A translation tool for face-to-face conversation.

Conversation has a workflow.

When two people don't share a language, the problem is bigger than translating a sentence. They need a clear way to take turns, know when to speak, and keep going while they wait for the result. PyxMint treats the whole exchange as the product: people, speaking actions, noisy surroundings and pacing.

Make each speaking turn explicit.

The product is built around push-to-talk rather than always-on listening. A participant speaks a turn in the browser, the translation completes, and the next participant continues. An optional transcript and a usage allowance based on active speaking time are part of the documented scope.

Product decisions, and the questions they leave.

• Browser-based: one shared surface for face-to-face exchange; device conditions still need a structured test matrix.
• Push-to-talk: an explicit speaking action; turn length remains a key usability question.
• Optional transcript: its value for continuity needs more testing.
• Five languages: not evidence of equal quality in every direction.

What the field test exposed.

Testing in a Gyumri café brought the MVP into a noisy, real setting. Noise was a practical condition; some Armenian → Persian attempts showed sentence-start issues; synthetic Armenian input did better than some live, noisy attempts; and longer turns made latency more noticeable. These are qualitative observations, not benchmarks.

Known limits and next tests.

Evidence is qualitative; no quantified accuracy or latency results are claimed; production-scale reliability isn't established. The proposed next step is a repeatable test matrix (language direction, quiet vs noisy, live vs synthetic, short vs long turns, browser and device) before stronger claims are made.