cillegio/az-asr-human-430h

Azerbaijani read speech (LocalDoc + FLEURS)

Azerbaijani read speech assembled from LocalDoc/azerbaijani_asr and LocalDoc/fleurs-azerbaijani-asr. 371,515 clips, 429.85 h, 16 kHz.

Clean, accurately labelled, and the wrong register for telephony. It is literature and schoolbooks read aloud, plus FLEURS sentences. Useful for vocabulary and general acoustics; it will not teach a model what a spontaneous phone conversation sounds like.

A related experiment is worth knowing about before weighting this heavily: adding 147.8 h of read broadcast (VOA) to a call model moved strict WER on human call audio from 44.23% to 56.69%. The labels were accurate and the register was wrong. This corpus is four times that size and the same kind of material.

Labelling

field	value
label_origin	script
speech_register	read
channel	wideband-16k
provenance	inferred

Classified by inference, not documentation. The upstream card states no methodology -- no sources, no annotation process, no quality control. The script classification rests on content signatures: 98.1% of segments end in terminal punctuation, 97.2% start uppercase, 5,600 source recordings average 35.7 sentence-clips each, and the most repeated strings are textbook furniture (Sual və tapşırıqlar. 131 times, numbered exercises, Gəlin yoxlayaq.).

If that reading is right the labels are exact, because the text is the script rather than a transcription of 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).}