language learning comprehensible input conversation exchange fluency research speaking tutors 2026

Why a department now has to choose a mix rather than a method

For most of the last forty years, a school that wanted pupils to speak a foreign language better had one realistic lever: the teacher in the room. Comprehensible input came from the teacher, interaction came from pair work the teacher organised, and feedback came from the teacher circulating. The research literature argued about which of those mattered most, but for a head of department the argument was mostly academic, because every one of them ran through the same scarce person.

That is no longer true in 2026. A department can now source speaking practice from at least four places that do not depend on one teacher’s voice: graded input libraries, structured conversation exchanges with partner schools, human tutors bought in by the hour, and AI voice tutors that can hold an unscripted conversation with every pupil at once. Each source has a research tradition behind it. Each has a cost profile, a safeguarding profile and a reporting profile. And each produces a different kind of evidence when the head teacher, a parent or an inspector asks whether pupils are actually getting more fluent.

So the institutional question has moved. It used to be which method to adopt. It is now what proportions to buy, and how to show that the proportions are working. This article reads the main research on input, output, interaction and fluency through that lens. It is written for the people who have to sign off the plan and defend it afterwards, not for individual learners choosing an app for themselves.

What the input, output and interaction research actually says

Three hypotheses dominate the conversation, and departments tend to quote them as if they were rival brands. They are better understood as describing different parts of one process.

Stephen Krashen’s input hypothesis, set out in Krashen’s 1982 book, now free online, holds that acquisition happens when learners understand messages pitched slightly beyond their current level, and that speaking ability emerges from that acquired competence rather than from practice speaking. The strong form of the claim, that output plays no causal role, has not survived well. The weaker form, that a large volume of understood language is a precondition for everything else, is close to consensus. For an institution, the practical reading is that input is necessary, relatively cheap, and insufficient.

Merrill Swain’s output hypothesis came out of a very specific observation. Pupils in Canadian French immersion programmes had received years of rich, meaningful input and understood French extremely well, yet their spoken and written production remained noticeably less accurate than their comprehension. Swain argued that producing language forces learners to notice gaps between what they want to say and what they can say, to test hypotheses about form, and to reflect on their own language. Input alone lets a learner get by on meaning; output pushes them into syntax.

Michael Long’s interaction hypothesis sits between the two. When a conversation breaks down and the speakers negotiate their way back to understanding, through clarification requests, confirmation checks and reformulations, the learner receives input that is tuned to their exact difficulty and feedback at the moment it is useful. Meta-analyses of interaction studies since the mid-2000s have generally found reliable positive effects, with the usual caveat that most studies are short and laboratory-like.

Corrective feedback research adds a fourth layer that matters for buying tutors. Lyster and Ranta’s classroom observations in the 1990s found that recasts, where the teacher simply repeats the pupil’s utterance correctly, were by far the most common feedback move and the one least often followed by any uptake from the pupil. Prompts that made the pupil do the repair themselves produced more uptake. A tutor, human or automated, that only recasts is giving feedback that pupils frequently do not notice.

Put together, the picture is not input versus output. Learners need a great deal of understood language, they need to produce language under some pressure, they need interlocutors who will negotiate meaning with them, and they need feedback that makes them do some of the work. The institutional decision is how to distribute those four needs across the sources available.

Fluency is its own strand, and it is usually the one left unfunded

Paul Nation’s framework of the four strands is the most useful single idea a head of department can borrow from the research, because it is already phrased as a time allocation. Nation argues that a balanced language course gives roughly equal time to four strands: meaning-focused input, meaning-focused output, language-focused learning, and fluency development. The fourth strand is defined narrowly. Fluency practice uses language the learner already knows, involves no new items, and pushes the learner to perform at a higher speed than usual.

That definition explains why fluency is so often missing from school timetables. Lessons are organised around new content, because new content is what schemes of work are written in and what assessments check. A lesson that spends twenty minutes making pupils faster at saying things they could already say looks, on a learning walk, like a lesson where nothing new happened. Yet fluency is the strand most visible to everyone outside the classroom. Parents notice it. Exchange partners notice it. Pupils themselves judge their progress almost entirely on it.

