About this guide: Our earlier piece, why AI phone agents sound robotic, covers the voice itself: flat delivery, lag and clumsy turn-taking. This one covers what is left after the voice is fixed, when an AI receptionist answers in a language other than English. It is Futuro’s language research, compiled from published literature on politeness and translation, and it explains why register, not accent, is often what gives a voice AI away.
Who this guide is for: owners and office managers of US service businesses whose callers speak Spanish, Korean, Japanese, Vietnamese, Tagalog or another language besides English, and who want to understand what a multilingual AI receptionist has to get right. It is not for readers looking for a ranked vendor list, a language course, or medical or legal interpreting.
Editorial disclosure: Futuro Corporation, which makes Futuro AI (our product), an AI receptionist that answers in 53 languages, wrote and maintains this page, so we have a commercial interest in the answer. We did not accept payment from anyone for this page, and nothing here is paid placement. Evidence level: peer-reviewed translation and sociolinguistics research, Japan’s official keigo guideline, the Real Academia Española, and U.S. Census Bureau and Bureau of Labor Statistics data; the link from those findings to “sounding robotic” is our argument. Limits are in the methodology section.
Why do AI receptionists sound robotic in other languages?
Often because they get the social register wrong: the wrong “you,” missing politeness words, or English-style cheer translated word for word. A voice can have a slight accent and still sound like a person. A receptionist that talks to an elder like a classmate sounds like a machine reading a script.
English hides the problem. It has one “you” for a child, a customer and a judge, so a script written in English never records which one the caller is. Rico Sennrich and colleagues at the University of Edinburgh stated the consequence in a 2016 NAACL paper on controlling politeness in translation: “In machine translation from a language without honorifics such as English, it is difficult to predict the appropriate honorific.” The authors of CoCoA-MT, a 2022 formality benchmark, put the problem plainly: “Specific problems arise when dealing with honorifics, particularly translating from English into languages with formality markers.” They also describe what the listener experiences: “Using wrong or inconsistent tone may be perceived as inappropriate or jarring for users of certain cultures and demographics.”
This matters to US businesses because these callers are common. The Census Bureau’s 2022 profile of languages spoken at home names Spanish, Chinese, Tagalog, Vietnamese and Arabic as the five most common non-English languages in US homes in 2019. Human professionals have long treated tone as part of the job: the Bureau of Labor Statistics’ handbook entry for interpreters and translators describes the work as to “Convert concepts, style, and tone in the source language to equivalent concepts, style, and tone of the target language.” A voice AI answering a business line is doing a live version of that, with no time to think.
Where does politeness live in each language?
In a different part of the sentence, which is why a single English-to-anything script cannot cover them all. Here are five languages US businesses hear, with the source for each.
| Language | Where respect lives | What a robotic agent gets wrong | Source |
|---|---|---|---|
| Spanish | Choice of “you” (usted / tú / vos) | Opens with tú, or mixes forms in one call | RAE DPD |
| Korean | Verb endings (speech styles) | Leaves the polite formal style used on service calls | Kim 2005 |
| Japanese | Honorific verb forms (keigo) | Too little keigo, or so much it sounds artificial | Keigo guideline 2007 |
| Vietnamese | Kinship terms chosen by age and status | One flat “you” for every caller | Farris 2012 |
| Tagalog | Particles such as po and opo | Drops po with older callers | Fontillas 2015 |
The pattern is the same in each row. The information a speaker needs to sound polite, such as the caller’s age, status or relationship to the business, is not in the English source. A system has to infer it from the call, or ask, or take a safe default and adjust. A system that simply translates does none of these.
How does Futuro AI use this research across its languages?
To decide how to address each caller in the caller’s own language, rather than translating an English script. Here is what that means at the level a caller hears, using only what our public engine page says.
The caller picks the language. Our VoiceAlive engine page says: “There is no ‘press 2 for Spanish,’ no language line, no preamble. The caller starts talking; the agent identifies the language and answers in it.”
Register adapts per language. The page lists “Formality conventions — usted vs tú -level decisions, honorifics, phone politeness” and “Vocabulary register — the right word for the caller, in their language” among what changes from one language to the next. This page, and our Spanish guide, are the research behind those lines.
Mixed-language callers are expected. “Callers code-switch,” the page says, and the agent “follows the caller wherever the conversation goes.”
