Dataset Viewer
Auto-converted to Parquet Duplicate
emilia_id
stringlengths
22
22
video_id
stringlengths
11
11
batch
stringlengths
10
10
duration_s
float32
1.04
30
logit
float32
-12
12
p_synthetic
float32
0
1
EN_tKvmUvxYZXI_W000013
tKvmUvxYZXI
EN-B000000
16.621
-8.548335
0.000194
EN_tKvmUvxYZXI_W000029
tKvmUvxYZXI
EN-B000000
12.7
-12
0.000006
EN_tKvmUvxYZXI_W000001
tKvmUvxYZXI
EN-B000000
25.884001
-10.862325
0.000019
EN_tKvmUvxYZXI_W000027
tKvmUvxYZXI
EN-B000000
10.868
-10.133783
0.00004
EN_tKETJ-iK8Qo_W000017
tKETJ-iK8Qo
EN-B000000
17.233999
8.269691
0.999744
EN_tKETJ-iK8Qo_W000015
tKETJ-iK8Qo
EN-B000000
7.994
11.708621
0.999992
EN_tKETJ-iK8Qo_W000020
tKETJ-iK8Qo
EN-B000000
11.188
9.562814
0.99993
EN_tKJWGCWi_zA_W000003
tKJWGCWi_zA
EN-B000000
5.399
-12
0.000006
EN_tKJWGCWi_zA_W000001
tKJWGCWi_zA
EN-B000000
14.332
-12
0.000006
EN_tKJWGCWi_zA_W000017
tKJWGCWi_zA
EN-B000000
15.976
-12
0.000006
EN_tL3CpgcWBsE_W000006
tL3CpgcWBsE
EN-B000000
11.772
-12
0.000006
EN_tLjUFf1KSvA_W000007
tLjUFf1KSvA
EN-B000000
3.02
-12
0.000006
EN_tLC0Zhad_Rc_W000009
tLC0Zhad_Rc
EN-B000000
12.547
-0.868386
0.29559
EN_tLC0Zhad_Rc_W000003
tLC0Zhad_Rc
EN-B000000
15.993
4.284678
0.986409
EN_tL3CpgcWBsE_W000020
tL3CpgcWBsE
EN-B000000
15.804
-9.834011
0.000054
EN_tL3CpgcWBsE_W000019
tL3CpgcWBsE
EN-B000000
6.638
-8.466993
0.00021
EN_tLC0Zhad_Rc_W000016
tLC0Zhad_Rc
EN-B000000
8.048
-0.902364
0.288565
EN_tLC0Zhad_Rc_W000034
tLC0Zhad_Rc
EN-B000000
6.757
-0.340243
0.415751
EN_tLC0Zhad_Rc_W000038
tLC0Zhad_Rc
EN-B000000
8.675
-1.292954
0.215353
EN_tLC0Zhad_Rc_W000017
tLC0Zhad_Rc
EN-B000000
10.407
-1.437649
0.19191
EN_tMK_mCzWfSI_W000008
tMK_mCzWfSI
EN-B000000
7.097
-12
0.000006
EN_tMK_mCzWfSI_W000018
tMK_mCzWfSI
EN-B000000
4.363
-10.718533
0.000022
EN_tMK_mCzWfSI_W000004
tMK_mCzWfSI
EN-B000000
8.506
-12
0.000006
EN_tMK_mCzWfSI_W000027
tMK_mCzWfSI
EN-B000000
12.105
-12
0.000006
EN_tLC0Zhad_Rc_W000027
tLC0Zhad_Rc
EN-B000000
8.812
-2.069963
0.112051
EN_tMK_mCzWfSI_W000039
tMK_mCzWfSI
EN-B000000
10.187
-12
0.000006
EN_tMK_mCzWfSI_W000047
tMK_mCzWfSI
EN-B000000
7.284
-10.105968
0.000041
EN_tMK_mCzWfSI_W000038
tMK_mCzWfSI
EN-B000000
17.454
-12
0.000006
EN_tMK_mCzWfSI_W000056
tMK_mCzWfSI
EN-B000000
4.024
-12
0.000006
EN_tMK_mCzWfSI_W000060
tMK_mCzWfSI
EN-B000000
4.19
-11.964956
0.000006
EN_tMK_mCzWfSI_W000066
tMK_mCzWfSI
EN-B000000
10.255
-10.7238
0.000022
EN_tMSiF3yP9-4_W000007
