Datasets:
Master dataset: YouTube speech, Google ASR labels. Since 2026-10-13 it also holds the new videos of the former 195 h pack (train only, 53,305 clips, 121.7 h); the rest of that pack duplicated this set.
Renamed 2026-10-13 from
az-asr-youtube-136h; the old name still redirects.
The master sets (2026-10-13): az-asr-master-calls-human (human, calls),
az-asr-master-calls-soniox (Soniox pseudo-labels, calls), az-asr-master-read-human
(read speech), az-asr-master-broadcast-human (VOA), az-asr-master-youtube-google
(Google ASR, YouTube). The former az-asr-calls-human-140h (a strict subset of az-asr-master-calls-human) and
az-asr-youtube-195h (merged into az-asr-master-youtube-google) were deleted on 2026-10-13.
Labelling
| field | value |
|---|---|
label_origin |
asr:google |
speech_register |
spontaneous |
channel |
wideband-16k |
provenance |
documented |
Google ASR output over YouTube audio. No human labels anywhere in it.
label_origin distinguishes text that existed before the audio (script,
exact by construction) from text written by a listener (human-transcript, high
but edited) from machine output (asr:<vendor>, bounded by that vendor's error
rate). The older label_type field is retained for compatibility; it recorded
human for both script and human-transcript, which is what made a read
broadcast corpus look like more call supervision than it was.
az-asr-master-youtube-google
257.75 hours of Azerbaijani speech from long-form YouTube interviews and
podcasts, segmented into 116,680 utterances with aligned transcripts
(136.09 h until 2026-10-13, when 53,305 train clips, 121.66 h, from the new videos
of the former 195 h pack were added as train/0100.parquet onward).
| split | utterances | hours |
|---|---|---|
| train | 113,377 | 249.95 |
| dev | 1,561 | 3.35 |
| test | 1,742 | 4.45 |
| total | 116,680 | 257.75 |
Read this before using the transcripts
The transcripts are machine-generated, not human-written. They come from
YouTube's automatic caption track, specifically the az-orig track, which is
YouTube's own ASR output for videos whose original language is Azerbaijani.
This matters twice over:
- A model trained on this corpus learns to agree with YouTube's recogniser, including where that recogniser is wrong. Scores measured against these transcripts are agreement, not accuracy.
- The
az-origdistinction is not cosmetic. A track labelled plainazis frequently YouTube's machine translation of some other language's ASR output, which does not align with the audio at all. On one channel we surveyed, 76% of captioned videos were in that category. Onlyaz-origtracks were used here.
Use it for acoustic and vocabulary coverage. Do not use it as a reference set for reporting accuracy.
Format
One row per utterance, audio inline.
| column | |
|---|---|
id |
utterance id, prefixed by source video |
audio |
16 kHz mono FLAC, inline bytes |
text |
transcript as captioned, punctuation and casing preserved |
text_normalised |
lower-cased, punctuation stripped |
seconds |
duration |
source |
youtube-az |
label_type |
google-asr |
text_group |
shared-transcript group id, for split safety |
speaker_id |
empty for this corpus |
Audio is normalised to 16 kHz mono FLAC. Nothing is codec-degraded; the corpus is wideband, and degradation belongs in augmentation where severity can be varied.
Splits
Splits are assigned over connected components rather than rows: utterances sharing a transcript or a source video land on the same side of the boundary. A row-level split would leak, since a single video contributes hundreds of consecutive utterances.
Azerbaijani text handling
str.lower() is wrong for Azerbaijani in both directions: İ lowercases to
i and I lowercases to ı. Code that normalises this text should fold case
explicitly.
- Downloads last month
- 480