If your goal is to speak French, a language partner beats every app, because speaking is a thing you do with another person and an app is a thing you do alone. An app can teach you to translate, to recognize words, and to keep a streak. It cannot hear your voice, tell you your “s” came out as a “z”, care what you did last weekend, or notice you avoided the subjunctive for six months. A native speaker who is learning your language does all of that, and does it for free, in exchange for the same service.
LanguageMate is a program for English speakers who want to speak French, built around one relationship rather than one teacher.
Let me start with what the apps get right.
What does Duolingo actually do well?
Quite a lot, and I say that as someone who downloaded it, used it, then wrote a long article taking it apart.
It is a brilliantly designed motivational tool. When you forget to open it, it reminds you. Not everyone loves that. Ellane, my Australian language partner, put it this way: “I used to find Duolingo rather too proactive in nagging me to practice. Their reminders were intrusive and guilt-inducing.” But the design works, and people keep coming back for years.
The interface is genuinely good. They have poured an enormous amount of research into making language practice enjoyable, from the skill trees to the satisfaction buttons to the steady rhythm of “You did it” and “Good job”. They succeeded.
Immediate feedback and spaced repetition are real learning principles. You find out straight away whether you were right, and items come back at spaced intervals. Both work, and both are in my method too.
And it is cheap. The free version is free. Hard to beat that.
I even send my learners to the Duolingo podcast. Their podcasts are worth a shot for beginners. Forget the app, their podcasts are better.
So the problem is not that Duolingo is badly made. The problem is that it is superbly made for a goal that is not yours.
Why won’t you learn to speak French with Duolingo?
Here is my actual list, the same one I give on a call.
Translation exercises don’t let you acquire a language. Almost every Duolingo exercise is a game with translations. Think about the word chien. Your brain pulls up dogs you have met, the words tail and paw and leash, your childhood pet, whatever you feel about dogs. Now compare that with the Hungarian word kutya. Unless you speak Hungarian, you have zero associations. If I tell you kutya means dog, you now have exactly one, and it is a translation. That is useful for translating a word in an app and useless for talking about dogs.
You memorize vocabulary you don’t need, so you forget it. Use it or lose it. You do not need the names of every piece of furniture in a French house in your first years. The day I actually needed the word “shelf” to explain Montessori education to my language partner, I made do, and then she gave me the word. That one stuck.
You memorize words out of context. When I told my language partner about the day my car slipped on the snow because I had no snow tires, I did not have the vocabulary. I made do, she gave me feedback, and I own those words now because they are attached to a real morning of my life. It should not be an app choosing your conversation topics.
You don’t learn to speak. When I ask learners what they want, nobody says “I want to get better at rearranging words in the right order”. They say they want to talk to their neighbors, their kids’ school, their clients. Translating words is not that.
You don’t improve your pronunciation. This is the big one and I will come back to it.
You build no social connection. We are social beings and we need connection with other humans. Yes, Duolingo gives you a small hit of dopamine when you finish an exercise. Nothing replaces the joy of talking with a human being who shares your interests.
You don’t immerse yourself in anything. Sitting next to the language for twenty minutes without ever producing a word of it is the worst of both worlds, and it is exactly what a silent hour with an app is.
None of this is your fault. You downloaded the app because it is the thing everyone downloads. All the Americans I talk to mention Duolingo. One of them, Amanda, told me: “I’ve got like a 1700-day streak on Duolingo.” One thousand seven hundred days in a row. She is incredibly consistent. Imagine what those 1,700 days could have bought her if they had been pointed somewhere else.
On calls, people tell me the same things in their own words. “I’ve done Duolingo almost religiously every day,” one man said. “I know you’re not a big fan of the apps. I think they serve a purpose.” Another told me he did a year of it and got tired of it, and that neither that app nor the paid one he moved to afterwards “really prepare you to speak French or understand French when you’re spoken to”. Carolyn was doing Duolingo daily and had no opportunity to speak at all.
The problem isn’t you. It’s what you’ve been handed.
Should you practice speaking with an AI instead of a person?
