CoolFace
Datasetpublic

bofenghuang/stt-pseudo-labeled-whisper-large-v3-multilingual

This collection includes over 189,000 hours of speech-to-text data in seven languages: English, French, Spanish, Portuguese, Italian, German, and Dutch All segments were initially sorted by their IDs (timestamps). Adjacent segments from the same source were concatenated into 30-second chunks before being decoded using Whisper-Large-V3. The only exception was Common Voice, where segments were decoded individually before concatenation. In total, over 288,000 hours of audio data were collected… See the full description on the dataset page: https://huggingface.co/datasets/bofenghuang/stt-pseudo-labeled-whisper-large-v3-multilingual.

sourceHugging Facecc-by-3.0updated 2y agoView on Hugging Face
4likes1.3kdownloads
Dataset Card

This collection includes over 189,000 hours of speech-to-text data in seven languages: English, French, Spanish, Portuguese, Italian, German, and Dutch

All segments were initially sorted by their IDs (timestamps). Adjacent segments from the same source were concatenated into 30-second chunks before being decoded using Whisper-Large-V3. The only exception was Common Voice, where segments were decoded individually before concatenation.

In total, over 288,000 hours of audio data were collected and processed. This dataset retains only segments with a word error rate (WER) below 20% after normalization. Users can apply stricter filters if needed for their specific use cases.

Usage

python
from datasets import load_dataset
from pprint import pprint

dataset = load_dataset(
    "bofenghuang/stt-pseudo-labeled-whisper-large-v3-multilingual",
    "en-ami-ihm",  # can also select other subsets
    split="train",
    trust_remote_code=True,
)

pprint(dataset[0])
# {'audio': {'array': array([-6.10351562e-05, -1.22070312e-04, -1.83105469e-04, ...,
#         6.40869141e-04,  6.40869141e-04,  6.40869141e-04]),
#            'path': None,
#            'sampling_rate': 16000},
#  'audio_filepath': '/home/bhuang/.cache/huggingface/hub/datasets--bofenghuang--stt-pseudo-labeled-whisper-large-v3-multilingual/snapshots/8f1b7bbae8f1b657d1a95b66dba9d2e7b3b86665//distil-whisper/ami-ihm/ihm/train_concatenated/EN2001a.zip:62624564:958444',
#  'duration': 29.950000762939453,
#  'prev_text': "OKAY DOES ANYONE WANT TO SEE UH STEVE'S FEEDBACK FROM THE "
#               "SPECIFICATION RIGHT NOT REALLY UM JUST WHAT HE'S TALKING ABOUT "
#               'LIKE DUPLICATION OF EFFORT AND LIKE DUPLICATION OF EFFORT AND '
#               'STUFF AND UM YEAH HE WAS SAYING THAT WE SHOULD MAYBE UH THINK '
#               'ABOUT HAVING A PROTOTYPE FOR WEEK SIX WHICH IS NEXT WEEK YEAH '
#               'SO WE SHOULD PROBABLY PRIORITIZE OUR PACKAGES MM YEAH YEAH HMM',
#  'prev_whisper_transcript': "<|0.00|> Does anyone want to see Steve's feedback "
#                             'from the specification?<|4.80|><|4.80|> Not '
#                             'really, just what he was talking about, like '
#                             'duplication of effort and '
#                             'stuff.<|11.20|><|11.20|> And saying that we '
#                             'should maybe think about having a prototype for '
#                             'week six, which is next week.<|21.00|><|21.00|> '
#                             'So we should probably prioritise our '
#                             'packages.<|28.34|>',
#  'text': 'HAS HAS ANYONE ACTUALLY LOOKED AT THE JAVA CODE FOR THE HUH HMM YEAH '
#          "I THINK SO YEAH I I DON'T KNOW ABOUT THE SEARCH FUNCTIONALITY THAT "
#          "MIGHT BE ONLINE DEPENDS HOW IT'S GONNA WORK YEAH MM-HMM YEAH THAT "
#          'MAKES SENSE HMM HMM YEAH YOU JUST CONCATENATE THEM TOGETHER HMM YEAH '
#          'IT JUST MEANS IT LOADS ON DEMAND IT ONLY LOADS WHEN IT NEEDS A '
#          "PARTICULAR TYPE OF FILE LIKE WHEN IT'S BEING ACCESSED YEAH I THINK "
#          "THAT'S THE IDEA IT JUST LOADS THE PARTICULAR ONES IT NEEDS BUT IF "
#          "YOU WERE DOING A SEARCH OVER THE WHOLE CORPUS YOU'D HAVE TO LOAD "
#          'THEM ALL HMM',
#  'text_norm': 'has has anyone actually looked at the java code for the huh '
#               'yeah i think so yeah i i do not know about the search '
#               'functionality that might be online depends how it is going to '
#               'work yeah yeah that makes sense yeah you just concatenate them '
#               'together yeah it just means it loads on demand it only loads '
#               'when it needs a particular type of file like when it is being '
#               'accessed yeah i think that is the idea it just loads the '
#               'particular ones it needs but if you were doing a search over '
#               'the whole corpus you would have to load them all',
#  'wer': 4.716980934143066,
#  'whisper_transcript': '<|0.00|> Has anyone actually looked at the Java code '
#                        'for the AMX?<|5.00|><|5.38|> Yeah, I think '
#                        "so.<|6.22|><|6.22|> Yeah, I don't know about the "
#                        'search functionality.<|8.28|><|8.28|> That might be '
#                        "online.<|10.20|><|10.20|> Depends how it's gonna "
#                        'work.<|11.92|><|11.92|> Yeah, that makes '
#                        'sense.<|13.22|><|13.22|> Yeah, you just concatenate '
#                        'them together.<|15.60|><|15.60|> It just means it '
#                        'loads on demand.<|17.42|><|17.42|> It only loads when '
#                        'it needs a particular type of file,<|22.24|><|22.24|> '
#                        "like when it's being accessed.<|23.40|><|23.40|> Yeah, "
#                        "I think that's the idea.<|24.40|><|24.40|> It just "
#                        'loads the particular ones it needs.<|26.96|><|26.96|> '
#                        'But if you were doing a search over the whole '
#                        "corpus,<|28.66|><|28.66|> you'd have to load them "
#                        'all.<|29.96|>',
#  'whisper_transcript_norm': 'has anyone actually looked at the java code for '
#                             'the amx yeah i think so yeah i do not know about '
#                             'the search functionality that might be online '
#                             'depends how it is going to work yeah that makes '
#                             'sense yeah you just concatenate them together it '
#                             'just means it loads on demand it only loads when '
#                             'it needs a particular type of file like when it '
#                             'is being accessed yeah i think that is the idea '
#                             'it just loads the particular ones it needs but if '
#                             'you were doing a search over the whole corpus you '
#                             'would have to load them all'}

