← All work
Feasibility study for a Czech-language voice AI advisor
A feasibility study for a Czech-language voice AI advisor: four documents on what latency and Czech speech quality allow, and what they do not.
The challenge
Do latency and Czech speech quality allow it at all?
A customer in agricultural inputs wanted a voice advisor that holds a genuinely natural expert conversation in Czech, reaches a concrete recommendation including dosage and price, and prepares a binding order. It was open whether the quality of Czech speech recognition and synthesis and the overall latency allow it at all, and how to stop the model inventing recommendations and prices.
Our approach
Four documents that say exactly what to build.
We produced a functional analysis that drew ten agent requirements out of the customer’s sample dialogue, a latency study, a proof-of-concept plan and a task-by-task implementation plan.
✓ Latency broken down by voice-pipeline layer at p50 and p95 for Czech, hosted in the EU
✓ Thresholds set for perceived silence
✓ In the proof-of-concept plan, order and VAT computed deterministically in code, not in the model
✓ An explicit line between what will actually be built and what will be fixtures