This is the real question in 2026, and it is more interesting than “apps bad”.
Duolingo and most of the big apps have gone the same way and put AI at the center, so a growing number of people now practice by talking to a machine. It is available at three in the morning, it never judges you, and it never cancels.
I deliberately did the opposite. The conversation stays human, and I use AI only where it is genuinely better than a person, which is the exhaustive analysis afterwards.
That is not a romantic position. It is a division of labor, and it is very precise.
The human owns pronunciation
A written analysis never hears your conversation. It only ever reads it.
Which means a transcript destroys pronunciation completely. Nothing written will ever show you that your dessus came out sounding exactly like dessous, or that your French u landed as ou, or that your “disabled” came out with a z instead of an s. The words sit there on the page looking perfectly correct, and the sound has simply gone.
Only an ear catches that, and only an ear that has been making those sounds since childhood. Your partner’s ear is the only instrument that catches your French, and yours is the only one that catches their English.
So in my method, the human partner has exactly one job during feedback: pronunciation. Five minutes at the end of the call, up to five words you mispronounced, each written inside the full sentence it appeared in. They show you the sentence first and let you try to self correct, then they model it, then they say exactly what was wrong. “The s is an s, not a z” teaches something. “No, it’s disabled” teaches nothing.
And they always explain why. Handing someone the right answer and nothing else has an effect size of 0.05, which is no measurable effect at all. Explaining why it is right takes it to 0.49. That is Steve Masson’s work, in Activer ses neurones, and it is the difference between feedback and noise.
The machine owns the fifty-four
Here is the number that made me change how I work.
I had thirty minutes of my own English analyzed. My language partner is attentive, generous and genuinely good at this. During that half hour she caught eight mistakes. When the recording was analyzed afterwards, there were fifty-four more.
That is not a failure on her part. It is what a human ear can do in real time while also holding a conversation and thinking about what to say next.
An AI analysis of the recording finds all of them, sorted by category and frequency. It even spots the one thing no human can, which is the structures you avoid entirely. If a tense feels risky and you quietly route around it every week, nobody notices, because absence is invisible in conversation.
So the machine gets the exhaustive part no human has the patience to do properly. Within twenty-four hours of the call, the transcript goes into Claude and comes back with three things: your uncorrected mistakes ranked by frequency, with three or four successes listed first; a two column CSV ready to import into Brainscape; and the whole conversation rewritten the way two natives would have had it. If you only ever do one thing from my method, import that CSV and work it ten minutes a day.
Why the conversation itself still has to be human
Accuracy is only half the argument. You will not keep talking to a chatbot for two years. You will keep talking to the person who actually wants to know how your week went.
An hour with somebody who is genuinely interested in your week is doing something no software does. It is also what makes a correction land, because being corrected by a friend and being corrected by a machine are not the same event. Krashen’s affective filter is not a metaphor: when you are anxious, the input does not reach the part of your brain that acquires language, even if you understood every word. Being tested raises that filter. Being liked lowers it.
There is also what only a native knows. An English speaker writes “Je suis très excité de vous rencontrer la semaine prochaine”. It is grammatically perfect, and no translator, app or AI check flags it. Only a French person will tell you that excité carries a sexual connotation and that what you wanted was j’ai hâte.
AI generated practice material has a specific failure mode I see constantly. Fabienne, one of the French speakers I work with, told me she did not feel like doing her flashcards, so I asked her to share her screen. Every new word from a podcast had gone to ChatGPT for four or five example sentences, and those had become her cards. The sentences were fictional. No memory attached, no emotion, no real context, and the brain does not retain that.
Use AI to correct the email you wrote in French. Do not write it in English and have it translated, because then you learn nothing. And never use it to check your daily sentence before you send it, because a corrected sentence you never got wrong teaches you nothing.
