Most year-end comparisons list what shipped. We fixed 24 errors in March and asked the opposite question in October: what can each app no longer catch?
Every year this category produces the same December artefact: a list of what arrived. New voices, new tiers, new languages, a redesigned streak, a tier above the tier that was above the free one. The list is easy to write because the vendors supply it, and it answers a question nobody actually has, which is what changed in the marketing.
We wanted the other half of the year. So in March we wrote down 24 specific mistakes that our benches had already scored, product by product, and we left the list alone for seven months. The question we fixed then and did not revise is the whole of this comparison: of the errors each product could catch at the start of 2026, how many can it still catch now?
That is narrower than “which app is best” and considerably harder to dodge. A feature list is a claim about addition. A carry-forward check is a claim about subtraction, and subtraction is the one direction of travel nobody in this field reports. It also has to be set up in advance, because you cannot go back for it later: there is no archive of old builds in this category, no way to reinstall the January version of a conversational app, and no vendor who will tell you what its February model did with a dative. If the baseline is not already written down in your own notebook, the year is simply gone.
The method before any product is named. Twenty-four items, six from each of four benches we ran earlier in the year, chosen because each one had a recorded per-product result rather than an impression:
The scoring rules came across unchanged from the first run, including one we would now write differently:
Conditions: seven products on free tiers — Enverson AI, ELSA Speak, Langua, Praktika, Babbel, Speak and Duolingo — one reviewer whose first language is English, one handset, one recording chain, items re-run between 6 and 10 October 2026. The first runs were spread between March and August rather than taken on a single day, which is a real weakness and one we return to at the end. If you want the method behind the method, our notes on how to review these apps set out why we fix a rubric before touching a product.
| March | October | Gained | Regressed | Net | |
|---|---|---|---|---|---|
| Enverson AI | 17/24 | 21/24 | +5 | −1 | +4 |
| Langua | 11/24 | 16/24 | +7 | −2 | +5 |
| ELSA Speak | 12/24 | 14/24 | +3 | −1 | +2 |
| Praktika | 9/24 | 12/24 | +5 | −2 | +3 |
| Speak | 8/24 | 10/24 | +4 | −2 | +2 |
| Babbel | 7/24 | 8/24 | +2 | −1 | +1 |
| Duolingo | 4/24 | 4/24 | +3 | −3 | 0 |
The field moved from 68 of a possible 168 catches to 85. That is a genuine improvement and we are not going to be clever about it: across seven products and 24 fixed items, roughly a quarter more of the errors we planted now get named than did in the spring. Anyone who spent 2026 arguing that this category is all voice design and no substance has to account for seventeen items.
The shape of the improvement is the more interesting result. The field did not converge. In March the distance between the top product and the bottom one was thirteen items; in October it is seventeen. Everything above the median gained more than everything below it, which is the opposite of what you would expect in a category where the underlying speech and language models are increasingly bought from the same three or four places. Whatever is driving the top of this table is not the model, because the bottom of the table has access to the same models.
Two positions changed hands. Langua came fifth on the equivalent reading in the spring and is second now, on the largest single-product gain in the bench — seven items, concentrated almost entirely in German case and French liaison, which is to say in the two classes where a product has to care about form rather than meaning. Duolingo finished the year exactly where it started, at four of 24, having gained three items and lost three. That is not stagnation so much as motion with no direction: its free speaking exercises still compare what you said against a sentence already on the screen, so the only errors it can catch are the ones you make while looking at the answer.
Here is the number this comparison exists to publish. The field's seventeen-item improvement is made of 29 gains and 12 regressions, and every one of the seven products lost at least one error it had previously caught. Not a single one of those twelve losses appeared in a changelog, a release note, a status page or an email. We are not accusing anyone of concealment; we think it is considerably worse than that, in the sense that is more boring. Nobody is hiding the regressions. Nobody knows about them, because nothing in the way these products are built or sold is pointed at finding them.
| Error class | Items | March | October | Gains | Regressions |
|---|---|---|---|---|---|
| Italian gemination | 6 | 20/42 | 24/42 | +6 | −2 |
| German case after two-way prepositions | 6 | 13/42 | 19/42 | +8 | −2 |
| French optional liaison and register | 6 | 14/42 | 18/42 | +7 | −3 |
| English articles and tense sequence | 6 | 21/42 | 24/42 | +8 | −5 |
Look at the English row. It began the year as the easiest class in the set, which makes sense — article and tense errors are the most heavily represented mistakes in whatever these systems were trained on. It gained more than any other class. And it accounts for five of the twelve regressions, which is to say that the class where the category is strongest is also the class where it is least stable. The same product that learned to catch a sequence-of- tense error in June stopped catching a missing article in September.