tMSiF3yP9-4
EN-B000000
4.992
-12
0.000006
EN_tMLfVjYxtg4_W000001
tMLfVjYxtg4
EN-B000000
10.186
2.10239
0.891135
EN_tMLfVjYxtg4_W000000
tMLfVjYxtg4
EN-B000000
14.245
4.673234
0.990744
EN_tMSiF3yP9-4_W000009
tMSiF3yP9-4
EN-B000000
16.444
-12
0.000006
EN_tMSiF3yP9-4_W000006
tMSiF3yP9-4
EN-B000000
3.883
-12
0.000006
EN_tM4b_UB4IX4_W000001
tM4b_UB4IX4
EN-B000000
8.846
-12
0.000006
EN_tNSbmdiOfN8_W000010
tNSbmdiOfN8
EN-B000000
3.494
-11.836942
0.000007
EN_tNSbmdiOfN8_W000002
tNSbmdiOfN8
EN-B000000
11.926
-10.848763
0.000019
EN_tNSbmdiOfN8_W000024
tNSbmdiOfN8
EN-B000000
8.812
-9.579478
0.000069
EN_tNSbmdiOfN8_W000003
tNSbmdiOfN8
EN-B000000
14.754
-12
0.000006
EN_tNSbmdiOfN8_W000027
tNSbmdiOfN8
EN-B000000
13.7
-11.047673
0.000016
EN_tNUF7r_0KTU_W000011
tNUF7r_0KTU
EN-B000000
11.33
7.745476
0.999568
EN_tNUF7r_0KTU_W000007
tNUF7r_0KTU
EN-B000000
6.248
6.545622
0.998566
EN_tNSbmdiOfN8_W000040
tNSbmdiOfN8
EN-B000000
11.307
-10.226794
0.000036
EN_tNUF7r_0KTU_W000026
tNUF7r_0KTU
EN-B000000
10.339
7.915174
0.999635
EN_tNSbmdiOfN8_W000026
tNSbmdiOfN8
EN-B000000
7.283
-12
0.000006
EN_tNUF7r_0KTU_W000002
tNUF7r_0KTU
EN-B000000
5.5
6.891573
0.998985
EN_tNUF7r_0KTU_W000029
tNUF7r_0KTU
EN-B000000
5.654
4.949137
0.99296
EN_tNUF7r_0KTU_W000043
tNUF7r_0KTU
EN-B000000
13.43
8.648124
0.999825
EN_tNUF7r_0KTU_W000049
tNUF7r_0KTU
EN-B000000
5.009
8.156931
0.999713
EN_tNUF7r_0KTU_W000040
tNUF7r_0KTU
EN-B000000
5.374
5.265912
0.994862
EN_tp3pBsbThZY_W000001
tp3pBsbThZY
EN-B000000
13.239
-12
0.000006
EN_tNUF7r_0KTU_W000058
tNUF7r_0KTU
EN-B000000
7.538
7.906015
0.999632
EN_tNUF7r_0KTU_W000004
tNUF7r_0KTU
EN-B000000
5.56
6.553924
0.998578
EN_tp3pBsbThZY_W000019
tp3pBsbThZY
EN-B000000
9.168
-12
0.000006
EN_tp3pBsbThZY_W000009
tp3pBsbThZY
EN-B000000
5.602
-12
0.000006
EN_tp3pBsbThZY_W000036
tp3pBsbThZY
EN-B000000
10.323
-12
0.000006
EN_tp3pBsbThZY_W000017
tp3pBsbThZY
EN-B000000
13.334
-11.296232
0.000012
EN_tp3pBsbThZY_W000011
tp3pBsbThZY
EN-B000000
4.711
-12
0.000006
EN_tp3pBsbThZY_W000052
tp3pBsbThZY
EN-B000000
6.16
-7.902936
0.00037
EN_tp3pBsbThZY_W000044
tp3pBsbThZY
EN-B000000
10.85
-12
0.000006
EN_tp3pBsbThZY_W000049
tp3pBsbThZY
EN-B000000
9.983
-8.834023
0.000146
EN_tp3pBsbThZY_W000060
tp3pBsbThZY
EN-B000000
4.844
-9.071377
0.000115
EN_tp3pBsbThZY_W000074
tp3pBsbThZY
EN-B000000
6.57
-12
0.000006
EN_tp3pBsbThZY_W000084
tp3pBsbThZY
EN-B000000
3.532
-11.690001
0.000008
