laion/voice-acting-edge-top3

Edge-case reward-Top-3 — voice annotations

1,494,503 synthetic expressive-speech utterances (the reward-Top-3 selection of the edge-case corpus), annotated with:

1. a CrisperWhisper-format transcript with corrected vocal bursts — each surviving burst carries a class name from laion/vocal-burst-detector-v2 at its original timestamp;

2. raw voice scores at two granularities — whole utterance and per sentence — from laion/Empathic-Insight-Voice-Plus (all 40 emotions), the 57 VoiceNet dimensions, genuineness, vocal-burst blend and 4 quality scores;

3. a procedural <GENERAL> + <SCRIPT> caption derived from those scores, with per-sentence delivery cues in (round brackets), burst class names in (round brackets) inline, and pauses in [square brackets].

4. the source audio, byte-for-byte, in WebDataset tar shards aligned to the metadata parts;

5. a second vocal-burst annotation (vb2_* columns, added 2026-09-10) from the current production detector on voiceclap-large-v2, over every located span, plus the bursts themselves cut out as a per-class WebDataset under bursts/ — see Vocal bursts — two annotations side by side.

If you are here for ASR: much of this audio is deliberately hard to understand, and the verbatim CrisperWhisper large-v2 transcripts plus corrected burst labels are what make that useful rather than useless. See Use as an ASR training set.

The MP3 payload is the original generation output — 48 kHz mono MP3, 160 kbps, copied unchanged, never decoded or re-encoded.

The scores are a first-class artefact, not an intermediate: the caption step (derive_captions.py) reads only data/scores and runs on CPU in minutes. Swap the baseline, re-run, get a different caption set — no audio, no GPU, no re-scoring.

---

What is in here

Numbers computed from the parquets of this release (not from the generation logs); the column each one comes from is named.

	value	from
utterances	1,494,503	data/transcripts id (unique across all four parquet families and the audio tars)
audio	11,840 h	Σ duration
clip length	median 28.5 s, p10 14.1 s, p90 41.0 s, mean 28.5 s	duration
prompt groups	503,660 (3 takes each, minus the 5,362 empty-transcript rows)	group_key
reference voices	1,264	not a column — each group was generated against one reference clip drawn from the enhanced voice pool; the ref_id is recorded in the generation manifest (all_groups.jsonl), the most frequent one covers 0.3 % of rows
format	48 kHz mono MP3, 160 kbps, unchanged generator output	audio tar members

What this corpus is

The edge-case corpus: 266 hand-written edge cases of expressive speech (speaking while climbing, out of breath, shivering, drunk, masking anger, …) in 7 categories, each turned into 2,000 acting prompts by google/gemma-4-E4B-it — 1,000 with a German dialogue and 1,000 with an English dialogue, the stage direction always in English, each with one mandatory word — 532,000 prompt groups in all. Every group was rendered 20 times (20 seeds) by the voice-acting model against one random reference voice; the 20 takes were scored with the VoiceCLAP-commercial heads and ranked inside the group by reward = norm(blend) + norm(genuineness) + norm(arousal) (min-max within the 20). The top 3 of every group are what is published here — the reward-Top-3 of the edge-case corpus, hence the name. The reward favours expressive delivery, not intelligibility, which is why a lot of this audio is hard to understand (see Use as an ASR training set).

edge-case category	utterances	share
Emotion, Masking & the Social Game	495,286	33.1 %
Voice-Craft, Bursts & Delivery	311,333	20.8 %
Body, Illness, Pain & Substances	185,362	12.4 %
Life, Language & Society	175,924	11.8 %
Physical Exertion & Environment	155,376	10.4 %
Consciousness, Shock & Altered States	124,673	8.3 %
Crisis, Emergency & Survival	46,549	3.1 %

(from case_category of the prompt files, joined on the ecNNN_MMMM part of group_key; 262 distinct cases)

Language

There is no language column. Three views, which do not agree, and the disagreement is the finding:

view	German	English	other	from
what the prompt asked for	49.9 % (746,092)	50.1 % (748,411)	—	prompt files, language, joined on group_key
what the transcript reads as	6.1 % (90,936)	93.1 % (1,391,640)	0.8 %	fastText lid.176 top-1 over text (burst tokens stripped)

Spoken language, measured on the waveform of a stratified sample (2,400 clips, 150 per prompt language from each of 8 shards, Whisper large-v3 language-ID head — 95.0 % on the project's own LID benchmark):

prompt language	clips	audio says German	audio says English	other
German	1,200	25.8 %	73.0 %	1.2 %
English	1,200	0.0 %	99.9 %	0.1 %

The prompts are half German; the audio is not. On the German-scripted half Whisper hears German in 26 % of the clips and English in 73 % (mean P(en) 0.68); on the English half it hears English in 99.9 %. By what is actually spoken the corpus is therefore roughly 87 % English / 13 % German: the generator rendered most German scripts as English-sounding speech, and the transcript LID's 6 % German is CrisperWhisper reading only part of that German audio as German. Treat the German subset as small, and do not infer the spoken language from the prompt or from the transcript alone — for a per-clip decision run a spoken-language model on the waveform.

