laion/dramabox-edge-top3-voice

DramaBox edge-case reward-Top-3 — voice annotations

1,494,503 synthetic expressive-speech utterances (the reward-Top-3 selection of the DramaBox 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.

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.

---

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
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("REPO_ID", "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 DramaBox 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).

---

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 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 DramaBox 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

The annotations were produced with LAION models from public repositories; the audio is synthetic DramaBox output. The license field is other — the usage terms for the synthetic audio have not been settled, so treat this as a research release and check with LAION before redistributing.

Citation

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