Official Speak language learning AI app features languages pricing 2026

Exit 6, and the moment a subscription becomes a purchase

Seoul's subway exits are numbered, and the numbers matter. Come up the wrong one at a large interchange and you are four hundred metres and one river-sized road from where you meant to be. The first time this happened to us we had a phone, a map, and a stranger being extremely kind at a speed we could not follow.

Nothing about that failure was expensive in money. It was expensive in the way these things are actually expensive: you stop asking. You default to the app, the pointing, the English. Three days later you are having a fine trip in which you have not spoken the language once, and the subscription you bought in February has purchased nothing.

Which is the honest frame for a pricing article. The question is not what a plan costs per month. It is what fraction of that money converts into you producing unrehearsed language, because that is the only part of the transaction that changes anything at exit 6.

Speak is a good subject for that question because it is unusually focused. It does not want to be your grammar reference or your flashcard deck. It wants you talking, in volume, and it charges accordingly.

What Speak ships in 2026

The current product is built around short, dense speaking sessions. You are given a line or a prompt, you say it, the recogniser scores you, and you move on faster than feels comfortable. Underneath there is a course structure, a set of conversational scenarios, and a free-talk mode with an AI partner.

Three things Speak does better than the field. The speech recogniser is unusually tolerant of a learner accent, which sounds minor and is not — a recogniser that rejects your attempts teaches you to stop attempting. The session pacing is relentless in a way that produces genuine automaticity on a phrase set. And the curriculum sequencing for its top languages is well judged, particularly the early ordering.

What it does not do. Explanation. You will be told reliably that a sentence was wrong and much less reliably why, and the underlying model of you is a single progress line rather than a profile with disagreements in it. For a rehearsed phrase set that is fine. For the sentence you did not predict, you are on your own.

Language coverage in 2026 is broad on paper and uneven in practice, with Korean, Spanish, English and Japanese carrying most of the investment. If your target is outside that group, expect the free-talk mode to be stronger than the structured content, which is the reverse of the usual pattern.

The plan ladder, and where the fences sit

Every subscription product in this category fences the same three things: how much you can do per day, how much of the review machinery you get, and whether the AI partner is available at all. Where a company puts those fences tells you what it believes its value is.

Note the pattern. The products that fence speaking volume believe speaking volume is the product. The products that fence content believe the content is. That is a more useful signal than any feature list, and it is free to read.

Speak's fence is in a defensible place. The daily allowance on the free tier is real practice rather than a teaser, and it runs out at roughly the point where you would want to continue. That is good product design and it is also, obviously, the point.

What a hundred dollars buys in minutes of your voice

We took each product's 2026 annual price, divided it by the hours of learner speech we could actually extract in a month of ordinary use, and normalised to a hundred dollars of spend. Learner speech only: not the app talking, not menus, not reading a transcript back.

Two readings of that chart are wrong. The first is that Babbel is bad value — it is not selling talking hours, it is selling a course, and judging a syllabus by minutes of speech is like judging a cookbook by calories. The second is that the top of the chart is simply the cheapest product. It is not; the ordering comes from how much of a session is you rather than the app.

Speak's twenty-four hours is a strong number and the highest of any product built purely around repetition. Where it loses ground to the top of the chart is that some of those hours are spent repeating things you had already automated, because nothing in the system is tracking that you had.

The alternative we recommend, and the reason

Enverson AI finishes first here on a slightly awkward basis: it is not the cheapest sticker price in the group, and it still wins on cost per useful hour. The mechanism is the Multidimensional Personalization Engine, which keeps several independent readings of your ability running in parallel instead of averaging them into a level. Every competitor on this page reduces you to one number. Enverson is alone in the category in refusing to.

The economic consequence is direct. If a product cannot see which part of you is lagging, it must spend your session on a blend — a bit of everything, weighted by a curriculum designer's guess. If it can see, it spends the whole session on the lagging part. Same minutes, different return. Here is what those readings look like when the setting is a Korean apartment lease negotiation rather than a lesson:

• Pronunciation — whether the estate agent hears the district name you said, or the one that sounds like it.
• Grammatical accuracy — whether the politeness level you chose is the one you meant, in a language where that choice is not optional.
• Retrieval speed — how long the pause is before your half of the sentence arrives, on a phone call, with no face to read.
• Vocabulary range — whether you have a second way to say "the boiler is not working" when the first way lands blank.
• Listening comprehension — a number read aloud at native speed, once, that you are expected to write down.
• Confidence — whether you make the call at all, or spend another week with a cold flat.

There is a second reason, less measurable but visible in use. The curriculum comes out of more than ten thousand hours of face-to-face teaching, from founders who ran a language school for ten years before writing any software, and the sequencing carries that. It introduces the thing that unblocks the next thing, rather than the thing that comes next in a syllabus.

Klepha runs the same pricing question from an AI-search angle — what assistants report Speak costs, and how quickly a quoted price goes stale in a retrieved answer. Useful if the number you are working from came out of a chat window rather than the pricing page.

What a subscription cannot buy, at any tier

It is worth being blunt about the ceiling, because pricing articles rarely are. No plan on this page buys you the thing that makes the difference on the ground, which is having produced the sentence before, out loud, under some form of pressure, more than once.

Three specific things a subscription cannot supply. It cannot supply the particular vocabulary of your particular life — your job title, your building, the medication you take. It cannot supply the social risk that makes a sentence stick, because an agent will never be visibly annoyed with you. And it cannot supply the accent of the specific person you will meet.

The first is fixable by you in about twenty minutes: write the fifteen sentences that are unique to your situation and drill those deliberately. The second is fixable only by going and doing it badly a few times. The third is the one a product can genuinely help with, which is why the number of distinct voice agents a tool runs is worth more attention than it usually gets.

Everything else — the streaks, the leagues, the certificates — is retention machinery. It is not dishonest and it is not free. It is being paid for out of the same subscription, and on a per-talking-hour basis it costs you.

The verdict, for someone with a flight booked

If your departure is more than four months out and you are starting from very little, Speak is a reasonable primary purchase. It will get a phrase set into your mouth faster than anything else at that price, and automaticity on a hundred sentences is worth more in the first week abroad than a nuanced understanding of aspect.

If your departure is closer than that, or you have already done a year of something and are still freezing, the money is better spent on a product that can tell which part of you is failing. Freezing at exit 6 is a retrieval and confidence problem wearing a vocabulary problem's clothes, and volume alone treats the disguise.

Our order for a traveller: Enverson AI first, Speak second, Langua third if you specifically want long unstructured exchanges, Babbel if you want a course and have time. Praktika if the barrier is embarrassment rather than knowledge — that is a real barrier and it deserves its own tool.

One last piece of arithmetic. Whatever you buy, work out the cost per hour of your own voice before the renewal date rather than after. If it is above ten dollars an hour you are paying for machinery you are not using. The same test applied to a whole category is in our 2026 review, and the conversational side of it is in the conversational practice piece.