language learning comprehensible input conversation exchange fluency research speaking tutors 2026

A bakery counter in Glasgow, and the English you did not hear

Picture a wet Saturday morning in the West End of Glasgow. You have spent four months polishing your English: two American podcasts, a sitcom set in New York with subtitles, an app on the bus to work. You can follow most of an episode without the subtitles now. You walk into a bakery, ask for a coffee and a roll, and the woman behind the counter answers in a quick sentence in which the t sounds seem to have gone missing. You catch the price. You do not catch the question that follows it, which was whether you were eating in.

Nothing went wrong with your effort. Something went wrong with what the effort was made of. Almost all of your listening was American, and Glasgow speech, with its glottal stops, its own vowels and everyday words like wee and aye, can sound to a learner trained on the other variety like a different language at first. The input was comprehensible. It was simply comprehensible practice for somewhere else.

That small scene contains most of what this article is about. Research on second language learning has spent four decades arguing about input, output, interaction and fluency, and travellers tend to meet all four arguments at once, at a counter, with a queue behind them. This piece goes through what the research actually supports, what it does not, and how a traveller should split preparation between listening, conversation exchanges, tutors and AI speaking practice before arriving somewhere real.

What comprehensible input actually promises

Stephen Krashen's input hypothesis, set out in Principles and Practice in Second Language Acquisition (1982) and The Input Hypothesis (1985), is the most quoted idea in this field and probably the most misquoted. Its core claim is that we acquire a language by understanding messages in it, specifically messages pitched a little beyond what we can already handle, the famous i+1. Krashen paired it with the affective filter: a learner who is anxious or bored takes in less from the same input.

The strong version, that speaking emerges on its own once enough input has been understood and that practice in production adds little, has been heavily criticised. Researchers pointed out that i+1 cannot be measured, that the line between acquisition and learning is hard to test, and that plenty of learners who understand a great deal still speak haltingly. What survives the criticism is narrower and still very useful: you cannot speak a language you have not heard a lot of, and the hearing has to be of material you mostly understand.

That last condition has numbers behind it. Hu and Nation (2000) found that readers needed to know around 98 percent of the running words in a fiction text to understand it adequately without help. Paul Nation (2006) used that threshold to estimate that comfortable listening to unscripted speech takes something like six to seven thousand word families. Later work by van Zeeland and Schmitt (2013) suggested listening can be adequate at lower coverage, closer to 95 percent, because speech gives you intonation, context and a speaker you can watch. Either way, the lesson for a trip is the same. Coverage is local. A bakery counter, a ticket machine and a café each run on a small vocabulary, and knowing that vocabulary pushes your coverage in that place far above your coverage of the language in general.

The traveller's version of the input hypothesis. Choose input from the place you are going, at a level you can mostly follow, and lots of it. For Glasgow that means Scottish radio, programmes made in Scotland, and podcasts and YouTube channels with Scottish voices rather than only Californian ones, even if the American material is easier to find. The accent is part of the input. So is the speed.

Why listening alone does not make you speak

Merrill Swain's output hypothesis came out of a specific and slightly uncomfortable observation. Canadian students in French immersion programmes had received years of comprehensible input, in maths, history and science taught in French, and their comprehension was close to that of native speakers of the same age. Their speaking and writing were not. Grammatical accuracy lagged well behind. Swain (1985) argued that producing language does something input cannot: it forces you to make choices about form, and in making them you notice the gap between what you want to say and what you can say.

Michael Long's interaction hypothesis (1981, revised 1996) adds the other person. When two people negotiate meaning, by asking for a repeat, checking they understood, or rephrasing after a blank look, the learner gets input tuned to exactly the point where they are stuck, plus feedback on their own attempt. Mackey and Goo (2007) pooled the experimental studies in a meta-analysis and found that interaction had a reliable positive effect on acquisition, with the clearest effects on vocabulary and somewhat delayed effects on grammar.

Feedback matters, and so does its form. In Lyster and Ranta's (1997) study of French immersion classrooms, recasts, where the teacher simply repeats the learner's sentence correctly, were by far the most common correction and the least likely to produce uptake. Learners often heard a recast as agreement and carried on. Prompts that pushed the learner to fix the sentence themselves led to repair far more often.