Language partner vs apps vs AI coaches: the comparison
| Duolingo, Babbel | AI speaking coach | LanguageMate language partner | |
|---|---|---|---|
| Who hears your voice | Nobody | A model reading a transcription of it | A native ear, every week |
| Pronunciation feedback | None you can use | Generic, and it never heard the z in your s | Five words per call, in full sentences, with a voice note model |
| The mistakes you make while fluent | Not measured | Some, in the moment | All of them, ranked, from the recording within 24 hours |
| Structures you avoid entirely | Invisible | Invisible | Visible in the analysis |
| Who chooses the topic | The app | You, from a menu | The two of you, from your actual lives |
| Cultural traps like excité | Passes without comment | Usually passes | Caught immediately |
| Why you show up on Thursday | A notification | Nothing happens if you skip | Someone is waiting for you |
| What you get after a year | A streak | Practice hours | A friendship, plus a deck of your own real sentences |
| Cost | Free to about 10 € a month | Subscription | 1,900 € (about $2,200) for three months, partner included |
A Duolingo notification is easy to ignore, because nothing real happens when you skip it. A language partner waiting for you at seven on Thursday evening is a completely different story.
Can’t I just find a language partner myself?
I’m not going to say it’s impossible. I did it!
I found Kate on italki and we talked every week for two years. I found Paulette by hearing her on a podcast and emailing the host for her contact details, and six years later she is one of my closest friends.
I also found a Canadian man on italki who sent me an article to read before our second call, and when I opened it, it was defending guns. I canceled. I gave Tandem a go too. When you are a woman on those apps you get a lot of messages from people whose goal is not language practice.
Rina spent several years on Tandem and HelloTalk. Every time it ended the same way. After one or two conversations the person disappeared, so she started again, introduced herself again, explained again why she was learning French, and a few weeks later the same silence. When an exchange did last, neither of them gave the other feedback, so they reinforced the same mistakes together. She is matched now, with Julie, and they later spent almost two whole days together in Paris.
Around 80% of the pairings I make work well. That’s my own estimate, from around 35 learners guided and more than 25 people matched since 2023.
So yes, it can work. It can also take months to find the right match. And finding someone nice, motivated and available is only the beginning, because after that you still have to structure the exchanges, give feedback that is actually useful, and reuse that feedback between conversations. That is the part I do.
Carolyn, who is based in Boston, put it better than I can: “My language partner and I communicate every day, not because we have to, but because we enjoy learning alongside each other. This is a benefit that you can’t get from any book or app.”
Frequently asked questions
Is Duolingo useless? No. It is a brilliantly designed motivational tool with a great interface, immediate feedback, spaced repetition and a very low price. It is simply built for translation and recognition, not for speaking. Keep it if you enjoy it, just don’t count it as speaking practice, and do try their podcast, which is better than their app.
Can an AI speaking partner replace a native speaker? Not for pronunciation, which is the thing most learners actually want fixed. A written analysis never hears your conversation, so it can never tell you your word came out with a z instead of an s. Use AI where it beats a human, which is the exhaustive analysis of the recording afterwards, and keep the conversation itself human.
How many mistakes does a human partner miss? I had thirty minutes of my own English analyzed. My partner caught eight mistakes during the call. The recording analysis found fifty-four more. That is normal, and it is exactly why the recording exists.
What does a LanguageMate week look like? One hour of conversation, split into two calls of thirty minutes so each language gets its own recording, with twenty-five minutes of talking and five minutes of pronunciation feedback in each. Then five minutes a day on WhatsApp, every day including weekends: two lines about your day in English, and one sentence in French your partner corrects in writing and reads aloud in a short voice note.
Do I need to be good at French already? You need to hold a basic conversation, roughly A2. Below that, build your listening first with comprehensible input on subjects you care about, thirty minutes a day.
How much does it cost, and what am I paying for? 1,900 € (about $2,200) for three months, with lifelong community access, in small cohorts. You are paying for the match, for the training of both people in the pair, and for someone to handle it when your partner goes quiet for three days.
What if my partner and I don’t click? Give a new pairing at least a month. If it still isn’t working, tell me and I find you someone else. That conversation is awkward coming from you and never awkward coming from me.
You already know an app is not going to make you speak. So what is actually stopping you from talking to a real French person every week?
Tell me in the comments, or send me a message and we’ll talk about it for thirty minutes.
Céline Your Language Mentor
