Dataset Viewer
Auto-converted to Parquet Duplicate
recording_id
stringlengths
32
32
language
stringclasses
13 values
minutes
float64
3
48.9
chunks
int64
1
5
max_speakers
int64
3
8
mean_speakers
float64
2.33
8
overlap_ratio
float64
0
0.44
de100/audio/00000145/11H4jRoOBYY
de
11.1
2
5
3.5
0.055
de101/audio/00000140/yD18rTVDiny
de
17.4
2
3
2.5
0.101
de101/audio/00000143/zXk3Hj-Dowo
de
12.3
2
5
4
0.07
de101/audio/00000252/nttlHkOfqcA
de
4.1
1
5
5
0.132
de101/audio/00000403/sHYCOYTGSLY
de
25.3
3
6
5.67
0.06
de101/audio/00000499/bzmE5-KhpBk
de
6.7
1
4
4
0.057
de102/audio/00000098/Rtn2Apt9Nnk
de
6.6
1
7
7
0.171
el000/audio/00000000/1H_WbyO6yoo
el
18.7
2
3
2.5
0.039
en000/audio/00000243/X09vpDKJ3l8
en
5.6
1
5
5
0.086
en002/audio/00000222/WsxoTWVxTQ8
en
3.3
1
3
3
0.131
en003/audio/00000429/Ps-HVAFQ3KI
en
10.3
2
5
4
0.113
en005/audio/00000173/RQlLBDCklrY
en
11.2
2
5
4
0.125
en006/audio/00000375/J3D_Z3JknZM
en
6.4
1
5
5
0.112
en100/audio/00000068/8WdcslHqr7U
en
11.4
2
4
3.5
0.209
en100/audio/00000472/2UOCheMcBxo
en
5.1
1
8
8
0.247
en101/audio/00000238/BzEC82g7QKA
en
8.7
1
8
8
0.061
en102/audio/00000454/bTn_G_jWa24
en
12.7
2
8
6.5
0.118
en103/audio/00000005/ik4BFAHUsbY
en
20
2
8
7
0.074
en103/audio/00000054/pBn_ZPfx9eu
en
23.2
3
6
5.67
0.137
en103/audio/00000433/fCJUuYnJx-u
en
8
1
8
8
0.145
en103/audio/00000459/j1A4LuGenIo
en
5.5
1
7
7
0.236
en105/audio/00000421/rcEqVSHy4X8
en
18.2
2
4
3
0.138
en106/audio/00000023/Evak_5Eg-jy
en
6.4
1
8
8
0.141
en106/audio/00000180/0Pzl4U27598
en
10
1
6
6
0.116
en106/audio/00000287/fdJPyUpmwgo
en
6.6
1
5
5
0.099
en108/audio/00000352/ZVTUxGoPPJ4
en
17.7
2
5
4.5
0.184
en109/audio/00000482/V-RshK8Kc6o
en
4.8
1
8
8
0.051
en110/audio/00000183/nghwLwNOFII
en
10.4
1
5
5
0.153
en114/audio/00000013/oF_mCsAF1fo
en
7.3
1
4
4
0.123
en116/audio/00000403/sK521mYcPKQ
en
11.6
2
6
5
0.141
en117/audio/00000004/DXYCFjSUYTg
en
3.9
1
6
6
0.074
en117/audio/00000432/BZujsWeW_0I
en
10.7
2
7
5
0.087
en118/audio/00000274/oMiYzTqajUc
en
7.8
1
4
4
0.105
en119/audio/00000136/b7qW-0q6fCA
en
6.3
1
5
5
0.084
en122/audio/00000494/i2GCveCZltc
en
5
1
7
7
0.067
en123/audio/00000076/tbAf5Sp8TOg
en
9.5
1
4
4
0.117
en125/audio/00000444/uqq2pf6OZhy
en
14.4
2
4
3.5
0.145
en126/audio/00000381/qepmkfvMG8g
en
6
1
4
4
0.145
en126/audio/00000398/tv82oyPaVwo
en
8.3
1
7
7
0.146
en127/audio/00000015/GN6P5a_QYB0
en
36.2
4
7
5.75
0.1
en127/audio/00000131/W-dGBDN1Fqk
en
4.4
1
4
4
0.115
en128/audio/00000394/AMxAF2vawS8
en
10.8
2
3
3
0.131
es103/audio/00000009/OiuMt2ggBV8
es
3
1
6
6
0.194
es103/audio/00000043/TytDCvVwfCy
es
4
1
3
3
0.283
es103/audio/00000406/-luwzAAILck
es
4.8
1
4
4
0.158
es103/audio/00000408/_ttTLVEYyG0
es
8.9
1
5
5
0.131
es104/audio/00000072/UyHKPM3QyNy
es
6.7
1
4
4
0.144
es105/audio/00000177/6taLBXF_Rlg
es