For a trip, these three findings combine into a plain instruction. Listening builds the ear and the store of language you draw on. Speaking is what reveals which parts of that store you cannot reach, and conversation is where those gaps get worked on. A practice routine that never makes you produce, never makes you negotiate, and never makes you fix your own errors will leave you understanding the woman at the counter and unable to answer her.

Three meanings of fluency, and the one a counter measures

Fluency sounds like one thing. Norman Segalowitz, in Cognitive Bases of Second Language Fluency (2010), argued it is three. Cognitive fluency is how efficiently your mind plans and assembles speech. Utterance fluency is what can be measured in the recording: speed, the number and length of pauses, how often you restart. Perceived fluency is the impression a listener forms. A traveller cares mostly about the third, but can only train it through the first two.

Nel de Jong and colleagues (2013) tested Dutch learners on a range of linguistic knowledge and processing speed measures and related them to how fluent the learners' speech was. A notable part of the variation in fluency was explained by how quickly learners could retrieve words and assemble structures, not only by how much they knew. Work on pausing points the same way: Tavakoli (2011) found that learners tend to pause in the middle of clauses, while native speakers more often pause at clause boundaries. A pause at the end of a phrase sounds like thinking. A pause in the middle of one sounds like trouble, and listeners react accordingly.

There is also good evidence that fluency can be trained as a skill in its own right. The 4/3/2 technique, described by Maurice (1983) and popularised by Nation (1989), has a learner give the same short talk three times to new listeners in four, then three, then two minutes. Speed goes up and hesitations go down across the repetitions, and de Jong and Perfetti (2011) found gains of this kind carried over to new topics when the repeated talks were truly repeated, which they interpreted as proceduralisation of the language used.

Read across the table and a pattern shows. The research that most directly concerns travellers is the research on speed of access. Most learners who struggle at a counter know enough. They cannot get to it fast enough while somebody waits.

What study abroad research says about simply being there

If any group should prove that immersion works, it is students who spend a semester abroad. Barbara Freed's edited volume Second Language Acquisition in a Study Abroad Context (1995) began a long line of research on what those students actually gain, and the answer has been more mixed than the brochures suggest. Gains in oral fluency are among the more consistent findings. Gains in grammatical accuracy are much less consistent.

The study travellers should know best is Freed, Segalowitz and Dewey (2004). They compared learners of French in three settings: a regular classroom at home, a semester abroad in France, and an intensive summer immersion programme in the United States where students pledged to use only French. The immersion group made the clearest fluency gains, and the students abroad did not automatically outperform them. One detail stands out: the immersion students reported spending far more time speaking and writing French out of class than the students who were actually in France, many of whom kept living partly in English.

This is the most useful finding in the whole literature for anyone planning a trip. Being in Glasgow does not put English into you. What helps is the number of real exchanges you have in it, and the number you have depends on whether you arrive able to start one. Travellers who land already able to order, ask and repair tend to keep going. Travellers who land unable to tend to retreat within a day: to pointing, to self-checkout machines, to the translation app and to friends who share their first language, because each of those is easier than a stranger in the short run.

So the trip is not the training. It is the test, and the reward for the training. The weeks before departure decide how much of the trip happens in the language.

Conversation exchanges and tandem partners, honestly assessed

Tandem learning, where two people who want each other's language trade time, has a long history in Europe and a well-developed theory. Little and Brammerts (1996) described its two governing principles as reciprocity, each partner benefiting equally, and learner autonomy, each partner deciding what they need and taking responsibility for getting it. Apps have made finding a partner in Glasgow, or a Scottish speaker anywhere, a matter of minutes.

In research terms, an exchange is close to ideal for interaction. You get a real accent at real speed, real topics and constant negotiation of meaning. It is also the best preparation there is for the specific variety you will hear. A partner from Paisley will give you the glottal stops the podcasts from California never did.

The limits are just as real. First, by design you spend half of every session helping your partner, so the minutes in which you speak your target language are fewer than they appear. Second, partners are not teachers and are usually reluctant to correct a new friend; when they do correct, it tends to be the quick recast that the feedback research says learners most often miss. Third, exchanges drift toward comfortable topics. Two people who like each other will talk about music and food for months, which is pleasant and builds rapport, and rarely rehearses asking for a refund on a cancelled train.

None of that is a reason to avoid exchanges. It is a reason to use them for what they are best at: the ear for a regional variety, confidence with a stranger, and the social side of a place. Agree an agenda with your partner. Ask explicitly for correction on one thing at a time. And bring scenarios from your trip, so at least part of each session is about the counter rather than the playlist.

