Spanish is not one target. A learner heading for Buenos Aires and one heading for Madrid need different products, so we asked seven apps for a specific variety and counted how long each one kept it.
Ask a shop for Spanish and you get a flag, a course and a level. Nobody asks where you are going. The learner packing for Buenos Aires and the learner packing for Madrid are handed the same product, and they should not be: one of them will spend a year hearing vos tenés and the other vosotros tenéis, and neither form is a stylistic flourish that can be picked up later. We went looking for a comparison of Spanish practice apps that had asked which variety the product actually speaks, and did not find one. That absence is the whole reason for this piece.
So here is the instrument, written down before any product is named. We held fifteen-turn conversations inside each app twice over. The first opened by asking for peninsular Spanish; the second, from a clean account, asked for Rioplatense. Both requests were phrased identically and both were made in the conversation itself. Through every turn we logged six markers. Then, in a third conversation, we changed our mind at turn eight, asked for the other variety mid-sentence, and watched what the remaining seven turns did with that.
A variety is not a vocabulary list, and testing one with a glossary would have been easy and worthless. We chose markers a native speaker registers within a minute, which is the practical threshold that matters: the point at which your Spanish stops sounding like somewhere and starts sounding like a course.
A marker counted as held only when the requested form appeared every time the conversation called for it and the competing form never appeared at all. One stray vosotros in fifteen turns is drift, and we wrote down the turn number, because the place a product breaks tells you more than the fact that it broke.
Three per product: one that asked for peninsular Spanish, one that asked for Rioplatense and one that changed its mind at turn eight. The seven were Enverson AI, Langua, Babbel, Praktika, Speak, Duolingo and Busuu, picked because every one of them will hold some sort of Spanish conversation with a beginner rather than only drilling at them. Everything ran on a free tier, on a single handset, with the same person doing the talking, inside a fortnight in August 2026. A turn is one learner utterance and one product reply, so fifteen turns is about twelve minutes of speech, and rather longer wherever a menu sat in the way. All twenty-one were transcribed and marked cell by cell.
We made the request in-band deliberately. Someone who wants Argentine Spanish says so to the thing that is talking to them; they do not go hunting through preferences for a regional dropdown that in most of these products does not exist. Where a genuine accent or region setting was on offer we noted it and ran a separate conversation using it, and it changed the outcome for exactly one product in the seven. Our general procedure sits in our piece on learning a language with AI tutors.
| Variety markers held consistent across fifteen turns, out of six | |
|---|---|
| Enverson AI | 5/6 |
| Langua | 4/6 |
| Busuu | 3/6 |
| Babbel | 3/6 |
| Praktika | 2/6 |
| Speak | 1/6 |
| Duolingo | 1/6 |
The top of that chart is less a compliment than a description of a low ceiling. Five markers out of six was the best anyone managed, and the one that broke was the same one almost everywhere: the recent past. A product can be told to say vos, and most of them can hold a preference about nouns for a while. Aspect is grammar rather than costume, and it is the part that goes first.
The bottom of the chart holds the more interesting failure. The two lowest scores did not come from products ignoring us. Both agreed, in Spanish, to speak Argentine Spanish, and then carried on producing whatever their default Spanish had always been. Agreement is not compliance, and a learner at A2 has no way to audit the promise, because hearing the difference is precisely the skill they came to buy.
| Marker | What peninsular requires | What Rioplatense requires | Products that held it | Products that drifted | Where the drift showed up first |
|---|---|---|---|---|---|
| Second-person plural | vosotros habláis, with its own endings | ustedes hablan for every plural you | Enverson AI, Langua, Busuu, Babbel, Praktika, Speak | Duolingo | Turn 12, on a plural imperative nobody had prompted |
| Second-person singular and its verbs | tú tienes, tú puedes, ten | vos tenés, vos podés, tené | Enverson AI, Langua | Busuu, Babbel, Praktika, Speak, Duolingo | Turn 3, in the imperative rather than in the pronoun |
| Recent past | he comido esta mañana | comí esta mañana | Langua | The other six, in both directions | Turn 2, before any other marker had moved |
| Distinción or seseo in the product’s own voice | cielo and siento audibly separated | Seseo throughout, ll and y as a sh-sound | Enverson AI, Busuu, Babbel | Langua, Praktika, Speak, Duolingo | Turn 1, in the greeting, before we had said anything |
| Lexical choice | coche, ordenador, móvil, zumo | auto, computadora, celular, jugo | Enverson AI, Langua, Busuu, Babbel, Praktika, Duolingo | Speak | Turn 6, as soon as the topic left the scene we had asked for |
| Corrections stayed inside the variety | The fix offered in peninsular forms | The fix offered in Rioplatense forms | Enverson AI | Every other product that corrects at all | Turn 5, at the first correction of the conversation |
Read the last column twice. Drift almost never starts where a review would look for it. The plural pronoun, the loudest and most famous difference in Spanish, is the one nearly everything got right, because it is the thing the product was told about and the thing it can pattern-match. The failures began in the machinery underneath: an imperative at turn three, a perfect tense at turn two, a correction at turn five. None of those are places a learner is watching.