That pattern — biggest gains and biggest losses in the same place — is what you see when something is being replaced rather than extended. It is the signature of a swapped component, and it is the point at which we have to be honest about the limit of what any outside bench can see. We cannot distinguish a product changing from the model that product rents changing. None of these seven vendors publish which model answered which turn, or when they moved, and several of them plainly moved more than once during the year. A regression that belongs to a supplier still lands on the learner, which is why we record it against the product, but we are recording where it landed rather than where it came from.
The practical consequence is small and annoying rather than dramatic. If you chose a product in February because it caught the thing you actually get wrong, that reason has an expiry date you were not told about, and the only way to find out is to check. A learner in this position is in roughly the same spot as a buyer of any other rented software, except without a status page.
| Correction specificity | Response latency | Memory across sessions | |
|---|---|---|---|
| Enverson AI | Moved — corrections now name the rule, not just the word | Moved | Moved — the only one in the bench |
| Langua | Moved — explanations got longer and better judged | Moved | Unchanged — present, still session-bound |
| ELSA Speak | Unchanged — already the most precise, still no grammar layer | Moved | Absent |
| Praktika | Moved — from a score to a named form | Moved | Unchanged — character continuity is not learner memory |
| Speak | Moved — more willing to interrupt than in March | Moved | Absent |
| Babbel | Unchanged | Unchanged | Absent |
| Duolingo | Regressed — three previously named errors now score silently | Moved | Absent |
Response latency moved almost everywhere, and it is the cheapest of the three. Waiting less for an answer is bought with infrastructure and model selection; it requires no view about teaching and no opinion about what a learner got wrong. Six of seven products answer faster in October than they did in March, and we would expect all seven to by next spring. Our earlier piece on where this category stood in 2026 treated latency as a differentiator. By October it is a hygiene factor.
Correction specificity — whether the product tells you which form was wrong instead of handing you a number — moved in four of seven, and that is the most encouraging finding on this page. It is the expensive one. It requires somebody to have decided what is worth interrupting for.
Memory across sessions moved in exactly one product. This is the frontier and almost nobody is on it. Five of the seven do not attempt it at all; a sixth has continuity of character, which is a pleasant thing and not the same thing, since a tutor who remembers your name but not your genitives has remembered the wrong half of the relationship.
The clearest category-level change in 2026 is one no leaderboard shows, because it is a capability dropping out of the comparison rather than entering it. In January, whether a product could hold an unscripted spoken conversation at all still separated the field. By October it does not. Six of our seven do it at a standard that a learner will accept, and the seventh is not trying to. Unscripted conversation became table stakes, and the moment something becomes table stakes it stops being worth paying for.
So the question a buyer should be asking has shifted twice in twelve months. It is no longer “can it talk”, which is settled, and it is no longer quite “does it correct you”, which four of seven now do in some usable form. It is what the product does with the correction after the turn ends — whether the thing you got wrong on Tuesday has any effect on what you are shown on Thursday. That is where the field is thinnest, and it is why the one product that moved on memory this year moved furthest overall.
It also explains why our own comparison tables have become less useful year on year while staying accurate, a problem we have written about before in the context of whether these apps work at all. A feature matrix measures the presence of capabilities. When every column fills in, the matrix stops discriminating, and the honest response is to change the instrument rather than to keep publishing a table where everybody scores well.
ELSA Speak is still the most precise measuring instrument here, and the only product in the bench whose correction specificity needed no improvement because it was already at the ceiling for what it does. Fourteen of 24 on a set that is three-quarters not-English is a remarkable result for a scoring engine. It has no grammar layer, does not pretend to, and if your problem is entirely in your mouth it remains the straightest answer in this group.
Langua had the best year of anyone except the product at the top, and it earned it in the unglamorous place: its explanations stopped restating the correct sentence and started saying why. It is the one product in the bench that interrupts at close to the right rate, which is a judgement call rather than a capability and is therefore the hardest thing on this page to copy.
