How Accurate Is AI Phone Call Translation? What the Research Actually Shows

AI translation is now accurate enough to rely on for everyday phone calls in major, well-resourced languages — published studies put machine translation in the low-to-mid 90s percent range for Spanish — but accuracy drops measurably for lower-resource languages, for idioms and jargon, and for anything high-stakes. No system is 100% accurate, every live call translator pauses for a second or more, and any product claiming either perfection or zero delay is overclaiming.

That's the honest answer, and the rest of this article is the evidence for it: the actual peer-reviewed studies with their numbers and years, what the research community considers a normal translation delay, where AI translation reliably fails, and when you should use a human interpreter instead. It also includes an honest accounting of our own product's limits — because a translation service that won't tell you where it breaks is not one you should trust with a call that matters.

Accuracy Is Not One Number

The question “how accurate is AI translation?” has no single-number answer, and anyone who gives you one is selling something. Measured accuracy swings by tens of percentage points depending on three things: the language pair (systems trained on oceans of English–Spanish data perform far better than on English–Armenian), the content (a delivery date translates near-perfectly; a joke, an idiom, or a legal term of art may not), and the stakes — because a 95% accurate translation means one sentence in twenty goes wrong, which is trivial when arranging a taxi and unacceptable when consenting to surgery.

So the useful question isn't “is it accurate?” but “is it accurate enough for this call, in this language?” The published research gives a surprisingly clear picture of where that line falls.

What the Published Research Actually Shows

The most rigorous accuracy studies come from medicine, where researchers had a concrete reason to check machine translation sentence by sentence: hospitals were using it for patient instructions. A study in JAMA Internal Medicine in 2019 (Khoong and colleagues) ran real emergency-department discharge instructions through machine translation and had bilingual reviewers grade every sentence: 92% came through accurately in Spanish, 81% in Chinese. More sobering, 2% of Spanish sentences and 8% of Chinese ones contained errors with the potential for clinically significant harm.

A 2021 study in the Journal of General Internal Medicine (Taira and colleagues) measured how much meaning survived translation across a wider spread of languages, and the gradient is the single most useful finding in this literature: Spanish 94%, Tagalog 90%, Korean 82.5%, Chinese 81.7%, Farsi 67.5%, Armenian 55%. Same engine, same source sentences — the only variable was the language. That forty-point spread between Spanish and Armenian is what “accuracy depends on the language pair” looks like in hard numbers: languages with vast training data behind them translate remarkably well, and lower-resource languages measurably worse.

For perspective on how fast this field moves, the historical baseline is a BMJ study from 2014 (Patil and Davies), which tested ten medical phrases across twenty-six languages on the machine translation of that era and found just 57.7% translated correctly overall— 74% for Western European languages versus 45% for African ones. Its most infamous result: “your child is fitting” (British English for having a seizure) rendered into Swahili as “your child is dead.” That was over a decade ago, and comparing it with the 2019 and 2021 numbers shows both how bad machine translation once was and how far it has come — while the per-language gap, note, never closed. It just moved up the scale.

How Much Delay Is Normal? (All of Them Pause)

Every live translation system on the market pauses, because translation requires hearing enough of a sentence to know what it means — many languages put the verb at the end, so translating word-by-word is impossible in principle. The research community is explicit about what “normal” looks like: the IWSLT 2025 simultaneous speech-translation evaluation — the annual academic benchmark for exactly this problem — brackets acceptable lag at roughly two to five seconds, with wider allowances for structurally distant pairs like English–Japanese. Meta's SeamlessStreaming research model, by the company's own 2023 figure, runs at about two seconds. That is the state of the art talking.

Hands-on testing matches the labs. When Tom's Guide tested iPhone and Galaxy call translation head-to-head in August 2025, neither was instant and neither was flawless: the iPhone flowed more naturally, beginning to speak after a couple of sentences, but stumbled on harder passages; the Galaxy translated more accurately — including passages the iPhone missed — but waited for the caller to finish speaking entirely, making conversation feel heavily delayed. Flow versus accuracy turned out to be a genuine trade-off, and the reviewer's pick went to flow. Two of the most advanced consumer implementations on earth, and both imperfect in different directions.

