For thirteen articles I have been making the same complaint about Japanese learning products. They teach you the forms and never make you produce them. You recognise, you repeat, you fill in a blank, and at no point does anything require you to open your mouth and commit.
AI conversation apps fix that. They are the first category in this entire market that genuinely, unambiguously makes you speak Japanese, out loud, in real time, with no scheduling and no social cost. I want to be clear about that up front, because the rest of this article is going to be critical and I do not want the criticism mistaken for dismissal.
I am a native speaker living in Tokyo, and this is the fourteenth article in a series testing Japanese learning products against what the language actually demands. Here is the finding, and it is the strangest one I have arrived at:
These products solve the problem I have been complaining about, and in doing so they delete the thing that makes Japanese hard.
The variable that goes missing
Every article on this site has circled the same fact. Japanese makes you encode your relationship to the listener into the sentence itself. You cannot produce a Japanese sentence without having decided who is in front of you. Plain or polite. Honorific or humble. Whether the person you are talking about is inside your circle or outside it.
I found no product that makes you choose the politeness level for a specific person. None that makes you get the direction of respect wrong in an interesting way. None that puts you in the listening half of a conversation and requires you to hold up your end.
Now consider what an AI conversation partner is.
It is nobody. It has no rank relative to you. It does not work at your company or at a client of your company. It is not older or younger than you. It has no colleagues who will hear about this later. It is not going to decide whether you get the project.
So you can talk to it in whatever register you like, forever, and nothing happens. The one input that Japanese demands you supply on every single sentence has been set to null.
This is not a bug that a better model fixes. It is what an assistant is.
Why the corrections do not save you
The obvious objection is that these apps correct you. Several of them correct you very well. So even if the AI is nobody, surely it tells you when your Japanese is wrong.
It tells you when your grammar is wrong. That is a different thing, and the difference is the whole argument.
Large language models are trained to be agreeable. This is not a criticism of any particular product, it is a well-documented property of how these systems are tuned, and it is exactly the wrong disposition for the specific failure mode of Japanese learners. If you address an AI in casual Japanese when the situation calls for polite Japanese, the AI does not experience that as rude, because it does not experience anything. It answers your question. It may even mirror you, because mirroring the user is what these systems do.
Now compare that to a Japanese listener, which is the thing you are training for. I made this point in the article on casual and polite Japanese and it is worth repeating because it governs everything here: a Japanese person will not correct your register either. They will understand you perfectly, register that something was off, and say nothing at all.
So the human and the machine produce the same silence. But the human silently reclassifies you, and the machine silently does nothing. From inside the conversation the two are indistinguishable, which means an AI conversation partner can train you for months into a habit that is quietly costing you, while giving you a stream of green ticks about your particles.
That is the core of it. The feedback profile is inverted. These tools correct the errors a Japanese listener would forgive and stay silent on the error a Japanese listener actually notices.
The second thing that goes missing: the timing
Japanese conversation has a rhythm that is not optional, and I wrote a whole article about it. In textbook Japanese versus real Japanese I laid out the research: Japanese speakers backchannel roughly twice as often as American English speakers by one count, fifteen to twenty times a minute by another, and in Komiya 1986 once every 6.1 seconds on the telephone. The listener has a job. Japanese sentences have slots in them where the speaker stops and waits for you to acknowledge before continuing.
An AI voice conversation does not work this way, and the reason is technical rather than cultural. The system waits for you to stop making noise, decides your turn has ended, and replies. Your silence is its cue.
Two consequences, and the second one is worse than the first.
One: it does not need your backchannels. You can stay completely silent while it talks and it will never once falter, never repeat itself, never conclude that you failed to understand. A Japanese speaker in that situation would start rephrasing within about ten seconds.
Two: it punishes you for speaking Japanese properly. If you talk the way a Japanese speaker actually talks, breaking after 〜んですけど or 〜でね and waiting for the listener to acknowledge, the system reads that pause as the end of your turn and starts talking over you. To be understood, you learn to deliver complete uninterrupted sentences at an even pace. That is a real behavioural change, it is trained by repetition, and it is the opposite of what a Japanese conversation requires.
