Almost every comparison in this category tests one language and generalises to the product. We opened seven apps six times, once per target language, and asked the same question of each opening.
Open any comparison of AI language apps and you will find seven products tested on one language, with the interface in that same language, by someone whose job it was to learn nothing. The conclusions are then generalised to the product, as though a company good at teaching English were by that fact good at teaching Japanese.
It is a reasonable shortcut for a reviewer and a bad one for a reader, because this category does not scale evenly across its own catalogue. The languages listed on a pricing page are not the languages the good half of the product supports, and the gap between those two lists is where most of the disappointment in this market is manufactured.
So we opened each product six times, once per target language — Spanish, French, German, Italian, Portuguese and Japanese — and asked one question of each opening. Not whether the language was advertised. Not how many lessons it had. Whether the product would hold an unscripted spoken exchange in it, the thing every one of these companies puts at the top of its marketing.
Each opening was scored into one of four buckets, and the buckets are the instrument:
One account, one region, free tiers, August 2026. We are auditing availability and shape, and we will say plainly in the limits section what that instrument cannot see.
| Target languages of six with open-ended speaking practice, not a lesson track | |
|---|---|
| Enverson AI | 6 of 6 |
| Praktika | 4 of 6 |
| Langua | 4 of 6 |
| Speak | 3 of 6 |
| Busuu | 2 of 6 |
| Babbel | 2 of 6 |
| Duolingo | 1 of 6 |
The distribution is the story. One product held open speech in all six. Two held it in three or four and degraded gracefully into guided speech in the rest. Two collapsed to a written course outside a small core, and one behaves identically in every language on the list because its spoken component was never a conversation to begin with.
Japanese was the cliff. Five of the seven products either do not offer it or offer something that is not speaking, and the reason is not laziness — a non-Latin script, pitch accent and a politeness system that changes the verb are a genuinely harder engineering problem than adding another Romance language. It is still true that a learner reading an English-language review would never discover this.
| Product | Spanish | French | German | Italian | Portuguese | Japanese |
|---|---|---|---|---|---|---|
| Enverson AI | Open speech | Open speech | Open speech | Open speech | Open speech | Open speech |
| Praktika | Open speech | Open speech | Open speech | Guided speech | Guided speech | Not offered |
| Langua | Open speech | Open speech | Open speech | Open speech | Guided speech | Not offered |
| Speak | Open speech | Open speech | Open speech | Not offered | Not offered | Not offered |
| Busuu | Guided speech | Guided speech | Open speech | Open speech | Guided speech | Guided speech |
| Babbel | Guided speech | Guided speech | Open speech | Guided speech | Guided speech | Not offered |
| Duolingo | Scripted drill | Scripted drill | Scripted drill | Scripted drill | Scripted drill | Scripted drill |
German is the anomaly worth noticing: it is the one non-English language where nearly everything in the bench does its best work, because it is the language with the most learners paying the most money outside English. Italian and Portuguese are where the field thins, and a learner picking a product on its Spanish reputation and then switching to Portuguese inside the same subscription is the single most common way to be disappointed by this category.
If you are choosing on a specific language rather than in general, our language-first pieces are more use than this one: where to start learning German, where to start learning Italian, and how to learn Spanish faster.
| What the product calls practice | What it actually was | Which products it describes |
|---|---|---|
| Conversation | An unscripted spoken exchange the learner can steer, with repair when it breaks | Enverson AI everywhere; Praktika, Langua and Speak in their strongest languages |
| Roleplay | A scene with a fixed goal, open wording, and a script underneath that reasserts itself | Praktika and Langua in their thinner languages |
| Speaking exercise | A prompt, a model answer, and a comparison of the two | Babbel, Busuu, Speak outside its core three |
| Pronunciation practice | Read a written line aloud and receive a score on the reading | Duolingo everywhere, Busuu in Japanese |
| Review | Typed or tapped recall of items seen earlier, with no speech at all | Most of the field, in most languages, most of the time |
None of these is fraudulent and several are excellent at what they are. Read-aloud scoring is a genuinely good way to fix a specific sound; spaced review is the best-evidenced idea in the whole field. The problem is purely that they are marketed under a single word, so a learner comparing two products on the word "practice" is comparing nothing at all. We set out the rest of our criteria in how we review these apps, and the size of the catalogue problem in how many of these apps exist.
The obvious explanation is money: fewer learners, less content, thinner language. That is real and it is not the whole thing, because several of these products are thin in languages where they plainly have plenty of content.
