Choosing a multilingual AI receptionist: test names, accents and language switching
Evaluate a multilingual AI receptionist with practical tests for names, numbers, accents, bookings and language switching. Includes a scorecard.
A multilingual AI receptionist should understand the languages your customers use, preserve their details and complete the intended task. A fluent greeting is only the first part of that job.
Before choosing a service, test it with a phone call in each required language. Give it a name, a callback number and a request. Then check what the business receives.
Use the protocol and blank scorecard below to collect evidence for your own callers. This article provides a testing method; it does not report a measured comparison of products.
Ask what “supports this language” means
Different parts of a phone system can have different language capabilities.
| Part of the experience | What to verify |
|---|---|
| Greeting and menu | The caller can understand the available choices. |
| Speech recognition | Names, numbers and requests are interpreted correctly. |
| Conversation | The assistant answers appropriately and asks for clarification when needed. |
| Generated speech | Pronunciation, pacing and number reading are understandable. |
| Business record | The saved request or booking preserves the confirmed details. |
| Human route | The fallback reaches someone who can help, or explains the available alternative. |
Ask the supplier to confirm the combination used by your account and plan. A voice that speaks Spanish does not establish that every other part of the setup handles Spanish equally well.
Vendors offer multilingual receptionist services. Zoom’s guide describes several supported languages, while Retell’s configuration documentation separates speech recognition and voice behavior. Treat those documents as configuration information, then test your use case.
Decide how to select a language for each call
There are three useful approaches.
A language chosen by number or menu. A caller reaches a dedicated number or selects a language before the conversation. This makes the choice explicit, but the greeting must be understandable to the caller.
A language selected during setup. A business whose callers mostly use one language can configure the agent for that language. Check what happens when someone uses another one.
Automatic detection and switching. The system identifies the language from speech or changes language during the conversation. This can be useful when customers regularly switch, but it needs testing with the actual language pair.
Retell states that its multilingual speech recognition has accuracy trade-offs compared with a single-language configuration and recommends selecting only the languages needed. That is guidance for Retell’s system, not a universal performance ranking. Retell multilingual documentation.
If callers only need English and Spanish, start by testing that pair. Adding languages unrelated to your customers creates more configurations to verify.
Define success before making calls
Pick the tasks the receptionist will be allowed to perform. An appointment request, a confirmed booking and a message for a person need different checks.
For a callback request, success might require the correct name, reachable phone number and accurate reason for calling. For a booking, also inspect the date, time zone, service and actual calendar entry.
Decide which variations are acceptable. If your system omits an accent in a name, does that affect the task? If an address is translated, does it still identify the correct location? Record that rule before comparing results.
A sample acceptance rule could read:
The confirmed callback number must match the controlled number in the test sheet. Any correction made by the caller must appear in the saved request.
That gives two reviewers something concrete to check.
Start with a small, repeatable test set
Choose two required languages and recruit three speakers for each. Ask each speaker to make ten scenario calls. That produces 60 test calls: two languages × three speakers × ten scenarios.
This is a proposed starting sample, not an industry standard or proof of general accuracy. Include speakers whose accents reflect your customers, and document which accents you did and did not cover. Choose speakers who are comfortable evaluating the assistant in the language they speak.
Keep the agent configuration fixed during the first round. Use controlled phone numbers, synthetic customer details and a test calendar. Do not create or alter real customer appointments.
For each call, record the configuration, speaker code, language, phone connection, noise condition and test time. Decide which details to give and what result to expect before you call.