Which makes vendor latency claims a useful honesty test. At least one call-translator app advertises sub-half-second translation — a self-reported figure, independently verified by no one, that would beat Meta's published research and undercut the academic benchmark's entire “low latency” band. When a product claims a delay dramatically below what the research community achieves, the likeliest explanation is marketing, not a breakthrough. A vendor honest about its pause is telling you something reassuring about its other claims too.

Where AI Translation Still Fails

The failure modes are consistent enough across studies and hands-on tests that you can plan around them:

• Idioms, jokes, and sarcasm.Figurative language is where literal translation goes wrong — the 2014 BMJ study's worst errors were idioms, and it remains the weakest category today. Mitigation: say what you mean plainly. “The price is too high” survives translation; “that's highway robbery” may not.
• Specialist jargon. Legal, medical, and technical terms of art often have no everyday equivalent, and a translator that renders them approximately can change the meaning. Mitigation: describe the thing instead of naming it, and confirm any term that matters.
• Lower-resource languages. The 94%-to-55% gradient above. If your call is in Farsi, Armenian, Swahili, or another lower-resource language, expect more rough sentences than a Spanish or German speaker would see. Mitigation: shorter sentences, more read-backs, and more weight on verifying afterwards.
• Proper names and numbers.Speech recognition mangles unfamiliar names, and a wrong digit in a reference number is worse than a wrong word. Mitigation: spell names letter by letter and say numbers digit by digit — “four-eight-two-one” translates perfectly; “forty-eight twenty-one” invites trouble.

When Not to Use AI Translation at All

Some calls should not run through any AI translator, ours included. The medical studies above found a small but real share of errors capable of clinically significant harm — 2% of sentences even in Spanish, the best-performing language. At those rates, AI translation is the wrong tool for medical consent and treatment decisions, legal proceedings, and contract negotiations— any conversation where a single mistranslated sentence carries consequences you can't undo. For those, use a professional interpreter: hospitals are typically required to provide medical interpreters, courts provide sworn interpreters, and no per-minute saving justifies skipping them. Our guide to calling a doctor abroad in another language draws exactly this line in practice: AI translation for the appointment booking and the prescription chase, a human medical interpreter for the diagnosis conversation.

Everything below that threshold is where AI translation has quietly become genuinely dependable: booking appointments, chasing deliveries, customer service, calling a landlord or a supplier, checking in with family. These calls are made of exactly the plain, concrete language machine translation handles best, and an occasional awkward sentence costs you a moment of clarification, not a legal problem.

How to Get the Most Accurate Translated Call

The same habits professional interpreters ask of their clients measurably improve AI translation, because both work sentence by sentence:

• One short sentence per thought. Translation systems handle complete, bounded sentences far better than winding sixty-word ones. Say it, pause, continue.
• Skip the idioms. Plain words translate; wordplay gambles.
• Read numbers back. Dates, amounts, reference numbers: repeat them digit by digit and ask the other side to confirm. Five seconds against the highest-stakes error class in any call.
• Spell proper names. Street names, surnames, product codes — spelled letters come through cleanly where a fast-spoken name may not.
• Review the transcript afterwards.If your service saves one, re-read the important parts once the pressure is off — it's where you catch the detail you missed live.

Our Own Numbers, Honestly

BubblyPhone's Live Translation is subject to everything above, and we'd rather say so than be caught pretending otherwise. It carries a Beta label because the translation layer is new and improving: expect a 1–2 second interpreter pause after each sentence, and occasionally a sentence that comes out awkward. It covers 62 languages, and the per-language quality gradient the research describes applies to us as much as anyone — Spanish and German calls run smoother than lower-resource languages. It costs $0.20/min on top of the normal call rate, with no subscription, and the person you call needs nothing: no app, no account, any phone including a landline.

One design decision is our direct answer to the trust question this article is about: during the call you see live captions of both sides in both languages, and the full bilingual transcript saves to your call history when you hang up. That exists precisely so you never have to take the translation on faith — you can verify what was said, in the moment on screen and afterwards in writing, and show it to a bilingual friend if something seems off. A translation you can check is categorically more trustworthy than a translation you must simply believe — and we think every product in this category should offer it. For how the alternatives compare on this and everything else, see our call-translator comparison — and if you were planning to hold Google Translate up to a speakerphone instead, here's why that doesn't work.

Frequently Asked Questions

Related Guides