Statistics

See below for the durations (in hours) after applying different WER filters to each subset.

English

Split20%10%5%0%
en-mcv1,571.181,527.701,181.29428.52
en-ls951.31932.31852.46450.89
en-voxpopuli494.10413.92260.3274.07
en-tedlium448.05416.95312.3678.16
en-peoples_speech-clean5,652.323,474.441,160.4673.02
en-peoplesspeech-cleansa955.14643.65260.2624.91
en-peoples_speech-dirty8,664.091,414.41181.965.35
en-peoplesspeech-dirtysa972.02206.2935.341.65
en-gigaspeech-l2,464.072,384.802,099.40901.77
en-ami-ihm50.8419.545.440.46
en-ami-sdm23.176.811.840.18
en-yodas-0003,699.622,902.961,891.37487.83
en-yodas-0013,693.852,896.021,887.55484.64
en-yodas-0023,687.302,890.381,895.27487.38
en-yodas-0033,650.572,843.521,841.51464.97
en-yodas-0043,710.202,907.121,890.28477.85
en-yodas-0052,936.862,302.641,496.23382.48
en-yodas-1003,831.712,692.861,496.27286.69
en-yodas-1013,816.332,689.931,497.48292.58
en-yodas-1023,826.862,701.171,501.30286.49
en-yodas-1033,825.102,698.541,498.18294.47
en-yodas-1042,449.361,717.38948.36184.66
en-yodas-1053,790.392,664.181,476.47285.18
en-yodas-1063,800.002,678.341,487.09287.32
en-yodas-1073,809.052,679.411,488.62289.25
en-yodas-1093,791.462,677.901,492.26290.38
en-yodas-1103,767.502,638.291,456.01281.24
en-yodas-1113,801.112,671.891,486.27287.74
en-yodas-1123,827.942,696.101,494.42285.60
en-yodas-1133,817.432,681.091,489.73289.74
en-yodas-1143,798.912,682.031,500.48296.98
en-yodas-1153,811.492,682.461,487.23288.86
en-yodas-1163,826.622,706.081,509.38292.36
en-yodas-1173,808.302,684.161,497.32293.84
en-yodas-1183,804.022,687.241,499.53292.64
en-yodas-1193,809.402,697.341,508.54296.01
en-yodas-1203,827.332,701.141,502.67287.27
en-yodas-1213,800.262,677.951,488.84290.16
en-yodas-1223,790.632,660.881,472.27285.75
en-yodas-1233,785.272,677.231,494.34289.47
en-yodas-1243,809.972,685.331,501.46293.05
en-yodas-1253,783.512,659.931,475.39288.66
en-yodas-1263,797.072,668.461,487.35289.60
en-yodas-1271,769.641,247.31699.89137.38
total143,001.3698,188.1156,190.4912,387.48

