Access request — LAION TTS Annotated v1, research subsets

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

The audio in this repository comes from podcasts that are openly available on the internet and consists of short snippets only. We cannot redistribute the audio itself, so it is made available here for non-commercial research use by collaboration partners within our TTS research. Please state who you are and what you intend to use the data for.

Log in or Sign Up to review the conditions and access this dataset content.

LAION TTS Annotated v1 — research subsets

29,739,936 annotated speech utterances across three subsets — with the audio, the codec tokens and the complete annotation stack.

The audio in this repository comes from podcasts that are openly available on the internet and consists of short snippets only. We cannot redistribute the audio itself, so it is made available here for non-commercial research use by collaboration partners within our TTS research.

The other six subsets of this corpus — 107,563,551 utterances — are openly published at laion/laion-tts-annotated-v1. Together the two repositories are the complete 137,303,487-utterance corpus. The schema, the layout and the join are identical in both, so code written against one reads the other with no change beyond the repository id.


What is in here

utterances 29,739,936
audio-hours 103,912
audio, shipped (MP3) 8.04 TB — 8.12 TB at the source
MOSS codec tokens 112 GB
VoiceCLAP embeddings 49.5 GB
annotation parquet 24.8 GB
WebDataset shards 8,234 tars
audio coverage 100.000 %
MOSS coverage 100.000 %
vocal-burst spans, re-classified 22.8 M (of 63.8 M corpus-wide)

Coverage was measured over every row of every shard, not sampled. The audio and MOSS sizes are sums of audio_bytes and moss_frames × 24 B over every parquet footer of these subsets; the VoiceCLAP size is rows × 1,664 B, exact by construction.

Full annotation stack, including the transcripts: word-level timings, 40 emotion intensities, 57 VoiceNet voice-character dimensions, four audio-quality heads, vocal-burst detections with timings and the procedural caption. Nothing is withheld at the column level.

Subsets

subset utterances hours shards source audio source format shipped as annotation parquet
podcast 26,606,892 96,838 6,496 7.00 TB MP3 48 kHz mono, 161 kbps source MP3, 1:1 22.4 GB
snippets 2,633,037 4,875 1,410 0.35 TB MP3 48 kHz mono, 161 kbps source MP3, 1:1 1.9 GB
evasnippets 500,007 2,199 328 0.76 TB WAV pcm_s16le 48 kHz mono MP3 128 kbps CBR, 48 kHz, mono 0.4 GB
total 29,739,936 103,912 8,234 8.12 TB 8.04 TB shipped 24.8 GB

Every shipped audio member is MP3shipped_audio_format is "mp3" on every row. evasnippets was uncompressed WAV at the source and was transcoded to MP3 128 kbps CBR, 48 kHz, mono: sample-exact at the start, at most +44 samples at the end (a deterministic LAME edge case), so every timestamp in words, burst_starts and burst_ends stays valid. audio_bytes is the source size (the WAV), shipped_audio_bytes the size of the MP3 that is actually in the tar; on podcast and snippets the two are equal. Every clip is mono.

evasnippets is, by material, 98.56 % podcast: its cond_podcastt_* uids all trace to clips that are themselves in the podcast subset, so the two subsets overlap in source audio and should not be treated as independent samples.

Sources

subset upstream source upstream licence
podcast TTS-AGI/podcast-tokenized-bg2.5-enj4.5 none declared upstream
snippets TTS-AGI/enhanced-audiosnippets-DACVAE none declared upstream
evasnippets TTS-AGI/Emotion-Voice-Attribute-Reference-Snippets-DACVAE-Wave declares CC-BY-4.0; a merge of two upstream repositories whose own sources are not documented

The annotations — the captions, scores and derived measurements — are LAION's own work and are offered under CC-BY-4.0. The audio is not covered by that.


