A conversation translator turns a two-way spoken exchange into each person's language, turn by turn, and some tools use earlier turns to help replies make sense. Turn handling matters, and so does whether a tool uses earlier context.
Microsoft helped name this category years ago with Lync Conversation Translator. Today its own Translator App Store listing carries a blunt note: "The multi-device conversation feature has been retired." Meanwhile, hands-free turn detection ships in free phone apps. If you've ever watched a translated reply come out grammatically perfect and completely wrong, you already know the problem this article is about.
Below you'll find what a conversation translator has to remember between turns, what two peer-reviewed research efforts found about it, a five-line test you can run on any tool in your own language pair, and how MirrorCaption handles it. If you came here for a ranked list of iPhone apps, our separate guide to picking a translator app for real-time conversation on iPhone covers that ground. This page is about what happens to meaning across turns.
Key Takeaways
- In an English–Russian study, about 7% of consecutive sentence pairs were individually good translations that turned wrong when read together; formal/informal "you" caused 37% of those errors.
- The WMT 2024 Chat Translation task found systems "excelled at translating individual turns," and that context from previous turns "almost always improved translation quality" (Mohammed et al., 2024).
- Hands-free turn detection is available in major apps; also test how a conversation translator handles context: register, pronouns, short answers and consistent terms.
- A five-line script, run in your own language pair, exposes those memory gaps in under two minutes.
- MirrorCaption Talk mode keeps one continuous session; use the five-line test to assess its output in your own language pair.
What is a conversation translator?
A conversation translator is a tool that translates a live, two-way spoken exchange between people who don't share a language. It listens to speakers, identifies the language of each turn when supported, translates into the other person's language, and shows or speaks the result. Some tools use recent turns for context, but accuracy still needs checking.
That last clause is the difference from a text translator. A conversation translator gets a stream of short, half-finished utterances from two directions: "Yes." "Him?" "Tomorrow, then." Lines like these carry little meaning on their own. The rest lives in the turns before them.
So a conversation translator has four jobs:
- Capturing speech from both people, ideally without anyone pressing a button per sentence
- Detecting which language each turn is in
- Translating each turn using the conversation so far, not just the sentence in isolation
- Delivering the result as text, and optionally as spoken audio for the listener
Hands-free turn detection is available in major apps
A few years ago, the first job was the hard one. Today major phone apps offer it. Google's own help page for bilingual conversation in Google Translate states: "The microphone automatically detects when one language stops and the other language starts." Google also introduced its Live translate conversation feature in August 2025, covering 70+ languages. Apple's Translate app offers an Auto Translate option with language detection for its conversation view.
Microsoft went the other way: its multi-device conversation feature is retired, yet several 2026 roundups still describe it as live. Check vendor documentation, not lists.
Turn detection alone does not settle the choice. Once a tool can hear the handover, ask: what does it remember from one turn to the next?
What a conversation translator has to remember between turns
Linguists have names for the things that break when a translator forgets. Four of them show up constantly in real conversations.
Who "you" is: formal versus informal address
English has one "you." Spanish has tú and usted, German has du and Sie, Russian has ты and вы. The choice is set early, often in the greeting, and then has to hold. A translator that picks the form fresh for every sentence will flip between polite and familiar mid-conversation. That's mildly confusing between friends and awkward with a new client. In the Voita, Sennrich and Titov study, this kind of error (deixis) made up 37% of all context-dependent mistakes.
The answer that leaves out half the sentence
"Did you sign the contract?" "Yes, yesterday." The reply has no verb and no object. Translated alone, "Yes, yesterday" is trivial. But in languages that mark tense, gender or case on the missing parts, the translator has to rebuild what was dropped from the question before it. Ellipsis accounted for 29% of the context errors in the same study.
Pronouns that speech doesn't mark
Written Mandarin distinguishes 他, 她 and 它. Spoken Mandarin says tā for all three. A conversation translator working from speech has to decide between "he," "she" and "it" from context alone. Japanese often drops the subject entirely: 昨日行きました. Who went? The sentence doesn't say. The previous turn usually does.
The same term, translated the same way every time
If the first mention of a product is translated as "invoice," the third mention shouldn't come back as "bill" and the fifth as "receipt." Lexical cohesion was 14% of context errors in the study. On a sales or support call, inconsistent terms make it sound as if three different things are being discussed.
