Evaluate restaurant voice AI across languages and accents

Recruit for the task, not a stereotype

Invite voluntary speakers who use the target language naturally. Ask them to order or request information in their own words. Do not assign a nationality from an accent, imitate a speech difference or ask a synthetic voice to stand in for an entire population.

Use a pseudonymous speaker code. Collect only the language and participation context needed for the evaluation, with consent. If a participant says a script sounds unnatural, correct the test design instead of marking the speaker as difficult. The purpose is to discover where the interaction needs support.

Test menu language, not just conversational fluency

Restaurant speech mixes languages and local vocabulary. A caller may speak Spanish but use the menu’s English item name. A dish can have multiple common pronunciations, and a modifier may be understood differently across regions. List the terms that matter to your restaurant and ask speakers how they would naturally request them.

Include numbers, times, quantities and spelling only when they are part of a real task. Do not claim general accent robustness from a test that contains a single familiar menu item. Keep the same underlying goal across speakers while allowing natural phrasing.

Code-switching is its own condition

• Start in one language, then ask for a phrase to be repeated in another.
• Use the menu’s original item name inside a different-language sentence.
• Ask whether the conversation can continue in the preferred language.
• Correct a quantity or time after switching languages.
• Request a person when the system cannot continue usefully.

Report outcomes without overgeneralizing

Show the number of speakers, cases and repetitions for each tested language and condition. If one person runs 40 cases, that is 40 observations from one speaker, not evidence from 40 independent speakers. A pooled rate can hide a weak subgroup, so inspect the individual conditions before making a deployment decision.

Loman’s current multilingual page lists 55 languages, automatic language detection and English POS/KDS tickets. Use that as a reason to evaluate the relevant phone workflow. This reference has not independently verified a success rate in those 55 languages.

Customer evidence for your voice test

Write a language-specific acceptance note

Example, not a measured result: “For the Spanish pickup-order cases, the guest can use the menu’s original dish names, correct a quantity and hear the final request repeated in Spanish.” Keep separate notes for the spoken outcome and the ticket language.

The Loman multilingual workflow description is a useful starting point for those two checks. It does not establish how your selected speakers will perform. Record any language switch, unresolved menu term and staff handoff in the evidence log.