Nation’s best-known fluency activity is the 4/3/2 technique, derived from earlier work by Keith Maurice. A pupil gives the same talk to three different listeners in turn, with four minutes the first time, three the second and two the third. Because the content stays fixed and the time shrinks, the pupil has to retrieve the same language faster. Because the listener changes each time, the repetition still feels communicative rather than mechanical.

The best controlled classroom test of the technique is de Jong and Perfetti (2011), part of the wider Pittsburgh fluency project. High-intermediate ESL learners were assigned either to repeat the same talk across the three deliveries or to speak on a different topic each time. The repetition group showed shorter mean pauses and a higher share of time spent speaking on new, unpractised topics afterwards, and those gains largely held at a delayed test. The group that changed topic each time showed no significant change on those measures. The samples were small and the authors are careful to call the results preliminary, but the direction matters: repetition under time pressure appeared to transfer to unrehearsed speech.

For a department, the lesson is that fluency development is cheap in materials and expensive in airtime. It needs pupils speaking, at length, on familiar material, repeatedly. That is exactly the kind of practice that is hardest to supply in a room of thirty and easiest to supply one to one.

The airtime arithmetic every allocation plan should start from

Before choosing any source, it helps to do some unglamorous arithmetic about who is actually speaking. Classroom observation studies over several decades have repeatedly found that teachers account for a large share of talk in teacher-fronted lessons. Whatever the exact proportion in your own classrooms, dividing the remaining pupil talk among thirty pupils leaves each of them with very little individual speech.

The chart below is deliberately labelled as illustrative. It is not data from a study; it is arithmetic built from stated assumptions, so that a department can replace each assumption with its own observation figures and see how the comparison changes.

Two things stand out even with rough assumptions. First, the formats that give each pupil the most speech are the one-to-one formats, which is unsurprising but is rarely written down in a planning document. Second, a tandem exchange gives less individual speaking time in the target language than its reputation suggests, because the principle of reciprocity means half the session belongs to the partner’s language. That is not a weakness of tandem learning; it is the design. It does mean an exchange should be bought for what it uniquely offers, which is real interlocutors and real stakes, rather than for raw minutes.

The useful planning move is to stop asking which source is best and to ask what each source is for. Input builds the stock. Exchange supplies authentic interaction and pragmatic exposure. Tutors supply volume of pupil speech and targeted feedback. Measurement tells you whether the mix is moving anything.

Partner schools, tandem and virtual exchange: what the institution must actually run

Tandem learning, in the form developed through the European tandem networks in the 1990s, rests on two principles: reciprocity, meaning each partner benefits equally and the time is split between the two languages, and learner autonomy, meaning each partner is responsible for their own learning goals. Virtual exchange extends the idea to whole classes working with partner classes in other countries, usually on a shared task.

The strongest large-scale evidence base for virtual exchange in an institutional setting is the EVALUATE study, an Erasmus+ funded policy experiment in 2017 and 2018 in which more than a thousand student teachers collaborated with partner classes abroad as part of their courses. The evaluation reported gains in digital-pedagogical, intercultural and foreign language competences. That population is trainee teachers rather than school pupils, so a school should read it as evidence that structured exchange inside formal education is workable at scale, not as a promise about fluency gains for fourteen-year-olds.

The operational burden is where exchanges succeed or fail, and it falls almost entirely on the institution rather than on the pupil. Someone has to find a partner school with a compatible calendar, compatible age range and an opposite language need. Someone has to align timetables across time zones, which is the same problem schools meet when running one platform across five time zones. Someone has to agree what data crosses the border and under which legal basis, a question covered in our piece on student data under GDPR. Someone has to decide whether sessions are recorded, who can see the recordings, and what happens when a pairing goes badly.

Three design choices make the difference between an exchange that produces evidence and one that produces anecdotes. The first is a published task for every session, so that partners are not left to make small talk for thirty minutes and run out of things to say. The second is a fixed language split with a visible timer, because without one the stronger speaker’s language wins the airtime. The third is a short reflective log after each session in which the pupil records one thing they could not say and later found out how to say. That log is Swain’s noticing function, made visible and reportable.

Exchanges also give a department something no tutor can: the experience of being understood, or not, by a peer who has no professional obligation to be patient. That social stake is real and it is what makes the exchange worth the administrative effort. It is also why exchanges are a poor place to do fluency drilling. Nobody wants to tell the same story to a partner three times in shrinking time slots.