Layout

<subset>/<shard>.parquet     the annotations — 226 or 234 columns, one row per utterance
<subset>/_sources.json       the upstream repositories the subset's rows came from
data/<subset>/<shard>.tar    the audio, the MOSS codes, the VCLAP embedding, the JSON record
COLUMNS.md                   the generated column reference
per_class_recall.json        per-class recall of the x2 re-classification head (vb2_*)
per_class_recall_x3.json     per-class recall of the x3 large-v2 head (vb3_*), with the x2 value on the same rows
per_class_recall_x3_commercial.json   the same for the x3 commercial head (vb3c_*)

<shard> is <subset>-NNNNN, and shard n of <subset>/ describes exactly the utterances in shard n of data/<subset>/.

Inside a tar — four members per utterance, one stem

member format detail
<uid>.mp3 MP3 on every row — see the subset table for source vs shipped source bytes 1:1 where the source was MP3, otherwise transcoded (below)
<uid>.json JSON the upstream record verbatim, plus the annotations and words
<uid>.moss.npy numpy uint16, shape [T, 12] MOSS-Audio-Tokenizer-v2 — 12 codebooks x 1024, 12.5 fps, one frame = 12 tokens = 80 ms
<uid>.vclap.npy numpy float32, shape [768] VoiceCLAP embedding, 1,664 bytes including the npy header

moss_frames == floor(dur_s * 12.5) holds corpus-wide. Checked on an individual clip: podcast/39805_00035956 is 18.792 s and its moss.npy is (234, 12); 234 / 12.5 = 18.72 s.

The join

shard is the tar basename without its extension. uid is the member stem.

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

row = pq.read_table("podcast/podcast-00000.parquet",
                    columns=["uid", "shard", "text", "caption_general",
                             "dur_s", "audio_bytes", "moss_frames", "moss_n_vq"]).slice(0, 1).to_pylist()[0]

with tarfile.open(f"data/podcast/{row['shard']}.tar") as tf:
    stem  = row["uid"]
    audio = tf.extractfile(stem + ".mp3").read()      # MP3 on every row, evasnippets included
    rec   = json.loads(tf.extractfile(stem + ".json").read())
    codes = np.load(io.BytesIO(tf.extractfile(stem + ".moss.npy").read()))    # uint16 [T, 12]
    vclap = np.load(io.BytesIO(tf.extractfile(stem + ".vclap.npy").read()))   # float32 [768]

assert len(audio) == row["audio_bytes"]
assert codes.shape == (row["moss_frames"], row["moss_n_vq"])

Checked per subset on 30 random uids from 6 random parquets: audio member 30/30, .moss.npy 30/30, .vclap.npy 30/30. And on 8 random shards per subset, parquet rows == audio members == .moss.npy members, 8/8 each.


Columns

226 columns for snippets and evasnippets, 234 for podcast (since 2026-09-11; 203 / 211 before the x3 columns) — the eight extra are the ref_* speaker-pointer block. The complete generated reference, with a one-line meaning for every column, all 40 emotion names and all 57 VoiceNet dimension names, is in COLUMNS.md.

group columns what it is
identity 5 uid, shard, dataset, src_shard, langthere is no speaker-key column; podcast and kartoffelphon carry ref_* same-voice pointers instead, vprof_vc carries voice
language detection 5 lang_iso (use this one), lang_orig (wrong on 4.7 M rows corpus-wide — do not use), lang_asr, lang_conf, lang_src
text and alignment 6 text, text_with_bursts, text_source, n_words, align_status, and words (word-level w/s/e timings — in the tar JSON, not in the parquet)
caption 2 caption_generalthe valid caption; caption_general_v1 — the previous version with inverted gender and background polarity, kept only for traceability, not an alternative
emotion 40 emo_*, Empathic-Insight-Voice intensities. Not on a shared scale
voice character 114 vn_<CODE>_reg + vn_<CODE>_bucket over 57 VoiceNet dimensions
quality 6 4 qual_* heads, genuineness_0_6, blend_0_10
vocal bursts 5 n_bursts, burst_labels, burst_starts, burst_ends, burst_placement
codec and audio metadata 4 dur_s, audio_bytes (source size), moss_frames, moss_n_vq — counts, not audio
top emotion 2 top_emotion, top_emotion_value
vocal-burst re-classification 13 vb2_* — the same spans re-labelled with the 17-class head; see below
shipped audio 2 shipped_audio_bytes, shipped_audio_format — what is actually in the tar (MP3 everywhere)
speaker pointers (podcast only) 8 ref_uid, ref_shard, ref_cos, ref_scope, ref_dur_s, ref_n_cand, ref_uids_topk, ref_cos_topk

ref_uid coverage in podcast is 15.4 % with median cosine 0.820 against a 0.80 bar. Treat ref_n_cand as a first-class filter, not as a guarantee.