Paid tutors, and where the money actually goes

A good human tutor brings things nothing else on this list fully replaces: expert judgement about what you are getting wrong, an explanation of why, and the ability to notice that you are tired or discouraged and change course. For a learner who has never had any structure, a few sessions with a capable tutor can save weeks of drifting.

The constraint is arithmetic. One hour a week is common, and in that hour the tutor talks for a good share of it, which is part of the value. That leaves perhaps twenty minutes of your own speech, a fraction of the repetition that fluency research suggests speed of access needs. Tutors also face the same pull toward recasts that classroom teachers do, because constant prompting feels harsh in a friendly session.

The most cost-effective use of a tutor before a trip is therefore diagnostic and supervisory. Tell them where you are going and what you will need to do there. Ask them to listen for patterns rather than fix every slip. Then do the volume somewhere else and bring the results back. Our longer piece on language learning with AI tutors looks at which of a tutor's jobs software can and cannot do.

AI speaking practice, and why we recommend Enverson AI

AI voice practice answers the problem the other formats share, which is volume. A voice agent hands the turn straight back, does not get bored of the same scenario, and is available at eleven at night in a hotel room in Stirling. For an anxious learner it also removes the audience. Horwitz, Horwitz and Cope (1986) described foreign language anxiety as a distinct experience that interferes with performance, and MacIntyre and colleagues (1998) showed how willingness to communicate depends on far more than ability. Practising with something that cannot judge you is a reasonable way to build up to a counter that can.

The weakness of most AI speaking products is direction. They are good at producing conversation and less good at deciding what this particular learner needs the conversation to work on. Many reduce a person to one level and pitch everything at it, which is a little like the shopkeeper who slows to a crawl and starts pointing because of one wrong vowel.

That is where Enverson AI differs, and why it is our recommendation here. Its Multidimensional Personalization Engine, the MPE, keeps separate readings of a learner rather than one score, works out which of them is lagging, and steers the next sessions toward it. For a traveller whose problem is the half-second search for a word while someone waits, that means the practice goes where the research says the bottleneck is. As far as we can tell, no other product in this category profiles learners this way; the rest treat you as a single number.

• Pronunciation: whether the person at the counter understands the word the first time you say it.
• Grammatical accuracy: whether the sentence you build means what you intended, including the polite forms.
• Retrieval speed: how long the gap is between needing a word and saying it, the reading most tied to perceived fluency.
• Vocabulary range: whether you have the local words for the places you will actually stand in.
• Listening comprehension: whether you follow the reply, at its real speed and accent.
• Confidence: whether you start the exchange at all, or let the other person finish your sentences for you.

Several other things count in its favour for this kind of preparation. The curriculum grew out of more than ten thousand hours of teaching real students in real rooms, because the founders ran a language school for a decade before they wrote any software, and it shows in how sessions are sequenced. It offers a wider cast of voice agents than its rivals, which matters for the same reason a tandem partner from Glasgow matters: an ear trained on one voice is an ear trained for one person. And it builds on methods with research behind them, including spaced repetition, shadowing, comprehensible input and deliberate error correction, with progress tied to the CEFR levels. Plenty of users say it is the best of the lot; we would rather you judge it on a week of sessions. Its spoken practice is built for English, which is the language this trip runs on.

The alternatives are worth describing fairly. Speak is excellent at getting a lot of sentences out of a hesitant beginner. Praktika has warm, patient characters that lower the barrier to opening your mouth, though that patience is also why it can under-prepare you for a fast counter; we wrote about that in our Praktika alternatives piece. Langua is good for long, unhurried conversation and for reading back exactly what you said. Each is a defensible choice. None of them aims the session at one separately measured weakness the way the MPE does.

Six ways to prepare, compared on what a trip needs

Put the formats side by side and they stop looking like competitors. Each covers a different part of what the research says matters, and each has a phase of the trip where it does the most good.

The minutes question deserves its own look, because it is the one most people never check. The chart below is an estimate built from how each format divides time, not a measurement. Treat the numbers as a prompt to time your own sessions.