The limit, in our own words. The same page says: “Language research was not uniform across 53. English and Spanish are still the deepest-tested pairing,” and “Any vendor claiming perfect parity across 53 languages is overselling.” Expect the mistakes below to be rarer in Spanish than in a language with fewer US speakers. On realism, in a company-run, double-blind study, 94% of 1,000 participants believed they were speaking with a person rather than AI (study details and caveats). It is not peer-reviewed, and as the engine page puts it, “The 94% study measured realism, not per-language perfection.”
What register mistakes make a voice AI sound robotic?
Eight mistakes, all versions of one problem: the system decided how to address the caller without the information it needed.
It uses the wrong “you”
Spanish separates usted, which the RAE says implies “cierto distanciamiento, cortesía y formalidad,” from the familiar tú. A receptionist that opens with tú claims a closeness the caller never offered. One that ignores a caller’s relaxed tú and stays stiff can sound cold. Our Spanish guide covers nine Spanish-call situations in detail, from vos to Spanglish.
It leaves the expected speech level
Korean marks politeness in the verb ending. In Kim Mijeong’s 2005 study of Korean service encounters, department-store staff answering customer queries by phone “mainly used the polite formal hapsyo style.” Market sellers in the same study used casual forms on purpose, to sound friendly. Register is a choice that depends on the setting, which is exactly what a translated script cannot make on its own.
It skips kinship terms where they are expected
Vietnamese speakers often address each other with family words. Michael Farris, in a 2012 study of Vietnamese pronouns, writes that kinship terms “are commonly used instead of pronouns, but they are also widely used with non-kin,” with anh (older brother) and chị (older sister) for adults of similar age and status. One neutral “you” translated from English skips that choice altogether and sounds like a form letter.
It drops politeness particles
In Tagalog, small words carry respect. Lemuel Fontillas’s De La Salle University dissertation on recorded Tagalog conversation reports that “Honorific markers such as po, opo, ho and oho were found to be dominant in the corpus,” and links their use to the speakers’ age and authority. Leave them out with an older caller and a grammatically correct sentence can still sound curt.
It is too polite
The fix is not maximum formality. Japan’s official keigo guideline (Council for Cultural Affairs, 2007) sorts honorific language into five categories and advises: “敬語は過剰でなく適度に使う,” use keigo in moderation, not excessively. A receptionist that stacks honorifics on a simple booking sounds as artificial as one that uses none.
It translates English cheer word for word
English service scripts lean on enthusiasm: “Awesome!”, “No problem!”, a bright yes to everything. Translated literally, that tone can sound odd or insincere in languages where service talk is more measured. This is our reading of the formality research above, not a measured finding, and it varies by caller.
It switches register mid-call
The CoCoA-MT authors flag inconsistent tone, not just wrong tone. An agent that says usted in the greeting and tú when confirming the booking sounds like two scripts stitched together, which is often what it is. Consistency is a separate skill from choosing well at the start.
It stalls on a mixed-language caller
Real callers mix languages. Pew Research Center found in 2023 that “A majority of U.S. Latinos (63%) say they use Spanglish at least sometimes, including 40% who say they do so often.” An agent that stalls, restarts in English, or corrects the caller’s English word turns a normal sentence into an error, and the caller hears a machine.
What does a register mistake cost a business?
Probably the call, sometimes the customer, but nobody has measured it, so read this as reasoning rather than data.
A caller who hears the wrong register learns two things at once: that they are talking to a machine, and that the business did not plan for them. The plausible results are a shorter, more guarded call, a request for “a real person,” or a hang-up and a call to a competitor whose staff speak like neighbors. Those outcomes match what CoCoA-MT calls “inappropriate or jarring,” but we found no published study that counts lost bookings from register errors, and we will not invent a figure. The scale of exposure is easier to state: the Census Bureau’s 2025 release on language tables found that 62% of people who spoke another language at home also spoke English “very well,” which means a large share of these callers can, and will, compare your AI receptionist against an English-speaking human at the next business on the list.
Where does the evidence cut the other way?
In four places, and a fair page has to say so.