I want to be fair about the state of the technology. The newest voice models are full duplex, meaning they can be interrupted mid-sentence, which is genuine progress and removes the worst of the old walkie-talkie feeling. What I have not seen is a system that requires your backchannels, or that notices their absence, or that produces them at Japanese frequency while you speak. Being interruptible and holding up the listener half of a Japanese conversation are not the same capability. I have not tested every model in Japanese and I would be pleased to be shown one that does this.
The third thing: do not shadow the robot
A smaller point but a practical one, because it costs people real accuracy.
I recommended shadowing at some length in the article on Japanese shadowing practice, and the whole method depends on the audio being correct. Japanese text to speech is very good and is not perfect. The failure cases are specific and predictable: proper nouns, personal names, place names, and compound words where the pitch pattern of the compound is not the pitch pattern of the parts. Kanji with multiple readings occasionally come out with the wrong one entirely.
You are a learner. You cannot hear which lines were correct.
I said in the article on pitch accent that native speakers do not perceive pitch as pitch, which is why we cannot teach it to you, and that the tools which do handle it are built by specialists on curated data. Synthetic voices are for conversation, not for imitation. Shadow recorded human audio. Talk to the AI.
And the fourth: the microphone is not a Japanese person
Every one of these products offers pronunciation feedback. Understand what is being measured.
Speech recognition is deliberately engineered to be robust to accented speech. That is the entire commercial value of it. Which means it will cheerfully transcribe Japanese produced with English stress timing, wrong mora lengths, and flat pitch, and then tell you that you were understood.
You were understood by a system built specifically to understand people who talk like you. In the shadowing article I argued that mora length is more consequential than pitch, because it distinguishes actual words: おばさん against おばあさん, ビル against ビール. An ASR engine will resolve those from context without noticing you got them wrong. A person on a phone line will not.
Treat a high pronunciation score as evidence that you were intelligible to a machine. It is not nothing. It is not the claim you want.
What these tools are genuinely good at, and it is a lot
I have now spent two thousand words on the limits, so let me be equally direct about the value, because I do recommend using them.
In the shadowing article I split the problem of understanding Japanese but not being able to speak it into three parts: retrieval speed, the physical business of running your mouth on a different rhythm, and fear. AI conversation is close to a complete answer to the third one and a very good answer to the first.
Fear. The reason most adult learners cannot speak is not that they lack the words. It is that a person is looking at them. An AI is not a person, and here that is the feature. There is no cost to freezing, no cost to asking it to repeat itself for the fifth time, no cost to your accent, no cost to a sentence that fell apart in the middle. For somebody who has studied for two years and has never once said a Japanese sentence to a living creature, this is the cheapest and most effective intervention available anywhere in this market, and I am not aware of a substitute.
Repetitions. Retrieval speed improves with the number of times you have retrieved. An AI will run the same restaurant scene forty times and be exactly as patient on the fortieth. No human teacher will do that, and you would feel bad asking.
Availability. Eleven at night, no booking, seven minutes.
Saying the thing you actually wanted to say. You can stop, describe in English what you were trying to express, and get the Japanese for it. That single loop is worth a lot and no textbook offers it.
If you use these tools for those four things and nothing else, they earn their money.
How I judged them
1. Can it be somebody? Can it hold a stated social position toward you, with stakes, rather than being a friendly nobody?
2. Does it react to register error, or only to grammar error? This is the one that matters and almost nothing scores on it.