Vocal bursts — two annotations side by side

The corpus carries three vocal-burst annotations on every row, and no newer one replaces an older one. (vb3_* / vb3c_*, added 2026-09-11, are described after this section.)

annotation head vocabulary columns
original (v2 detector) vocal-burst-detector-v2 83 classes, of which 21 are observed on this corpus n_bursts, burst_labels, burst_starts, burst_ends, burst_placement, text_with_bursts
re-classification (2026-09-09) laion/vocal-burst-detector-x2, production head: VoiceCLAP-large-v2 (3584-d, frozen) + 5 MLPs averaged in probability space 17 classes = 16 bursts + no_burst vb2_n, vb2_labels, vb2_probs, vb2_no_burst, vb2_families, vb2_top3, vb2_top3_p, vb2_dip_db, vb2_loc_conf, vb2_flag_short, vb2_missing, vb2_model, vb2_version

The re-classification labels the same spans: vb2_n == n_bursts, and vb2_labels[i] is the new reading of burst_labels[i]. vb2_labels is never no_burst — the reject probability is in vb2_no_burst, and no gate was applied (P(no_burst) >= 0.5 on 0.03 % of spans). Every parquet also carries vb2_model, vb2_version and the 17 class names (vb2_classes) in its file metadata. The 16 burst classes, grouped by vb2_families:

family classes
breath Sharp Inhale, Deep Breath, Heavy Breathing, Panting
sigh Exasperated Sigh, Relief Sigh, Wistful Sigh
groan Affirmative Grunt, Exhausted Groan, Frustrated Groan
laugh Chuckle, Breathy Giggle
hum Humming, Soft Hum
scream Scream
yawn Yawn

Why both annotations are kept. The vocabularies are not nested: of the 21 labels the original detector emitted on this corpus, only 10 exist in the 17-class vocabulary, and 83.8 % of the original events carry a label the new head does not know — Low Mumble and Ahem alone are about half of all spans. On those spans the head must choose something, and it mostly chooses Sharp Inhale: 70 % of all vb2_labels are Sharp Inhale, and roughly two thirds of those are Ahem / Low-Mumble spans the new head cannot name — not 45 million inhalations. Measured on 12,293 spans from 4 shards, confidence by original label:

original label n vb2_probs mean / median vb2_dip_db median new label
Low Mumble 6,611 0.561 / 0.55 -4.0 Sharp Inhale 61 %, Frustrated Groan 16 %
Ahem 4,038 0.684 / 0.71 -3.6 Sharp Inhale 77 %, Affirmative Grunt 10 %
Surprised Gasp 348 0.697 / 0.73 -3.8 Sharp Inhale 79 %
Contented Sigh 140 0.661 / 0.68 -7.3 Sharp Inhale 76 %
Wistful Sigh 135 0.576 / 0.56 -3.4 Sharp Inhale 50 %, Exhausted Groan 13 %
Chuckle 252 0.779 / 0.83 -4.5 Chuckle 87 % (confirmed)
Sharp Inhale 14 0.745 / 0.80 -8.5 Sharp Inhale 93 %

Outside the vocabulary the confidence is only moderately lower (median 0.55-0.73 against 0.83 on real chuckles), so vb2_probs alone is not a clean filter. Read vb2_labels on those spans as "the nearest nameable class — breath-like, a quiet transition", not as a verified inhalation, and read the original label and vb2_dip_db beside it. 42 % of the original spans sit in speech at full level (vb2_dip_db > -3 dB, no level dip at all); vb2_dip_db is the column that lets you filter them. The original columns stay exactly as they were, text_with_bursts continues to be rendered from the original labels, and where the two annotations disagree on a span that is information, not an error to be resolved.