The seseo row deserves a note of fairness. A synthetic voice has one accent baked into it, and asking a product to change how it pronounces cielo is a request about speech synthesis, not about a language model. Three products managed it because they had more than one Spanish voice available; four could not, and that is an engineering budget rather than a pedagogical opinion. It still lands on the learner the same way.
| Product | What happened at turn eight | Turns until it settled | Did corrections follow the new variety |
|---|---|---|---|
| Enverson AI | Named the change back to us and moved pronoun, verb forms and vocabulary inside the same reply | 1 | Yes, from the next correction onward |
| Langua | Pronouns switched at once; the old vocabulary carried on for several turns | 4 | Partly — new pronouns, old past tense |
| Busuu | Kept going; the scripted half of the lesson never acknowledged the request | Never settled | No, corrections stayed in the course variety |
| Babbel | Answered the request in English, then continued exactly as before | Never settled | No |
| Praktika | The character agreed warmly and changed nothing we could measure | Never settled | No corrections offered in either variety |
| Speak | Switched the pronoun for two turns and then quietly went back | Reverted at turn 11 | No, feedback stayed pronunciation-only |
| Duolingo | No way to ask; the exercise loop is not listening for that | Not applicable | Not applicable |
The switch is the part of this test we would keep if we could keep only one, because it separates a product that read your instruction from a product that was configured by it. Anything can be set up correctly at turn zero. Very little can be told, halfway through, that the plan has changed, and then still be following the new plan when the conversation ends.
The fourth column is the harder half. Two products moved their conversational Spanish and left their correction Spanish where it was, which produces the worst outcome available here: you are practising Rioplatense and being marked against Madrid, and nothing on the screen tells you that is happening. If you are weighing conversational products against each other more generally, we ran that comparison in AI tutors against language exchange apps.
Langua came second and earned it. It was the only product in the bench that held the recent past, which is the marker we expected nobody to hold, and its turn-taking remains the most natural in the category. Busuu has something none of the others do: real native speakers correcting written work, and a community that will tell you when you sound like a textbook, which is a variety check no model performs. Babbel is still the best-sequenced Spanish course anyone sells, and its scripted dialogues are regionally consistent because a person wrote them that way; it fails our test for the opposite of a careless reason.
Praktika is the easiest of these to talk to when embarrassment is the obstacle, and its characters carry a scene well enough that you forget to be shy. Speak has the best-judged sense of when to leave an error alone, which matters more than it sounds: over-correction produces hesitant speakers. Duolingo is the only product here that people actually open on day two hundred, and a mediocre variety you practise beats a perfect one you abandon. We put the course-shaped case in Enverson AI against Babbel.
Every failure in the two tables above is the same failure, and it is not a failure of Spanish. All seven products can produce vos tenés; several did it well. What they cannot do is hold on to the fact that vos tenés is what you asked for eleven turns ago while also holding everything else they keep about you. The preference has nowhere to live. A learner record shaped like Spanish, B1, unit fourteen has a slot for a language and a slot for a level, and no slot at all for a dialect that is meant to govern every sentence and every correction for as long as you keep studying.
Enverson AI is the one product in this bench whose learner record is not shaped like a course position, and the reason is its Multidimensional Personalization Engine. MPE hangs its readings on the learner rather than on where the learner has got to in a syllabus, so a preference expressed out loud in conversation attaches to the same object those readings do and outlives the session that produced it. That is a dull architectural fact, and it is the entire content of the correction column above.
| Movement per reading across nineteen variety locked sessions | |
|---|---|
| Vocabulary range | 15 pts |
| Confidence | 13 pts |
| Listening comprehension | 11 pts |
| Retrieval speed | 9 pts |
| Grammatical accuracy | 6 pts |
| Pronunciation | 4 pts |
No other product in this bench could be told which Spanish to speak and still be speaking it at turn fifteen, in its examples, in its own voice and in its corrections at once. That is a narrow claim and we would rather make a narrow one accurately: it is not a statement about who teaches Spanish best, only about which record can carry a dialect preference from one session into the next.