The supplier call that switched to du
This is a constructed example to show the mechanism, not a customer story. Lena, a buyer in Hamburg, starts a call with a new Spanish supplier using Sie. Twenty minutes in, the supplier's sales lead, Javier, suggests they drop the formality, and Lena agrees: "Gern, ich bin Lena." From that turn on, every "you" Lena says should come out as tú in Spanish, not usted. A translator that resets per sentence may drift back to usted on the next question, which reads as a polite retreat from the offer Lena just accepted. A translator that carries the conversation forward keeps the new register.
What the research says about conversation-level translation
Two independent research efforts measured what this article has been describing.
Voita, Sennrich and Titov (ACL 2019). The authors had professional translators look at pairs of consecutive sentences from English–Russian subtitles. In roughly 7% of pairs, each sentence was "individually good, but bad in context of each other." By cause, those errors split into deixis (37%), ellipsis (29%), lexical cohesion (14%), ambiguity (9%) and anaphora (6%). They then built test sets for each category. A context-agnostic baseline scored 50.0 on deixis and 28.4 on verb-phrase ellipsis. Their context-aware model scored 81.6 and 80.0. On lexical cohesion the gain was smaller, 45.9 to 58.1, which suggests consistent terminology is still the hardest of the four.
The 7% figure comes from this English–Russian subtitle study, not a measured error rate for live speech or any current app. It shows why a fluent-looking reply still deserves a context check.
WMT 2024 Chat Translation shared task. The Conference on Machine Translation ran a task on bilingual customer-support conversations, following its first chat edition in 2020. The 2024 findings report that participating systems "excelled at translating individual turns," but that there is "room for improvement in overall conversation-level translation quality." They also found that "incorporating contextual information from previous turns almost always improved translation quality."
The study found a gap between turn-level and conversation-level quality. Test that gap with your own speakers, setting and language pair. Spend your evaluation time there.
| Context problem | Share of context errors (Voita 2019) | What to listen for |
|---|---|---|
| Deixis (formal/informal "you") | 37% | The register flips mid-conversation |
| Ellipsis (short answers) | 29% | "Yes, yesterday" comes out with the wrong tense or gender |
| Lexical cohesion (repeated terms) | 14% | The same thing gets three different names |
| Ambiguity | 9% | A word picks the wrong sense despite earlier clues |
| Anaphora (pronouns) | 6% | "It" or "he" points at the wrong thing |
Want to see context carried across turns in practice? Open MirrorCaption in your browser and start a Talk mode session. Check the current free allowance in the app before starting.
A five-line test for any conversation translator
You don't need a benchmark to check a conversation translator. You need two people, two languages and about two minutes. Write these five lines in your own language pair, then say them aloud in order, alternating speakers:
- A formal greeting that sets the register. For example, a polite opening that uses Sie, usted, вы or keigo.
- A question that names an object. "Did you receive the revised contract?"
- A reply that refers to it only by pronoun. "Yes, I read it last night."
- An elliptical answer. "Did you sign it?" "Not yet. Tomorrow."
- A follow-up that reuses an earlier term. "Should I send the contract to your lawyer?"
Then read the translated output and check four things:
- Register holds. Lines 3 to 5 keep the form of address set in line 1.
- The pronoun resolves. "It" in line 3 is translated with the gender or form that matches "contract."
- The ellipsis expands correctly. "Tomorrow" reads as "I'll sign it tomorrow," with the right tense and object where the target language needs them.
- The term stays consistent. "Contract" in line 5 is the same word as in line 2.
If your pair involves Mandarin, add a sixth line that uses tā for a woman mentioned earlier. If it involves Japanese, drop the subject in the reply and see whether the translation recovers it. Run the same script on every tool you're considering. The results will tell you more than any ranking.
Who is tā? A hiring conversation
This is a constructed example to show the mechanism, not a customer story. Priya, a recruiter in Singapore, is talking with Mr. Zhou, a hiring manager in Shenzhen, about a candidate named Anna. Earlier in the call, Mr. Zhou mentioned his own manager, a man. Then he says, in Mandarin: "tā 下周可以来面试吗?" Is he asking whether Anna can interview next week, or whether his manager can join? The pronoun is identical in speech. The answer sits two turns back, where Priya had just asked about Anna's availability. A translator that sees those turns can render "Can she come in for an interview next week?" One that sees only the sentence has to guess.
How MirrorCaption handles a conversation
MirrorCaption is built around the idea that a conversation is one continuous thing, not a series of separate translation requests.
- One continuous Talk mode session. On your phone, you start Talk mode once and the microphone stays active until you stop it. Both people speak in turns inside that one session. It isn't push-to-talk, and nobody restarts capture for each phrase.
