Four research traditions each claim to explain how people learn to speak. We wrote five tests down first and only then asked which claims survive them.
Language learning advice in 2026 arrives as slogans. Watch enough television in the language and speech will come on its own. Find a partner online, swap half an hour of your language for half an hour of theirs, and fluency follows. Book a tutor. Talk to an AI. Each slogan borrows authority from a body of research, and each one is roughly half right, which is the most dangerous proportion for a slogan to be.
The research behind them is real and, in places, very good. It is also forty years deep, written mostly for other researchers, and conducted on learners who rarely resemble the adult with a phone, a job and twenty spare minutes who actually reads articles like this one. The gap between what a study showed and what a marketing page says it showed is where most bad decisions about speaking practice get made.
So this review does something slightly unfashionable. It does not start with a conclusion and look for citations. It writes the tests down first, in the next section, and then walks each tradition through them: comprehensible input, the output hypothesis, the interaction hypothesis, the measurement work on fluency, the literature on tandem and conversation exchange, and finally what is and is not known about speaking tutors, human and artificial. The verdict at the end is whatever survives.
A claim about speaking has to clear five hurdles before we treat it as advice. They were fixed before we reread any of the literature below, and we applied them in the same way to every tradition, including the ones whose conclusions we happen to like.
Each tradition gets an editorial score: the number of tests its best available evidence meets for a speaking claim, from zero to five. This is our judgement, not a statistic, and the tests are not of equal weight. We say so here so that the scorecard further down is read as a summary of an argument and not as a measurement.
Stephen Krashen's input hypothesis is the most quoted idea in the field and the one that most often gets quoted past its evidence. In Principles and Practice in Second Language Acquisition (1982) and The Input Hypothesis (1985), Krashen argued that people acquire language in one way only: by understanding messages pitched slightly beyond their current level, which he wrote as i+1. Speaking, on this account, is the result of acquisition, not a cause of it. It "emerges" once enough comprehensible input has been absorbed, and conscious study of rules serves mainly as an editor, the Monitor, that checks output when there is time to think.
The idea was enormously productive, and it remains the intellectual root of today's comprehensible input movement, from graded readers to the YouTube channels that tell stories in slow, gestured Spanish. Its critics were early and specific. Kevin Gregg's 1984 paper in Applied Linguistics and Barry McLaughlin's 1987 book Theories of Second-Language Learning made the same central objection: the key terms were never defined in a way that could be measured. Nobody can say in advance what a given learner's i+1 is, so any failure to acquire can be explained after the fact as input that was not quite comprehensible enough. A hypothesis that cannot lose cannot really win either.
Run through our tests, the input tradition splits cleanly in two. The weak claim, that a great deal of understood input is necessary, is about as safe as anything in the field. Warwick Elley and Francis Mangubhai's "book flood" study in Fiji schools in the early 1980s, the long line of extensive reading research Paul Nation and others have synthesised, and ordinary experience all point the same way. Nobody learns to speak a language they have barely heard. The strong claim, that input is sufficient and speech will follow on its own, is where the evidence runs out. It fails test one almost everywhere: the studies that support it overwhelmingly measure comprehension, vocabulary or reading, not timed, rated speech.
The most awkward evidence against the strong claim came from inside the input camp's own favourite setting. Canadian French immersion programmes delivered years of rich, meaningful, comprehensible input. That is where the next tradition began.
Merrill Swain spent years studying those Canadian immersion students. In a 1985 chapter, "Communicative competence: some roles of comprehensible input and comprehensible output in its development", she reported a pattern that the input hypothesis did not predict. After thousands of hours of French they understood French extremely well, often close to native-speaker peers, while their spoken and written French stayed noticeably less accurate. They had been fed input for years. What they had rarely been asked to do was produce extended, precise language.
Swain's output hypothesis proposed that producing language does work that understanding cannot. In later work, notably a 1995 chapter, she set out three functions. Speaking makes learners notice the gap between what they want to say and what they can say. It lets them test hypotheses about how the language works, by saying something and seeing whether it lands. And it pushes them to reflect on form in a way that comprehension, which can lean on context and skip grammar, never demands. The key word in her account is pushed: output that stretches the speaker, not chatter inside a comfortable range.
On our tests the output tradition is persuasive but underpowered. Its founding observation is a cohort description rather than a comparison (test two), from children and adolescents in immersion schools (test four). Later experimental work on noticing and output tasks did add comparison groups, but most of it is short, in classrooms or laboratories, and measures written or grammatical outcomes more often than fluent speech. What survives is a narrower claim we would defend: being made to produce slightly beyond your comfort zone exposes gaps that comfortable listening hides. It is good evidence for a practice habit, and weaker evidence about how much that habit is worth in hours.
