Talk time: the metric that predicts progress
The one number worth reporting
Language products report a lot of things: streaks, words learned, lessons completed, levels unlocked. Almost none of it predicts whether a learner will be able to hold a conversation, because almost none of it measures the activity that produces that ability.
For a speaking lesson there is one honest input metric, and it is minutes spoken. Not minutes attended. Not exercises finished. The number of minutes during which your own voice was making sentences in the target language, under time pressure, with someone waiting for the answer.
Alongside it sits the ratio: of the total conversation, what share was you. A lesson where you talked for a third of the time and a lesson where you talked for two thirds are not the same lesson, even though both last the same half hour and both feel like they happened.
Why the ratio is the diagnostic, not the total
Total minutes can be inflated trivially — take a longer lesson. The ratio cannot, and it exposes the failure mode that actually occurs.
The failure mode is a tutor that is too interesting. A well-informed conversation partner has a great deal to say, explains thoroughly, gives context, offers three alternative phrasings, and tells you about the cultural background of the idiom you just used. The learner has an enjoyable time and leaves having produced very little language. In a human classroom this is the standard critique of teacher talk time, and it is the reason experienced teachers deliberately shut up. A model with no such discipline is far worse, because its default behaviour is to be helpful, and being helpful reads as talking.
So the ratio is not a vanity statistic. It is the measurement that catches the most common way a speaking lesson quietly stops being one.
Designing the tutor to say less
If the ratio is the thing to protect, it has to be protected in the design rather than left to good manners.
The most direct lever is a hard ceiling on how much the tutor produces per turn. It is a crude lever and it works: a ceiling forces the tutor to choose the single most useful thing to say instead of covering everything it could say. The side effect is that explanations get shorter and better, because nothing edits prose like running out of room.
The second lever is one question at a time. Stacked questions — “So where do you work, and how long have you been there, and do you enjoy it?” — are a habit of enthusiastic conversationalists and a disaster for a learner. Faced with three, a learner answers the last one, or the easiest one, and the other two are lost. One question produces one full answer, and one full answer is more speaking than a scattered response to three.
The third is a correction budget. At most one correction per turn, chosen for what recurs rather than for what is nearest. Correcting everything is not thoroughness; it is a reliable way to make someone stop talking. A learner who is interrupted on every third word will start producing shorter, safer, duller sentences, which is the opposite of the goal — and their talk ratio will fall while their apparent accuracy rises, which is the worst possible trade.
Waiting properly
There is a fourth lever that is invisible until it goes wrong: how long the system waits before deciding you have finished your turn.
Get it too short and the tutor talks over you while you are still assembling a sentence — which is exactly the moment a language learner most needs the room, because assembling a sentence in a second language takes visibly longer than in a first. Get it too long and every exchange has a hole in it, and the conversation feels laggy and dead.
Neither setting is universally right. The pause a beginner needs mid-sentence is longer than the one an advanced speaker needs, and the difference between two learners at the same level can be just as large. So it is worth treating turn patience as a per-learner setting, tuned from how that person actually speaks, rather than as a constant chosen once by whoever built the thing. It is a small parameter with an outsized effect on whether a lesson feels safe enough to be slow in.
Length, frequency, and the marathon trap
A related question is how long a lesson should be. Longer looks like more practice and often is not: attention degrades, and the last stretch of a long session tends to be lower-quality production than the first.
The shape used here is a half-hour core lesson with a hard cap at 45 minutes, and a standing preference for frequent short sessions over rare long ones. That preference is about more than attention — it is the spacing argument from the forgetting curve. Three sessions across a week put three spacings into your schedule; one session of the same total length puts in one.
There is also a fairness argument for measuring minutes, and it is the reason the billing rules can be stated as plainly as they are. A lesson credit is consumed only after 10 minutes of session time; a connection that drops before that refunds the credit automatically. Rejoin within 5 minutes of a disconnect and you resume the same session, in the same place, with no second credit charged. Neither rule is implementable unless minutes are being counted in the first place — a metric earning its keep twice.
What raises your own talk ratio
Some of this is on the learner’s side, and it is worth being direct about the moves that work.
Answer past the question. “Did you have a good weekend?” invites a one-word reply, and taking the one-word reply is a wasted turn. Add the reason, the detail, the complication. You are not being asked for information; you are being given an opportunity to produce language.
Do not repair silently. When you cannot retrieve a word, the instinct is to stop and hunt for it. Talking around it — describing the thing, using a longer phrase, asking what it is called — keeps you producing language and practises the paraphrase skill that upper-intermediate speakers are explicitly assessed on.
Pick scenarios that demand output from you. A role-play where you are the candidate in an interview, the person giving a status update, or the one making a case for a raise puts you structurally in the talking seat. A role-play where you are asking a receptionist for information puts you in the listening seat. Both are useful, but only one of them raises the ratio.
And resist the tutor’s explanations when you do not need them. “I understand, let us keep going” is a legitimate sentence, and it is more practice than listening to a grammar note you already knew.
What the number cannot tell you
Talk time is an input metric, and it should be read as one. Minutes spent producing fluent, wrong, unchallenged English are still minutes, and they will show up in the total looking identical to minutes of hard, corrected, structure-targeted production. A high ratio in an unstructured conversation is a good sign about engagement and a weak sign about learning.
That is why it belongs on a report card next to the qualitative material — the corrections against your own sentences, the new vocabulary, the progress against specific objectives — rather than on its own. As a headline it would invite the wrong optimization. As one line among several, it answers a question none of the others can: did this lesson actually consist of you speaking.
The method page covers what the rest of the report card holds, and how it works sets out the shape of a session end to end.