Emotion

Model scores, not human ratings: data/scores global_emonet holds the 40 Empathic-Insight-Voice-Plus dimensions per utterance (order in schema.json). Top-1 per clip = argmax over the 40:

top-1 emotion	utterances	share	mean score over all clips
Interest	1,352,076	90.5 %	2.322
Concentration	38,242	2.6 %	1.316
Confusion	11,200	0.7 %	0.578
Amusement	11,117	0.7 %	0.736
Disappointment	10,470	0.7 %	0.617
Intoxication Altered States of Consciousness	8,945	0.6 %	0.516
Distress	7,515	0.5 %	0.149
Teasing	6,264	0.4 %	0.797
Hope Enthusiasm Optimism	6,245	0.4 %	0.807
Impatience and Irritability	6,210	0.4 %	0.620
Fatigue Exhaustion	5,840	0.4 %	0.157
Relief	5,109	0.3 %	0.398
Contemplation	4,515	0.3 %	0.828
Astonishment Surprise	2,709	0.2 %	0.490
Thankfulness Gratitude	2,402	0.2 %	0.379

Interest wins the argmax on 90.5 % of clips — it is the dimension with the highest baseline on acted speech, so the argmax is a poor summary; the captions use z-scored deviations from the in-domain baseline instead (see How it was built). Other global scores: vocal-burst blend mean 4.05/10 (median 3.90), genuineness mean 2.46/6; Empathic-Insight gender gate: 699,454 masculine (< −0.5), 466,937 feminine (> 0.5), 328,112 in between (global_ei_gender).

Vocal bursts, in one line each

• original annotation (data/bursts, class_v2 where kept): 2,053,108 kept spans of 5,401,483 located; 947,767 utterances (63.4 %) have at least one; top labels Low Mumble 26 %, Ahem 25 %, Contented Sigh 21 %, Surprised Gasp 5 %.
• re-classification (vb2_*, all 5,401,483 spans): Sharp Inhale 74 %, Chuckle 6 %, Affirmative Grunt 6 %, Frustrated Groan 3 % — see the next section for why that is not what it looks like.

---

Layout

data/audio/audio-{000..127}.tar                   # the audio, ~6.2 GiB per shard
data/transcripts/transcripts-{000..127}.parquet   # corrected CrisperWhisper transcripts
data/scores/scores-{000..127}.parquet             # raw scores, global + per sentence
data/bursts/bursts-{000..127}.parquet             # every original span, kept and dropped
data/captions/captions-{000..127}.parquet         # the derived captions
bursts/<class>/<class>-NNNNN.tar + .parquet   # every burst cut out, one config per class (vb_<class>)
per_class_recall.json                         # held-out recall of the vb2 head per class
captioner/                                    # caption.py + both baselines + schema.json
derive_captions.py                            # regenerates data/captions from data/scores
schema.json                                   # column orders and model↔column bindings
provenance.json                               # models, settings, captioner commit, baseline hash
stats.json                                    # the numbers quoted below

Total size 802G (796 GiB of it audio), 128 shards per family. All five families join on id ("{group_key}.s{seed}"), and audio-007.tar holds exactly the utterances in scores-007.parquet — the shards are aligned, so a consumer never needs a global index.

from datasets import load_dataset
ds = load_dataset("laion/voice-acting-edge-top3", "captions", split="train")     # metadata only

Joining audio to metadata

The audio member name is the join key plus .mp3 — nothing else. One sequential pass over a shard, no index, no random access:

import io, tarfile, soundfile as sf
import pyarrow.parquet as pq

meta = {r["id"]: r for r in pq.read_table("data/scores/scores-007.parquet").to_pylist()}

with tarfile.open("data/audio/audio-007.tar", "r|") as tf:      # streaming
    for m in tf:
        key = m.name[:-len(".mp3")]                             # == the metadata id
        row = meta[key]
        wav, sr = sf.read(io.BytesIO(tf.extractfile(m).read()))
        # row["global_emonet"], row["sentences"], ... line up with this waveform

There is deliberately no audio entry under configs: in the card header, for the same reason described next: an automatic WebDataset loader would mis-group these keys. Load the audio with the recipe above.

One caveat if you use the webdataset library. Its default key splitter takes everything before the first dot, and these keys contain one (..._0000.s1.mp3), so the three seeds of a prompt group would be folded into a single sample with extensions s1.mp3, s2.mp3, s3.mp3. Pass a key function that splits at the last dot instead — lambda p: p.rpartition(".")[::2] — through webdataset.tariterators.group_by_keys(keys=...). (Stated from the library's documented base_plus_ext behaviour; webdataset was not installed in the environment used to build this dataset, so this particular snippet is untested. The plain-tarfile recipe above is tested.)

---

Fields

transcripts — artefact (a)

column	type	meaning
id	string	{group_key}.s{seed}, the join key
group_key, seed	string, int32	prompt group and generation seed in the source corpus
source_tar, member	string	where the audio lives, relative to the generator's outputs/ root
duration	float32	decoded clip length in seconds
reward, rank_in_group	float32, int32	the reward-model score that put this take in the Top-3
text	string	transcript, burst tokens rendered as [Class Name]
words	list	w (token), start, end (seconds), is_burst, sent (sentence index)
bursts	list	start, end, type — type is the v2 class name, at the original CrisperWhisper timestamp

The old CrisperWhisper tags (UH, laughter, …) are not in this family. They are preserved in bursts for comparison, and nowhere else.