Timing is more universal than people think. Stivers and colleagues (PNAS, 2009) compared 10 languages and found “all of the languages tested provide clear evidence for a general avoidance of overlapping talk and a minimization of silence between conversational turns.” Claims that one culture “likes silence” and another does not are weaker than they sound. The voice still matters. Lag, flat delivery and interruptions can make an agent sound robotic in any language; register is an extra layer, not a replacement, as our earlier piece explains. Casual can be right. Kim’s market sellers used casual Korean on purpose, and a 2022 study of New York City Spanish by Fernández-Mallat and Newman found “greater use of the familiar tuteo variant over the formal ustedeo” among younger speakers. A rigid formal default can misfire too. The research is mostly text. Sennrich and CoCoA-MT studied written translation, not phone calls; we are applying their findings to voice, which is our inference.
What can’t an AI receptionist do in another language?
It is a receptionist, not an interpreter, and it does not know your regulars.
It should not stand in for medical or legal interpreting. The BLS notes that “Healthcare or medical interpreters must be sensitive to patients’ personal circumstances and must maintain confidentiality and ethical standards,” and that “most states require certification for court interpreters.” It cannot know which long-time customers prefer an informal “you” until someone tells it. And no AI receptionist is equally good in every language; ours included, by our own published account.
Who should not rely on an AI receptionist for non-English calls?
A business with a fluent receptionist from its callers’ own community who answers every call; keep them, and use AI for nights, weekends and overflow. A business whose non-English calls are mostly clinical, legal or immigration conversations that need a qualified interpreter. And a business whose callers speak a language its AI receptionist has not been tested in deeply; ask the vendor which languages are deepest, and test them yourself.
How did we source the facts on this page?
Language findings come from peer-reviewed papers (Sennrich et al., NAACL 2016; Nădejde et al., Findings of NAACL 2022; Kim, 2005; Farris, 2012; Stivers et al., PNAS 2009; Fernández-Mallat and Newman, 2022), a De La Salle University dissertation (Fontillas, 2015), Japan’s 2007 official keigo guideline, and the Real Academia Española. Population figures come from the U.S. Census Bureau and Pew Research Center; job-duty language comes from the Bureau of Labor Statistics. Futuro AI is described only as our public engine page describes it. What we did not do: we did not run a listening test, record real calls, measure hang-ups, or score any vendor, including ourselves, on politeness in any language, and no native-speaker reviewers have checked this page yet. All links were verified on November 15, 2026.
Citable facts from this page
- Sennrich et al., 2016: from English, “it is difficult to predict the appropriate honorific.”
- CoCoA-MT, 2022: wrong or inconsistent tone “may be perceived as inappropriate or jarring.”
- Kim, 2005: Korean department-store staff on the phone “mainly used the polite formal hapsyo style.”
- Japan’s 2007 keigo guideline: use honorifics in moderation, not excessively.
- Stivers et al., 2009: all 10 languages studied avoid overlap and minimize silence between turns.
Frequently Asked Questions
Often because of register, not accent: the wrong “you,” missing politeness words, or English cheer translated word for word. A 2016 University of Edinburgh paper noted it is difficult to predict the right honorific when translating from English, and the 2022 CoCoA-MT benchmark warns wrong or inconsistent tone can feel jarring.
Languages that mark politeness in grammar, such as Korean verb endings, Japanese keigo, Vietnamese kinship terms and Tagalog particles. Three of these are in the Census Bureau’s list of the five most common non-English languages in US homes in 2019: Spanish, Chinese, Tagalog, Vietnamese and Arabic.
Both matter, but a slight accent can still sound human while the wrong register sounds scripted. Stivers et al. found that all 10 languages they studied avoid overlapping talk and minimize silence between turns, so timing is more universal than politeness forms are.
Our public engine page lists formality conventions, including usted vs tú-level decisions, honorifics and phone politeness, among what adapts per language. It also says research was not uniform across all 53 languages.
Yes. Japan’s 2007 keigo guideline advises using honorifics in moderation, not excessively.
The level of formality and social distance a speaker chooses for a listener, shown through word choice, forms of address and grammar.
Not always. Formal is the safer opening with a stranger, but casual speech is sometimes right, and younger New York Spanish speakers use the familiar form more often.
No. The BLS describes medical interpreters as bound by confidentiality and ethical standards, and most states require certification for court interpreters.
A well-built agent follows the caller. Pew found 63% of US Latinos use Spanglish at least sometimes.
Not in published research we could find. The translation studies measure text quality, not lost calls.