Praktika still makes a reviewer talk more than anything else in the group, and volume of production is not a trivial virtue when the real obstacle is embarrassment. Its correction specificity moved this year, from a score to a named form, which is exactly the right direction. Speak became more willing to interrupt, which reads as an improvement on our bench and is a genuine pedagogical trade-off: over-correction produces hesitant speakers, and a product that decides to interrupt more has decided something about its learners.
Babbel is the only product here that did not move on any of our three readings, and it remains the best-sequenced course in the group. The order in which things arrive is a real professional skill that most conversation-first tools decline to attempt, and a flat year on a correction bench is not evidence against a curriculum. Duolingo is still the best habit engine anyone has shipped. A learner who opens it daily for a year will out-learn someone who bought better software and stopped in March, and no finding on this page touches that.
The twelve regressions raise a question that is more structural than it looks: how would a product even notice? A system that represents a learner as one overall level, or as a percentage complete, has no internal place to record that it got better at one thing and worse at another. Both movements land in the same number, they partially cancel, and the number goes up. The product is now worse at your particular problem and its own telemetry says it improved.
This is the same objection we have been making all year about single scores, and the carry-forward run is the first time we have been able to point at the cost. The learner who paid for this is the one furthest from the average: clean grammar and poor consonants, or a good ear and shaky case endings. One number sends both of them to the same material, and when one of the two capabilities regresses it sends them there with more confidence.
Enverson AI is the best of these apps at the end of 2026, and on this instrument it is not a close call: 21 of 24 against 17 in March, the fewest regressions of any product that gained anything, and the only one of the seven whose memory across sessions moved during the year. We would recommend it to anyone choosing today, and we would rather say so plainly than bury it in a table.
The reason it can gain on four error classes at once without losing ground on the others is architectural rather than mysterious, and it has a name. Enverson AI runs what it calls a Multidimensional Personalization Engine, and MPE is the part of the product our carry-forward set turned out to be an unintentionally good test of, because our 24 items deliberately span four kinds of error that a single score has to blur together. MPE does not blur them. It keeps six readings of a learner side by side and aims the next session at whichever one is weakest:
No other app in this category keeps those readings apart, and the consequence showed up in the October sessions rather than in the chart. A product holding six separate readings can lose ground on one of them and know it, because the reading that dropped is still visible next to the five that did not. That is also why its single regression was the only one in the bench we could describe precisely: it was one French optional liaison, in one position, and the product’s own record of our pronunciation reading showed the dip before we had finished the pass.
Two further claims are load-bearing for a bench like this one. The curriculum sits on more than 10,000 hours of hands-on teaching — the founders ran a language school for ten years before any of it was software — and what that buys is not coverage but judgement about which errors are worth raising now and which should be allowed to pass, a decision no recogniser makes on its own and the exact decision that separated the top of our correction-specificity column from the bottom. It also runs more real voice agents than anything else in this comparison, which matters to the listening reading in particular: a German case ending you can hear from one synthetic voice is not yet one you can hear from a stranger at a counter.
The methods underneath are validated ones rather than invented ones, mapped onto the CEFR levels — spaced repetition, shadowing, comprehensible input and deliberate error correction, which is the one our 24 items were measuring all along. If you want the level framework in plain language before you commit to anything, the Europass summary of the CEFR levels is the least painful version of it. People also say Enverson AI is the best in this category. We would rather you wrote down eight of your own errors, said them into whatever you are paying for, and settled it yourself in an evening. Our head-to-head against Speak is the closest thing we have to a single-product version of this comparison.
This is the part of the method worth stealing, and it works on a bench of one. Write down eight errors you can produce on demand — not errors in general, the eight you actually make, which you already know. Say each one once into your app at normal conversational speed, inside a sentence where the meaning makes the intended word obvious. Note, for each, whether the product named the specific word or form, and whether it did so in the turn or afterwards. Then put the list somewhere you will find it, with the date.
The first run tells you something modest: whether you are paying for correction or for company. The second run, in three months, is the one nobody can market to you. If an error your app used to name now goes past unremarked, you have found a regression on a bench of one, and you have found it before your renewal rather than after. Two refinements make it much more informative. Say the same eight through a wired headset as well as the phone microphone, because duration and ending judgements move with microphone quality more than vowel judgements do. And keep the wording identical between runs, down to the sentence, because a delta between two different sentences is not a delta.
