A soft sigh, a sharp inhale and a breath are low-energy by nature — exactly the classes in play — so an energy heuristic is biased against precisely the bursts it most needs to exclude.

Negative verification — the number, whatever it is

Each arm was re-sent to gemini-3.8-flash as its own clip, with the byte-identical system instruction, schema and temperature of the pass that produced the labels. This is deliberately the same re-segmentation condition §61 found unstable, because that is the condition a consumer of these files puts them in.

arm	n	Gemini returned a burst	by half	most common label it returned
speech negatives (shipped)	300	4.3 %	dramabox 4.0 % (n=150), real 4.7 % (n=150)	Sharp Inhale 2, Cough 1, Effort Grunt 1, Contented Sigh 1
silence negatives (shipped)	300	99.0 %	dramabox 100.0 % (n=206), real 96.8 % (n=94)	Surprised Gasp 41, Cough 40, Sharp Inhale 35, Snicker 21
naive-gap control (not shipped)	300	16.7 %	dramabox 19.8 % (n=222), real 7.7 % (n=78)	Cough 8, Sharp Inhale 6, Surprised Gasp 5, Sniff 3
burst positive control	150	91.3 %	dramabox 94.7 % (n=75), real 88.0 % (n=75)	Relief Sigh 21, Panting 16, Exhausted Groan 14, Sharp Inhale 14

The sample is stratified by half, not proportional — the DramaBox negative pool is ~20× the real one and a proportional draw of 300 silence windows returned one real row, which would have lost the per-half number the cross-source arms need. The rate of the shipped pool is the per-half rates re-weighted by the pool sizes in metadata/stats.json:

pool	contamination rate
speech negatives as shipped	4.1 %
silence negatives as shipped	100.0 % — this number measures the instrument, not the data; see the null control below

naive_gap is the control: the same 0.5 s guard and none of the other conditions — the negative a naive pipeline ships. burst_positive re-sends segments this pass did annotate: if the model called those empty too, the excerpt-level instrument would be worthless.

The silence arm measures the instrument, not the data — and here is the proof

98 % looks like catastrophic contamination. It has two readings that predict the same number: the windows really contain quiet bursts, or the annotator confabulates on near-empty input. The verification pass cannot separate them, so synthetic audio was sent through the identical call.

synthetic arm	n	burst returned	mean confidence	labels it returned
zeros	40	100.0 %	0.884	Sharp Inhale, Surprised Gasp, Sniff, Cough
noise_-60dB	40	100.0 %	0.889	Cough, Sharp Inhale, Hiccup, Surprised Gasp
noise_-45dB	40	100.0 %	0.882	Sniff, Sharp Inhale, Cough, Snort
noise_-30dB	40	100.0 %	0.878	Cough, Sniff, Surprised Gasp, Sharp Inhale

A vocal burst cannot exist in a buffer of zeros. At 100.0 % on exact digital silence, with mean confidence 0.88 and confident prose descriptions of coughs and sniffs, the excerpt-level instrument is unusable below speech level. So:

• the speech arm is meaningful — the model declines to find a burst on ~95 % of those excerpts, so it is not answering "burst" unconditionally, and that number is a real check;

• the silence arm is not a contamination rate. It cannot be verified by this instrument at all. What can be said about those windows is physical, measured over every shipped negative:

  pool	n	peak dBFS p10 / median / p90	RMS dBFS median	peak < −50 dBFS
  dramabox/speech	30,000	-13.8 / -8.0 / -2.9	-22.6	0.0 %
  real/speech	6,400	-10.4 / -6.5 / -3.4	-20.1	0.0 %
  dramabox/silence	10,000	-98.0 / -52.3 / -39.0	-67.6	58.0 %
  real/silence	94	-79.7 / -42.4 / -30.2	-59.7	38.3 %

  A burst whose peak is 50 dB below full scale is not an audible burst. The honest qualification is the p90: a minority of silence windows do reach −39 dBFS (DramaBox) / −30 dBFS (real), so "effectively empty" describes the bulk of the sub-type, not all of it;

• the naive_gap → speech gap is almost entirely the filter avoiding the region where the instrument breaks. Bucketing both arms by the excerpt's own peak level separates the two effects cleanly:

  arm	< -50 dBFS	-50 to -30	-30 to -15	>= -15 (speech level)
  speech (shipped)	—	—	14.3 % (n=7)	4.1 % (n=293)
  naive gap (control)	100.0 % (n=11)	95.2 % (n=21)	25.0 % (n=28)	5.0 % (n=240)
  silence (shipped)	100.0 % (n=162)	98.4 % (n=123)	93.3 % (n=15)	—
  burst positive control	—	100.0 % (n=14)	100.0 % (n=45)	85.7 % (n=91)

  At matched level the filter buys almost nothing — naive-gap windows that happen to sit at speech level come back at 5.0 %, against 4.1 % for the filtered ones. The filter is still the right thing to ship, because it guarantees the negatives sit at the level where the detector actually operates, but its measured value over a naive gap-miner is a level effect and not a burst-detection effect. That is a smaller and more specific claim than "16.7 % → 4.3 %";

• and this re-frames §61's own control (Scream/Shriek on 11 of 20 unannotated excerpts): some of that is likely the same confabulation rather than evidence of missed bursts.

