laion/vocal-bursts-gemini-utterances

burst_gemini_utterances (Dataset A)

3598 whole utterances from real audio, each carrying a blind gemini-3.8-flash vocal-burst annotation and the general + script training conversion built from it.

WebDataset: data/vbg-utt-*.tar, 4 shards. Per sample:

• <key>.ogg — the whole utterance, 48000 Hz mono OGG/Vorbis. Decoded once from the corpus MOSS codes with the half-speed canary (samples_per_frame == sr/12.5 = 3840) asserted on every row; a row that failed was dropped, not repaired. 0 rows failed it.
• <key>.json — Gemini's events (spans, 1-3 labels, confidence, free-text description), the corpus's own class and spans, the source dataset, the verification level, and the converted general + script.

Duration: median 11.8 s, mean 11.723 s, max 32.0 s.

The general + script conversion, and its round trip

script is this project's training format, produced by anno.render from the word alignment and the burst spans, with Gemini's spans substituted for the detector's. general is the row's own voice/style caption and is not rewritten — Gemini was never asked about voice or style, and putting its prose there would smuggle model text into a field the corpus builds from a different instrument. Everything Gemini contributed is in the script.

The clock is the audio, not the row: dur_s is replaced by frames / 12.5 (13.6 % of corpus rows state a burst longer than their own clip because those two disagree) and every span is clamped into it. Overlapping Gemini events are merged first, since the renderer subtracts a burst's overlap with speech but not with another burst.

Every emitted script is parsed back with the same parser the GRPO reward uses and the row is dropped unless the labels return in the same order, no burst runs past the audio, the spoken text survives verbatim, and the emitted numbers sum to the clip within 0.50 s. That tolerance is calibrated, not chosen: rendering the corpus's own detector spans through the identical path over 792 rows gives a residual of p50 0.00, p99 +0.26, range -1.20..+0.40.

The check was validated against itself. Running the whole 3667-row selection through the identical round trip on the corpus's own spans converts 3,255/3,280 = 99.2 %, so a failure on Gemini's spans is attributable to Gemini and not to the check.

Per-class ladder

Target per class: >= 200 verified rows available -> take 200; 100-199 -> take 100; fewer than 100 -> take everything there is. "Available" is counted after the publication and manufactured-row exclusions, because a row that cannot ship is not available.

tier	classes	meaning
full-200	11	T1 >=200 verified available -> capped at 200
reduced-100	1	T2 100-199 available -> capped at 100
below-100	5	T3 <100 available -> took everything (strict only)
below-100+relaxed	18	T3r <100 strict -> took everything strict, then family-relaxed rows
relaxed-only	14	T3x no strict row exists -- certified entirely by the relaxed gate

Read the capped? column. A class sitting at 100 because the target stopped it is a different fact from a class sitting at 82 because 82 is everything that exists, and both appear below. 23 of 49 classes are exhausted — they contain every publishable verified row this corpus has.

