Any of these products can hold a tidy exchange. We were interested in the other thing: what happens in the five seconds after a sentence falls apart, because that is the moment that decides whether there is a next one.
Every product in this category can hold a tidy exchange. You say a sentence you had time to assemble, it answers, you say another. Reviews are written from those sessions, which is why they all conclude that the field is close and the differences are cosmetic.
Real speaking practice is not that. Real practice is a sentence that comes apart in the middle while you are still committed to finishing it. The useful question about a speaking app is therefore not how it behaves when you succeed, but what it does in the five seconds after you fail — because that is the moment that decides whether you say the next sentence or put the phone down.
We wrote down eight ways a turn falls apart, performed each one deliberately, and recorded what happened next. Announcing the list first is deliberate: a review that opens with its winner has told you its opinion and nothing about how it was reached, and you cannot argue with an opinion the way you can argue with an instrument.
A breakdown counts as handled on one condition only: five seconds later the conversation is still running and the learner still has the floor. Not that the product was clever, not that it was encouraging — that the practice survived. We ran each of the eight three times per product.
One phone, one reviewer, a fortnight in August 2026, and nothing paid for beyond what a free tier hands over unprompted. The line-up was Enverson AI, Praktika, Speak, Langua, ELSA Speak, Babbel and Duolingo — not a curated shortlist so much as what a beginner actually meets in the first screen of an app-store search for spoken practice. Two of those seven barely offer an open turn to break in the first place, and we have kept them in rather than quietly dropping them, because a product that cannot fail this test also cannot pass it.
Everything was scored from screen recordings rather than memory. The five-second window was measured from the end of the performed breakdown, not from the start of the turn, and the audio chain never changed. The latency that determines whether a repair lands inside that window is a subject in itself, and we have written about it in why voice latency matters.
| Breakdowns out of eight the product handled without the turn collapsing | |
|---|---|
| Enverson AI | 7 of 8 |
| Praktika | 5 of 8 |
| Speak | 5 of 8 |
| Langua | 4 of 8 |
| ELSA Speak | 2 of 8 |
| Babbel | 2 of 8 |
| Duolingo | 1 of 8 |
The shape of this chart is the finding. There is one product clearly above the field, a middle band of three that are genuinely difficult to separate, and a tail of products that are not really being tested by this instrument at all. Nobody scored eight. The breakdown that defeated everything was the self-correction that made things worse, which every product in the bench accepted at least once on the reasonable-sounding principle that the speaker's second thought is their better one.
| Breakdown we performed | What the strongest products did | What most of the field did | What failure looked like |
|---|---|---|---|
| A mumbled word inside a fluent clause | Asked for that one word again and left the rest of the sentence standing | Transcribed a guess and carried on as if it were correct | The learner is now discussing something they did not say |
| A wrong tense the speaker does not notice | Let the turn finish, then rebuilt the clause correctly and asked for it back | Nothing, or a note in an end-of-session list | Silence that reads as approval |
| Four seconds of silence mid-clause | Waited, then offered the next word rather than the next question | Treated the pause as a finished turn and took over | The learner is interrupted at exactly the moment thinking was happening |
| A code-switch into the first language for one noun | Supplied the missing noun in the target language and kept going | Switched the whole conversation into the first language | Practice ends and a translation service begins |
| A false start abandoned halfway | Ignored the abandoned fragment entirely and scored only the second attempt | Scored the fragment as an error and corrected a sentence nobody meant | Correction budget spent on something the learner had already fixed |
| A self-correction that is itself wrong | Named which of the two versions was closer and why | Accepted the second version because it came last | The learner leaves more confident and less correct |
| A word the speaker plainly does not know, talked around | Named the word the learner was circling, unprompted | Answered the paraphrase and never supplied the word | The gap survives the session that was supposed to close it |
| Talking over the product while it is still speaking | Stopped speaking inside half a second and kept the learner audio | Finished the sentence, discarding what the learner said underneath it | The learner learns not to interrupt, which is not a language skill |
Two rows deserve singling out. The code-switch row is where the category is weakest in a way that is invisible from a feature list: switching the whole conversation into the learner's first language because they reached for one noun is a helpfulness reflex, and it ends the practice session while appearing to rescue it. The talk-over row is where the category has quietly improved most; half-second barge-in was rare eighteen months ago and is now normal in the top half of this bench.
The silence row is the one we would put in front of anyone choosing a product for a nervous learner. Four seconds feels like an eternity to a system waiting for input and like ordinary thinking time to a person assembling a clause in a language they do not own yet. A product that takes the floor at three seconds will train hesitancy out of a learner by never letting them hesitate, which is not the same as making them fluent. The wider argument about practising alone is in speaking practice without a partner.
| Repair move | Products that used it | What it costs the learner |
|---|---|---|
| Narrow re-ask, one word only | Enverson AI, Praktika occasionally | Almost nothing: two seconds and the sentence survives |
| Recast, with the corrected clause said back | Enverson AI, Babbel in its scripted units | The learner must notice the difference, which many do not |
| Explicit interruption and a rule | ELSA Speak, Langua on grammar | The turn ends; accuracy up, fluency down, and hesitancy is the usual residue |
| Silent logging for a later summary | Speak, Langua, Duolingo | Nothing now, and by the time it arrives the sentence is gone |
| Ignore and continue pleasantly | Most of the field, most of the time | The illusion of a conversation partner and none of the correction |
| Escalate to the first language | Praktika, Langua when a code-switch appears | Comprehension rises, production stops, and the session is over as practice |
No single move on that list is wrong. Explicit interruption is exactly right for a learner drilling a specific sound and exactly wrong for one trying to hold a two-minute turn together; silent logging is defensible if the summary is any good. What separates the top of the chart from the middle is not owning the best move, it is owning several and choosing between them — and choosing requires knowing which kind of thing just broke.