Practical advice. Train on speech negatives. Use silence only if you want the trivial negative, keep it separable, and never report a pooled negative accuracy.

The balanced set

metadata/balanced_train.parquet materialises what was asked for: equal no-burst and burst material, roughly equal per burst class.

classes (present in both halves at ≥ 100)	16
target per burst class	1,302
burst rows	20,832
no-burst rows	20,832
speech	15,624
silence	5,208

Shortfalls are reported, never topped up from the other half — topping up would quietly make a class single-source and break the cross-source design: none.

The shipped detector head was not trained from this manifest: it balances in the sampler (K draws per class per epoch, no_burst at the same 75/25 mix), because truncating every class to the smallest throws most of the data away. The two are equivalent in expectation; the manifest exists so a consumer can reproduce the balance without the sampler.

The headline: cross-source 2×2

A within-source split cannot answer "how robust is it". The training labels are Gemini's and the test labels would be Gemini's too, so the model is graded by the standard it was trained on. What is informative is training on one generator and testing on the other. Real speech and DramaBox TTS have nothing acoustic in common; a head that transfers between them has learned the burst rather than the generator.

768-d embedding (the shipped detector's own frozen extractor) → 256 → 17 classes, 5 seeds, grouped splits (real: by speaker; DramaBox: by prompt, so all three seeds of one sentence move together). The test set for a source is fixed per seed and reused by every arm, so the cells differ only in what was trained on.

train → test	balanced acc	burst vs no-burst	neg speech	neg silence	shipped (restricted)	shipped (83-way)
real->real	43.4 % ± 0.4	97.6 %	96.1 %	70.4 %	26.6 %	14.8 %
real->dramabox	34.2 % ± 1.5	92.8 %	95.3 %	83.9 %	25.5 %	14.8 %
dramabox->real	34.3 % ± 1.0	94.7 %	95.8 %	26.4 %	26.6 %	14.8 %
dramabox->dramabox	50.4 % ± 0.3	97.4 %	97.9 %	82.0 %	25.5 %	14.8 %
both->real	38.2 % ± 1.7	96.3 %	96.7 %	28.0 %	26.6 %	14.8 %
both->dramabox	51.2 % ± 1.6	97.2 %	98.2 %	79.7 %	25.5 %	14.8 %

Chance = 5.9 % over 17 classes (including no_burst).

shipped (restricted) is laion/vocal-burst-detector-v2's argmax limited to these classes — the fair comparison, since the new head cannot emit the other classes. shipped (83-way) is what it actually does in the pipeline. Reporting only one of them would flatter one side.

Three things to read out of that table before using this data

1. The head transfers. Both cross-source cells (34.2 % and 34.3 %) sit near 6x chance and above the shipped detector restricted to the same classes on the same test sets. A head that has never seen the test generator still out-scores the detector this project has been using.
2. Adding the dramabox half makes the real-speech head worse. both->real 38.2 % against real->real 43.4 %. DramaBox outnumbers the real half more than ten to one in every class, and even a class-balanced sampler cannot stop it dragging the head towards its own distribution. If you want the best real-speech head, train on the real config alone. This is the opposite of what "more data" predicts and it is the most actionable number in this card.
3. The silence negatives do not transfer at all -- dramabox->real scores 26.4 % on real silence against 82.0 % on DramaBox silence, while the speech negatives transfer at 95-98 % everywhere. "Silence" is physically a different object in the two corpora. Never pool the two sub-types into one negative accuracy.

The encoder matters more than the data mix

Every number above uses the shipped detector's own frozen 768-d extractor, so that the head stays a drop-in replacement for it. Re-running the identical experiment — same segments, same splits, same head shape, same seeds — through laion/voiceclap-large-v2 (3584-d) changes only the encoder, and it wins every cell:

train → test	FastScorer 768-d	VoiceCLAP 3584-d	Δ
real->real	43.4 %	45.6 %	+2.2 pts
real->dramabox	34.2 %	40.6 %	+6.4 pts
dramabox->real	34.3 %	35.7 %	+1.4 pts
dramabox->dramabox	50.4 %	56.0 %	+5.6 pts
both->real	38.2 %	44.9 %	+6.8 pts
both->dramabox	51.2 %	57.6 %	+6.3 pts

It also removes most of the penalty for mixing the two halves: both→real against real→real goes from a 5.2-point loss with the 768-d extractor to under a point with VoiceCLAP. So the "train on real alone" advice above is advice about that encoder, not about this data. Both heads are published at laion/vocal-burst-detector-x2.