scores — the re-derivation input

column	type	meaning
id, duration	string, float32
global_voicenet	list[57]	VoiceNet regression, 0–6, in schema.json:voicenet_dims order
global_emonet	list[40]	Empathic-Insight emotions, in schema.json:emonet order
global_quality	list[4]	content_enjoyment, overall_quality, speech_quality, background_quality
global_genuineness	float32	0–6
global_blend	float32	vocal-burst blend, 0–10
global_ei_gender	float32	Empathic-Insight Gender expert, −2 masculine … +2 feminine (gate only)
words	list	same as in transcripts — pauses and inline burst placement come from here
sentences	list	one entry per sentence, see below

sentences[]: idx, text, start, end, n_words, scored, then voicenet (57), emonet (40), quality (4), genuineness, blend, plus bursts (class names falling in this sentence) and burst_starts.

scored is false for sentences under 8 words, and their score lists are null. That threshold is inherited from the previous run and kept: below ~8 words a sentence is often under a second of audio, and both scorers are 30 s-window models whose per-dimension agreement degrades sharply on very short segments. 4,588,699 of 6,284,328 sentences (73.0 %) are scored.

bursts — every original span, including the dropped ones

column	meaning
id, burst_idx	join key + index into the original span list
start, end, duration	the CrisperWhisper span, unmodified
crisperwhisper_type	the original coarse ASR tag — kept only here
class_v2, prob	the v2 classifier's top-1 over the recognised burst classes, and its probability
p_no_burst	P(no_burst) — the gate value
top5, top5_prob	the top-5 of the masked softmax
kept	whether the span survived into transcripts / captions
drop_reason	"" (kept), under_200ms, no_burst, bad_timestamp
sentence_idx	which sentence the span was assigned to

Storing the full top-5 and p_no_burst for dropped spans means a different duration floor or a different no_burst gate can be applied without re-running anything — except for spans under 200 ms, which were never embedded (see below).

captions — artefact (b)

id, general, script, caption (the assembled <GENERAL>…</GENERAL><SCRIPT>…</SCRIPT> block), n_sentences, n_sentences_cued, synonym_seed. The parquet key-value metadata carries caption_provenance: baseline file + SHA-256, captioner commit, templates and k values.

Example:

<GENERAL>
A voice that is very thin in head resonance; very organic and soft; notably thin in throat
resonance; very decelerating in pace; adult; strongly masculine, deep and low-pitched; low and
bassy in register; notably carrying exhaustion; only slightly genuine, somewhat performed.
</GENERAL>
<SCRIPT>
(very soft-onset, very organic, very flat, very detachment) "I must these packets of everything
up" [pause 1.6s] (Ahem) [pause 0.5s] "I'm lacked."
(very chesty, very formal, very throaty, detachment) [pause 1.5s] "The Alton sympathisants is here"
</SCRIPT>

Round brackets carry two different things and are distinguishable by position: the first bracket on a line is the sentence's delivery cue; brackets inside the quoted run are burst class names. Square brackets are pauses, in seconds, computed from consecutive word timestamps (threshold 0.35 s).

---

Vocal bursts — two annotations side by side

This repository carries two vocal-burst annotations, and the newer one does not replace the older one.

annotation	spans	classifier	vocabulary	where
original	CrisperWhisper-large-v2 spans, ≥ 200 ms, p(no_burst) < 0.5	laion/vocal-burst-detector-v2 on voiceclap-commercial	82 classes + no_burst	data/bursts (class_v2, prob, p_no_burst, top5, kept), data/transcripts (bursts, text)
re-classification (vb2_*)	the same spans, all of them — including the 3,348,375 the original pipeline dropped as under_200ms and never classified	laion/vocal-burst-detector-x2 production head on voiceclap-large-v2 (3584-d, frozen) + 5 × MLP 3584→256→17, averaged in probability space	16 burst classes + no_burst	data/bursts (vb2_label …, per span), data/transcripts (vb2_labels …, per utterance), bursts/ (the cuts)

Every span was cut out (± 0.05 s of context, the production detector's own pad) and only the cut was classified — the head never sees the surrounding speech. No gate was applied: vb2_labels is always a burst class, never no_burst; the reject probability is delivered per span in vb2_no_burst and the consumer decides.

The rule: the new columns are added, nothing is overwritten. The original columns are byte-identical to the previous release (verified per file after writing, see *.verify.json in the build). The two vocabularies are not nested — of the 56 labels the original detector emitted here, only 15 exist in the 16-class vocabulary, and 81.3 % of the kept original events carry a label the new head cannot say (Low Mumble, Ahem, Contented Sigh, Surprised Gasp …). Where the two disagree on a span, that is information, not an error to be resolved.

Columns

data/bursts/bursts-NNN.parquet — one row per span, the new values on the same row as the old ones:

column	type	meaning
vb2_label	string	new top-1 among the 16 burst classes (never no_burst)
vb2_prob	float	its ensemble probability
vb2_no_burst	float	P(no_burst) for this span, no gate applied
vb2_family	string	breath / sigh / groan / laugh / hum / scream / yawn
vb2_top3, vb2_top3_p	list<string>, list<float>	the three strongest classes
vb2_dip_db	float	level dip of the span against ± 2 s of context, dB; NaN without context
vb2_flag_short	bool	span < 0.10 s or past the clip end — classified, not reliable
vb2_missing	bool	audio could not be decoded → the other vb2_* are null (this release: 0 rows)
vb2_model, vb2_version	string	stamps
vb3_label, vb3_prob, vb3_no_burst, vb3_family, vb3_top3, vb3_top3_p	string, float, float, string, list<string>, list<float>	(x3, 2026-09-12) the same span read by vocal-burst-detector-x3 production head — 18 burst classes (16 + Ahem, Low Mumble); vb3_no_burst ungated, recommended gate 0.9; families add throat
vb3c_label, vb3c_prob, vb3c_no_burst, vb3c_family, vb3c_top3, vb3c_top3_p	as above	the x3 commercial head; vb3c_no_burst gate 0.5
vb3_missing	bool	cut not decodable → the vb3_*/vb3c_* values are null (this release: 0 rows)
vb3_model, vb3_version, vb3c_model, vb3c_version	string	stamps