The head's per-class recall on held-out real speech — strict and by family, with 95 % intervals and the top confusions — is in per_class_recall.json; a measured hit rate for a class cannot meaningfully exceed its recall there. Strict recall spans 0.90 (Sharp Inhale) to 0.06 (Relief Sigh); no_burst 0.98; Heavy Breathing, Soft Hum and Wistful Sigh have fewer than 30 held-out examples and are marked unreliable.

The third annotation — vb3_* / vb3c_* (x3, 2026-09-11)

The same spans a third time, with laion/vocal-burst-detector-x3, in both of its heads: vb3_* = production/ (VoiceCLAP-large-v2, 3584-d) and vb3c_* = commercial/ (VoiceCLAP-commercial, 768-d). 19 classes = the 16 of x2 + Ahem (throat-clear; Gemini's Clears Throat folded in) + Low Mumble + no_burst. Index-aligned like vb2_*: vb3_n == vb3c_n == n_bursts, vb3_labels[i] is the x3 reading of burst_labels[i]. Nothing was removed or overwritten; vb2_* is exactly as before.

Why x3 exists, and what it changed. The x2 section above reads its 70 % Sharp Inhale as "Ahem / Low-Mumble spans the head cannot name". That was checked against 3,598 utterances that carry both the original spans and a blind whole-clip Gemini annotation, and it is wrong in an important way: a Gemini event overlaps an original Ahem span in 16.5 % of cases and an original Low Mumble span in 7.3 % (against 86.5 % for original Chuckle spans and 91.5 % for Sharp Inhale). The original detector's two commonest labels are mostly locator false alarms in running speech, not throat-clears or mumbles — and x2 never rejects them because its no_burst training rows were long speech/silence windows, never a 0.3 s cut inside speech (x2's P(no_burst) ≥ 0.5 on 0.03 % of spans). x3 therefore (a) adds the two classes and (b) was trained with a third no_burst sub-type, fp_span — exactly such spans (1,115 real + 1,200 DramaBox). Details, ablations and the gate sweep are on the model card.

Read vb3_no_burst first, then the label. P(no_burst) is delivered ungated, as in x2. The recommended gates are 0.9 for vb3_* and 0.5 for vb3c_*: at those, the large head rejects 91 % of the original Ahem/Low Mumble spans and 6 % of Gemini-verified bursts, the commercial head 84 % and 3 %. The two heads reject differently — vb3 broadly (at 0.5 it also drops 14 % of verified bursts), vb3c selectively — so both are shipped and the disagreement is information.

Sharp Inhale is no longer the story. With the gate, the spans that x2 called Sharp Inhale become no_burst; they do not become Ahem / Low Mumble (a real throat-clear is rare in this corpus). Balance x2 → x3 on every span of this repository (≥ 0.9 = share of all spans with vb3_no_burst ≥ 0.9; "kept" = the rest). The single Sharp Inhale percentage uses all original spans as its denominator; the top-class lists are normalized over retained spans:

subset spans x2: Sharp Inhale x3 (vb3_*): rejected at P(no_burst) ≥ 0.9 ≥ 0.5 x3 retained Sharp Inhale (% of all spans) x3 top classes (% of retained spans) x3 Ahem / Low Mumble kept x3c (vb3c_*): rejected ≥ 0.5 x3c top classes (% of retained spans)
podcast 21,793,545 70.2 % 81.8 % 89.4 % 5.5 % Chuckle 32 %, Sharp Inhale 30 %, Exhausted Groan 10 %, Breathy Giggle 7 % 169,712 / 96,312 72.3 % Sharp Inhale 23 %, Affirmative Grunt 21 %, Chuckle 19 %, Frustrated Groan 12 %
snippets 621,686 73.5 % 76.6 % 86.7 % 9.5 % Sharp Inhale 41 %, Chuckle 21 %, Exhausted Groan 8 %, Frustrated Groan 6 % 5,336 / 3,119 74.1 % Sharp Inhale 27 %, Affirmative Grunt 19 %, Chuckle 14 %, Frustrated Groan 10 %
evasnippets 349,224 61.8 % 73.3 % 80.6 % 5.4 % Chuckle 51 %, Sharp Inhale 20 %, Breathy Giggle 9 %, Frustrated Groan 6 % 2,040 / 605 65.5 % Chuckle 36 %, Sharp Inhale 20 %, Affirmative Grunt 11 %, Frustrated Groan 11 %
all nine subsets (both repos) 63,838,964 70.3 % 83.8 % 91.5 % 5.3 % Sharp Inhale 33 %, Chuckle 23 %, Exhausted Groan 10 %, Affirmative Grunt 7 % 473,528 / 528,909 78.4 % Sharp Inhale 26 %, Affirmative Grunt 24 %, Chuckle 15 %, Frustrated Groan 13 %

What the new classes are worth. Ahem was learned from 44 real + 1,055 DramaBox (TTS) segments and transfers to real speech at 0.23–0.31 recall (25 held-out rows; chance 0.05). Low Mumble does not hold on real speech: 14 real training segments, 0 % recall on real held-out rows — treat Low Mumble in vb3_labels as a DramaBox-only class and read vb3_no_burst / vb3_dip_db beside it. On the 16 old classes x3 loses ~6 points of balanced accuracy on real speech against x2 on the same held-out rows (mostly Frustrated Groan, Breathy Giggle, Sharp Inhale, which move to a neighbour class), and gains nothing on the prior-weighted sets; per-class recall of both x3 heads, with the x2 value on the same rows, is in per_class_recall_x3.json (large-v2) and per_class_recall_x3_commercial.json beside this file.

Columns (23, appended after shipped_audio_format): vb3_n, vb3_labels, vb3_probs, vb3_no_burst, vb3_families (now also throat for Ahem; Low Mumble is hum), vb3_top3, vb3_top3_p, vb3_dip_db, vb3_loc_conf, vb3_flag_short (span properties, identical to the vb2_ ones by construction), vb3_missing, vb3_model, vb3_version; and vb3c_n, vb3c_labels, vb3c_probs, vb3c_no_burst, vb3c_families, vb3c_top3, vb3c_top3_p, vb3c_missing, vb3c_model, vb3c_version. File metadata: vb3_model, vb3c_model, vb3_version, vb3_classes. Every parquet was re-checked after the merge: row order and every old column value-identical.


Captions were regenerated on 2026-08-23 — read this if you used an earlier copy

Two defects were found in caption_general and fixed corpus-wide.

1. Two ladders were rendered backwards. High vn_GEND is masculine but was rendered "feminine"; high vn_BKGN is cleaner but was rendered "noisy". The numbers were always right; only the prose was inverted.

2. The emotion gate used an absolute threshold. The 40 emotion heads are not on a shared scale, so a fixed cutoff reported the head's scale rather than the clip's emotion. It is now a pooled, tie-aware mid-rank ECDF gate (U >= 0.90, at most 3 emotions named, otherwise the caption says "no dominant emotion").

caption_general is the valid caption. caption_general_v1 is the previous version and carries the inverted polarity — measured: at vn_GEND_bucket 6 (vn_GEND_reg 5.27) caption_general says "strongly masculine", which is right, and caption_general_v1 says "feminine". It is kept verbatim on every row so the change is auditable row by row, not as an alternative caption. Do not condition on it, and do not treat it as a paraphrase. The fix recomputes the caption from the numeric bucket columns rather than string-editing the old caption, so it is idempotent.

If you cached captions before 2026-08-23, rebuild rather than patch.


Related

Citation

@misc{laion_tts_annotated_v1_research,
  title  = {LAION TTS Annotated v1 --- research subsets},
  author = {LAION},
  year   = {2026},
  url    = {https://huggingface.co/datasets/laion/laion-tts-annotated-v1-research}
}
Downloads last month
61