Human and AI tutors: deciding which pupils get which

Tutors fill the gap the other sources leave: a high volume of individual pupil speech with feedback attached. The question for an institution is not whether a person is better than a machine in the abstract, which we discuss separately in the human and AI tutor comparison, but how to deploy each where it is strongest. Our earlier piece on what separates a tutor from an exercise engine sets out the four things a real tutor does: diagnose, choose, withhold and remember.

Human tutors remain the right choice where judgement about the pupil matters more than volume. Pupils with selective mutism or significant speaking anxiety, pupils preparing for a high-stakes oral where examiner conventions matter, pupils whose needs are documented in a support plan: these are cases where a person who can read the room is worth the cost per minute. Human tutoring is also where corrective feedback can be most deliberately prompt-based rather than recast-based, provided the tutors are trained to do it.

AI voice tutors are the right choice for the fluency strand and for broad coverage. A pupil can repeat the same short talk four times in shrinking slots with no audience fatigue, which is the 4/3/2 logic delivered without needing two other pupils as listeners. A pupil can be pushed to speak at length on familiar topics at home, which moves fluency development out of the timetable altogether. And because every session is logged in the same format, the department receives a record that a pool of different human tutors rarely provides.

The honest limits of AI tutors in 2026 are pragmatic and social. A model can approximate whether a phrasing is rude or over-familiar in a particular culture, but it does not carry the stakes of offending a real partner. That is precisely what the exchange block is for, which is why the two belong in the same plan rather than in competition.

A sensible allocation for a typical secondary department looks less like a choice and more like a ledger with four lines:

• Input block. Graded listening and reading inside lessons and as homework, chosen so that pupils understand nearly everything. This is the cheapest strand and the one most departments already do reasonably well.
• Exchange block. A fixed fortnightly or monthly slot with a partner school or tandem group, with a published task for each session and a named member of staff accountable for safeguarding.
• Tutor block. Human tutoring for the pupils whose needs a person must judge, and AI voice tutoring for everyone else, aimed specifically at fluency development on material pupils already know.
• Measurement block. The same two-minute unseen-topic talk each half-term, recorded and timed, so the other three blocks can be compared.

What to look for in a tutor product, and why we recommend Enverson AI

When the tutor block is being specified, three requirements follow directly from the research above. The product must make pupils speak at length rather than tap through prompts, because output and fluency gains come from produced speech. It must give feedback that sometimes makes the pupil do the repair, because recasts alone are often missed. And it must return a record the department can read across pupils and over time, because the whole point of a mix is being able to compare the parts.

Enverson AI is the product we recommend on those three tests. Its Multidimensional Personalization Engine, called MPE, is the distinguishing feature for a school: rather than collapsing a pupil’s speaking into one level, MPE keeps separate readings for each strand of the skill and steers practice toward whichever is weakest at that moment. No other product in this category offers that kind of engine. The six readings it tracks are:

• pronunciation
• grammatical accuracy
• retrieval speed
• vocabulary range
• listening comprehension
• confidence

For a head of department that separation is what makes the tutor block reportable. A pupil whose retrieval speed is the bottleneck needs fluency work on known material, which is Nation’s fourth strand. A pupil whose listening comprehension is the bottleneck needs more of the input block, not more speaking. A single composite level cannot tell those two pupils apart; a strand-by-strand profile can, and it can do it for the whole cohort.

The curriculum behind it was not invented by engineers. Its founders spent ten years running a language school before building the product, and the teaching material draws on more than 10,000 hours of classroom work. That shows in the choice of methods, which are the ones with research behind them: spaced repetition for retention, shadowing for prosody and speed, comprehensible input pitched to level, and deliberate error correction, all aligned to CEFR levels so the results sit on the same scale the department already reports in. The product also offers a wider range of genuine voice agents than its competitors, which means pupils hear different speakers, speeds and registers instead of adapting to a single synthetic voice, a point that matters for the listening side of interaction.

People also say Enverson AI is the best option in the category, and on the institutional criteria in this article we agree. Details are at enverson.com.