data/transcripts/transcripts-NNN.parquet — one row per utterance, list columns index-aligned to the row's own bursts list (the kept spans, in order), exactly the column set of laion/laion-tts-annotated-v1:

column	type	meaning
vb2_n	int32	== len(bursts)
vb2_labels, vb2_probs, vb2_no_burst, vb2_families, vb2_top3, vb2_top3_p, vb2_dip_db, vb2_flag_short	lists	the per-span values above, one entry per kept span
vb2_loc_conf	list<float>	NaN throughout — these spans come from the ASR, not from a locator
vb2_missing	bool	audio not decodable → empty lists
vb2_model, vb2_version	string	stamps
vb3_n, vb3_labels, vb3_probs, vb3_no_burst, vb3_families, vb3_top3, vb3_top3_p, vb3_missing, vb3_model, vb3_version	int32 / lists / bool / string	(x3, 2026-09-12) the production head's reading of the same kept spans, same alignment as vb2_*; vb3_no_burst gate 0.9
vb3c_n, vb3c_labels, vb3c_probs, vb3c_no_burst, vb3c_families, vb3c_top3, vb3c_top3_p, vb3c_missing, vb3c_model, vb3c_version	as above	the commercial head; vb3c_no_burst gate 0.5

The dropped (under_200ms) spans are classified too, but only in data/bursts — transcripts.bursts never contained them, so the transcript lists do not either.

vb2_model = laion/vocal-burst-detector-x2 production: voiceclap-large-v2 (3584-d, frozen) + 5x MLP 3584>256>17 averaged in probability space, vb2_version = dbx-edge/spans:crisperwhisper/pad0.05/gate:none/2026-09-10/bs_score@9c127343cfdb/hf@eee775ffd2. The same strings sit in the parquet key-value metadata together with vb2_classes (the 17 names in checkpoint order). per_class_recall.json in the repository root is the head's held-out recall per class — a measured hit rate for a class cannot meaningfully exceed its recall there.

Read this before using vb2_labels

74.5 % of the spans are named Sharp Inhale. That is not 4,023,929 inhalations. The original detector's most frequent labels — Low Mumble, Ahem, Contented Sigh, Surprised Gasp — do not exist in the 16-class vocabulary, so on those spans the new head must choose something, and for a short, breathy, low-level transition between words the nearest nameable class is Sharp Inhale. Its confidence there is only moderately lower than on real bursts (median vb2_probs 0.73 overall), so it is not a clean filter either.

original label (kept spans)	n	what vb2_labels says
Low Mumble	537,298	Sharp Inhale 55 %, Affirmative Grunt 12 %, Frustrated Groan 11 %
Ahem	511,277	Sharp Inhale 70 %, Affirmative Grunt 9 %, Chuckle 9 %
Contented Sigh	439,481	Sharp Inhale 63 %, Exasperated Sigh 23 %, Relief Sigh 6 %
Surprised Gasp	110,991	Sharp Inhale 65 %, Frustrated Groan 11 %, Exasperated Sigh 5 %
Breathy Giggle	105,813	Chuckle 33 %, Sharp Inhale 29 %, Breathy Giggle 28 %
Wistful Sigh	94,737	Sharp Inhale 31 %, Exasperated Sigh 22 %, Frustrated Groan 12 %
Chuckle	80,273	Chuckle 78 %, Sharp Inhale 9 %, Breathy Giggle 3 %
Exhausted Groan	62,849	Frustrated Groan 45 %, Exhausted Groan 18 %, Exasperated Sigh 16 %
Childlike Giggle	54,094	Sharp Inhale 43 %, Chuckle 21 %, Breathy Giggle 20 %
Sharp Inhale	32,500	Sharp Inhale 94 %, Panting 2 %, Breathy Giggle 1 %
Resonant Hum	10,325	Soft Hum 42 %, Affirmative Grunt 22 %, Humming 16 %
Scream	2,653	Scream 58 %, Frustrated Groan 30 %, Sharp Inhale 8 %

Label agreement on the 2,053,108 kept spans where both annotations exist: 6.8 % same string — mostly Chuckle, Sharp Inhale, Exhausted Groan, the classes both vocabularies share.

How to read it: for a span whose original label is outside the new vocabulary, vb2_labels is the nearest nameable class, not a verified event of that class. Read the original label and vb2_dip_db next to it: vb2_dip_db is the quietest 100 ms inside the span minus the median level of ± 2 s of context, in dB. A genuine breath or pause dips (median here -2.6 dB, p10 -18.3); a span inside loud speech does not (51.4 % of spans are above −3 dB). vb2_no_burst ≥ 0.5 on 5,559 spans (0.10 %) — the head almost never rejects an ASR-located span outright, as the original no_burst gate did not either. vb2_flag_short is set on 1,762,885 spans (32.6 %) shorter than 0.10 s: classified, but not to be trusted.

