laion/dramabox-burst-audio

DramaBox vocal-burst scenes — audio, annotated

73,500 generated scenes across 49 vocal-burst classes that our burst adapters handle badly or not at all — 500 prompts per class, each rendered at 3 seeds, so every prompt appears three times and a class effect can be told apart from a sampling accident.

Every clip now carries a Gemini 3.8 Flash annotation: where the bursts are and what they are. 123,004 events over 73,500 clips (1.674 per clip), 97 distinct labels.

Generated with DramaBox TTS at its default settings on 32 GH200 nodes (1 h 49 m, 128 ranks, zero failures). Prompts published separately at laion/dramabox-burst-prompts.

Why this exists

Our own burst detector emitted Shriek zero times across a 60-clip audit, used 8 of its 83 labels, and put the requested burst in its top-3 on 3 of 60 clips. Every burst adapter this project ships was trained on rows that detector labelled. This corpus gets burst audio from a different generator, densely and on purpose, and labels it with a different annotator.

The annotation is a second model's opinion, not verified ground truth. No human has listened to these. A prompt asking for a shriek is not evidence a shriek was produced, and neither is a label saying so.

The annotation

clips annotated	73,500
successful responses	73,500
errors / unparsable	0
clips with no response	0
burst events found	123,004
events per clip	1.674
clips where the model heard no burst	394
distinct labels used	97
median event length	1.27 s
spans clamped to the clip	0

Model gemini-3.8-flash, thinkingLevel: "low", temperature 0, JSON schema enforced, run through the Batch API. Measured cost: 100.8 M input and 12.6 M output tokens, 2,362,228 thinking tokens.

Each clip is shown to the model whole, never pre-cut: this model's labels are context-dependent, and on excerpts it had itself left unannotated it returned Scream or Shriek eleven times out of twenty. A span reaching past the end of its own audio is clamped and marked out_of_range, never dropped, so the counts above reconcile exactly.

The 25 most common labels

label (first choice)	events
Deep Breath	18,121
Exasperated Sigh	17,771
Humming	10,449
Chuckle	9,893
Relief Sigh	6,043
Sharp Inhale	5,937
Yawn	4,089
Scream	3,637
Cackle	2,434
Hiss	2,199
Displeased Grunt	1,991
Frustrated Groan	1,810
Surprised Gasp	1,760
Affirmative Grunt	1,699
Hiccup	1,691
Growl	1,633
Contented Sigh	1,621
Guffaw	1,602
Breathy Giggle	1,601
Soft Hum	1,527
Coughing	1,522
Wistful Sigh	1,478
Exhausted Groan	1,440
Childlike Giggle	1,423
Sniff	1,416

Files and layout

81 files at the repository root.

path	count	what it is
shard-00000.tar … shard-00073.tar	74	the audio, plus two JSON sidecars per clip
metadata.parquet	1	one row per clip, 7 columns — filter the corpus without opening a tar
stats.json	1	corpus totals and the per-class / per-gender / per-age counts
annotation_stats.json	1	the Gemini annotation run: label histogram, event counts, token usage
vocal_burst_groups.json, GROUPS.md	2	the 23-group scheme over 118 burst label strings
README.md	1	this card

1000 samples per shard, 500 in the last. Shard order is a fixed seeded permutation, so one shard already contains the whole distribution — no large shuffle buffer needed.

Inside a tar — three members per clip, one stem

member	format	detail
<key>.mp3	MPEG-1 Layer III, 48 kHz, 160 kbps CBR, mono	as written by the generator — not re-encoded, and not re-encoded again when the annotations were added
<key>.json	JSON	the generation record
<key>.gemini.json	JSON	the Gemini annotation

Keys are <class>__<prompt index>__s<seed>, e.g. affirmative_grunt__0378__s5678.

Read back from the shipped bytes, not assumed: every file's MPEG frame header was checked during packing; all 73,500 passed and all 73,500 keys round-tripped out of the written shards exactly once. The clips' own <key>.json records the same thing as {"codec": "mp3", "bitrate_kbps": 160, "channels": 1, "sr": 48000}.

There are no codec tokens here

This repository ships no MOSS codes, no .moss.npy, no embeddings and no .vclap.npy — audio and two JSON sidecars, that is the whole payload. If you need MOSS tokens for this material you must encode it yourself with OpenMOSS-Team/MOSS-Audio-Tokenizer-v2. There is no transcript column either: the spoken line is inside prompt / spec, as it was asked for, not as it came out.