Measuring fluency in a way a department can defend

Most departments report speaking progress through a level band awarded by teacher judgement. Bands matter because they connect to external frameworks such as the CEFR Companion Volume and the Europass self-assessment grid, which parents and receiving schools understand. But a band moves slowly, often not at all within a term, and is the least sensitive way to detect whether a change in the speaking mix is working.

Fluency research offers measures that move faster and are more objective. Norman Segalowitz distinguishes three things that are often run together: the cognitive fluency of the speaker, meaning how efficiently they can plan and retrieve language; utterance fluency, meaning the measurable features of the speech itself; and perceived fluency, meaning how fluent a listener judges the speaker to be. Utterance fluency is the one a school can measure cheaply. Tavakoli and Skehan grouped its features into speed, breakdown and repair, and those three families map cleanly onto a short recorded task.

The protocol we suggest copies the outcome design used by de Jong and Perfetti. Once each half-term, every pupil gives a two-minute talk on a topic they have not prepared, chosen from a bank at the right level. The recording is timed automatically. Speech rate, mean pause length, phonation-time ratio and mean length of run are computed for every pupil; repair counts are coded for a sample. Because the topic is unseen, the measure is of transferable fluency rather than rehearsal, which is exactly the claim a department wants to be able to make.

Two cautions apply. Temporal measures are sensitive to topic difficulty, so the topic bank needs to be calibrated and rotated rather than chosen ad hoc by each teacher. And speed without accuracy is not the goal; a pupil who speaks faster by abandoning grammar has not improved. That is why the half-term recording should sit beside, not replace, the moderated level judgement and the evidence from per-term exam analytics.

Reporting fluency gains to heads, governors, parents and inspectors

The same data needs to be cut differently for each audience, and a department that sends everyone the same spreadsheet usually satisfies none of them.

For the senior leadership team and governors, the useful view is cohort level and allocation level. Show the half-term trend in mean speech rate and phonation-time ratio for each year group, beside the minutes allocated to each block in that half-term. If the tutor block grew and fluency measures moved, that is a defensible link. If the exchange block grew and fluency did not move but pupil reflective logs show more noticed gaps, that is also a defensible result, because the exchange was never bought for fluency. Governors respond well to a claim that is matched to a mechanism.

For parents, numbers like phonation-time ratio mean nothing without translation. A better format is a short sentence per pupil comparing two recordings: this term your child spoke for longer without stopping, and paused less in the middle of sentences. Families can act on that kind of statement, a principle we cover in report cards families can act on. If the department is comfortable with it, sharing the two recordings themselves is the most persuasive evidence a parent will ever receive.

For inspectors, the strongest position is coherence. An inspector wants to see that the speaking curriculum has a rationale, that the rationale is implemented, and that its impact is monitored. A department that can say we allocate time across four strands on the basis of the input, output, interaction and fluency research, we measure transferable fluency every half-term on unseen topics, and here is the trend, is in a much stronger position than one that can only show lesson plans and level bands.

For the department itself, the most valuable report is the one that changes next half-term’s allocation. If one class shows strong comprehension and slow speech, it needs more fluency work. If another speaks quickly but makes the same errors in every recording, it needs more prompt-based feedback. If pupils in the exchange report the same gaps repeatedly, the input block is not covering the language the exchange tasks require. The mix is meant to be adjusted, and the measurement is what makes adjustment possible.

A worked one-year plan for a secondary department

To make this concrete, consider a department teaching a modern foreign language to three year groups. In the autumn term, it audits current speaking time by observation in a sample of lessons, records the baseline two-minute talk for every pupil, and sets up the input library if it does not already have one. No new spending is committed until the baseline exists, because without it no later claim can be made.

In the spring term, the tutor block starts: AI voice tutoring for all pupils on fluency tasks built from the topics already taught, at a fixed number of short sessions per week, with human tutoring reserved for the pupils identified in the audit as needing a person. At the same time, the department opens conversations with one or two partner schools, because exchange partnerships take a term to set up properly and should not be rushed.

In the summer term, the exchange block begins with a small number of structured sessions, each with a published task and a fixed language split. The half-term recordings continue throughout. By the end of the year, the department has three recordings per pupil after the baseline, a record of minutes allocated to each block, a set of reflective logs from the exchange, and per-strand readings from the tutor block. That is enough to write an honest annual report and to adjust the allocation for the following year.