Family-relaxed, because the failure is granularity and not deafness

Scoring the same predictions at burst-family level (sigh, groan, laugh, breath, hum) gains 16-21 points in every cell -- dramabox->real goes 34.3 % -> 55.7 %. The confusion is almost entirely within family: Breathy Giggle -> Chuckle 43 %, Exhausted Groan -> Frustrated Groan 62 %, Deep Breath -> Sharp Inhale 48 %, Humming -> Soft Hum 29 %, Relief Sigh -> Sharp Inhale 27 %.

Per class, because a mean hides the interesting failure

§58 and §61 both found that the failure mode is granularity — Shriek called Scream, Soft Hum called Humming — not deafness. A mean cannot see that.

class	accuracy (both→real)
no_burst	82.4 %
Scream	76.8 %
Panting	72.8 %
Sharp Inhale	60.8 %
Chuckle	60.0 %
Affirmative Grunt	53.6 %
Frustrated Groan	53.6 %
Soft Hum	38.4 %
Breathy Giggle	33.6 %
Yawn	26.4 %
Exhausted Groan	21.6 %
Humming	18.4 %
Wistful Sigh	15.2 %
Heavy Breathing	12.0 %
Exasperated Sigh	11.2 %
Relief Sigh	8.8 %
Deep Breath	3.2 %

Confusion matrices for every cell are in train_report_x2.json (cells[*].cm_sum), summed over seeds.

Layout

real/data/vbs-real-*.tar            burst segments, 1000 per shard
real/negatives/vbs-real-neg-*.tar   no-burst segments
dramabox/data/vbs-db-*.tar
dramabox/negatives/vbs-db-neg-*.tar
metadata/segments.parquet           one row per burst segment
metadata/negatives.parquet          one row per no-burst segment
metadata/balanced_train.parquet     the balanced manifest
metadata/stats.json  metadata/balanced_train.json  metadata/negative_verification.json

import webdataset as wds
ds = wds.WebDataset("dramabox/data/vbs-db-{00000..00123}.tar").decode()
for r in ds:
    audio, label, meta = r["ogg"], r["txt"], r["json"]

Fields worth knowing

field	meaning
labels	all 1–3 labels the annotator returned, most likely first. Use the set, not labels[0] — §61 measured soft_hum at 14.9 % top-1 against 95.5 % anywhere in the three
speech_overlap_frac_word	fraction of the span covered by Parakeet-TDT word extents, each capped at 0.30 s
speech_overlap_frac_asr	the same with raw TDT spans. A TDT token's duration runs to the next token, so the raw version measures "inside the spoken region", not "overlapping a word". Both ship; the capped one is what the filters use
det_overlap_frac	fraction of the span covered by the shipped locator+detector — an independent second opinion on where
nucleus_start_s / nucleus_end_s	energy-tightened window, computed and never applied
requested_rank	position of the class the DramaBox prompt asked for within this event's labels; −1 if absent
sub_type	speech or silence, negatives only. Never pool them

Span width — the weak part of this release

The annotator is better at what and worse at where: §55 measured its spans at 38.2 % clip coverage for a 1.27× lift over chance at containing the clip's loudest moment, against the shipped locator's 6.5 % for 3.57×. Nothing here is silently narrowed; nucleus_* is shipped as metadata so narrowing is deliberate and reversible, and speech_overlap_frac_word is shipped because a wide span sitting over speech is the specific way this data is wrong.

What this dataset cannot tell you

• whether any label is correct — that is a listening question, and nobody has listened;
• whether the two annotators share a prior. Both are downstream of models trained on expressive speech; a shared prior would inflate every agreement number here and nothing measured can see it;
• whether the negatives are clean in an absolute sense. The verification arm above is the same model that produced the labels, asked again;
• anything about boundaries. Use it for the label.

Class groups

vocal_burst_groups.json and GROUPS.md carry a 23-group scheme over 117 burst label strings, grouping names that denote the same or a very similar sound (snicker/chuckle, shriek/scream, cough/coughing). Scoring the same predictions at group level raises the mean generation hit rate from 0.302 to 0.537; a random grouping with identical group sizes reaches 0.355, so +0.182 of it is the grouping being right and the rest is arithmetic. Groups were checked with directed lift rather than raw confusion, because two labels account for 29 % of all annotator top-1 calls whatever was requested and merging on raw confusion books a generation failure as a hit.

For training the classifier, keep the fine classes: collapsing them raises raw accuracy only because chance rises with it. Group at evaluation time — that can be done at any point, the reverse cannot.