You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

sl-tundra — Russian-language broadcast audio from Tajik state media

22 episodes · 4.98 hours · 16 kHz mono opus.

What this is

Tajik state broadcasters publish in Russian as well as Tajik. While assembling a Tajik ASR corpus from the same archive, these episodes were identified as Russian and separated out. They are real, clean broadcast speech — simply not Tajik — so they are published here rather than discarded.

Sources: the presidential channel (10 episodes) and Телевизиони Тоҷикистон (12 episodes). Content is diplomatic and state-visit coverage: bilateral meetings, official receptions, press statements.

Why it exists as a separate dataset

Russian and Tajik share the Cyrillic alphabet, so script-based filters cannot tell them apart, and an ASR model will transcribe Russian audio confidently without any sign that something is wrong. Russian audio silently entering a Tajik training set is a real and costly failure mode. Isolating it is worth doing regardless of whether the Russian half is ever used.

Identification used three independent signals — Tajik-only letters (ӯ ҳ қ ҷ ғ ӣ) per episode, Whisper audio language ID, and Russian/Tajik function-word frequency — and an episode is included here only when Russian function words outnumber Tajik ones by at least 3×. Borderline and code-switched episodes were excluded and are not in this dataset.

Contents

  • audio/ — one opus file per episode, 16 kHz mono, named <channel>__<playlist>__<video_id>.opus
  • metadata.jsonl — one row per episode: duration, source channel, and segment-level timings

⚠️ About the transcripts

metadata.jsonl carries segment text, but it was produced by a Tajik ASR model reading Russian audio. It is a rough guide for navigation and filtering, not a label. Anyone training on this should re-transcribe with a Russian model.

Licence and use

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 If you hold rights to this material and want it withdrawn, open a discussion.

Downloads last month
8