Class distribution

vb2_labels	family	spans	share
Sharp Inhale	breath	4,023,929	74.5 %
Chuckle	laugh	344,117	6.4 %
Affirmative Grunt	groan	325,140	6.0 %
Frustrated Groan	groan	180,568	3.3 %
Exasperated Sigh	sigh	175,675	3.3 %
Breathy Giggle	laugh	108,691	2.0 %
Exhausted Groan	groan	96,846	1.8 %
Soft Hum	hum	63,114	1.2 %
Relief Sigh	sigh	32,180	0.6 %
Humming	hum	17,044	0.3 %
Scream	scream	11,387	0.2 %
Wistful Sigh	sigh	8,812	0.2 %
Panting	breath	6,735	0.1 %
Deep Breath	breath	4,816	0.1 %
Yawn	yawn	1,737	0.0 %
Heavy Breathing	breath	692	0.0 %

For comparison, the original kept labels (top 12 of 56):

class_v2 (kept)	spans	share
Low Mumble	537,298	26.2 %
Ahem	511,277	24.9 %
Contented Sigh	439,481	21.4 %
Surprised Gasp	110,991	5.4 %
Breathy Giggle	105,813	5.2 %
Wistful Sigh	94,737	4.6 %
Chuckle	80,273	3.9 %
Exhausted Groan	62,849	3.1 %
Childlike Giggle	54,094	2.6 %
Sharp Inhale	32,500	1.6 %
Resonant Hum	10,325	0.5 %
Scream	2,653	0.1 %

The third annotation — vb3_* / vb3c_* (x3, 2026-09-12)

The same spans a third time, with laion/vocal-burst-detector-x3 in both of its heads: vb3_* = production/ (VoiceCLAP-large-v2), vb3c_* = commercial/ (VoiceCLAP-commercial). 19 classes = the 16 of x2 + Ahem + Low Mumble + no_burst; same column layout and alignment as vb2_* (data/bursts: per-span scalars; data/transcripts: list columns aligned to the row’s kept bursts; both include the production and commercial heads); vb2_* unchanged. P(no_burst) is delivered ungated; recommended gates 0.9 for vb3_* and 0.5 for vb3c_*. x3 was trained with locator false-alarm spans as no_burst negatives (the original detector's Ahem/Low Mumble spans in real speech overlap a blind Gemini event in only 16.5 % / 7.3 % of cases), which x2 never saw — so the reject output is now usable. Over the 5,401,483 spans here:

	x2 (vb2_*)	x3 vb3_*, gate 0.9	x3c vb3c_*, gate 0.5
spans rejected (P(no_burst) ≥ gate)	—	64.8 % (at 0.5: 78.2 %)	39.7 %
Sharp Inhale, share of all spans	74.5 %	19.6 %	34.9 %
top classes (share of all spans)	Sharp Inhale 74 %, Chuckle 6 %, Affirmative Grunt 6 %, Frustrated Groan 3 %	Sharp Inhale 20 %, Chuckle 3 %, Exasperated Sigh 3 %, Frustrated Groan 2 %	Sharp Inhale 35 %, Affirmative Grunt 7 %, Soft Hum 4 %, Chuckle 4 %
Ahem / Low Mumble among kept spans	—	76,765 / 1,375	99,057 / 31,200

Low Mumble does not hold on real speech (0 % recall on real held-out rows); on this synthetic corpus it is the head's own reading of DramaBox-like audio and should still be read with vb3_no_burst beside it. Ahem transfers at 0.23–0.31 recall. Per-class recall of both heads (with the x2 value on the same rows): per_class_recall_x3.json, per_class_recall_x3_commercial.json.

The cut-out bursts — bursts/, one config per class

Every classified span is also shipped as its own audio file, so a burst can be listened to or trained on without decoding the 28-second clip around it:

bursts/<class>/<class>-NNNNN.tar        WebDataset, ≤ 10,000 members (5,000 cuts × {.ogg, .json})
bursts/<class>/<class>-NNNNN.parquet    index of that tar, same fields as the json
bursts/stats.json                       counts per class

The class directory is the new top-1 (vb2_label); the original label travels in the json. The audio is the exact window the classifier saw: the span ± 0.05 s, mono, OGG/Vorbis at the original 48 kHz (not MP3 — the MP3 encoder leads the input by 1,105 samples and these cuts carry timestamps). Member stems are the source key with . replaced by - (WebDataset splits keys at the first dot) plus __<span_idx>; the verbatim source key is source_key in the json.

json fields: key, source_key, span_idx, start, end (the span in the source clip, s), cut_start, cut_end (the window actually cut), sr, audio_format, span_source, old_label, old_kept, old_cw_type, loc_conf, vb2_label, vb2_prob, vb2_no_burst, vb2_family, vb2_top3, vb2_top3_p, vb2_dip_db, vb2_flag_short, shard.