class	tier	strict avail.	rows	strict	family-relaxed	capped?	Gemini top-1	top-3	family
chuckle	full-200	370	199	200	0	yes, by the 200 target	82.4 %	88.9 %	88.9 %
low_mumble	full-200	318	199	200	0	yes, by the 200 target	0.5 %	1.5 %	10.5 %
sharp_inhale	full-200	300	198	200	0	yes, by the 200 target	28.3 %	48.0 %	93.9 %
contented_sigh	full-200	553	197	200	0	yes, by the 200 target	3.0 %	12.7 %	70.0 %
deep_breath	full-200	537	197	200	0	yes, by the 200 target	42.6 %	66.0 %	72.1 %
exhausted_groan	full-200	272	197	200	0	yes, by the 200 target	3.0 %	17.8 %	41.6 %
breathy_giggle	full-200	253	196	200	0	yes, by the 200 target	12.2 %	74.5 %	78.1 %
resonant_hum	full-200	414	196	200	0	yes, by the 200 target	2.5 %	45.4 %	78.1 %
scream	full-200	443	196	200	0	yes, by the 200 target	42.4 %	53.1 %	56.1 %
yawn	full-200	434	195	200	0	yes, by the 200 target	32.8 %	34.9 %	34.9 %
wistful_sigh	full-200	336	193	200	0	yes, by the 200 target	13.5 %	24.9 %	71.0 %
childlike_giggle	reduced-100	106	93	100	0	yes, by the 100 target	3.2 %	8.6 %	83.9 %
surprised_gasp	below-100	82	79	82	0	no — 82 is all there is	8.9 %	15.2 %	38.0 %
ahem	below-100	79	78	79	0	no — 79 is all there is	1.3 %	1.3 %	1.3 %
hiss	below-100	6	6	6	0	no — 6 is all there is	0.0 %	0.0 %	0.0 %
sharp_whistle	below-100	1	1	1	0	no — 1 is all there is	0.0 %	0.0 %	0.0 %
soft_whistle	below-100	1	1	1	0	no — 1 is all there is	0.0 %	0.0 %	0.0 %
exasperated_sigh	below-100+relaxed	72	100	72	28	yes, by the 100 target (needed relaxed rows)	29.0 %	51.0 %	64.0 %
soft_hum	below-100+relaxed	46	98	46	54	no — 116 is all there is at either level	18.4 %	64.3 %	74.5 %
cackle	below-100+relaxed	22	96	22	78	no — 120 is all there is at either level	13.5 %	31.2 %	90.6 %
shriek	below-100+relaxed	48	77	48	31	no — 79 is all there is at either level	9.1 %	59.7 %	72.7 %
mournful_wail	below-100+relaxed	53	57	53	5	no — 58 is all there is at either level	10.5 %	17.5 %	26.3 %
guffaw	below-100+relaxed	3	55	3	52	no — 55 is all there is at either level	5.5 %	12.7 %	80.0 %
sniff	below-100+relaxed	1	47	1	46	no — 47 is all there is at either level	10.6 %	10.6 %	95.7 %
effort_grunt	below-100+relaxed	5	33	5	28	no — 33 is all there is at either level	0.0 %	3.0 %	93.9 %
frustrated_groan	below-100+relaxed	1	32	1	31	no — 32 is all there is at either level	21.9 %	56.2 %	84.4 %
purr	below-100+relaxed	3	26	3	29	no — 32 is all there is at either level	0.0 %	0.0 %	53.8 %
growl	below-100+relaxed	6	25	6	19	no — 25 is all there is at either level	4.0 %	4.0 %	68.0 %
deep_breathing	below-100+relaxed	6	19	6	13	no — 19 is all there is at either level	0.0 %	5.3 %	47.4 %
trembling_whimper	below-100+relaxed	16	18	16	2	no — 18 is all there is at either level	0.0 %	5.6 %	66.7 %
affirmative_grunt	below-100+relaxed	13	17	13	4	no — 17 is all there is at either level	11.8 %	11.8 %	52.9 %
pleasure_moan	below-100+relaxed	3	16	3	13	no — 16 is all there is at either level	0.0 %	0.0 %	81.2 %
coughing	below-100+relaxed	7	14	7	7	no — 14 is all there is at either level	0.0 %	7.1 %	7.1 %
heavy_breathing	below-100+relaxed	1	14	1	13	no — 14 is all there is at either level	0.0 %	35.7 %	92.9 %
cough	below-100+relaxed	6	10	6	4	no — 10 is all there is at either level	30.0 %	60.0 %	60.0 %
nervous_giggle	relaxed-only	0	87	0	88	no — 88 relaxed rows, 0 strict	0.0 %	2.3 %	79.3 %
fearful_gasp	relaxed-only	0	61	0	61	no — 61 relaxed rows, 0 strict	3.3 %	31.1 %	42.6 %
humming	relaxed-only	0	53	0	56	no — 56 relaxed rows, 0 strict	54.7 %	81.1 %	90.6 %
snicker	relaxed-only	0	52	0	54	no — 54 relaxed rows, 0 strict	1.9 %	48.1 %	71.2 %
displeased_grunt	relaxed-only	0	42	0	42	no — 42 relaxed rows, 0 strict	4.8 %	30.9 %	83.3 %
relief_sigh	relaxed-only	0	37	0	37	no — 37 relaxed rows, 0 strict	46.0 %	59.5 %	73.0 %
fast_breathing	relaxed-only	0	24	0	24	no — 24 relaxed rows, 0 strict	0.0 %	58.3 %	95.8 %
panting	relaxed-only	0	19	0	19	no — 19 relaxed rows, 0 strict	47.4 %	52.6 %	94.7 %
normal_breathing	relaxed-only	0	16	0	16	no — 16 relaxed rows, 0 strict	0.0 %	12.5 %	87.5 %
pain_moan	relaxed-only	0	10	0	10	no — 10 relaxed rows, 0 strict	0.0 %	0.0 %	80.0 %
slow_breathing	relaxed-only	0	8	0	8	no — 8 relaxed rows, 0 strict	0.0 %	0.0 %	75.0 %
snort	relaxed-only	0	5	0	5	no — 5 relaxed rows, 0 strict	0.0 %	20.0 %	40.0 %
whispered_mumble	relaxed-only	0	5	0	5	no — 5 relaxed rows, 0 strict	0.0 %	0.0 %	80.0 %
clears_throat	relaxed-only	0	4	0	4	no — 4 relaxed rows, 0 strict	0.0 %	50.0 %	50.0 %

