Restaurant voice AI test-pack builder

What the generator holds constant

Each selected scenario is repeated across languages, conditions and speaker labels. Case IDs make it possible to tie a recording and a result to the planned test. Speaker labels are pseudonyms, not instructions to collect names, nationality or other personal information.

Keep the vendor version, voice configuration, menu and phone path fixed during a comparison. If you change a setting, start a new run or label it clearly. Do not quietly mix a repaired configuration into an older result set. The generator caps a pack at 300 cases to keep accidental configurations manageable.

Adapt the scripts without losing comparability

Replace the sample menu term with something from your real menu. A fluent speaker should adapt the intention naturally in each language; word-for-word translation can create an unnatural task. Preserve the intended request, the correction and the expected conversational behavior.

“Pass” should mean the predefined interaction goal was met without an unresolved misunderstanding. It should not mean the voice sounded pleasant, the order reached the kitchen or a reviewer liked the vendor. Define downstream transaction checks separately.

Prepare recording responsibly

• Use informed participants and test identities. Obtain any required recording consent before the session.
• Keep payment credentials and unrelated guest information out of the exercise.
• Assign a recording reference or local filename to each case and store audio in an access-controlled location.
• Record the exact setup, including noise and connection conditions. Do not invent a decibel value you did not measure.
• Run cases in varied order and preserve failures, interruptions and unavailable tests.

A small starter pack you can reproduce

Illustrative setup: 2 languages × 2 speakers × 2 conditions × 4 scenarios × 1 repetition produces 32 planned cases. Run the quiet baseline and background-noise condition with the same menu version. A planned case is not an observation until a person performs and reviews it.

If the run is too large, reduce conditions or repetitions rather than dropping difficult results afterward. Compare the language labels with Loman’s stated language support, then ask your speakers to phrase each intention naturally.