EN_tp3pBsbThZY_W000085
tp3pBsbThZY
EN-B000000
8.794
-12
0.000006
EN_tp3pBsbThZY_W000088
tp3pBsbThZY
EN-B000000
9.866
-12
0.000006
EN_tLC0Zhad_Rc_W000028
tLC0Zhad_Rc
EN-B000000
6.588
0.0089
0.502225
EN_tp3pBsbThZY_W000104
tp3pBsbThZY
EN-B000000
10.832
-11.261031
0.000013
EN_tp3pBsbThZY_W000113
tp3pBsbThZY
EN-B000000
9.888
-12
0.000006
EN_tp3pBsbThZY_W000112
tp3pBsbThZY
EN-B000000
9.16
-12
0.000006
EN_tp3pBsbThZY_W000129
tp3pBsbThZY
EN-B000000
10.68
-8.558441
0.000192
EN_tp3pBsbThZY_W000128
tp3pBsbThZY
EN-B000000
3.293
-12
0.000006
EN_tp3pBsbThZY_W000137
tp3pBsbThZY
EN-B000000
5.45
-12
0.000006
EN_tp3pBsbThZY_W000116
tp3pBsbThZY
EN-B000000
8.762
-12
0.000006
EN_tp3pBsbThZY_W000120
tp3pBsbThZY
EN-B000000
6.211
-11.29828
0.000012
EN_tp3pBsbThZY_W000139
tp3pBsbThZY
EN-B000000
18.615999
-12
0.000006
EN_tp3pBsbThZY_W000145
tp3pBsbThZY
EN-B000000
3.311
-12
0.000006
EN_tp3pBsbThZY_W000162
tp3pBsbThZY
EN-B000000
18.011999
-12
0.000006
EN_tp3pBsbThZY_W000166
tp3pBsbThZY
EN-B000000
8.556
-12
0.000006
EN_tp3pBsbThZY_W000175
tp3pBsbThZY
EN-B000000
4.686
-10.543199
0.000026
EN_tp3pBsbThZY_W000187
tp3pBsbThZY
EN-B000000
10.442
-11.362967
0.000012
EN_tp3pBsbThZY_W000189
tp3pBsbThZY
EN-B000000
18.709999
-12
0.000006
EN_tp3pBsbThZY_W000172
tp3pBsbThZY
EN-B000000
16.011
-12
0.000006
EN_tpu3nakk9JQ_W000006
tpu3nakk9JQ
EN-B000000
4.652
-9.028908
0.00012
EN_tpu3nakk9JQ_W000012
tpu3nakk9JQ
EN-B000000
6.524
-6.064447
0.002319
EN_tpu3nakk9JQ_W000027
tpu3nakk9JQ
EN-B000000
4.737
-8.911864
0.000135
EN_tpu3nakk9JQ_W000014
tpu3nakk9JQ
EN-B000000
8.455
-7.072043
0.000848
EN_tpu3nakk9JQ_W000032
tpu3nakk9JQ
EN-B000000
21.775
-8.039803
0.000322
EN_tpu3nakk9JQ_W000029
tpu3nakk9JQ
EN-B000000
8.727
-9.010598
0.000122
EN_tpu3nakk9JQ_W000021
tpu3nakk9JQ
EN-B000000
14.115
-9.439511
0.00008
EN_tpu3nakk9JQ_W000044
tpu3nakk9JQ
EN-B000000
7.45
-12
0.000006
EN_tpu3nakk9JQ_W000048
tpu3nakk9JQ
EN-B000000
3.565
-10.82921
0.00002
EN_tpu3nakk9JQ_W000037
tpu3nakk9JQ
EN-B000000
4.957
-10.133532
0.00004
EN_tpu3nakk9JQ_W000050
tpu3nakk9JQ
EN-B000000
18.862
-10.714559
0.000022
EN_tpu3nakk9JQ_W000058
tpu3nakk9JQ
EN-B000000
5.162
-6.891226
0.001016
EN_tpu3nakk9JQ_W000056
tpu3nakk9JQ
EN-B000000
19.612
-11.560647
0.00001
EN_tpu3nakk9JQ_W000039
tpu3nakk9JQ
EN-B000000
14.33
-12
0.000006
EN_tpu3nakk9JQ_W000060
tpu3nakk9JQ
EN-B000000
9.779
-6.309494
0.001816
End of preview. Expand in Data Studio