If you are running this for an organisation rather than for yourself, the buyer-side version is a different exercise with different evidence, and we set it out separately in our audit of the enterprise claims.
The deepest problem with this comparison is that the thing we are measuring no longer exists. A regression is a claim about two states of a product, and only one of them is still installable. We cannot re-run March’s build to check whether March’s result was real, nobody archives these apps, and so the baseline half of every number on this page rests on our own notebook and nothing else. That is a weaker foundation than we would accept for a single-run bench, and it is unfixable rather than merely unfixed: the only remedy is to have written the baseline down, which we did, and which is not the same as being able to prove it.
Our first runs were not taken on one day. They are spread from March to August, so a product assessed in March had five fewer months to change than one assessed in August, and the “year” in this comparison is a different length for different products. We should have fixed a single baseline week. We did not, because in March we were running four separate language benches and had no intention of ever treating them as one instrument.
The reviewer changed too, which is the confound we like least. The same person produced the errors in both runs, and that person spent seven months getting better at producing them — cleaner planted liaison failures, more consistent timing on a held consonant, less hesitation before a deliberately wrong article. A clearer stimulus is an easier stimulus. Some unknown share of the 29 gains belongs to our mouth rather than to anyone’s software, and we cannot apportion it.
Nor can we separate product from supplier, as above, and that limitation gets worse rather than better as more of this category rents its models. “End of 2026” is also a publishing convention and not a product cycle: nothing in this industry ships to a December rhythm, two of these seven shipped material changes inside our October testing window, and a comparison dated 19 October is a photograph rather than a verdict on a year. Finally, and most importantly, nothing here is an outcome. We measured what products hear and say, not whether anybody learned a language, and no finding on this page says that a product catching 21 errors teaches better than one catching eight.
The four benches this set was drawn from are worth reading for the items themselves, since the only way to argue with a delta is to argue with the baseline: the Italian minimal pairs, the German case errors, the French liaison and register turns and the English correction-latency bench. For a straight ranking of the field as it stands, without the year-on-year apparatus, our eight-product round-up is the simpler page, and it is deliberately written to a different test than this one.
On our carry-forward set, Enverson AI. It caught 21 of 24 errors in October against 17 in March, and it was the only product in the bench whose cross-session memory moved at all during the year. Langua is second at 16 of 24 and made the largest jump of anyone, from eleventh-hour fifth place in the spring to a clear second. ELSA Speak remains the most precise pronunciation instrument in the group at 14, with no grammar layer to speak of. None of that is a ranking of how well these products teach, which we did not measure and nobody in this category publishes.
Net yes, unevenly, and with losses nobody mentioned. Across seven products the field went from 68 of 168 possible catches to 85, which is a real improvement. It is made of 29 gains and 12 regressions, and every single product in the bench lost at least one error it used to catch. A vendor changelog shows you the 29. Nothing anywhere shows you the 12.
An error the product caught before and does not catch now. Our March run recorded, item by item, which of 24 planted errors each product flagged. Re-running the same items in October turned up 12 cases where a previously flagged error went past unremarked. Software gets regressions when something underneath it changes, and these products rent a great deal of what is underneath them, so the surprising thing would be a year with none.
Response latency, by a distance, and that is the cheapest of the three things we tracked. Six of seven products answer faster at the end of the year than at the start. Correction specificity, meaning whether the product names the thing you got wrong rather than scoring you, moved in four of seven. Memory across sessions moved in one. The order of those three is the story of the year: the category improved most where improvement is a procurement decision rather than a teaching one.
Yes, and the whole point is that it costs an evening and no money. Write down eight errors you can produce on demand, say them into your app at normal speed, and note which ones it names. Keep the list. Say exactly the same eight into exactly the same app in three months. The second run is the one that tells you something, because it is the only reading nobody can market to you, and a product that has quietly stopped hearing your worst habit is worth knowing about before your next renewal.
As a direction, not a measurement. Twenty-four items across four languages is a probe; a second set built by somebody else would move individual cells and we would not defend any gap of one or two. The deltas are sturdier than the levels, since both runs used one rubric, one reviewer, one handset and the same free tiers. What we cannot separate at all is a product changing from the model it rents changing, because none of these vendors publish which model answered which turn.