5.2
1
8
8
0.085
es105/audio/00000181/6TySfeub7zg
es
14.5
2
4
4
0.157
es105/audio/00000181/6VOIQ82AaAM
es
7.4
1
4
4
0.172
es106/audio/00000313/k5-H6VI8CjY
es
4.8
1
3
3
0.166
es106/audio/00000458/-p2ToYR5dP0
es
16.8
2
3
2.5
0.304
es107/audio/00000070/M1sDppe51Hk
es
7.3
1
5
5
0.14
es107/audio/00000365/mSP16Vy3nju
es
20
2
4
4
0.143
fr100/audio/00000394/AYJhfa7jnqk
fr
9.9
1
4
4
0.435
fr100/audio/00000445/OD1hCiK2D0U
fr
3.5
1
4
4
0.116
fr101/audio/00000074/oAfW1ZefD0g
fr
16.1
2
6
5
0.12
fr102/audio/00000021/ggOJfl9baTE
fr
15.7
2
5
5
0.083
fr102/audio/00000185/Vvdh5Pgsqe8
fr
6
1
7
7
0.185
fr102/audio/00000439/2FOf1wHxa2A
fr
14.5
2
6
6
0.078
fr103/audio/00000114/Lr9noIzyNiM
fr
17.8
2
6
5.5
0.228
fr103/audio/00000381/TwMsnngj6Qy
fr
5.7
1
4
4
0.077
hu000/audio/00000005/jZFDYM4t5Sy
hu
16.3
2
4
3
0.01
hu000/audio/00000027/LsPlqAQZnXU
hu
21.2
3
3
2.33
0
it000/audio/00000109/btCzxrsDdD4
it
4.3
1
3
3
0.073
it100/audio/00000097/NIBJR2HS8Yg
it
8.6
1
5
5
0.066
it100/audio/00000162/ohpxYJ5HFzc
it
4.1
1
3
3
0.107
it100/audio/00000193/CoS3hWBjVxA
it
5.3
1
3
3
0.116
it101/audio/00000342/e-ST0ewtsWc
it
7.9
1
3
3
0.174
it101/audio/00000428/VhGGIhg_9eu
it
5.1
1
3
3
0.284
it101/audio/00000430/X5XxLOVOJ8k
it
13.8
2
3
3
0.106
nl000/audio/00000051/WuY_nChE310
nl
13.6
2
6
4.5
0.036
nl100/audio/00000015/u2tvkeEFvyg
nl
3.7
1
4
4
0.13
nl100/audio/00000089/FkPlxFbg3qu
nl
3.8
1
4
4
0.055
nl100/audio/00000095/SI9yIAcmMYg
nl
6.5
1
4
4
0.042
pl000/audio/00000058/O98lrlu7keU
pl
18.6
2
3
2.5
0.05
pl000/audio/00000113/BA9e-ZYdlSY
pl
16
2
3
2.5
0.03
pl000/audio/00000123/KB2GGZwxjgQ
pl
30
3
3
2.33
0.032
pt100/audio/00000309/k1vdvO3bXPI
pt
8.5
1
4
4
0.103
pt100/audio/00000379/_eac57GB38y
pt
48.9
5
4
3.4
0.115
pt100/audio/00000442/RLoz-O5gfeM
pt
10
1
6
6
0.119
pt101/audio/00000064/pcaGmvKyhnA
pt
6.1
1
4
4
0.169
pt101/audio/00000289/nSOESHCxzIg
pt
12.4
2
5
4.5
0.132
pt101/audio/00000477/b1ADa-7GilE
pt
7.6
1
5
5
0.081
pt102/audio/00000137/QoVtf13t2XU
pt
4.3
1
4
4
0.104
pt102/audio/00000139/Qp8VwtxIx9g
pt
4.3
1
4
4
0.151
ru001/audio/00000261/QNdvBi0QOmc
ru
6
1
4
4
0.208
ru100/audio/00000278/BSXzPEc4ybI
ru
16.3
2
5
4.5
0.133
ru101/audio/00000154/v_DaX88ENro
ru
4.8
1
4
4
0.185
ru102/audio/00000169/EloqGXAPGME
ru
8.9
1
4
4
0.197
ru102/audio/00000354/814O7czZYDI
ru
5.1
1
6
6
0.089
ru104/audio/00000368/uN41b_0GvRk
ru
16.3
2
7
7
0.1
ru105/audio/00000160/9lO7SecOELA
ru
7
1
6
6
0.166
ru105/audio/00000237/kffHj6qC3ZA
ru
28.4
3
7
6.33
0.168
ru105/audio/00000402/H1BJX_JuFQI
ru
3.2
1
3
3
0.291
ru106/audio/00000158/kQmt3qpqc7I
ru
19.4
2
6
5
0.206
sv000/audio/00000005/BmckyH5bUcE
sv
30.2
3
4
2.67
0.028
uk000/audio/00000002/4C9m3p_qEXQ
uk
12.3
2
8
5
0.006
uk000/audio/00000054/I-ZArph2cjc
uk
8
1
4
4
0.015
uk100/audio/00000000/6rjEp4E7RkA
uk
3.5
1
3
3
0.019

