Mixed recordings: Kazakh and Russian in one interview
Who this is for
That is how people actually talk here: the question in Kazakh, half of the answer in Russian, the technical term in English. Most interviews recorded in Kazakhstan sound like that, and the first worry before cutting them up is always the same — will the service get confused?
Here is how it works on our side, with nothing promised that you cannot check.
What happens to mixed speech
We recognise speech in 10 languages, Kazakh among them. The language is decided word by word, not once for the whole file: in a single conversation some words come back as Kazakh and some as Russian, and that is correct work, not a glitch. "87 % Kazakh, 13 % Russian" is what a real conversation looks like. A genuine error looks different: Kazakh was ordered and 96 % came back Russian.
We keep the language of every word together with the recognition confidence. Not for decoration: if the subtitles are not what you expected, the conversation is not a matter of opinion — there is something to look at (offer, clause 5.15).
"Spoken language" is a hint, not a subtitle setting
When you place an order you can name the spoken language or leave "detect automatically" — that is the default.
Three things worth knowing about that field:
- it does not choose the subtitle language. Subtitles come out in the language that was spoken. Tag a Russian recording as English and you get no translation: you get Russian speech recognised as English — gibberish in Latin letters;
- pick Kazakh and you get Russian with it. A Russian hint is added to Kazakh automatically, because Russian words inside Kazakh speech are the norm for our audience. Without that pair, recognition starts spelling Russian words with Kazakh letters;
- "detect automatically" is not "no hint". It also goes out as the Kazakh + Russian pair: it is our default, and most orders placed on it are mixed recordings.
A hint is not an order. Language detection overrides it: in live order #232, Ukrainian speech sent with a "Kazakh, Russian" hint came back as Ukrainian throughout, with confidence 0.9897. Getting the field wrong for an unfamiliar language is not fatal.
What to pick when there are two languages: pick Kazakh, or leave "detect automatically" — both give the same pair of hints. Do not pick Russian: it has no companion language, so the Kazakh half of the conversation goes in with no hint at all.
Letters on screen are a separate problem
Recognising a Kazakh word and drawing it in the frame are two different jobs. Nine letters — ә, ғ, қ, ң, ө, ұ, ү, һ, і — are missing from many fonts, and when they are missing they do not turn into boxes. They simply vanish, leaving a hole inside the word.
On our side the font is chosen for the letters that end up in the frame, and coverage is verified by reading the font file itself. How to test your own clip — in a separate article.
In a mixed recording this matters more than in pure Kazakh: the Cyrillic of Russian words survives almost always, so the holes appear only in the Kazakh words. The text looks whole until you try to read it.
If you need one language instead of two
Subtitles can be translated into another language, and then a mixed conversation reaches the screen in a single one. Translation is free and costs no minutes.
After the clips are ready
Subtitle text is editable in your account, and rebuilding a clip is free and costs no minutes: one misheard word takes a minute to fix, not a new order.
Prices in tenge and what the free first order includes — in a separate article. What video clipping costs in Kazakhstan.