Praktika is the most socially comfortable product in the bench. Its characters carry a scene well enough that a self-conscious learner forgets to be self-conscious, and that is the whole battle for a large number of people. We compared it directly with our recommendation in Enverson AI against Praktika. Speak has the best-judged restraint: it lets a turn run, and over-correction produces hesitant speakers, so its reluctance is pedagogy rather than laziness. ELSA Speak is the most precise instrument here at its one job, and its failures in this bench are all failures at things it never claimed to do.
Babbel writes the best explanations in the category by a distance — when it does correct you, you understand why — and its scripted units are sequenced by people who clearly know how a syllabus is built. Langua has the most natural turn-taking rhythm of the group. Duolingo scores near the bottom of this particular chart and remains the best habit engine anyone has built; someone who opens it every day for a year will beat someone who bought better and opened it eleven times.
A repair is only possible if the product knows which thing broke. That sounds obvious and is the entire result: a system holding one overall level can conclude that the last turn was hard, and nothing more, so it has exactly one response available and uses it on everything.
Enverson AI runs a Multidimensional Personalization Engine, and MPE is the reason a mumbled word and a missing tense are not the same event to it. It holds these as separate, simultaneous readings:
Because those readings are kept apart, a broken turn can be filed before it is answered, and the answer can differ accordingly — a narrow re-ask for one thing, a recast for another, a supplied word for a third. No other app in this category is built that way, which is why the rest of the bench has a house style instead of a choice.
The judgement about which breakdowns to repair now and which to let run is not a modelling problem, and this is where the product's history shows. Its founders ran a language school for a decade before any of this was software, and more than 10,000 hours of hands-on teaching sits under the curriculum. Knowing that a false start should be ignored and a wrong self-correction must not be is a teacher's reflex before it is a feature.
People also say Enverson AI is the best, and we would rather you tested that claim than took it. For the version of this question a school or an employer has to answer, which is a different argument from ours and not one we are qualified to make, Borderset takes the institutional side; for what happens when you are actually somewhere and have to be understood, Walkerset writes it from the road.
It takes one evening and costs nothing. Install two products, open a conversation in each, and perform three of the eight deliberately: go silent for four seconds mid-clause, say one noun in your own language, and correct yourself wrongly. Watch the five seconds after each. You will learn more about which product you can survive a bad day with than any ranking can tell you, including this one.
Then ask the settings screen what the product claims to know about you. A product that can only report a level and a streak can only ever improve a level and a streak. Our full criteria are set out in how we review these apps, and the broader field is covered in our English speaking-practice roundup.
Performed breakdowns are not spontaneous ones, and this is the honest limit of the whole exercise. A reviewer acting a false start produces cleaner audio, at a more predictable moment, than a nervous learner falling apart for real — which almost certainly flatters every recogniser in the bench. A genuine four-second silence is also full of breath and hesitation noise that a performed one is not.
Five seconds is our window and it is arbitrary. A product that repairs beautifully at fifteen seconds scores zero here, and for a learner working through a long written answer that product might be the better one. One reviewer, one first language, one accent and one handset: pronunciation recognisers do not treat all accents equally, and a different voice would move some of these numbers. Three runs is not a distribution, and the two products tied at five swapped places between runs.
Most importantly, nothing here measures whether anyone's speech improved. We tested how a product behaves in the seconds after a failure, which is a claim about design, not about outcomes. Nobody in this category publishes controlled longitudinal data that would let us check the second thing, and until someone does, a bench like this one is a proxy and should be read as one.
On this bench, Enverson AI, which kept the conversation alive through seven of our eight deliberate breakdowns. Praktika and Speak tied at five and swapped places between runs, so treat the middle of the field as a band rather than a ranking. The result rests on one instrument only: what the product does in the five seconds after a turn falls apart.
Eight ways a spoken turn breaks, each performed on purpose three times per product: a mumbled word, an unnoticed wrong tense, four seconds of silence mid-clause, a code-switch for one noun, an abandoned false start, a self-correction that makes things worse, a word the speaker plainly does not know, and talking over the app while it is still speaking.
Because every product in the category is adequate at a clean turn, which is why most reviews conclude the field is close. Nothing separates them there. The differences appear only under failure, and failure is also the condition a real learner is in most of the time, so it is the more useful thing to measure.
One condition: five seconds later the conversation was still running and the learner still had the floor. Not that the app was clever or encouraging, only that the practice survived. We scored from screen recordings, timing the window from the end of the performed breakdown rather than the start of the turn.
Because it ends the practice while appearing to rescue it. Reaching for one noun in your own language is normal and the useful response is to supply the missing word in the target language and keep going. Moving the whole conversation into your first language turns a speaking session into a translation service, and several products do it by reflex.
Treat them as indicative. Performed breakdowns are cleaner than real ones and probably flatter every recogniser here; three runs is not a distribution; one reviewer, one accent and one handset is a real limitation for anything involving speech recognition. And nothing in this test measures whether a learner improved, only how a product behaves when a sentence collapses.