3. Does the conversation have Japanese timing? Does it need your backchannels, or notice their absence?
4. Is the audio safe to imitate? Synthetic or human.
The comparison
| Product | Cost | Can hold a social position | Handles register | Can it pay me? |
|---|---|---|---|---|
| Gemini Live and ChatGPT voice | Free tiers exist. Paid tiers from roughly $20 a month | Yes, if you instruct it. The only ones here that let you | Only if you demand it | No |
| Langua | $19.99 a month standard, $29.99 unlimited, less on annual | Roleplay scenarios, not arbitrary instruction | Not mentioned on the official page | Yes. $30 per referral or 8.5 percent lifetime. The highest here |
| Pingo AI | Free tier, paid Pro | Roleplay scenarios | Advertises です and ます, keigo, and casual against formal register. The only one that does | No program found |
| Speak | Roughly $80 to $200 a year | Structured lessons, not free instruction | Not advertised | No program found |
| Talkpal | Subscription | Roleplay scenarios | Not advertised | Yes, 30 percent — I applied and was declined |
| Hanashi | Not published | Characters with fixed personalities | Not advertised | No program found |
| SakuraSpeak | Conversation is behind the paywall | Roleplay and story mode | Not advertised | No program found |
| NinjaTalk | $99, $299 and $599 in-app purchases | Scenarios, JLPT N5 to N1 claimed | Not mentioned anywhere | No program found |
| italki and Langua human tutors | $10 to $40 an hour | Yes. Actually yes | Yes, if you ask | Yes. At least $10 for italki |
Look at the last column. The thing I would put first is free and pays me nothing. The product with the biggest commission on offer is second, I am not in that programme, and I have things to say about it. That has been the shape of every article in this series, and this time I did not have to arrange it.
The general assistants — my first recommendation, and they pay me nothing
Gemini Live works on a free Google account. ChatGPT has a free voice tier with limits and better models above it. If you have a subscription to either for other reasons, you already own the best Japanese conversation partner discussed on this page.
The reason is not that they are better at Japanese than the dedicated apps. It is that they are the only ones you can instruct.
Everything I have said about the missing social position is a prompting problem in disguise. A dedicated app hands you a menu of scenarios written by somebody who does not know you. A general assistant will take a paragraph of instruction in Japanese and hold it for the whole conversation. That is the difference between practising with a friendly nobody and practising with a specific person who has a rank.
Here is what I would actually paste in. Adjust the details to your own situation.
あなたは私の会社の課長です。私は入社1年目の部下です。
私が友達に話すような言葉づかいをしたら、その場では直さずに、
上司として自然に反応してください。
会話が終わったあとで、言葉づかいについてだけまとめて教えてください。
The important clause is the middle one. Left to itself the assistant will break character to be helpful, which destroys the exercise. Telling it to stay in role and save the feedback for the end is what produces something resembling a real interaction.
A second one, for the situation I described in the article on business Japanese, which is the single highest-value thing you can drill this way:
あなたは取引先の担当者です。私の上司の高橋部長について私に質問してください。
私が自分の上司に敬称や尊敬語を使ったら、そこだけ最後に指摘してください。
And a third, which is the closest I have found to practising the listener half:
私が話しているあいだ、日本語の会話らしく相槌を打ってください。
私が話し終わるまで、内容には答えないでください。
Two warnings, and I would rather you heard them from me than discovered them after a month.
It will still be too nice. Test this before you trust it. Deliberately use casual Japanese with the boss character and watch what happens. If it sails past without a flicker, the role is decorative and you should say so more forcefully in the prompt.
Its Japanese is not neutral. Assistants trained mostly on text produce written Japanese read aloud, which comes out as complete, well-formed sentences with every particle in place. Real speech drops particles, trails off, and leaves half the sentence to the listener. That is exactly the textbook problem I wrote about in textbook Japanese versus real Japanese, reappearing in a new technology. Use the AI to build fluency. Get your model of how people actually talk from people.
Langua — the best of the dedicated products, and the biggest commission I am not collecting
I have to declare this plainly: Langua pays affiliates $30 per referral or 8.5 percent lifetime commission, which their own comparison page correctly notes is three to five times what italki pays. It is the largest commission of anything on this page. As of 4 September 2026 I hold four approved affiliate programmes, one with JapanesePod101, two with knife retailers, and one with The Japan Shop, a bookshop approved on 4 September 2026. None of them is a product on this page, which carries no tracked links and earns nothing, but you should still weigh what I say here accordingly.
