AI Translation Quality Benchmark
How well InterMIND actually translates — measured by our automated test suite over a fixed FLORES-200 benchmark set, per language pair, every month. Synthetic and reproducible, not cherry-picked averages.
typed messages, translated inline
one score for speech recognition + translation, end-to-end
Score = semantic fidelity to a reference translation (0–100). Higher is better.
Text chat translation
Typed messages translated inline as they're sent.
Average quality per language · Q3 2026 vs Q2 2026
Each bar pools every run translating to or from that language this quarter. The dashed notch marks the previous quarter.
Faded bars have a thin sample (<3 runs) — read them as provisional.
Voice translation
Live speech, transcribed and translated in real time — the harder end-to-end problem.
Average quality per language · Q3 2026 vs Q2 2026
Each bar pools every run translating to or from that language this quarter. The dashed notch marks the previous quarter.
Faded bars have a thin sample (<3 runs) — read them as provisional.
Quality over time
Monthly median per modality, since we started publishing.
These numbers come from real sessions
Every run on the live demo is scored automatically and folded into next month's benchmark. A session looks like this: