When your teacher should switch to your language
Two bad answers to a real question
Ask whether a language lesson should ever use the learner’s own language and you get two confident, opposite answers.
The immersion position says never. The target language only, from the first minute, because every second spent in your first language is a second not spent in the one you are learning. There is real evidence behind this instinct — production time is the scarce resource, and a lesson conducted mostly in the learner’s own language is not a lesson.
The translation position says constantly. Explain everything in the language the learner already has, because comprehension is the bottleneck and an unexplained rule is a wasted one. This also has something to it: a beginner facing an unfamiliar structure explained only in the language they cannot yet parse learns nothing except that they are behind.
Both answers are wrong in the same way. They treat the question as a policy to be decided once, when it is a parameter that should have a different value for a nervous beginner in their first session and for an upper-intermediate speaker preparing for a negotiation.
The model: a course is a pair of languages
The cleanest way to think about it is that a course is not “English.” It is a pair: the language being taught, and the language used to support the learner while teaching it. Call them the target language and the support language.
Framing it as a pair rather than as a product makes the asymmetry obvious, and the asymmetry is the whole design. The target language needs everything: accurate speech recognition, natural-sounding synthesis, and genuine teaching quality from the model driving the conversation. The support language needs much less — enough recognition to understand you when you switch into it, enough generation to explain something clearly, and occasional synthesis for a reassuring phrase or a gloss.
That difference is why supporting many pairs is more tractable than it sounds, and it is also why the two roles should never be conflated. The support language is a scaffold. It is present so the target language can be practiced, and it should be sized accordingly.
The support level should decay
Rather than a global policy, the useful setting is a per-learner support level that starts high and falls as the level rises. Concretely, across the bands:
At the beginner level, support is high and expected. Glosses of new words, reassurance, procedural instructions, and the explicit permission to answer in your own language when you are stuck. A learner at this stage who is forced into target-language-only will spend the session in a state of low-grade panic, and panic is a poor condition for producing language.
Through elementary, support decays from high toward medium across the band. The learner starts handling routine exchanges without a safety net, and the scaffold is removed roughly as fast as it stops being needed.
At intermediate, support is medium and narrowly scoped: the first language is for genuinely new or abstract items, not for anything the learner could work out from context. The default has flipped — the target language is the working language of the lesson, and switching is an exception with a reason.
At upper-intermediate and above, support is low. The operating rule is target-language-only unless the learner switches first. That last clause matters: it is not a prohibition on the learner, it is a prohibition on the tutor pre-emptively translating. If you reach for your own language, the tutor meets you there and then comes back.
At advanced, the support language is essentially unused, and its remaining role is a fallback that almost never fires.
Why the first session is the exception
The intake conversation deserves separate treatment, and it is the one place where generous support is unambiguously correct.
The purpose of a first conversation is to find out how someone actually speaks. That is impossible if they are too anxious to speak at all, and a first call in a language you are insecure about is a genuinely stressful event for a lot of adults — particularly adults who read well and know their speaking lags behind, because they are acutely aware of the gap.
So the correct move is to say explicitly, at the start, that answering in your own language is fine. Not as a fallback the learner has to ask for, but stated up front. It costs a little target-language production in the one session where production was never the point, and it buys an honest sample instead of a defensive one. A level estimated from a conversation someone was too frightened to have is not an estimate of anything.
The same reasoning applies to the written form that precedes the first conversation. A short set of questions — your goal, your timeline, how much time you have each week, what you are interested in — is much easier to face than a cold voice call, and it gives the conversation something to start from.
Switching cleanly, mid-sentence
There is a mechanical problem hiding under all this that is worth naming, because it is where implementations usually fail.
If a tutor’s reply contains a target-language sentence and a short gloss in the learner’s language, the speech synthesis has to handle both, and a single voice trained on one language producing a phrase in another sounds wrong in a way that undermines the lesson. It reads as sloppy, and worse, a mispronounced gloss is not a gloss.
The workable approach is to have the model tag the language of each segment of its own output, and route each segment to the appropriate voice. The learner hears their own language pronounced properly by a voice suited to it, and the target language pronounced properly by the voice they associate with their teacher. This has to exist from the beginning rather than being retrofitted, because the set of languages a synthesis engine handles well is never the set you need, and per-segment routing is the only honest way to cover the gap.
The same requirement applies in reverse, on the way in. A learner mid-sentence will code-switch — start in the target language, drop a word from their own, carry on. Recognition that assumes a single language per utterance will mangle that, and the mangled version is what the tutor responds to. For a language learner, mixed-language input is not an edge case to be tolerated — it is ordinary, it happens in most sessions, and the recognition side has to be built for it.
What this looks like from the learner’s side
The practical version of all of this: you should be able to say “I do not understand” in your own language and get an explanation you can actually use, without that becoming the mode the lesson stays in. And you should notice, over months, that you need to do it less — not because a rule was imposed, but because the scaffold came down as you stopped leaning on it.
You should also be able to ask for the opposite. Wanting an explanation in the target language even though your band allows the easier route is a good instinct at any level, and a tutor should take that request as a standing preference rather than something you have to repeat each session.
What you should not experience is a tutor that pre-emptively translates things you would have understood. That is the most common failure at the intermediate bands, it feels like helpfulness, and it quietly removes the moments where comprehension gets built.
Support in your own language is not a compromise on immersion. It is what makes the immersive part survivable long enough to work — which is the same argument, in a different register, as the one in why speaking is the skill that does not transfer, and it is why the support level lives in your profile alongside everything else the tutor remembers about you.