Until the benchmark is done, the best way to check is to upload your own recordings: the first review is free.

How checking works

• Every mark — “done”, “not done” or “not applicable” — comes with the model’s confidence.
• Confident marks are set automatically, the rest go to a person as “needs review”. A disputed “not done” is sent to review by us.
• The day’s score is shown as a “from – to” range until the disputed items are checked.

What we will publish

100 recordings in Kazakh, Russian and mixed — calls and in-store conversations; transcribed by two people; recognition errors by language and mark accuracy on a held-out set; the files and the comparison script.

Good to know

• The quality of the marks depends on how the checklist is worded — the studio suggests how to phrase items so they can be checked.
• The model answers “yes/no” based on what was said and doesn’t rate tone on a scale.
• Speech recognition works for Kazakh and Russian today; English is coming soon.

Recognition compared with other systems

A comparison with Yandex SpeechKit and Google — after a benchmark on 100 recordings with human transcripts.