Source datasets

Every row is attributed from uid + src + voice_key; 0 rows are unattributable and none is guessed.

source	rows	share
emolia	1038	28.8 %
voice_profile/emolia	835	23.2 %
voice_profile/the_wild_audio	627	17.4 %
voice_profile/kseries	523	14.5 %
vocal_bursts_clean	352	9.8 %
voice_profile/refvoice	118	3.3 %
voice_profile/anime	77	2.1 %
kartoffelphon	28	0.8 %

emolia, kartoffelphon and vocal_bursts_clean rows are recordings of people (1418 rows). voice_profile/* rows are LAION voice profiles: DramaBox-written text spoken by a synthetic voice built from a real speaker, where the burst was scripted as a cue and then confirmed by the detector in the produced audio. They are labelled human_recording: false on every sample so a consumer can split on it.

Nothing here is spliced, concatenated or voice-converted. The one manufactured row that survived the filter (vb:improved_synthetic) was excluded by name.

What was excluded, and why

reason	rows
not_publishable:podcast	2094
manufactured	3

• Podcast rows are not published. 2094 rows matching (EN|DE|ZH)_<11-char youtube id>_W###### were dropped before selection. Their transcripts are not open, and laion/tts-realspeech-sft-en-de — the public release these rows come from — ships emolia + kartoffelphon + mls and deliberately omits them.
• Mediathek: the restriction does not bite here, and this is worth stating precisely. The brief flagged ~937 the_wild_audio rows. Those rows are src == 0 voice profiles, not the_wild_audio recordings and not the_wild_audio transcripts: the text is DramaBox-written and the voice is a synthetic one built from a the_wild_audio speaker. laion/laion-voice-profiles-sft (public, CC-BY-4.0) already ships exactly these — its part-00377 is 100 % the_wild_audio_# voices. They ship here too. This corpus contains no real the_wild_audio recordings or transcripts, so there was nothing of the restricted kind to exclude.
• MLS: no MLS row survived the strict burst filter, so the source is absent rather than excluded.
• Attribution is an allowlist: an unrecognised uid is not publishable.

Two verification levels, never mixed unlabelled

Every row carries verification, one of:

• strict (2727 rows) — the gate that built $SC/buckets_strict: every target span carries a same-class detection within 1.5 s and the weakest clears theta = 0.174.
• family_relaxed (871 rows) — identical in every other respect (same theta, same tolerance, same geometry gate, no new compute, scored from the same stored detector output) except that a detection from the same burst family counts. This is the metric the owner accepted on 2026-09-02.

Family-relaxed rows were admitted only for classes that could not reach their ladder tier on strict rows alone. The strict filter was never loosened to hit a target: a class short of 100 is reported short. The per-class table below gives both counts for every class, so the whole relaxed set drops out with one filter.

What Gemini said, against the label the corpus already carried

Pooled over 3598 rows, 5161 events (1.434 per row):

level	per row (clip scope)	restricted to events overlapping the corpus span
top-1 label == corpus class	19.0 %	17.0 %
corpus class in the 1-3 labels	37.3 %	33.4 %
same burst family	64.8 %	59.0 %
Gemini returned no burst at all	8.2 %	—

70 distinct taxonomy labels were used in some slot and 59 as a top label, against the detector's 8. 0 labels outside the taxonomy were returned.