config	class	cuts	tars
vb_sharp_inhale	Sharp Inhale	4,023,929	805
vb_chuckle	Chuckle	344,117	69
vb_affirmative_grunt	Affirmative Grunt	325,140	66
vb_frustrated_groan	Frustrated Groan	180,568	37
vb_exasperated_sigh	Exasperated Sigh	175,675	36
vb_breathy_giggle	Breathy Giggle	108,691	22
vb_exhausted_groan	Exhausted Groan	96,846	20
vb_soft_hum	Soft Hum	63,114	13
vb_relief_sigh	Relief Sigh	32,180	7
vb_humming	Humming	17,044	4
vb_scream	Scream	11,387	3
vb_wistful_sigh	Wistful Sigh	8,812	2
vb_panting	Panting	6,735	2
vb_deep_breath	Deep Breath	4,816	1
vb_yawn	Yawn	1,737	1
vb_heavy_breathing	Heavy Breathing	692	1

from datasets import load_dataset
ds = load_dataset("laion/voice-acting-edge-top3", "vb_chuckle", split="train")   # one class
r = ds[0]; r["json"]["source_key"], r["json"]["old_label"], r["ogg"]

import pyarrow.parquet as pq
b = pq.read_table("data/bursts/bursts-007.parquet").to_pylist()
for r in b[:5]:
    print(r["id"], r["burst_idx"], r["crisperwhisper_type"], "|", r["class_v2"], r["kept"],
          "->", r["vb2_label"], round(r["vb2_prob"], 2), "dip", round(r["vb2_dip_db"], 1), "dB")

t = pq.read_table("data/transcripts/transcripts-007.parquet").to_pylist()
r = t[0]
for old, new, dip in zip(r["bursts"], r["vb2_labels"], r["vb2_dip_db"]):
    print(old["start"], old["end"], old["type"], "->", new, dip)

How it was built

CrisperWhisper-large-v2 had already produced the transcript, the word timestamps and the burst spans for this corpus. Neither ASR nor a burst locator was re-run here. What this pipeline did:

1. Duration floor. Every span shorter than 200 ms is dropped. This is a property of the span, not of the classifier, so it is applied first — a sub-200 ms span is discarded whatever it would have been called.
2. Revalidation. Each surviving span is cut (± 50 ms of context), embedded with laion/voiceclap-commercial and classified by laion/vocal-burst-detector-v2. The five folded classes (Blowing a Kiss, Finger Snaps, Hand Scratching Head, Hand Slaps, Slap Face) are masked to −∞ before the softmax, matching the model repo's own inference.py. Masking after argmax instead — remapping a folded win to no_burst — throws the burst away rather than falling back to the next-best real class. A span with p(no_burst) ≥ 0.5 is vetoed.
3. Rescoring. Every utterance and every sentence of ≥ 8 words is scored on its own audio with laion/Empathic-Insight-Voice-Plus — not the distilled VoiceCLAP attribute heads, which were measured at a median Spearman ρ of 0.38 (utterance) / 0.32 (sentence) against it on 72,809 matched pairs from this same corpus. A caption only uses the top-k dimensions by |z|, so rank agreement is exactly the question that matters, and at ρ ≈ 0.35 the two scorers do not pick the same dimensions.
4. Captioning. caption.py at commit 4064baf10704 of LAION-AI/procedural-voice-captions, with the in-domain baseline_stats.json and reliability weighting.

Models

role	model
transcript, word timestamps, burst spans (upstream, not re-run)	CrisperWhisper-large-v2
burst re-classification (vb2_*, 2026-09-10)	laion/vocal-burst-detector-x2 production head on voiceclap-large-v2
burst classifier	laion/vocal-burst-detector-v2 — vocal_burst_mlp_v2.pt, 82 classes + no_burst
embedder	laion/voiceclap-commercial
57 voice dimensions	laion/voicenet-dimension-predictors-commercial
genuineness (0–6)	laion/voiceclap-commercial-genuineness
vocal-burst blend (0–10)	laion/voiceclap-commercial-vocalburst-blend
40 emotions + 4 quality + gender gate	laion/Empathic-Insight-Voice-Plus on laion/BUD-E-Whisper
captioner	LAION-AI/procedural-voice-captions@4064baf10704, baseline_stats.json (sha256 b35c8ea74e2157fd…)

All 40 emotions are present. A previous annotation of this corpus stored only 39 — the Jealousy_&_Envy head was missing from the scorer's emotion list and therefore never loaded. Heads are keyed by checkpoint name, not by index, so the other 39 were correctly labelled; the 40th was simply never predicted. Everything here was recomputed rather than reused.

---

Re-deriving the captions

The captions are a pure function of data/scores + a baseline + a template. Nothing else.

# reproduce the shipped captions exactly
python derive_captions.py --scores data/scores --out /tmp/again

# the same clips against the previously published (out-of-domain) baseline
python derive_captions.py --scores data/scores --out /tmp/emolia \
       --baseline captioner/baseline_stats_emolia.json

CPU only, stdlib + pyarrow, no audio, no model download. provenance.json and each caption parquet's caption_provenance metadata record the baseline hash and captioner commit, so a regenerated set is always distinguishable from the shipped one.

To vary the wording instead of the baseline — 11 surface templates, synonym rotation, dim shuffling, all from the same stored scores — see augment.py in the captioner repo. That is the intended use for training-time text augmentation: score once, caption many times.

---

Numbers

utterances	1,494,503
audio	796 GiB, 11,840 h, 48 kHz mono MP3 @ 160 kbps
mean clip length	28.5 s
sentences	6,284,328 (scored: 4,588,699, 73.0 %)
burst spans in	5,401,483
dropped, under 200 ms	3,347,356 (62.0 %)
dropped, unusable timestamp	0
revalidated by the classifier	2,054,127
vetoed, p(no_burst) ≥ 0.5	1,019 (0.05 % of revalidated)
kept	2,053,108 (38.0 % of the input spans)

Burst duration (all input spans): p10 0.02 s · p25 0.08 s · p50 0.16 s · p75 0.26 s · p90 0.36 s.