French

Split20%10%5%0%
fr-mcv689.80663.32439.6193.34
fr-mls1,042.59936.38703.29260.22
fr-voxpopuli191.70146.5184.1521.91
fr-mtedx146.09100.6757.2212.98
fr-yodas-0001,497.83912.55445.3271.56
fr-yodas-1001,860.75606.79149.2313.01
fr-yodas-1011,857.40612.54151.9614.09
fr-yodas-1021,850.93610.35152.3413.33
fr-yodas-1031,172.29390.8998.229.21
total10,309.394,979.992,281.33509.65

Spanish

Split20%10%5%0%
es-mcv446.01435.19350.02145.81
es-mls844.79722.04535.54210.70
es-voxpopuli139.52112.1870.9920.37
es-mtedx150.89114.3968.6016.38
es-yodas-0002,408.251,592.19851.84180.98
es-yodas-1002,982.871,610.76667.60104.08
es-yodas-1012,987.421,584.88647.20100.99
total9,959.766,171.633,191.77779.30

Portuguese

Split20%10%5%0%
pt-mcv21.7521.4219.1310.46
pt-mls147.01113.4169.6620.76
pt-mtedx131.7194.5149.579.32
pt-yodas-000859.83453.14211.4242.00
pt-yodas-1001,853.79549.13140.0322.36
pt-yodas-1011,849.31552.92141.3421.55
pt-yodas-1021,871.89560.40143.6322.47
pt-yodas-1031,288.90383.8698.7115.83
total8,024.192,728.80873.48164.75

Italian

Split20%10%5%0%
it-mcv232.83229.54187.9670.52
it-mls232.95185.26113.8735.30
it-voxpopuli58.2341.6423.066.31
it-mtedx88.7273.8449.4713.02
it-yodas-000952.76600.62317.5885.28
it-yodas-1002,664.501,242.66453.3170.46
it-yodas-1012,277.281,062.83387.9960.38
total6,507.273,436.381,533.24341.27

German

Split20%10%5%0%
de-mcv875.27862.86720.29324.03
de-mls1,919.131,736.711,315.35661.96
de-voxpopuli232.27146.6070.0117.32
de-mtedx8.395.703.280.81
de-yodas-0001,607.82925.94476.64128.54
de-yodas-1002,304.63856.85260.5641.08
de-yodas-1012,343.04875.50265.7440.33
de-yodas-102426.58156.5147.437.58
total9,717.135,566.673,159.301,221.65

Dutch

Split20%10%5%0%
nl-mcv41.0240.8135.8112.01
nl-mls1,455.241,133.02692.95253.02
nl-voxpopuli38.8623.209.501.79
nl-yodas-000215.26108.5944.076.59
nl-yodas-100512.37127.1631.755.48
total2,262.751,432.78814.08278.89

Code-switching

Split20%10%5%0%
cs-mcv3,346.803,333.613,192.751,428.04
cs-yodas6,465.596,434.344,967.96572.10
total9,812.399,767.948,160.722,000.14
bofenghuang/stt-pseudo-labeled-whisper-large-v3-multilingual · CoolFace