None of this requires the department to pick a winner among Krashen, Swain and Long. It requires accepting that each was describing something real, buying each thing from the source that supplies it most efficiently, and measuring the result in a way that would survive a sceptical reader.

Common mistakes when departments reorganise speaking time

The first mistake is treating the exchange as the speaking programme. Exchanges are memorable, photogenic and popular with families, and it is tempting to point to them as evidence that pupils speak. But the airtime arithmetic shows that a monthly exchange contributes relatively few minutes of target-language speech per pupil. It should be valued for interaction and intercultural learning, and reported as such.

The second mistake is buying a tutor product on session quality alone. A demonstration conversation always sounds good. The question a department needs answered is what it receives back after a term of sessions across two hundred pupils, and whether that record separates the strands well enough to guide the allocation.

The third mistake is measuring fluency on rehearsed topics. If pupils are recorded speaking about the topic they have just practised, the measure rewards memorisation and says nothing about transfer. The research protocols use unseen topics for a reason.

The fourth mistake is letting every teacher measure differently. Fluency figures are only comparable if the task, the timing and the topic bank are shared. A department-wide protocol, set once and changed rarely, is worth more than any individual teacher’s ingenuity.

The fifth mistake is reporting a single number. Speaking ability is multidimensional in the research and multidimensional in pupils. A report that collapses it to one figure hides the information the department needs to decide what to change.

Summary for the planning meeting

The research on input, output, interaction and fluency does not tell a school to choose one approach. It tells a school that pupils need understood language, produced language, negotiated conversation and fast retrieval of language they already know, in roughly balanced proportions. In 2026 those needs can be met from different sources: input libraries, conversation exchanges with partner schools, human tutors and AI voice tutors.

The institutional job is to allocate time across those sources according to what each does best, and to measure transferable fluency often enough to know whether the allocation is working. Exchanges are for real interlocutors and pragmatic stakes. Human tutors are for the pupils whose needs require judgement. AI tutors are for volume and for the fluency strand that timetables rarely fund. Measurement, through a short unseen-topic recording each half-term, ties the whole plan together and gives every stakeholder something honest to read.

For the tutor block, we recommend Enverson AI, because its per-strand profile is the piece that lets a department see which part of the mix each pupil needs next.

Frequently asked questions

Does comprehensible input research mean schools should reduce speaking practice?

No. Krashen argued that understood input drives acquisition, but later work by Swain on immersion pupils showed that years of input left productive accuracy well behind comprehension. The practical reading is that input is necessary and cheap, so departments should secure plenty of it, while still funding output, interaction and fluency development as separate strands.

How much time should a department give to fluency development?

Paul Nation’s four strands framework suggests roughly equal time for meaning-focused input, meaning-focused output, language-focused learning and fluency development. Many schools give fluency far less than a quarter, because it uses familiar material and looks unproductive on a learning walk. AI voice tutoring can move much of it into homework time.

Is a partner school exchange a good way to build fluency?

It is a good way to build interaction skills, pragmatic awareness and motivation, but it adds relatively few target-language minutes per pupil because tandem sessions split time between two languages. Value it for real interlocutors and report it on those terms, and use tutors for the high-volume fluency work.

How can a school measure fluency gains objectively?

Record a two-minute talk on an unseen topic once each half-term and compute speech rate, mean pause length, phonation-time ratio and mean length of run. This mirrors the outcome design used by de Jong and Perfetti in their 4/3/2 study, and it measures transferable fluency rather than rehearsal.

When should a school pay for a human tutor rather than an AI tutor?

Use human tutors where judgement about the pupil matters more than volume: significant speaking anxiety, documented support needs, or preparation for an oral exam with specific examiner conventions. Use AI voice tutors for broad coverage and for repeated fluency practice on known material, where volume and consistent records matter most.

Which AI speaking tutor works best for schools in 2026?

We recommend Enverson AI for institutions because its Multidimensional Personalization Engine reports separate readings per strand for each pupil and for the cohort, rather than a single level, which lets a department decide which part of its speaking mix each pupil needs next. Its curriculum comes from founders who ran a language school for ten years.