Say this plainly: the 200 ms floor removed 3,347,356 of 5,401,483 spans, and the median input span is 0.16 s — shorter than the floor. The cut therefore does not trim a tail, it goes through the middle of the distribution, and the shipped burst set is biased towards longer bursts. Short events — clicks, lip smacks, quick breaths, single-frame hesitations — are systematically absent from transcripts and captions. They are all still in data/bursts with drop_reason = "under_200ms" if you want them back, but they were never classified, so recovering them means re-running the classifier on those spans.

Class distribution of the kept bursts (top 25 of 82)

class	n	share
Low Mumble	537,298	26.2 %
Ahem	511,277	24.9 %
Contented Sigh	439,481	21.4 %
Surprised Gasp	110,991	5.4 %
Breathy Giggle	105,813	5.2 %
Wistful Sigh	94,737	4.6 %
Chuckle	80,273	3.9 %
Exhausted Groan	62,849	3.1 %
Childlike Giggle	54,094	2.6 %
Sharp Inhale	32,500	1.6 %
Resonant Hum	10,325	0.5 %
Scream	2,653	0.1 %
Soft Hum	2,341	0.1 %
Exasperated Sigh	1,562	0.1 %
Tsk	1,276	0.1 %
Yawn	1,075	0.1 %
Cackle	703	0.0 %
Sniff	615	0.0 %
Cough	614	0.0 %
Deep Breath	569	0.0 %
Effort Grunt	469	0.0 %
Whispered Mumble	437	0.0 %
Coughing	415	0.0 %
Snort	185	0.0 %
Purr	107	0.0 %

CrisperWhisper tag vs. v2 class

CrisperWhisper emits ~11 coarse tags; the classifier emits 82 fine ones. There is no ground truth here, so "agreement" is measured against a hand-written, deliberately generous mapping from each coarse tag to the set of taxonomy classes that would count as the same event (e.g. laughter → all 8 laughter classes, breath → the 7 breathing + 3 gasp/inhale classes).

UH, UM and noise have no counterpart in the taxonomy at all — the first two are filled pauses, which are speech, not vocal bursts — so for those spans agreement is undefined, not zero. They are the majority of the input.

CrisperWhisper tag	kept spans	agreement	most common v2 classes
laughter	374,342	38.3 %	Contented Sigh (27.1 %), Breathy Giggle (17.1 %), Chuckle (15.1 %)
yawn	202,689	0.2 %	Contented Sigh (73.3 %), Surprised Gasp (10.0 %), Wistful Sigh (7.4 %)
breath	59,106	5.0 %	Contented Sigh (75.1 %), Ahem (7.4 %), Wistful Sigh (4.2 %)
sigh	35,389	93.6 %	Contented Sigh (89.3 %), Wistful Sigh (4.3 %), Surprised Gasp (2.0 %)
lipsmack	12,155	1.9 %	Ahem (32.2 %), Contented Sigh (26.5 %), Low Mumble (12.8 %)
throatclearing	37,655	39.3 %	Ahem (39.3 %), Contented Sigh (12.0 %), Surprised Gasp (10.1 %)
sniff	28,016	0.6 %	Contented Sigh (39.9 %), Sharp Inhale (36.3 %), Ahem (8.3 %)
cough	37,175	1.8 %	Breathy Giggle (20.3 %), Surprised Gasp (16.3 %), Ahem (14.5 %)
UH	996,293	n/a	Low Mumble (37.2 %), Ahem (33.7 %), Contented Sigh (8.6 %)
UM	265,013	n/a	Low Mumble (52.3 %), Ahem (34.7 %), Wistful Sigh (3.1 %)
noise	3,971	n/a	Ahem (38.5 %), Low Mumble (26.9 %), Surprised Gasp (7.4 %)

Over the mappable tags only: 24.9 % (195,864 / 786,527). That is low, and it is not uniform: sigh agrees 93.6 %, but yawn 0.2 %, sniff 0.6 % and cough 1.8 %. Which of the two labellers is wrong on those spans cannot be determined from this data — there are no human labels here. Read the number as "the two models disagree", not as "the classifier is 25 % accurate".

---

Known limitations

Stated plainly, because they change what these annotations can be used for.