Ten scenarios worth repeating
| # | Scenario | Example instruction for the tester | What to inspect |
|---|---|---|---|
| 1 | Name and surname | Give a test name naturally, then spell the surname if asked. | Both the confirmed name and the saved name match your acceptance rule. |
| 2 | Corrected callback number | Give a controlled number, then correct its final digits. | The stored number contains the correction. |
| 3 | Street and unit | Give a fictional address with a building and unit identifier. | The essential location details survive the conversation. |
| 4 | Ambiguous date | Ask for “next Friday morning.” | The assistant resolves the intended date and offers only authorized availability. |
| 5 | Spoken time | Give a time, then clarify whether you mean morning or afternoon. | The final time and time zone match the confirmation. |
| 6 | Changed request | Start with a new booking, then explain you need to change a test appointment. | The final action reflects the revised request. |
| 7 | Interruption | Correct the assistant while it is reading details back. | It acknowledges the correction and updates the result. |
| 8 | Realistic background sound | Repeat a routine request with moderate, safe background noise. | The assistant completes the task or asks a useful clarifying question. |
| 9 | Language switch | Switch between the configured languages during the call. | The response and saved details follow the documented switching policy. |
| 10 | Unresolved misunderstanding | Repeat a detail the system cannot capture, then request help. | It reaches the approved alternative without inventing the missing detail. |
Use the same scenario set for every candidate service. If language switching is outside a product’s documented scope, record that limitation explicitly rather than marking the language-switching test as complete.
Example English and Spanish prompts
These original prompts use a fictional customer. Have the participating speakers adapt them naturally to the regional language you need.
For the name test:
My name is Elena Muñoz. I’d like someone to call me about a service appointment.
Me llamo Elena Muñoz. Quisiera que alguien me llamara para concertar una cita de mantenimiento.
For a change of intent:
I already have an appointment. I need to change it, not make another one.
Ya tengo una cita. Necesito cambiarla, no reservar otra.
Give the tester the controlled callback number separately. Ask them to read it in their usual style, then repeat the test with a correction.
For the language-switch test, choose a policy in advance. You might expect the assistant to continue in Spanish after the caller requests Spanish. Another setup might offer a dedicated Spanish route. Test the behavior the service promises.
Listen to the call and inspect the saved result
Review both outputs. A conversation can sound correct while the saved number is wrong. A summary can look tidy while hiding repeated misunderstandings that frustrated the caller.
Ask a fluent reviewer to assess whether the assistant understood the request, clarified ambiguity and used understandable language. Have the operational reviewer inspect the actual record or calendar.
Keep task correctness separate from voice preference. You can prefer one accent or speaking style while finding that another configuration captures details more reliably.
If the assistant reads back a number correctly but saves something else, that is a failed task. If it cannot capture the number and provides the approved human alternative, record that as a successful fallback and a failure to capture the number automatically.
Those are different results, and both matter.
Report counts alongside the test conditions
Useful measures include:
- Requests with every required field correct.
- Bookings confirmed in the correct test-calendar slot.
- Calls completed without a person taking over.
- Calls that reached the intended human or callback fallback.
- Calls with an unsupported promise or invented detail.
- Calls abandoned during the test.
Report the numerator and denominator for each. “27 of 30 callback requests contained the correct confirmed number” is interpretable when you also give the language, conditions and configuration. Do not present that illustrative wording as an observed result unless you have run the calls.
A small sample cannot establish a precise general failure rate. Repeating the same script is useful for detecting regressions, but real callers introduce more variation.
Rerun failed scenarios after changes and keep the earlier results. If you change the voice or speech provider, record it as a new configuration.
Make the fallback work in the caller’s language
The route to help needs the same attention as the greeting.
Check whether the person taking the call can help in the caller’s language. If they cannot, explain the alternative accurately. Do not promise a Spanish-speaking colleague when the current rota has none.
Verify that an English summary preserves names, numbers and uncertainty from a Spanish conversation. Keep the original information available where your privacy policy and system permit it.
For European deployments, check the AI introduction and applicable privacy information in each supported language. The privacy and disclosure guide covers the questions to resolve.
Use the human-handoff checklist to test unavailable staff and failed transfers.
Choose a service based on tests in both languages
Begin with the callers and tasks your business already has. Compare evidence from that configuration, then expand coverage when there is a need.
For Double Menu, discuss the languages, voices and route you intend to use before activation. Its demo page contains prepared menu samples, not a measured multilingual conversation benchmark. Listen to samples and test the actual phone path.
Bring a short test script and your acceptance rules to a setup discussion. Ask to hear a name corrected, a date clarified and the fallback completed in the languages your customers use.