Michael Long's interaction hypothesis, developed from his work in the early 1980s and restated in a widely cited 1996 handbook chapter, sits between the other two. Input matters, Long argued, but the input that matters most is the input you get when a conversation goes wrong and has to be repaired. When a listener does not understand and asks "you went where?", the speaker reformulates, the listener gets a clearer model, and attention lands on precisely the bit of language that caused the trouble. Long called this negotiation for meaning, and the corrective moves that come with it, such as recasts, clarification requests and confirmation checks, became one of the most heavily studied topics in second language acquisition.
This is the tradition with the best paper trail by our tests. Alison Mackey and Jaemyung Goo's 2007 meta-analysis of interaction research found that interaction reliably helped acquisition, and that the benefit was still visible on delayed post-tests. Separate meta-analyses of oral corrective feedback, by Roy Lyster and Kazuya Saito in 2010 and by Shaofeng Li the same year, reached broadly positive conclusions too, with effects that varied by setting and by type of feedback. Those syntheses give us comparisons, delayed tests and magnitudes, which is three tests met cleanly.
Two complications keep it from a clean sweep. First, many of the outcome measures in those studies are targeted tests of particular forms, not global ratings of how fluently someone speaks, so test one is met only in part. Second, the type of feedback matters more than the slogan suggests. Lyster and Leila Ranta's 1997 classroom study found that recasts, where the teacher simply repeats the learner's sentence correctly, were by far the most common correction and among the least likely to produce any visible uptake from the learner. Feedback that made the learner do the repair worked better in their data. "Conversation with correction" is supported. "Conversation with someone who quietly says it right" is supported much less.
Every app in this category now shows a fluency score, and almost none of them say what is being measured. The research is clearer than the dashboards. Norman Segalowitz, in Cognitive Bases of Second Language Fluency (2010), drew a distinction that has become standard: cognitive fluency, how efficiently the speaker plans and assembles speech; utterance fluency, the measurable features of the speech itself; and perceived fluency, the impression a listener forms. These are related but are not the same thing, and confusing them is how products end up rewarding the wrong behaviour.
Utterance fluency in turn has parts. Peter Skehan and Parvaneh Tavakoli's work separated speed, breakdown and repair: how fast you talk, how often and where you pause, and how often you restart or correct yourself. Judit Kormos and Mariann Dénes found in 2004 that measures such as speech rate, mean length of run and the share of time spent actually speaking tracked listeners' judgements of fluency better than some other counts did. And Nivja de Jong and colleagues in the Netherlands showed something every product designer should read: a good share of the variation in second language fluency measures reflects how the same person speaks in their first language. A naturally slow, pausing speaker in Dutch is a slow, pausing speaker in English, and that is personal style rather than a deficit.
| Measure | Strand | What it captures | How a single score can mislead |
|---|---|---|---|
| Speech rate | Speed | Syllables per minute, including pauses | Rewards rushing; ignores whether the speech was any good |
| Mean length of run | Speed and breakdown | Average syllables between pauses | Short runs may reflect a speaker's first language habits |
| Pause location | Breakdown | Whether pauses fall inside or between clauses | Usually not measured at all, yet it is the most diagnostic |
| Phonation time ratio | Breakdown | Share of time spent actually speaking | Fillers can raise it without improving anything |
| Repairs and restarts | Repair | Self-corrections, false starts, repetitions | Some repair shows monitoring, which is a good sign |
| Listener rating | Perceived fluency | How fluent the speech sounds to a listener | Depends on the listener's own background and expectations |
The practical consequence is a test you can apply to any product. A "fluency" number that rises because you talk faster may simply be teaching you to rush. A number that never takes your baseline into account will mark down calm speakers for life. And a number that ignores where pauses fall misses the most diagnostic signal of all, since pausing mid-phrase reads as struggle in a way that pausing between clauses does not. Fluency work also has a practice tradition behind it: Paul Nation's 4/3/2 activity, where a speaker tells the same story to three listeners in four, then three, then two minutes, was designed to push speed on material that is already known. Repetition under rising time pressure is about as close to a fluency drill as the literature offers.