<key>.json — the generation record

key	meaning
key	the clip key; also the tar member stem
cls	the vocal-burst class this clip was generated for (one of 49)
idx	prompt index within the class, 0..499
seed	generation seed; each prompt is rendered at 3 seeds
prompt	the full text prompt handed to the generator
gender, age_band	the requested speaker gender and age band
spec	the structured prompt specification: age, register, place, topic, trigger, phon (the burst as written), position, return_to_calm, intensity
settings	generator settings: cfg_scale, stg_scale, duration_multiplier, gen_duration, ref_duration
audio	{codec, bitrate_kbps, channels, sr} of the file beside it
dur_s, peak	duration in seconds and the peak sample value

<key>.gemini.json — the annotation

key	meaning
no_burst	the model heard no vocal burst at all in the clip
events	list of {start_s, end_s, labels[1-3], confidence, description} — labels[0] is the first choice
overall	one sentence describing the whole clip
usage	token counts for the call
model, thinking_level	gemini-3.8-flash, low

A span reaching past the end of its own audio is clamped and marked out_of_range, never dropped, so the counts in this card reconcile exactly.

metadata.parquet — the column reference

7 columns, 73,500 rows — one per clip, read from the shipped schema:

column	type	meaning
key	string	the clip key. This is the tar member stem.
cls	string	the vocal-burst class the clip was generated for. 49 distinct values, 1,500 clips each, exactly
seed	int64	generation seed. Exactly three values — 1234, 5678, 9012 — 24,500 clips each
gender	string	requested speaker gender — male / female, 36,750 each
age_band	string	requested age band — young_adult 24,696 / older_adult 24,402 / middle_aged 24,402
shard	string	the tar file name, with the .tar extension, e.g. shard-00000.tar. 74 distinct values
prompt	string	the full text prompt that produced the clip

Note the spelling of shard: it carries the extension. Use it as the file name directly, do not append .tar a second time.

The join

import tarfile, io, json
import pyarrow.parquet as pq

meta = pq.read_table("metadata.parquet")
row  = meta.slice(0, 1).to_pylist()[0]

with tarfile.open(row["shard"]) as tf:            # "shard-00000.tar" — extension included
    stem  = row["key"]                            # the member stem
    audio = tf.extractfile(stem + ".mp3").read()
    gen   = json.loads(tf.extractfile(stem + ".json").read())
    ann   = json.loads(tf.extractfile(stem + ".gemini.json").read())

print(row["cls"], ann["events"][0]["labels"], gen["audio"])

Verified on shard-00000: the parquet's first rows for that shard name members that are in the tar, in the tar's own order.

Or stream the whole thing and ignore the parquet:

import webdataset as wds
ds = wds.WebDataset("shard-{00000..00073}.tar").decode()
for r in ds:
    audio, meta, ann = r["mp3"], r["json"], r["gemini.json"]

Balance

Held by construction, not by luck:

male / female	36,750 / 36,750
young / middle-aged / older adult	24,696 / 24,402 / 24,402
per class	1,500 (500 prompts × 3 seeds)
audio	32.8 GB

Audio format

160 kbit/s mono 48 kHz MP3, as written by the generator — not re-encoded, and not re-encoded again when the annotations were added. Every file's MPEG frame header was read back out of the bytes during packing; all 73,500 passed and all 73,500 keys round-tripped out of the written shards exactly once.

Related

• laion/dramabox-burst-prompts — the prompts
• laion/vocal-bursts-gemini-segments — real-speech burst segments, CC BY 4.0
• laion/vocal-burst-classifier-gemini — the classifier trained on those

Licence and provenance

The audio was generated by DramaBox TTS; check that model's licence before training on this corpus. The prompts, the annotations and the packaging are Apache-2.0.

---

Correction (2026-09-04)

An earlier version of this card said the shipped vocal-burst-detector-v2 "used 8 of its 83 labels" and "emitted Shriek zero times". That was measured on a 60-clip audit and does not generalise. Run over both full corpora (72,500 DramaBox clips and 3,598 real utterances), the same detector emits 41 distinct labels on the DramaBox half and 36 on the real half, and it does emit Shriek — on 50 DramaBox clips and 49 real ones.

What the 60-clip audit actually saw is how concentrated the distribution is: Contented Sigh alone accounts for 16,694 of 72,500 DramaBox clips, Surprised Gasp 7,451, Wistful Sigh 5,437. The correct statement is that the detector's effective vocabulary is small, not that its emitted vocabulary is 8. The conclusions drawn elsewhere in this card — that the detector is the bottleneck, and that its failure is granularity rather than deafness — are unaffected and are independently supported by the cross-source results in §62.

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.