YODAS2 multi-talker pseudo-labels

4.6k hours of multi-speaker YouTube conversations from YODAS2, cut into chunks of up to 10 minutes and transcribed speaker by speaker with DiCoW v3.3 conditioned on an automatic diarization. This is the pseudo-labeled training stage of Pandemonium, described in section 3.1.1 of the paper.

The audio and the labels are not redistributed. Because of possible licensing issues with the source recordings, this dataset is shared as the scripts that mine it (recipe/), its statistics and a list of 100 sample recordings (sample_pool.tsv).

Statistics

Of the corpus Pandemonium was trained on. dashboard.html has the same numbers as charts (download it and open it in a browser).

audio 4,627.2 h in 32,865 chunks from 11,862 recordings (3,867.5 h of speech)
languages 23, those of Parakeet-v3 except Estonian and Maltese
speakers per chunk 2–8, 3.2 on average; 1.2 % of the audio has a single speaker
overlapped speech 148.2 h; median 1.6 % of a chunk's speech, 8.5 % of chunks have none
words 45.9 M
chunk length up to 10 min; 53 % of chunks are 9–10 min long, the rest are recording tails

How it was built

step recordings chunks hours
YODAS2 recordings with diarization statistics 514,406
selected: random, stratified by language 15,086 5,528.8
chunked and pseudo-labeled (79 languages) 14,682 38,399 5,362.6
in Parakeet-v3's languages: the training set 11,862 32,865 4,627.2

Hours per language

en es ru pt fr de it nl pl uk cs hu
2,114.8 842.7 595.9 337.3 283.6 181.3 171.9 56.2 15.0 10.5 5.0 3.0

Under 3 h each: bg 2.2, fi 2.1, sv 1.9, el 1.6, da 0.5, sk 0.5, sl 0.4, ro 0.3, lv 0.2, lt 0.1, hr 0.1. 589.6 h come from YODAS2's manually captioned splits, 4,037.6 h from the automatic-caption splits (the captions are not used).

Hours by speakers per chunk

1 2 3 4 5 6 7 8
53.6 1,957.2 1,093.6 692.5 379.5 218.9 135.3 96.6

Hours are chunk durations; speech, overlap and speaker counts come from the diarization the labels were conditioned on; the language is the YODAS2 shard's.

