sl-juniper — Kyrgyz-language audio found in a Tajik broadcast archive

1 episode · 5 minutes · 16 kHz mono opus.

What this is

A single Kyrgyz-language item found while assembling a Tajik ASR corpus from Tajik state media: a military honour-guard address during a state visit, delivered in Kyrgyz ("…сиздин келишимизге улуттук гвардиянын… ыстык саламыңызды билдиремиз").

This is five minutes of audio. It is an artefact worth keeping, not a training set. It is published for completeness and provenance, so the material is not simply lost when it is removed from the Tajik corpus.

Why finding it was hard

Kyrgyz, Tajik and Russian all use Cyrillic, so script filters cannot separate them. Two things failed on this episode and are worth recording:

• Whisper's audio language ID called it "Latin" at p=0.51 — it does not reliably identify Kyrgyz.
• A letter-frequency test called it Russian, because ы is common in Kyrgyz and Russian, and the transcript came from a Tajik ASR model whose output carries no Kyrgyz-specific letters (ң ө ү) at all.

What identified it correctly was Kyrgyz function words (жана, менен, үчүн, сиздин, улуттук) in the transcript. If you are separating Cyrillic languages, function words beat both alphabets and off-the-shelf language ID.

Contents

• audio/ — the opus file, 16 kHz mono
• metadata.jsonl — duration, source channel, segment timings

⚠️ About the transcript

Produced by a Tajik ASR model reading Kyrgyz audio. It is readable enough to identify the language and the subject, and it is not a usable label. Re-transcribe with a Kyrgyz model.

Licence and use

Archival state-broadcaster material. Released for research use only. Saidzoda Lab claims no rights over the underlying recording and no licence here extends to it. Access is gated and granted manually. If you hold rights to this material and want it withdrawn, open a discussion.