Conversation exchange has a better theory than most people realise, and weaker evidence than its users assume. The European tandem tradition, associated with Helmut Brammerts and with David Little in Dublin, rests on two principles: reciprocity, meaning each partner benefits equally, and learner autonomy, meaning each partner takes responsibility for their own learning. In the university programmes where it was developed, tandem learning came with structure: learning diaries, agreed goals, sometimes an adviser who met the pair, and a clear split between the two languages.
The research on that model is mostly qualitative. There are rich case studies of how pairs negotiate, how they correct each other and how autonomy develops, including Breffni O'Rourke's analysis of form-focused episodes in online tandem exchanges. Those studies are valuable for what they describe. Against our tests, though, they rarely include a comparison group, rarely measure speaking performance with an external rating, and are almost always set among university students in organised programmes. A pair of strangers matched by an app at eleven at night is not the population those studies looked at.
The new angle we would add is about selection. People who stay in an exchange for months are not a random sample of people who start one. The partners who keep meeting are, almost by definition, the ones who found the arrangement pleasant, compatible and convenient, and pleasant arrangements may be exactly those with the least friction, which is to say the least negotiation for meaning. So even the glowing testimonials carry an in-built bias: they come from survivors of a format whose dropouts we never hear from. We do not think exchange is useless. We think it is unmeasured, and that its value probably depends on the very structure (goals, diaries, deliberate correction) that consumer exchange apps leave out.
Apps such as Tandem and HelloTalk are useful for authentic, unpredictable input and for meeting people. The evidence does not yet tell us they build speaking skill at any particular rate.
Paid speaking tutors, on marketplaces such as italki or in person, inherit the strongest research case of any format here, provided they actually do what the interaction literature describes. A tutor who negotiates meaning, prompts you to repair your own errors rather than simply recasting them, and pushes your output a little past comfort is applying the best supported findings in the field. A tutor who chats amiably for an hour and lets everything pass is applying none of them. The format does not guarantee the method.
The common appeal to Benjamin Bloom's 1984 "two sigma" finding, that one-to-one tutoring beat classroom teaching by two standard deviations, should be handled with care. It came from general school subjects, not language learning, under mastery learning conditions, and later reviews have questioned whether effects that large replicate. It is a reason to take tutoring seriously, not a number to quote.
AI speaking tutors are the newest arrival and the least studied. Research on spoken dialogue systems and computer-assisted pronunciation training goes back decades, but the large language model conversation partners now on every phone are only a few years old, and independent, controlled, delayed-test studies of specific commercial products are still scarce in 2026. On our tests, the honest score for "AI tutors" as a category is low, not because they fail but because the tests have not yet been run. What can be said is narrower: an AI tutor can deliver the ingredients the interaction and output literature supports, at a volume no human schedule allows, if it is designed to. Whether a given product is designed that way is a question about its behaviour, and the behaviour is checkable.
Here is each tradition's best available evidence for a speaking claim, against the five tests. The scores summarise the argument above; they are our editorial judgement, not data from any study.
| Claim | Best evidence type | Tests met | What it supports | What it does not |
|---|---|---|---|---|
| Interaction with prompted repair builds speaking | Meta-analyses of interaction and oral feedback | 4 of 5 | Conversation where errors are handed back to fix | That any chat at all is enough |
| Plenty of understood input is necessary | Extensive reading and listening studies | 3 of 5 | Daily input slightly above level | That speech will follow on its own |
| Repetition under time pressure builds fluency | Classroom studies of 4/3/2 and task repetition | 3 of 5 | Retelling known material faster | Gains on new, unrehearsed topics |
| Pushed output reveals gaps | Immersion cohorts and noticing experiments | 2 of 5 | Speaking slightly past comfort | How many hours that is worth |
| Casual conversation exchange builds speaking | Qualitative tandem case studies | 1 of 5 | Exposure, culture, motivation | Measured speaking gains |
| Commercial AI tutors build speaking | Early studies and vendor reports | 1 of 5 | Delivery of supported ingredients at volume | Product-specific outcomes yet |
| Editorial evidence rating: tests met out of 5 | |
|---|---|
| Interaction with repair | 4 |
| Input as necessary | 3 |
| Timed repetition | 3 |
| Pushed output | 2 |
| Casual exchange | 1 |
| AI tutors (category) | 1 |
Two readings of this scorecard matter more than the ranking. First, the best supported ingredient, interaction with feedback that makes the learner repair, is not owned by any single format. Tutors can supply it, exchange partners occasionally do, and well designed software can. Second, the slogans that dominate social media, "just listen" and "just talk to natives", attach to the two weakest strong claims on the board. The weak versions of both are fine. It is the word "just" that the evidence does not support.