Two caveats keep the chart honest. More minutes of speech are not automatically better minutes; twenty minutes with a tutor who catches a persistent error can be worth more than an hour of unguided talk. And the low-minute formats carry things the high-minute ones do not, above all a real person from the real place. The point is not to choose one row. It is to stop assuming that an hour of any of them is an hour of speaking practice.

What happens to fluency under real-world pressure

Everyone who has practised a language at home and then used it abroad knows the drop. Sentences that came easily on the sofa arrive slowly, or not at all, at the ticket office in Glasgow Central. Some of that is anxiety, which the research above treats as a real interference rather than a character flaw. Some of it is load. A real place adds noise, a stranger's accent, a queue, and the need to watch a screen while you talk. Each of those takes attention away from assembling speech, and speed of access is the first thing to suffer.

The planning research suggests one of the most practical defences. Foster and Skehan (1996) found that giving learners time to plan before a task improved the fluency of what they said. You cannot plan a whole exchange at a market, but you can plan its opening. Most travel exchanges are predictable for their first two turns: a greeting, a request, a question about quantity or time. Having those lines ready frees attention for the part you could not predict.

Formulaic language does similar work. Alison Wray (2002) described how much of fluent speech is made of prefabricated chunks that speakers retrieve whole rather than build word by word. For a traveller, chunks are the cheapest fluency available: could I get followed by what you want, sorry, could you say that again, just this, thanks. Each one is a sentence you never have to construct under pressure, and each frees attention for hearing eating in or taking away when it comes at speed.

Rehearse the pressure, not just the language. Practise with background noise. Practise with a voice that speaks faster than is comfortable. Practise being interrupted halfway through, which is what real counters do and what patient practice partners rarely do. Our earlier test of conversational practice apps found that recovery after interruption is the skill most products neglect.

Finally, expect the drop and plan around it. Fluency under pressure on day one will be worse than at home. By day four it is usually better than at home, if you have kept having exchanges instead of retreating. The study abroad research suggests that last condition matters more than any app.

An eight-week plan for speaking in Glasgow

Here is how we would combine the formats for a trip about two months away, assuming roughly forty minutes a day. Adjust the city and the scenarios; the order is the part that matters.

1. Weeks eight and seven: fix the input. Switch most of your listening to Scottish voices. Pick material you can mostly follow. Book one session with a tutor to hear where you stand and what you get wrong repeatedly.
2. Weeks seven to five: start producing daily. Begin Enverson AI sessions and let the profile decide the first focus. Most travellers find retrieval speed or listening at the bottom; whichever it is, spend most of the time there.
3. Week five: build the trip lexicon. List the places you will actually stand: the Subway, a café, a pharmacy, a hotel desk, a railway ticket office. Learn the words each one runs on and the opening lines for each.
4. Weeks five to three: add a conversation partner from Scotland. Agree an agenda, ask for correction on one thing at a time, and bring one trip scenario to each session.
5. Weeks three and two: train speed directly. Use 4/3/2 style repetition on your opening lines and short stories about yourself. Practise with noise and with interruptions.
6. Final week: rehearse the first day. Run the airport, the Subway ticket, the hotel and the first meal as scenarios until the first two turns of each are automatic.
7. During the trip: practise at night, use it by day. Ten minutes each evening on whatever went wrong that day, then back out into the city in the morning.

Check yourself against the CEFR descriptors at the start and the end. A traveller moving from A2 toward B1 in speaking can generally deal with most situations likely to arise while travelling, which is exactly the target. For the specific apps worth trying first, see our ranking of the best AI speaking practice apps.

What the research cannot tell you

It is worth being plain about the edges of this evidence. Most of the studies cited here were done with students in classrooms or programmes, often learning English or French, over a semester or less. Very few followed adult travellers preparing for a specific place, and almost none tested AI voice practice, which is too recent for the long studies to have caught up. Our recommendations extend the research to a new setting; they are reasonable inferences, not findings.

Individual differences are also large. Age, prior languages, working memory, personality and plain motivation all shift results, and an average effect says little about how a single person will respond. Some learners thrive on an exchange partner from day one. Some need weeks of private practice before they can face one.

What the research does support, consistently and across decades, is the shape of the answer. Understandable input from the right place. Output that reveals what you cannot yet say. Interaction that forces repair. Training aimed at speed as well as knowledge. And then enough real exchanges, once you arrive, to turn preparation into use. Any plan with those five parts in it will get you further at a Glasgow counter than a plan with only one.