Span width — the weak part of this release

§55 measured this model's spans at 38.2 % clip coverage for a 1.27x lift over chance at containing the clip's loudest moment, against our detector's 6.5 % for 3.57x. It is better at what and worse at where. Measured again on this release:

	median	mean	p90
Gemini event duration (s)	0.76	0.969	1.6
corpus detector span (s)	0.36	0.479	0.936
fraction of the clip Gemini covers	0.1	0.176	0.479

How this was handled: by not hiding it. Nothing is silently narrowed. Dataset B cuts the span Gemini actually asserted, plus 50 ms. An energy-tightened window is computed and shipped as metadata (nucleus_start_s, nucleus_end_s) so a consumer can narrow deliberately and reversibly, and every segment carries speech_overlap_frac — the fraction of the span covered by an aligned word — because a wide span sitting over speech is the specific way this data is wrong. Median speech overlap is 0.0, p90 0.595.

Use Dataset B for the label. Do not use it to train a locator.

These labels are a second opinion, not verified ground truth

Every agreement number in this README is between two models: this project's burst detector, which produced the class each row was already filed under, and gemini-3.8-flash, which was asked blind. Both are downstream of models trained on expressive speech. A shared prior would inflate every one of these figures and nothing measured here can see it. What this release adds is a measurably broader second opinion — 70 of the 83 taxonomy classes used against the detector's 8 — not a verification. Whether a given label is correct is a listening question, and $SC/pages/gemini_traindata/ exists to answer it.

How the annotation was done

• Blind. The request never named the class the row was filed under, never said a burst had been found, and never carried the transcript. §55 established the reason: told to expect a shriek, this model finds one. The corpus label is joined back on only in scoring, after every response is on disk.
• Whole utterances only. On 20 excerpts cut out of audio the same model had itself left unannotated, it said "no burst" once and returned Scream/Shriek eleven times. Its labels are context-dependent and not stable under re-segmentation, so nothing shorter than a whole utterance was ever sent. Dataset B's segments are cut out of those whole-utterance spans afterwards; they were never annotated on their own.
• Closed 83-class taxonomy in the system instruction, structured JSON response schema, temperature 0, 1-3 labels per event most-likely-first, and an explicit licence to answer "nothing here" (8.2 % of rows did).
• 3598 calls, 3,977,142 prompt + 577,683 output + 2,369,979 thinking tokens.

Audio format: OGG, and why not MP3

laion/vocal-bursts-clean is 48 kHz mono MP3 and matching it was the plan. MP3 encoding works here — lameenc 1.8.4 round-trips at corr 0.9998 — so this is not a missing-encoder story. It is a measurement: on all six clips tested, the decoded MP3 leads the input by exactly 1,105 samples = 23.0 ms at 48 kHz and runs 32–40 ms long at the tail. That is LAME's encoder delay; a player cancels it using the Xing/LAME info frame, and lameenc exposes no API to write one. Read with soundfile or librosa — which is how this data will be read — every span in this release would sit 23 ms off its own audio, and spans are the entire content.

OGG/Vorbis through libsndfile is sample-exact on the same clips (0 samples added, 0 lag) and about 40 % smaller, so the release ships OGG at the same 48 kHz mono. gt_audio.to_mp3 is kept and tested; MP3_DELAY_SAMPLES_48K = 1105 records the number an MP3 release would have to compensate.

Licence and provenance

Rows derive from laion/tts-realspeech-sft-en-de, laion/laion-voice-profiles-sft and laion/vocal-bursts-clean (all CC-BY-4.0). Annotations are gemini-3.8-flash output. Not uploaded — staged only.

Licence

Per source, not one licence for the whole set. These are complete utterances of real speech, so each row carries the terms of the corpus it came from — read the source field and follow that corpus. This is deliberately unlike the segment release (laion/vocal-bursts-gemini-segments, CC BY 4.0), which holds only sub-second non-verbal fragments.

The Gemini annotations themselves — spans, labels, confidences — are CC BY 4.0.