Google ASR output over YouTube audio. Largely a fuller repack of az-asr-youtube-136h from the same nine channels, so the two overlap heavily and should not be summed.

_{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.}

Azerbaijani YouTube ASR — 194.55 h

Azerbaijani speech from 9 public YouTube channels, segmented into 87,893 clips.

The labels are machine-generated

Every transcript in this dataset is Google ASR output. There is no human transcription anywhere in it (label_type is google-asr for all 87,893 rows).

This is the single most important fact about the dataset, so it is stated first. A model trained on these labels learns to agree with Google ASR, and evaluating it against these labels measures agreement, not accuracy. We have measured that gap directly on our own audio: a commercial ASR system scored 46.65 % WER against human transcripts of the same recordings, while models trained on its output reported ~19.6 % against that system's own labels and ~49.9 % against the humans. Treat any WER computed against this dataset as a similarity score.

Use it for pre-training, augmentation, or semi-supervised work. Do not use it as an evaluation set, and do not quote a WER against it as a quality figure.


Clips	87,893
Hours	194.55
Videos	263
Channels	9
Clip length	1.03 – 18.43 s (mean 7.97 s)
Audio	FLAC, embedded in the parquet rows
Splits	`train` only

There are no dev or test splits by design. Evaluation for this project runs against separately frozen, human-transcribed splits; carving an eval set out of pseudo-labelled data would only measure agreement with Google ASR.

Per channel

source	clips
youtube-podkastkimiadam	24,766
youtube-parvizi	16,719
youtube-aynur	12,162
youtube-huseyn	9,802
youtube-mailyaqub	7,343
youtube-astar	5,692
youtube-gozgoze	4,026
youtube-2x	3,907
youtube-elihekim	3,476

Fields

field	type	notes
`id`	string	`<video_id>_<nnnnn>`
`audio`	struct	`bytes` (FLAC) + `path`
`text`	string	Google ASR output, as returned
`text_normalised`	string	lowercased, punctuation stripped
`seconds`	double	clip duration
`source`	string	channel tag
`label_type`	string	always `google-asr`
`text_group`	string	empty in this dataset
`speaker_id`	string	empty in this dataset

Splitting without leakage

text_group and speaker_id are present for schema compatibility with the other datasets in this collection but are empty here. Clips from one video share a speaker and overlapping content, so a random split leaks. Group on the video id — the id prefix before the final underscore — when splitting:

video_id = row["id"].rsplit("_", 1)[0]

Normalisation

text_normalised is lowercased and stripped of punctuation. Azerbaijani casefolding is not str.lower(): İ→i and I→ı. Applying Python's default lowercasing to Azerbaijani text corrupts both letters.

Digits are left as digits. Grapheme-based tokenisers generally want them spelled out; subword tokenisers trained with digits present do not.

Provenance and licence

The audio is derived from publicly available YouTube videos and the transcripts are third-party ASR output. Neither is ours to relicense, which is why this repository is private and marked license: other. It is an internal research asset. Do not redistribute the audio, and check the rights on any individual channel before using clips outside research.

Downloads last month: 9

Total file size:

24.4 GB

You need to agree to share your contact information to access this dataset

Labelling