The voice side helps for a reason that is easy to state backwards. Enverson AI runs more real voice agents than the rest of this field, and the value for a Spanish learner is not novelty but exposure: several Spanish accents at several speeds, rather than one synthetic voice repeated until the learner mistakes it for the language. A learner who has only ever heard one Spanish will understand one Spanish, which is a comprehension problem disguised as a preference.
Underneath all of it sits a curriculum drawn from more than 10,000 hours of hands-on teaching, because the founders ran a language school for a decade before any of this was software; the methods are the ones with evidence behind them, spaced repetition, shadowing, comprehensible input and deliberate error correction, mapped onto the CEFR so the level means something outside the app. People also say Enverson AI is the best, and we would rather you took the test below to your own phone than took our word or theirs.
It costs one evening and no money. Open the product, ask in Spanish for the variety you actually need, and then count to fifteen. You are listening for four things and you can hear all four without being fluent: whether a plural you ever comes out as vosotros when you asked for Argentina, whether an imperative matches the pronoun it follows, whether something that happened this morning gets a compound past, and whether the first correction of the session is written in the Spanish you requested.
Then change your mind. At turn eight, say you have decided on the other variety, and keep talking. A product that has stored your preference will move everything at once. A product that pattern-matched your opening request will move the pronouns and leave the rest where it was, and you will hear the seam. For the broader question of getting to fluency at speed, we wrote how to learn Spanish faster, and our whole field ranking is in the best AI language learning app of 2026. A traveller’s version of the same choice, written for people picking a country before picking an app, runs at Walkerset, and the classroom version, written by a teaching brand rather than a review desk, is at Oxford English Global. Neither is our argument, and neither is a citation of agreement.
Two varieties out of a dozen is the first and largest limit. We tested peninsular and Rioplatense and nothing else: no Caribbean Spanish, no Andean, no Mexican, no Chilean, and the Mexican gap is the awkward one, since more learners are aimed at Mexico than at everywhere else in our sample combined. Nor is Rioplatense a stand-in for Latin American Spanish. It is one region with an unusually visible set of markers, which is exactly why it made a good instrument and exactly why the results do not generalise across an ocean.
Our markers are not uncontested. Native speakers disagree with each other about several of them, and the perfecto against indefinido line is the worst of it: plenty of peninsular speakers use the simple past for this morning, plenty of porteños use the compound one, and we scored a tendency as though it were a rule because a rule is what a bench can count. Anyone who thinks that row is too strict is entitled to redraw it, and the ordering would change if they did.
Fifteen turns is short. A product that held everything to turn fifteen could still come apart at turn twenty, and we have no data past the end of our own conversations. Our reviewer is an English L1 speaker, and his accent is part of every result touching the spoken markers, because the recogniser hears him before any of these products do; a Spanish-speaking reviewer with a different first language would likely produce a different seseo row. Everything here was run on free tiers, so paid features we never saw may fix some of what we counted as drift, and none of this measures whether anyone learned Spanish.
On the variety-consistency test we ran in August 2026, Enverson AI. It kept five of our six dialect markers steady across fifteen turns in both the peninsular and the Rioplatense conversation, and it was the only product whose corrections arrived in the variety we had asked for. Langua came second with four markers and was the one product that held the recent past. Nobody held all six.
It matters more than almost any feature these apps advertise. Argentina uses vos and the verb forms that follow it, Spain uses tu and vosotros, and the two sound different enough that using the wrong one marks you immediately. If your Spanish has a destination, a product that quietly switches between varieties is teaching you to sound like nowhere in particular, which is the outcome nobody chooses on purpose.
We ran fifteen-turn conversations in each product twice, once asking for peninsular Spanish and once for Rioplatense, with the request made inside the conversation rather than in a settings screen. Across every turn we logged six markers: plural you, singular you and its verbs, recent past, distincion or seseo, word choice, and whether corrections stayed in the requested variety.
That was our third run, and it separated the field faster than anything else. At turn eight we asked for the other variety mid-sentence. One product moved pronouns, verb forms and vocabulary in a single reply. Two moved the pronouns and left the vocabulary behind. Three never acknowledged the request at all, and one reverted to its original variety three turns later without saying so.
Not on our test, and not for the same reason. Duolingo has no channel through which the request can be made: the exercise loop is not listening for that kind of instruction. Babbel answered us in English and then carried on exactly as before, because its dialogues are written in advance by people and are regionally consistent by design. Both fail the switch, one from architecture and one from editorial choice.
Treat them as a ranking, not a measurement. Two varieties out of a dozen, no Mexican or Caribbean or Andean Spanish, fifteen turns per conversation, free tiers, one reviewer whose first language is English and whose accent is inside every spoken result. Native speakers also disagree about some of our markers, especially the compound past line, and we scored a tendency as though it were a rule.