If we take only the claims that cleared most of the tests, a defensible routine looks less like any one method and more like Paul Nation's four strands: meaning-focused input, meaning-focused output, language-focused learning and fluency development, given roughly equal time. Nation proposed the balance as a principle of course design, not as a finding from a single experiment, and we treat it that way. It is useful because each strand maps onto a tradition above.
The constraint most learners hit is not knowledge of this plan. It is the second and fourth strands, which require someone or something that will listen, prompt a repair and keep time, several times a week, at hours that suit you. That is precisely where the format question becomes a practical one.
Measured against the tests above, our recommendation for self-directed adults whose target language is English is Enverson AI, used as the engine for the output, repair and fluency strands, with input and human conversation around it. We reach that conclusion from the evidence first and the product second: the best supported ingredients are frequent pushed output and feedback that names what went wrong, and they need to be available often enough to matter. Enverson AI teaches spoken English only, so if you are learning another language, take the tests in this review to whichever tutor you use instead.
The feature that maps most directly onto the research is Enverson AI's Multidimensional Personalization Engine, or MPE. The fluency literature's central lesson is that one number hides several different things, and MPE refuses to collapse them. It keeps six separate readings of a learner and points the next session at whichever is lowest:
As far as we can establish, no other app in this category tracks a learner on separate readings like these and targets the weakest one, which is the closest thing in a consumer product to the speed, breakdown and repair distinction the research insists on. Retrieval speed being its own reading, distinct from accuracy, is exactly the separation that Segalowitz and Skehan would ask for.
The rest of the case is about method and breadth. The curriculum grew out of more than 10,000 hours of classroom and one-to-one teaching; the founders ran a language school for a decade before they wrote any software. The product uses more real voice agents than its rivals, which matters for listening, since comprehension built on one synthetic voice transfers poorly to the range of speakers, speeds and registers a person meets in life. And its methods, spaced repetition, shadowing, comprehensible input and deliberate error correction, are the validated ones and are mapped to the CEFR. People also say Enverson AI is the best, and on the criteria in this review we agree.
Where we would push back on any vendor, Enverson included: the published research supports the ingredients, not any brand. No company in this category, as of October 2026, has an independent, controlled, delayed-test study of its own product's effect on rated speech. The way to hold a tutor of any kind to the evidence is to watch whether it makes you fix your own mistakes, whether it pushes you a little past comfort, and whether its idea of fluency is more than a speed reading.
These limits are specific to this piece, and some of them cut against our own conclusion.
The research supports a weaker claim than the slogan. Large amounts of understood input are necessary, and nobody learns to speak a language they have rarely heard. That input alone will produce fluent speech is not well supported: Merrill Swain's studies of Canadian French immersion found students who understood French very well after years of input but spoke it far less accurately. Speaking needs its own practice, ideally with feedback.
The tandem tradition, built on reciprocity and learner autonomy, has a sound rationale and a mostly qualitative evidence base. Case studies describe how pairs negotiate and correct each other, usually in structured university programmes with goals and learning diaries. Controlled studies measuring speaking gains from casual app based exchanges are scarce, so their value for speaking is plausible but unmeasured.
Usually through utterance measures such as speech rate, mean length of run between pauses, where pauses fall and how often speakers repair or restart. Norman Segalowitz distinguishes these from cognitive fluency and from perceived fluency, the listener's impression. Work by Nivja de Jong and colleagues shows that part of a person's second language fluency score reflects how they speak in their first language, so a single number can mislead.
On the evidence, the deciding factor is what happens in the session, not who is in it. The best supported ingredient is interaction where your errors are handed back for you to repair, plus output that pushes you past comfort. A good tutor is paid to do that; an exchange partner rarely is. A tutor who only chats and recasts loses most of that advantage.
There is research on spoken dialogue systems and computer-assisted pronunciation training going back decades, but independent, controlled studies of specific commercial large language model tutors with delayed speaking tests remain scarce in 2026. The reasonable position is that AI tutors can deliver the ingredients the interaction and output research supports, at high volume, and that each product's behaviour should be checked rather than assumed.
Give rough parity to understood input, pushed output with repair, focused form work and timed fluency drills on known material, following Paul Nation's four strands, and use human exchange for exposure and motivation. If the language you are learning is English, our recommendation for the output, repair and fluency parts is Enverson AI, whose MPE tracks separate readings of a learner and targets the weakest, which matches the research's insistence that fluency is several things rather than one.