Dataset Viewer
Auto-converted to Parquet Duplicate
video_id
string
language
string
metric
string
system
string
reference_units
int64
hypothesis_units
int64
edits
int64
error_rate
float64
translated_caption
int64
caption_coverage_fail
int64
--DKkzWVh-E
en
WER
whisper-large-v3_two-pass
2,307
2,300
30
0.013
0
0
--DKkzWVh-E
en
WER
whisper-large-v3_two-pass_batched
2,307
2,315
42
0.0182
0
0
--DKkzWVh-E
en
WER
whisper-large-v3_version-c
2,307
2,315
42
0.0182
0
0
--DKkzWVh-E
en
WER
commercial-api
2,307
2,327
50
0.0217
0
0
--esqOD-uds
en
WER
whisper-large-v3_two-pass
310
395
160
0.5161
0
0
--esqOD-uds
en
WER
whisper-large-v3_two-pass_batched
310
401
159
0.5129
0
0
--esqOD-uds
en
WER
whisper-large-v3_version-c
310
401
159
0.5129
0
0
--esqOD-uds
en
WER
commercial-api
310
480
222
0.7161
0
0
--sr64z-RcM
en
WER
whisper-large-v3_two-pass
883
861
76
0.0861
0
0
--sr64z-RcM
en
WER
whisper-large-v3_two-pass_batched
883
870
79
0.0895
0
0
--sr64z-RcM
en
WER
whisper-large-v3_version-c
883
870
79
0.0895
0
0
--sr64z-RcM
en
WER
commercial-api
883
938
116
0.1314
0
0
-0Ca8M_xttQ
en
WER
whisper-large-v3_two-pass
1,017
1,023
105
0.1032
0
0
-0Ca8M_xttQ
en
WER
whisper-large-v3_two-pass_batched
1,017
1,022
90
0.0885
0
0
-0Ca8M_xttQ
en
WER
whisper-large-v3_version-c
1,017
1,022
90
0.0885
0
0
-0Ca8M_xttQ
en
WER
commercial-api
1,017
1,096
145
0.1426
0
0
-0h5Ge_azGY
en
WER
whisper-large-v3_two-pass
2,428
2,395
116
0.0478
0
0
-0h5Ge_azGY
en
WER
whisper-large-v3_two-pass_batched
2,428
2,334
152
0.0626
0
0
-0h5Ge_azGY
en
WER
whisper-large-v3_version-c
2,428
2,334
152
0.0626
0
0
-0h5Ge_azGY
en
WER
commercial-api
2,428
2,476
146
0.0601
0
0
-0nWOVH4h0U
en
WER
whisper-large-v3_two-pass
258
519
271
1.0504
0
1
-0nWOVH4h0U
en
WER
whisper-large-v3_two-pass_batched
258
523
271
1.0504
0
1
-0nWOVH4h0U
en
WER
whisper-large-v3_version-c
258
523
271
1.0504
0
1
-0nWOVH4h0U
en
WER
commercial-api
258
536
301
1.1667
0
1
-1DjTR1sXc0
en
WER
whisper-large-v3_two-pass
2,151
2,126
54
0.0251
0
0
-1DjTR1sXc0
en
WER
whisper-large-v3_two-pass_batched
2,151
2,116
61
0.0284
0
0
-1DjTR1sXc0
en
WER
whisper-large-v3_version-c
2,151
2,116
61
0.0284
0
0
-1DjTR1sXc0
en
WER
commercial-api
2,151
2,165
75
0.0349
0
0
-1cHf3qHIrg
en
WER
whisper-large-v3_two-pass
2,217
2,216
40
0.018
0
0
-1cHf3qHIrg
en
WER
whisper-large-v3_two-pass_batched
2,217
2,206
58
0.0262
0
0
-1cHf3qHIrg
en
WER
whisper-large-v3_version-c
2,217
2,206
58
0.0262
0
0
-1cHf3qHIrg
en
WER
commercial-api
2,217
2,211
24
0.0108
0
0
-1fsX_dxqic
en
WER
whisper-large-v3_two-pass
1,279
1,266
155
0.1212
0
0
-1fsX_dxqic
en
WER
whisper-large-v3_two-pass_batched
1,279
1,253
141
0.1102
0
0
-1fsX_dxqic
en
WER
whisper-large-v3_version-c
1,279
1,253
141
0.1102
0
0
-1fsX_dxqic
en
WER
commercial-api
1,279
1,348
209
0.1634
0
0
-2_InyFhiSg
en
WER
whisper-large-v3_two-pass
2,634
2,600
162
0.0615
0
0
-2_InyFhiSg
en
WER
whisper-large-v3_two-pass_batched
2,634
2,537
222
0.0843
0
0
-2_InyFhiSg
en
WER
whisper-large-v3_version-c
2,634
2,537
222
0.0843
0
0
-2_InyFhiSg
en
WER
commercial-api
2,634
2,786
284
0.1078
0
0
-2crDXSOuXg
en
WER
whisper-large-v3_two-pass
615
342
370
0.6016
0
0
-2crDXSOuXg
en
WER
whisper-large-v3_two-pass_batched
615
526
180
0.2927
0
0
-2crDXSOuXg
en
WER
whisper-large-v3_version-c
615
526
180
0.2927
0
0
-2crDXSOuXg
en
WER
commercial-api
615
637
91
0.148
0
0
-33asB9Qau4
en
WER
whisper-large-v3_two-pass
1,753
1,744
39
0.0222
0
0
-33asB9Qau4
en
WER
whisper-large-v3_two-pass_batched
1,753
1,756
26
0.0148
0
0
-33asB9Qau4
en
WER
whisper-large-v3_version-c
1,753
1,756
26
0.0148
0
0
-33asB9Qau4
en
WER
commercial-api
1,753
1,741
69
0.0394
0
0
