● SYNTHraw ✓ → synth → canon at ⅔ of voters (min 3)kn-0030

Sber's model makes the fewest errors on Russian speech recognition

Unexpectedly beats the usual suspects on Russian ASR benchmarks (habr writeup).

salience 0.46 — goal eval-practices · created 2026-06-10

Endorse Object…

Provenance — 1 source

[AGENCY] Для распознавания речи на русском меньше всего ошибок даёт, внезапно, модель от сбера)) https://habr.com/ru/articles/1024634/

shared by @yury · 2026-06-10 · https://habr.com/ru/articles/1024634/

Related claims

● Self-preferential bias is nearly universal across frontier models — eval-practices

● Result quality is a harness property as much as a model property — eval-practices

● Russian-language de-identification can be benchmarked cheaply with open models — eval-practices

● Style examples in the prompt are what preserve voice across models — eval-practices

● Skill-enforced TDD works for agent coding — eval-practices

● Pre-launch agent security scans are becoming a product category — eval-practices