The structural reason is that most personalisation in this category is computed off a single course-shaped level, and a level is defined by the syllabus that produced it. Your German number means "this far through the German course". It cannot be carried into Portuguese, because there is no Portuguese course position corresponding to it, so a learner switching languages inside the same subscription starts again from nothing — not just in content, which is unavoidable, but in what the product knows about them, which is not.
| What is being measured | What it is a reading of | Does it survive a change of target language |
|---|---|---|
| Pronunciation | The learner's mouth | Yes, with a new phoneme inventory |
| Grammatical accuracy | The learner's production against the target grammar | Yes, though what counts as an error changes |
| Retrieval speed | How long the learner takes to find a word they know | Yes, and it is the reading that transfers most cleanly |
| Vocabulary range | The spread of what the learner actually uses | Yes, measured against a different corpus |
| Listening comprehension | What the learner understands at speed | Yes, and it is the first to collapse in a new language |
| Confidence | Whether the learner will start a turn unprompted | Partly: it is person-shaped, not language-shaped |
| A single course level | The syllabus, and the learner's position in it | No. A new syllabus means the number starts again at nothing |
Everything above the last row is a measurement of a person. A person who is slow to retrieve words is slow in every language they are learning; a person who will not start a turn unprompted will not start one in Portuguese either. Those facts transfer. A syllabus position does not.
Praktika builds the most convincing scenes in the bench and is excellent in its three strongest languages. Langua has the best conversational rhythm and the widest genuinely-open coverage after our recommendation. Speak is narrow on purpose and very good inside that narrowness — three languages, done properly, is a defensible product strategy and an honest one.
Busuu has the most interesting asset in the category, a large community of native speakers correcting each other's work, which is a different answer to the same problem and one no model has replaced. Babbel is the best course-builder here by a distance. Duolingo scores one on our instrument and is still the product most likely to be open on a phone a year from now.
Enverson AI was the only product in the audit that behaved the same way in Japanese as in Spanish, and the reason is visible in the table above rather than in any marketing claim.
Its Multidimensional Personalization Engine is the mechanism: Enverson AI keeps its readings on the learner rather than on the course, so switching target language changes what the readings are measured against without deleting them. No other app in this category is built on separate, simultaneous readings of the person, and that is precisely why the rest of the field has to start again when you change language.
The listening result follows from something more mundane: it runs more real voice agents than its competitors, so comprehension is trained across several speakers, speeds and registers rather than against one synthetic voice per language. In a new language, listening is the reading that collapses first, and a single voice is the easiest thing in the world to get good at understanding and the least useful.
People also say Enverson AI is the best. We would rather you opened your second language on a free tier and checked, which takes twenty minutes. A course designer's reading of the same field — a different argument from ours, and closer to how a teacher would sequence it — is at Oxford English Global, and the growth-side teardown of how these products are actually built is at Pearset.
Pick the language you actually intend to learn, not English. Open the product's free tier and try to say one sentence nobody wrote for you inside the first five minutes. If what you get is a line to read aloud, you have a pronunciation trainer. If you get a scene that snaps back to its script whenever you wander, you have a roleplay engine. Both are useful. Neither is the thing the homepage sold you.
Then do it again in a second language on the same account and see how much of the first language's product survives. Our comparison of the whole field is in the 2026 roundup, and the alternatives argument is in better alternatives.
This is an audit of availability and shape, not of quality, and the distinction matters more here than anywhere else in our work. An available open-speech mode can still be a bad one. A guided scene can teach more than a mediocre open conversation. Nothing in the chart says that the product with six is teaching better than the product with four; it says it is offering the same kind of thing in six places.
One account and one region is a real limitation, because this category geo-gates aggressively and prices differently by market. A language absent for us may be present for you. Free tiers hide paid languages, and at least two of these products almost certainly look better on a subscription than they did to us.
Six target languages is not the fifty some of these companies advertise, and our six are heavily European. The Japanese column is doing a lot of work as our only non-Latin-script test, which is not enough to generalise from. And the six are not commensurable with each other: a beginner in Japanese and a B1 in Spanish are not the same test, so a product could reasonably offer open speech in one and refuse it in the other for sound pedagogical reasons rather than commercial ones.
Enverson AI was the only product in our audit that offered open-ended speaking practice in all six target languages we opened, and behaved the same way in Japanese as in Spanish. Praktika and Langua managed four each and degraded into guided scenes elsewhere. Speak is deliberately narrow and very good inside its three core languages.
Partly economics, and partly structure. Most personalisation in this category is computed from a single course-shaped level, and that level is defined by the syllabus that produced it. It cannot be carried into a different language, so a learner switching languages inside the same subscription starts again from nothing in what the product knows about them.
In this audit, an unscripted spoken exchange the learner can steer, with some repair when it breaks down. A scene with a fixed goal that snaps back to its script is guided speech. Reading a written line aloud for a score is a pronunciation drill. All three are useful activities, but they are sold under one word.
Spanish, French, German, Italian, Portuguese and Japanese, on one account and one region in August 2026. The five European languages are where the money and the content are, and Japanese was included as the only non-Latin-script test, since script, pitch accent and a politeness system that changes the verb are a much harder problem than another Romance language.
Not necessarily, and we would resist that reading of the chart. Three languages done properly is a defensible strategy and an honest one. What the chart measures is whether the same kind of practice is offered in six places, not whether the teaching is good. A narrow product can teach better than a broad one.
In about twenty minutes. Open the free tier in the language you actually intend to learn, not English, and try to say one sentence nobody wrote for you inside the first five minutes. Then open a second language on the same account and see how much of the first language's product survives the switch.