- Check later replies. The live transcript lets you compare each translated reply with what was said earlier; test context handling rather than assume complete recall.
- Original and translation side by side. You see what was said and how it was translated. Compare the translation with the original transcript when a pronoun or term looks off.
- Speak Translations. When captions aren't enough, MirrorCaption can read your translated speech aloud in the other person's language, through your laptop speaker or a phone paired by QR code. The other side hears the translation and keeps the exchange moving.
- Multiple selectable languages. Check the currently supported speech and translation directions for your pair before relying on a live session.
- Open it in a browser. No install is needed to try it: Talk mode works best in Chrome on mobile. Native mobile apps are also available; check the current platform listing for your device.
For conversations on a laptop, Meet mode captures meeting-tab audio in desktop Chrome or Microsoft Edge, so live captions and translation work on browser-based Zoom, Microsoft Teams, Google Meet and Webex calls without a bot joining.
There are limits. MirrorCaption needs an internet connection, because live speech transcription and translation run online. It translates speech, not documents. If you need to translate a typed paragraph, a PDF or a photo of a sign, a text translator is the right category of tool. For live, spoken, back-and-forth conversation, that's what MirrorCaption is for.
「それでいいです」 at the end of a long call
This is a constructed example to show the mechanism, not a customer story. Daniel, a product manager in Toronto, spends 30 minutes in Talk mode with Ms. Sato, a manufacturing partner visiting from Osaka, going through three packaging options. Near the end he summarizes: option B, shipped in March. Ms. Sato replies: 「それでいいです」. On its own, that line is a short, flexible phrase. Inside the conversation, it's agreement to a specific plan. Because Daniel's session has been continuous, the translation sits next to the turns that give it meaning, and he can tap through to the original wording before he writes the follow-up email.
Run the five-line test on MirrorCaption
Start one Talk mode session, speak in turns, and check whether turn five still knows about turn one. Check the current free allowance in the app.
Start Free TrialWhen to bring in a human interpreter
A conversation translator is a strong fit for everyday cross-language talk: supplier calls, visits with relatives, team check-ins, a chat with a new neighbor. Some conversations carry consequences that call for a qualified human interpreter instead:
- Legal proceedings, depositions and anything that will be entered into a record
- Medical consent, diagnosis and medication instructions
- High-stakes negotiations where a single misread phrase can cost a deal
There, a professional interpreter brings accountability, ethics training and the standing to ask a speaker to clarify. A conversation translator can help you follow along, but it shouldn't be the only line of understanding. Our comparison of real-time translation vs human interpretation covers where that line falls.
FAQ
What is a conversation translator?
A conversation translator is a tool that translates a spoken, two-way exchange between people who don't share a language. It listens to both sides and translates turns into the other person's language. Check whether it uses earlier turns to improve replies, pronouns and forms of address.
Can a conversation translator work hands-free?
Yes. Hands-free turn detection is now common: Google's help documentation says its conversation mode detects when one language stops and the other starts. MirrorCaption Talk mode runs as one continuous session, so both people speak in turns without restarting the microphone for every phrase.
Does a conversation translator remember what was said earlier?
It depends on the tool and language pair. MirrorCaption Talk mode runs a continuous live session, but you should test later replies rather than assume it always retains earlier context.
Is Microsoft Translator's conversation feature still available?
Not the multi-device version. Microsoft's App Store listing for Translator states that the multi-device conversation feature has been retired, even though several 2026 roundups still describe it as current.
Can a conversation translator speak the translation out loud?
Many can. In MirrorCaption, Speak Translations reads your translated speech aloud in the other person's language. Playback can use your laptop speaker or a phone paired by QR code, so the other side hears the translation during the live exchange.
Do conversation translators work offline?
Some do, with limits. Google documents offline live translate only on certain Pixel phones and English-paired languages. MirrorCaption needs an internet connection, because live speech transcription and translation run online.
The second sentence is the test
Major apps can detect when speakers take turns, but performance still depends on audio and language pair. The open one, measured in ACL and WMT research, is meaning across turns: keeping the register, resolving the pronoun, rebuilding the short answer and holding the same term throughout.
So don't judge a conversation translator on its first sentence. Judge it on the reply that only makes sense because of the first one. Write your five lines, run them in your own language pair, and read the output closely.
When you're ready, run the test on MirrorCaption. Talk mode runs as a continuous session, and Speak Translations can voice your side so the other person hears it in their language.
Test a conversation translator across turns
Check the current free allowance before starting. Open it in your browser or the Android app.
Start Free Trial