The recipe

The recipe uses only public models: Nemotron-3-Diarization for the diarization and DiCoW v3.3 for the transcripts. The paper's corpus was labeled with in-house diarization, so the recipe gives a comparable corpus, not an identical one.

recipe/diarize.py              Nemotron-3-Diarization over YODAS2 shards       -> RTTMs
recipe/analyze_diarization.py  per-recording statistics, selection             -> recording list
recipe/plan_jobs.py            split the selection into GPU jobs
recipe/pseudo_label.py         10-min chunks, DiCoW v3.3 per speaker           -> <chunk>.{wav,rttm,json}
recipe/export_shar.py          Parakeet languages -> Lhotse Shar training set
optional:  filter_hallucinations.py, filter_leakage.py, lid_check.py          (mark only)
           window_30s.py                                                      (30 s windows for DiCoW)

diarize.py runs in a NeMo environment (as for the Pandemonium demo, demo/requirements.txt); the rest in recipe/requirements.txt with DiCoW's inference code (BUTSpeechFIT/DiCoW, tested at 801e213) on PYTHONPATH. Every step skips work already done, so a killed job can simply be restarted. Run the scripts from recipe/.

1. Diarize. One job per GPU; only the shards of Parakeet-v3's languages are needed for a Pandemonium-like set.

python diarize.py --shards en000,en001,... --out rttm/ --part 0 --nparts 8
cat rttm/*/*.rttm > all.rttm

2. Select recordings with at least 2 speakers, 5 s of overlapped speech and 15 s from the quietest speaker, and draw a random 5,500 h from them, preserving each language's share.

python analyze_diarization.py stats --rttm all.rttm --output stats.csv
python analyze_diarization.py select --stats stats.csv \
    --min-speakers 2 --min-overlap-sec 5 --min-speaker-sec 15 --target-hours 5500 --stratify-by lang \
    --output-list pool.lst --rttm all.rttm --output-rttm pool.rttm --summary

Nemotron-3-Diarization finds more brief speakers than the paper's diarization, so the 15 s rule removes more recordings; expect to diarize more audio per selected hour.

3. Pseudo-label. Chunks of at most 10 minutes, cut at silences; chunks with more than 10 speakers are dropped; DiCoW transcribes each speaker in the shard's language.

python plan_jobs.py --list pool.lst --rttm pool.rttm --stats stats.csv --num-jobs 16 --output-dir jobs/
PYTHONPATH=/path/to/DiCoW python pseudo_label.py \
    --recordings-file jobs/job_00.lst --rttm jobs/job_00.rttm --output-dir yodas_pseudo --tmp-dir tmp/job_00

4. Export the chunks in Parakeet-v3's languages, the only filter behind Pandemonium's training set, as a Lhotse Shar directory (16 kHz FLAC, one supervision per DiCoW segment).

python export_shar.py --data-dir yodas_pseudo --output-dir yodas_shar

Optional filters

The paper's DiCoW study (Table 2) compared three label filters; Pandemonium was trained without them. Each script only marks the chunk JSONs, and --drop-flagged of export_shar.py or window_30s.py applies the marks. A chunk with flagged segments is cut down to the windows between them that are at least 2 minutes long.

script marks on the training set
filter_hallucinations.py boilerplate phrases ("thanks for watching", ...), a 2–4-word phrase repeated ≥ 6 times over ≥ 60 % of a segment, ≥ 5 consecutive identical segments 5,625 chunks flagged, 246 unsalvageable
filter_leakage.py two speakers sharing ≥ 6 consecutive words within 1 s: one voice transcribed under two labels 1,885 chunks flagged, 71 unsalvageable
lid_check.py transcripts whose dominant language is not the shard's, judged by an LLM through any OpenAI-compatible API (needs an API)

The paper used Kimi K3 for the language check, within a longer prompt, so lid_check.py will not reproduce its decisions exactly. window_30s.py builds the 30 s windows of the DiCoW study: utterance groups split at 2 s pauses, long groups cut outside overlapped speech, windows with one speaker or with leakage dropped.

Sample recordings

sample_pool.tsv lists 100 recordings of the training set (18.2 h, 13 languages) for trying the recipe end to end and comparing with the statistics above: recordings with 3–8 speakers and more overlap than usual (median 11.7 %), stratified by language, with statistics of the paper's labels (labeled minutes, chunks, maximum and mean speakers per chunk, overlapped share of speech).

python recipe/diarize.py --recordings sample_pool.tsv --out rttm_sample/

The ids are YODAS2 recording ids (<shard>/audio/<tar>/<YouTube video id>); no audio or text is included.

License

This repository (statistics, sample list and scripts) is CC-BY-4.0. YODAS2 is distributed under CC-BY-3.0; the rights in the underlying recordings remain with their owners, which is why no audio or transcripts are shared.

Downloads last month
68