-3PiTEVRkig
en
WER
whisper-large-v3_two-pass
887
841
156
0.1759
0
0
-3PiTEVRkig
en
WER
whisper-large-v3_two-pass_batched
887
784
217
0.2446
0
0
-3PiTEVRkig
en
WER
whisper-large-v3_version-c
887
784
217
0.2446
0
0
-3PiTEVRkig
en
WER
commercial-api
887
923
153
0.1725
0
0
-3R2n3aW7ME
en
WER
whisper-large-v3_two-pass
652
693
250
0.3834
0
0
-3R2n3aW7ME
en
WER
whisper-large-v3_two-pass_batched
652
519
292
0.4479
0
0
-3R2n3aW7ME
en
WER
whisper-large-v3_version-c
652
519
292
0.4479
0
0
-3R2n3aW7ME
en
WER
commercial-api
652
746
272
0.4172
0
0
-3cX5Z7AeHE
en
WER
whisper-large-v3_two-pass
914
919
14
0.0153
0
0
-3cX5Z7AeHE
en
WER
whisper-large-v3_two-pass_batched
914
914
19
0.0208
0
0
-3cX5Z7AeHE
en
WER
whisper-large-v3_version-c
914
914
19
0.0208
0
0
-3cX5Z7AeHE
en
WER
commercial-api
914
907
27
0.0295
0
0
-44WxUPwKxc
en
WER
whisper-large-v3_two-pass
385
405
70
0.1818
0
0
-44WxUPwKxc
en
WER
whisper-large-v3_two-pass_batched
385
404
67
0.174
0
0
-44WxUPwKxc
en
WER
whisper-large-v3_version-c
385
404
67
0.174
0
0
-44WxUPwKxc
en
WER
commercial-api
385
445
101
0.2623
0
0
-4FFYD8pufc
en
WER
whisper-large-v3_two-pass
1,566
1,563
54
0.0345
0
0
-4FFYD8pufc
en
WER
whisper-large-v3_two-pass_batched
1,566
1,564
61
0.039
0
0
-4FFYD8pufc
en
WER
whisper-large-v3_version-c
1,566
1,564
61
0.039
0
0
-4FFYD8pufc
en
WER
commercial-api
1,566
1,563
70
0.0447
0
0
-55YUGdah8g
en
WER
whisper-large-v3_two-pass
768
311
463
0.6029
0
1
-55YUGdah8g
en
WER
whisper-large-v3_two-pass_batched
768
311
463
0.6029
0
1
-55YUGdah8g
en
WER
whisper-large-v3_version-c
768
311
463
0.6029
0
1
-55YUGdah8g
en
WER
commercial-api
768
327
469
0.6107
0
1
-568yRON6F8
en
WER
whisper-large-v3_two-pass
300
333
103
0.3433
0
0
-568yRON6F8
en
WER
whisper-large-v3_two-pass_batched
300
293
99
0.33
0
0
-568yRON6F8
en
WER
whisper-large-v3_version-c
300
293
99
0.33
0
0
-568yRON6F8
en
WER
commercial-api
300
360
141
0.47
0
0
-5mQ2XsOX_U
en
WER
whisper-large-v3_two-pass
711
730
59
0.083
0
0
-5mQ2XsOX_U
en
WER
whisper-large-v3_two-pass_batched
711
717
65
0.0914
0
0
-5mQ2XsOX_U
en
WER
whisper-large-v3_version-c
711
717
65
0.0914
0
0
-5mQ2XsOX_U
en
WER
commercial-api
711
759
81
0.1139
0
0
-5uBrOdKXgE
en
WER
whisper-large-v3_two-pass
2,375
2,128
313
0.1318
0
0
-5uBrOdKXgE
en
WER
whisper-large-v3_two-pass_batched
2,375
2,080
342
0.144
0
0
-5uBrOdKXgE
en
WER
whisper-large-v3_version-c
2,375
2,080
342
0.144
0
0
-5uBrOdKXgE
en
WER
commercial-api
2,375
2,530
229
0.0964
0
0
-6OlWqCqxLM
en
WER
whisper-large-v3_two-pass
3,072
3,049
316
0.1029
0
0
-6OlWqCqxLM
en
WER
whisper-large-v3_two-pass_batched
3,072
3,013
364
0.1185
0
0
-6OlWqCqxLM
en
WER
whisper-large-v3_version-c
3,072
3,013
364
0.1185
0
0
-6OlWqCqxLM
en
WER
commercial-api
3,072
3,462
569
0.1852
0
0
-6jpvJo-eK4
en
WER
whisper-large-v3_two-pass
685
683
20
0.0292
0
0
-6jpvJo-eK4
en
WER
whisper-large-v3_two-pass_batched
685
683
16
0.0234
0
0
-6jpvJo-eK4
en
WER
whisper-large-v3_version-c
685
683
16
0.0234
0
0
-6jpvJo-eK4
en
WER
commercial-api
685
689
24
0.035
0
0
-6o8GtZXeWI
en
WER
whisper-large-v3_two-pass
1,687
1,638
93
0.0551
0
0
-6o8GtZXeWI
en
WER
whisper-large-v3_two-pass_batched
1,687
1,624
97
0.0575
0
0
-6o8GtZXeWI
en
WER
whisper-large-v3_version-c
1,687
1,624
97
0.0575
0
0
-6o8GtZXeWI
en
WER
commercial-api
1,687
1,793
236
0.1399
0
0
-6qiO49Q180
en
WER
whisper-large-v3_two-pass
1,516
1,470
267
0.1761
0
0
-6qiO49Q180
en
WER
whisper-large-v3_two-pass_batched
1,516
1,453
242
0.1596
0
0
-6qiO49Q180
en
WER
whisper-large-v3_version-c
1,516
1,453
242
0.1596
0
0
-6qiO49Q180
en
WER
commercial-api
1,516
1,420
288
0.19
0
0
End of preview. Expand in Data Studio