Your TTS is learning from other TTS

About 1 in 5 hours of Emilia-YODAS English is synthetic speech. Of the 21,206 hours we scored (8.1M clips from 448k YouTube videos), 4,251 hours (20.0%) come from videos narrated by a TTS voice. Almost all of it is Wikipedia articles read aloud by TTS; one voice alone narrates more than ten thousand videos.

Emilia-YODAS is one of the most widely used open corpora for training speech generation models. A model trained on it learns, in part, to imitate other TTS systems: their prosody, their pronunciation errors, their flatness. This dataset gives a synthetic-speech score for every clip we scored, and a per-video decision, so you can filter it out.

The detector that produced the scores is released as cloud0day3/tts-contamination-detector (Apache-2.0); use it to score the clips this table does not cover.

The numbers

Videos Clips Hours Share of hours
Scored 448,394 8,091,785 21,205.8 100%
Flagged as TTS (threshold 0.32) 62,538 1,588,547 4,251.2 20.05%
Flagged as TTS (threshold 0.54) 61,219 1,571,555 4,210.8 19.86%

The scorer samples at most 40 clips per video in each shard part, so long videos are represented by a sample. Counting every clip of the scored videos instead (8,899,008 clips, 23,263.1 h), the flagged videos hold 4,725.3 h: 20.31% (20.13% at 0.54). The headline does not depend on the sampling.

The scores are strongly bimodal, so the threshold hardly matters: 76.6% of the scored hours are in videos whose mean probability is below 0.05, and 18.9% in videos at 0.95 or above. Only 1.9% of the hours sit between 0.05 and 0.54.

Checked by ear, blind. A listener heard clips drawn at random, weighted by hours, from flagged and unflagged videos. The clips were shuffled together, played at one loudness, and shown with no scores or text. The listener answered 44 of the 90 clips in the sample (the first 44 in a page order that does not depend on the score):

Heard as TTS Heard as human Can't tell TTS share (Wilson 95%)
Flagged videos 25 0 1 100% (86.7%–100%)
Unflagged videos 0 13 5 0% (0%–22.8%)

The protocol and the reading rule were committed before any clip was heard. The sample is small; the precision of the flags is well supported, and the share of TTS the detector misses is only loosely bounded.

One voice, ten thousand videos. In an early 3,881-hour run of our pipeline, a speaker-embedding join across videos found a single voice holding 161,034 clips from 10,628 videos, 11.9% of the clips that run kept. Its per-video voice embeddings are nearly identical (mean cosine 0.922) and its clips read encyclopedia prose. These are Wikipedia audio-article videos, read by TTS.

A text-based rule agrees. Our corpus build on 2,195 of these units also flags Wikipedia audio-article videos from their transcripts (the narration reads out section markup), and the voices heard in them. About 98% of the clips the detector flags are flagged by that rule too (all but 35,388 of 1,599,348), although the detector never sees the text. The two are not fully independent: the detector's training positives included Wikipedia-reader voices that this rule found in a different YouTube set (YODAS2), never in Emilia.

Use it

Drop every clip of a flagged video. Filter on the video id, not on the scored clip ids: the table holds a sample of each video's clips, and the decision applies to all of them.

import pandas as pd

videos = pd.read_parquet("hf://datasets/cloud0day3/emilia-yodas-tts-contamination/data/videos.parquet")
tts_videos = set(videos.loc[videos.flagged_032, "video_id"])

def video_id(emilia_id: str) -> str:  # "EN_tKvmUvxYZXI_W000013" -> "tKvmUvxYZXI"
    return emilia_id[len("EN_"):].rsplit("_W", 1)[0]

# ...then skip any Emilia-YODAS sample whose id (its file stem / webdataset __key__) gives a video in `tts_videos`.

Decide per video, not per clip: a single clip's score is noisier than a video's mean, and a TTS-narrated video is TTS throughout. Videos with fewer than 3 scored clips (94,683 videos, 360.5 h) are never flagged by the rule; use clips.p_synthetic directly for them if you want to be stricter.

