Labelling

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

Editorial broadcast text aligned to VOA audio. Terminal punctuation at 77.4% (against 98.1% for LocalDoc) is consistent with segments cut from continuous broadcast rather than at sentence ends.

Adding this to a call model made it worse. Chinar-F8 v4 mixed in 147.8 h of it and strict WER on human-transcribed calls moved 44.23% -> 56.69%. Register, not label quality: the text is accurate and the speaking style is wrong. See runs/f8v4-voa/RESULT.md in az-asr-training.

_{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-voa-305h

305.43 h of human-transcribed Azerbaijani broadcast speech — news and interviews from Voice of America — repackaged into our corpus schema at 16 kHz mono FLAC.

This is a derivative of disco-eth/WorldSpeech (arXiv:2605.09167), config az_az. We did not collect or transcribe this material; we converted the audio to 16 kHz FLAC and kept the columns our pipeline uses.

Licence, carefully

Two different things are in play and conflating them would be a mistake.

The compilation is CC-BY-NC-4.0. That is WorldSpeech's licence and it governs this repackaging. Non-commercial use only, attribution required. This copy exists for research and evaluation.

The underlying material is public domain. WorldSpeech's own per-source table records az_az as "Public Domain (17 USC §105) — US federal agency content (VOA)". Works of the US federal government are not subject to copyright, and Voice of America is a federal agency.

So the audio and VOA's transcripts are not themselves restricted; the NC clause attaches to the curation. For commercial use, re-derive rather than reuse this repo: every row carries source_url, source_start_s and source_end_s, which is enough to re-fetch the same segments from the original VOA articles independently of the NC-licensed packaging. That path was the reason those columns were preserved. Take legal advice before relying on it.

Contents

split	clips	hours
train	99,446	290.23
test	5,211	15.19
total	104,657	305.43

Audio: 16 kHz mono FLAC (source was 48 kHz opus). Mean clip 10.5 s.

Columns

column	meaning
id	WorldSpeech segment_id
audio	16 kHz mono FLAC
text	the human transcript
asr_transcript	what a generic multilingual recogniser produced — a free baseline
cer	disagreement between the two. High usually means a misaligned segment, not hard audio
snr, dnsmos_ovr	per-row quality, for filtering
source	voa_az_gold or voa_az_silver
source_url, source_start_s, source_end_s	provenance, and the re-derivation path

On the gold/silver labels

They do not indicate transcript quality. Measured across the whole subset, silver is marginally better on both metrics:

source	rows	hours	median CER	median SNR
voa_az_gold	40,236	68.57	0.074	7.9 dB
voa_az_silver	64,421	236.86	0.070	8.6 dB

The real difference is segment length — gold averages 6.1 s, silver 13.2 s. The labels are kept as given, but do not filter on them expecting a quality gain.

Measured properties

• Zero empty transcripts, zero rows without audio, zero audio that failed to convert
• Median CER between human and ASR transcript 0.071, p90 0.209
• Median SNR 8.4 dB, median DNSMOS-P.835 overall 2.83 — studio broadcast speech
• Correct Azerbaijani orthography: 1.24 M ə, 464 k ı, 216 k q, 90 k x
• 65 Cyrillic characters in 305 hours
• 59,886 digit characters — numbers are written as digits here, unlike our call corpora where annotators wrote them as words. Canonicalise before scoring or training against either.

What it is good for, and not

Broadcast and read speech at 48 kHz source quality. It roughly doubles the human-labelled wideband material available to us, and is the natural training data for the 16 kHz models.

It is not telephone speech. Our weakest condition is spontaneous 8 kHz telephony, and this corpus is about as far from that as Azerbaijani audio gets. Do not expect it to move call-domain error rates on its own.