Whisper large-v3 on YouTube audio, scored against uploaded captions

Error rates of open speech-to-text models on 234 public YouTube videos in 9 languages. The reference for each video is the caption track the channel uploaded in the video's own language, not an auto-generated track and not any system's output, so no system under test is graded against itself.

This dataset holds scores only: per-video error rates keyed by YouTube video ID, per-language summaries, the scoring script and the compute-cost method. It contains no audio, no caption text and no transcripts.

Systems

system what it is
whisper-large-v3_two-pass faster-whisper 1.1.1, large-v3, beam 5, Silero VAD; a second pass without VAD when the first returns under 30 characters per minute
whisper-large-v3_two-pass_batched the same, first pass through faster-whisper's batched pipeline (batch 16)
whisper-large-v3_version-c batched, language detected at three points in the audio, low-threshold VAD on the second pass
kotoba-whisper-v2.0_two-pass Japanese only, first 180 videos only
reazonspeech-nemo-v2 Japanese only, first 180 videos only
commercial-api a commercial speech-to-text API, not named

All open models ran on an NVIDIA T4 (Modal).

Selection and filters

  • Videos: public, 3 to 15 minutes, with an uploaded caption track in the language under test; 20 per language picked first, then 6 more per language that required the video's declared audio language to match.
  • Filter 1, translated captions: a video is dropped when the speech is not in the caption's language (language as settled by Whisper large-v3 two-pass). Many non-English uploaded captions are translations.
  • Filter 2, caption coverage: a video is dropped when its caption is under half the shortest system output or over twice the longest, i.e. every system disagrees with the caption in the same direction (partial or unrelated caption). One system failing alone never removes a video.
  • scores_per_video.csv keeps every scored video, with translated_caption and caption_coverage_fail flags; summary_per_language.csv uses only videos that pass both filters.

Metric

CER for ja, zh, th, ko; WER for the others. NFKC, lower case, punctuation and symbols removed, whitespace collapsed, Chinese folded to Simplified on both sides. score.py reproduces the per-video rates given reference and hypothesis text files. Summary: median per-video rate, and corpus rate (total edits / total reference units) with a 95% bootstrap interval over videos.

Files

  • scores_per_video.csv: video_id, language, metric, system, reference_units, hypothesis_units, edits, error_rate, translated_caption, caption_coverage_fail
  • summary_per_language.csv: language, metric, system, videos, median_error_rate, corpus_error_rate, ci95_low, ci95_high
  • score.py: the scorer (MIT)
  • COMPUTE_COST.md: how the GPU cost per audio hour was measured

Limits

  • 13 to 26 videos per language after filtering; medians are indicative.
  • Uploaded captions are not always verbatim (edited, abridged), which raises every system's error alike.
  • Videos can be removed or made private after publication; the scores stay but cannot be re-derived.

License and rights

Scores and documentation: CC BY 4.0. score.py: MIT. Rights holder: Glasp. The underlying videos and captions belong to their uploaders and are not included.

To have a video removed from this dataset, open a discussion in the Community tab with the video ID.

Downloads last month
123