Files

data/clips.parquet: one row per scored clip (at most 40 per video per shard part, chosen by hash).

Column Meaning
emilia_id The Emilia-YODAS sample id (EN_<video>_W<segment>)
video_id The YouTube video id
batch The Emilia-YODAS tar the clip is in (EN-B000000 …)
duration_s Clip duration in seconds
logit The detector's logit (clipped at −12)
p_synthetic The detector's probability that the clip is synthetic speech

data/videos.parquet: one row per video.

Column Meaning
video_id The YouTube video id
clips, hours Scored clips and their hours
p_video Duration-weighted mean of p_synthetic (durations floored at 0.1 s)
flagged_032, flagged_054 At least 3 scored clips and p_video ≥ 0.32 (or 0.54)

Coverage

The table covers 21,206 of Emilia-YODAS English's 92.2k hours (23%): the clips that our data pipeline had processed by 2026-10-11, spread over parts of 1,170 of its tar batches (EN-B000000 to EN-B001361). They were processed in the pipeline's own order, not drawn at random, so treat the 20% as a measurement on this subset. Clips the pipeline dropped before scoring (for example non-English or untranscribed audio) are not in the table.

How the scores were made

  • Detector: frozen HuBERT-Large features (facebook/hubert-large-ll60k), mean- and std-pooled over eight layers, and one logistic layer. Each clip is trimmed of silence and its middle 10 s is scored at 16 kHz.
  • Trained only on permissively licensed audio: renders of eight open TTS systems (Apache-2.0 or MIT) and YouTube Wikipedia-reader voices as positives; LibriTTS-R, VCTK, People's Speech and human YouTube videos as negatives. No commercial TTS audio was used for training. See the model card for its held-out evaluation.
  • Decision: a video is flagged when the duration-weighted mean probability of its clips reaches the threshold (at least 3 clips). 0.32 was chosen to flag at most 1% of each certainly-human evaluation set. A stricter re-evaluation with leak-free folds moves that point to 0.54: at 0.32 it flags 2.0% of LibriVox books, while every YouTube and conversational human set stays at 0. Both thresholds are provided.

Limitations

  • The detector was trained on eight open TTS systems' renders and on the Wikipedia-reader voices. It may miss TTS that sounds unlike any of them, such as newer commercial voices; the listening check bounds the misses only loosely (up to about 23%).
  • A flagged video may contain some human speech (an intro, a guest), and an unflagged video may contain short TTS segments.
  • It is a training-data filter. It is not a deepfake or forensic detector, and a single clip's score is not a verdict on that clip.

Licence and attribution

The scores are released under CC BY 4.0. They describe Emilia-YODAS, which is itself CC BY 4.0 and drawn from YouTube via YODAS; no audio is redistributed here. Please cite Emilia alongside this dataset.

Made by the Spero TTS project, an upcoming Apache-2.0 English text-to-speech model.

@misc{spero2026emiliatts,
  title  = {Emilia-YODAS TTS contamination: per-clip synthetic-speech scores},
  author = {{Spero TTS}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/cloud0day3/emilia-yodas-tts-contamination}
}

@inproceedings{emilialarge,
    author={He, Haorui and Shang, Zengqiang and Wang, Chaoren and Li, Xuyuan and Gu, Yicheng and Hua, Hua and Liu, Liwei and Yang, Chen and Li, Jiaqi and Shi, Peiyang and Wang, Yuancheng and Chen, Kai and Zhang, Pengyuan and Wu, Zhizheng},
    title={Emilia: A Large-Scale, Extensive, Multilingual, and Diverse Dataset for Speech Generation},
    booktitle={arXiv:2501.15907},
    year={2025}
}

@inproceedings{emilia,
    author={He, Haorui and Shang, Zengqiang and Wang, Chaoren and Li, Xuyuan and Gu, Yicheng and Hua, Hua and Liu, Liwei and Yang, Chen and Li, Jiaqi and Shi, Peiyang and Wang, Yuancheng and Chen, Kai and Zhang, Pengyuan and Wu, Zhizheng},
    title={Emilia: An Extensive, Multilingual, and Diverse Speech Dataset for Large-Scale Speech Generation},
    booktitle={Proc.~of SLT},
    year={2024}
}
Downloads last month
36

Paper for cloud0day3/emilia-yodas-tts-contamination