Russian and Central Asian Speech Data: One Language, Two Scripts, and a Neighbour

A Central Asian speech order usually names three languages: Uzbek, Kyrgyz and Russian. The three are not independent choices. Russian is a working language across the region, Uzbek and Kyrgyz are the state languages of their countries, and everyday speech mixes them. A dataset designed as three separate monolingual corpora will not match how people actually talk.

Writing systems add a second layer. Uzbek is officially written in Latin but a large share of existing text is in Cyrillic. Kyrgyz uses Cyrillic. Russian uses Cyrillic. The script has to be decided per language before recruitment.

Uzbekistan: two scripts in circulation at once

Uzbek switched to a Latin alphabet officially, but Cyrillic text did not disappear. Both are read and written, and conversion between them is not a clean one-to-one mapping for every letter. Converting existing Cyrillic text into Latin introduces errors precisely at the letters without a clean equivalent.

For a speech dataset, the practical consequence is that the text layer has to be prepared in the target script from the start, not converted afterwards. If the buyer needs both, two parallel text layers from the same audio are more useful than one converted list.

Kyrgyzstan: Cyrillic, with Russian alongside

Kyrgyz is written in Cyrillic and Russian is widely used in professional and urban settings. Code-switching between Kyrgyz and Russian is ordinary, particularly in the capital. A Kyrgyz speech dataset collected only in Bishkek will look different from one collected in the south.

The regional split is real: northern and southern Kyrgyz varieties differ, and a corpus that mixes them without recording where each speaker came from cannot be separated later.

Russian is not a foreign language here

In Central Asia, Russian functions as a lingua franca rather than a foreign language. Treating it as a separate monolingual corpus misses the mixed speech that a regional product will actually encounter. The same loanword boundary problem that appears in Kazakh appears here too: an integrated borrowing is not the same thing as a switch into another language, and the distinction has to be written down before annotation.

The annotation convention matters more than the model. Two annotators without a rule will disagree about where Kyrgyz stops and Russian begins, and the disagreement is systematic rather than random.

Checklist for a Central Asian speech dataset

• Which languages, and whether each is monolingual or mixed.
• Script per language: Uzbek Latin, Uzbek Cyrillic, Kyrgyz Cyrillic.
• Whether to deliver parallel text layers for Uzbek.
• Mixed-speech convention, written down and applied consistently.
• Region per speaker: north versus south, capital versus other cities.
• Acceptance check per language and per region.

The decision that matters is which language, which script and which region, not how many hours. All three are cheap to specify and expensive to reverse.

For related pages, see Russian, Uzbek, Kyrgyz, the full dataset briefs, and the technical reference.