• Burst timestamps come from an ASR model, not from human annotation. Every start/end in this dataset is CrisperWhisper-large-v2's, carried over unchanged. They were never verified against a human. A boundary that is 100 ms off is invisible here.
• Burst class labels are a classifier's top-1. laion/vocal-burst-detector-v2 reports 58.1 % argmax accuracy on its own held-out validation set (1,940 clips, 83 classes). Roughly two in five labels are wrong at the class level, and the model's own held-out set is easier than in-the-wild ASR-cut spans. Use prob, top5 and the group-level taxonomy in schema.json if you need something more robust than the top-1 string.
• The class distribution collapses onto a handful of labels. Only 56 of the 82 classes are ever predicted, and the top three — Low Mumble (26.2 %), Ahem (24.9 %) and Contented Sigh (21.4 %) — cover 72.5 % of all kept bursts. Much of that is the input distribution — 64.6 % of the spans CrisperWhisper found are filled pauses with no taxonomy counterpart, and they land mostly on Low Mumble and Ahem — but it means the fine-grained tail of the taxonomy is not usefully populated here. Treat the labels as a coarse event type, not as 82-way ground truth.
• The no_burst gate is effectively inert on this corpus. It vetoed 1,019 of 2,054,127 classified spans (0.05 %). Whether that means the ASR-located spans really are almost all vocal events, or the classifier is simply reluctant to emit no_burst on speech-adjacent audio, is not established here. p_no_burst is stored per span so a stricter gate can be applied without recomputation.
• The caption baseline is in-domain for expressive acted speech. baseline_stats.json was measured on 256,000 clips of this corpus. "Average" therefore means average for a dramatic synthetic performance. Applied to calm read speech the same captioner will overstate deviations — a neutral audiobook voice will be described as extremely flat, extremely low-arousal and so on, because it sits far from this baseline's centre. Regenerate against a broader baseline for such material; that is what derive_captions.py --baseline is for.
• 11 of the 40 emotion baselines, and the blend baseline, are not in-domain. They were carried over from the older 722-clip Emolia measurement because the run that built the in-domain baseline produced no values for them. Jealousy & Envy is one of them — the very emotion recovered here — and its carried-over spread (0.19) is narrower than the in-domain emotion spreads, so its z-scores run large and it will be selected for captions more often than it should be. The affected keys are listed in captioner/baseline_stats.json under _meta.imported_from.carried_over, and each carries "source": "carried_over_from_emolia_baseline".
• Scores cover the first 30 seconds. Both scorers are 30 s-window models; longer clips are truncated, not chunked. A small number of clips in this corpus run past 30 s, and their global scores describe only the beginning. Per-sentence scores are unaffected in practice (no sentence is that long).
• Sub-200 ms spans were never embedded. They are in data/bursts with drop_reason = "under_200ms" and null classifier fields. Lowering the floor later requires re-running the classifier on those spans; every other threshold can be changed from the stored data alone.
• <GENERAL> can say "clean of non-verbal vocal bursts" on a clip that has bursts. That clause comes from the vocal-burst blend regression head, which scores how burst-inflected the delivery sounds overall — it is a different measurement from the located spans and the two can disagree.
• The transcripts are of synthetic speech, and much of it is deliberately degenerate. This is the edge-case selection: prompts were chosen to stress the TTS model, so transcripts contain disfluencies, invented words and broken syntax. That is the corpus, not an annotation error. See Use as an ASR training set below — this property is why the transcripts are interesting, not a reason to discard them.
• id is unique, but only after a repair. Six of the 128 source annotation shards (17, 52, 74, 75, 77, 79) contained 7,657 duplicated records - the same (group_key, seed) written twice, pointing at the same source audio, with identical duration and reward. The scoring pipeline carried them through, so the first build of this package had 1,502,160 rows for 1,494,503 distinct ids and 7,657 tar members with a duplicated name. They were byte-identical, so the extra copy was removed from all five families; nothing was lost. id is now a primary key in every family and every audio tar - verified, 0 duplicates, 0 orphans in either direction. If you also consume the raw top3_ann jsonl, be aware the duplication is still there.
• 5,362 utterances of the 1,507,522 in the source annotations are absent. Their CrisperWhisper word list was empty (no speech transcribed), so there was nothing to segment, place bursts in or caption. That is 0.36 % of the corpus, and it is the only source of row loss — no utterance was dropped for a decode or read error.
• source_tar / member point at paths that are not public. They record where each utterance came from in the generation output (edgecases_audio/**/*.tar). The audio itself is included here, byte-for-byte, under data/audio/ — those two columns are provenance, not a retrieval path.

Use as an ASR training set

A large share of these utterances is hard to understand — slurred, shouted, whispered, sobbed, strained, cut off mid-word. That is by construction: this is the edge-case selection, generated from prompts written to push the TTS model past clean read speech, and kept by a reward that favours expressive delivery rather than intelligibility. Judged as a text-to-speech corpus, a lot of it would be rejected.

Judged as an ASR corpus, that is exactly what makes it valuable. Clean read speech is abundant; speech that is expressive to the point of breaking down is not, and it is where recognisers fail.

The transcripts hold up under that stress. They come from CrisperWhisper large-v2, which is built for verbatim transcription — it keeps disfluencies, repetitions and false starts instead of tidying them away, and gives word-level timestamps. On top of that, every vocal burst carries a corrected class label: the original burst annotations were re-classified with laion/vocal-burst-detector-v2 at their original timestamps, so a laugh, gasp or sigh in the middle of a line is named rather than transcribed as a word or dropped.

So the intended reading is: degraded, expressive audio paired with accurate verbatim text and labelled non-verbal events. Useful for training or evaluating recognisers on expressive and disfluent speech, for verbatim-vs-clean transcription styles, and for models that must handle vocal bursts inside an utterance instead of ignoring them.

Two honest caveats. The audio is synthetic, so acoustic conditions are those of the generator, not of a microphone in a room — expect it to complement real-world ASR data rather than replace it. And CrisperWhisper is a model, not a human annotator: on the most degraded utterances its output is a strong transcript, not ground truth. data/bursts carries prob and p_no_burst per span if you want to filter on classifier confidence.

License

CC BY 4.0 — audio and annotations alike. The annotations were produced with LAION models from public repositories.

Citation

Models: see the table above. Captioner: LAION-AI/procedural-voice-captions. Burst taxonomy: LAION-AI/voice-taxonomies.

---

The audio in this dataset was generated with a voice-acting AI model.