What it does well. Of the dedicated products this is the most complete. Conversations in several formats including roleplay and debate, corrections you can set to subtle or explicit, transcripts, saved vocabulary that goes into spaced repetition, feedback reports after a call, voices cloned from native speakers, and a fallback to your own language when you are stuck. If you want the conversation practice packaged rather than assembled, this is the package.
What I have to report. The official Langua page for the Japanese AI tutor makes no mention of keigo, of politeness levels, or of casual against formal register. Not a word. For most of the twenty-plus languages it supports that omission would be unremarkable. For Japanese it is the omission. I also saw a secondary source describe Japanese support as beta and I could not confirm that on the official site, so treat it as unverified rather than as fact.
Pricing is $19.99 a month for the standard plan with a daily cap of about 30 minutes of calls, and $29.99 for unlimited, with substantial discounts annually. There is a free trial and a money-back guarantee on web subscriptions.
My honest read: worth the money if you will use it daily and want structure. Not worth it over a general assistant if what you want is a partner you can instruct.
Pingo — the only one that names the problem
Pingo is the single product in this comparison whose own marketing mentions です and ます, keigo, and the distinction between casual and formal register. Everybody else is silent on it.
That deserves credit and it deserves a caveat of exactly equal size. Advertising that you cover register and requiring the learner to choose it under pressure are different things, and the difference is precisely what this entire series has been about. I have not paid for Pingo and I have not verified what it does with register in practice, so I am naming it as the one to look at first and explicitly not telling you it works. If you try it, the test is simple: address it wrongly on purpose and see whether anything in the conversation changes.
The rest, briefly
Speak has the most serious pedagogy of the products here, built around a learn, drill, then apply loop, and Japanese is supported. Roughly $80 to $200 a year depending on plan and region. No affiliate programme I could find.
Talkpal pays 30 percent, confirmed on their own affiliate page, which is the second largest commission here. I applied to that programme through Impact and was told on September 3, 2026 that my application had been declined, so it pays me nothing. It is a general multi-language conversation product with no Japanese-specific claims that I could verify.
Hanashi is made by Studio Mirai Limited, founded by somebody who moved to Tokyo in 2025 and could not find speaking practice. Characters with distinct personalities and voices. No published pricing and nothing about register. I mention it because a small team that lives here is more likely to fix the Japanese-specific problems than a platform serving twenty languages, and I would keep an eye on it.
SakuraSpeak puts the conversation feature behind the paywall, which makes it hard to evaluate before paying.
NinjaTalk advertises in-app purchases of $99, $299 and $599. For that money you can buy roughly forty to sixty hours with an actual Japanese teacher who can be your boss, correct your register, and remember you next week. I would not.
What I would actually do
If you have never spoken Japanese to anything: Gemini Live, free, ten minutes a day, for two weeks. The only goal is that speaking Japanese out loud stops feeling like an event. Do not evaluate your progress. Do not pay for anything yet.
If you can already hold a slow conversation: keep the AI for volume, and add one human hour a week. That is the split I would defend against any other allocation of the same money. The AI gives you the repetitions. The human gives you the one thing the AI structurally cannot, which is a person with a position who reacts to how you addressed them.
If you are learning Japanese for work: use the role prompts above every single day, with your real job in them. Your manager, your client, your team. This is the highest-value use of AI conversation I know of, because the business situations are exactly the ones where the register decision is scored and where you get no practice reps in real life, since every rep at work is live.
If you want one paid product and no assembly: Langua. Declared interest above.
The bottom line
AI conversation apps are the first products in this market that make you produce Japanese, and I have been asking for that for thirteen articles. They deserve the credit.
But the thing that makes Japanese difficult is not producing sentences. It is that every sentence carries a claim about who you are to the person in front of you, and a machine that has no position and takes no offence cannot ask you to make that claim. It can only ask you for sentences.
So use them for what they are extremely good at, which is turning silence into speech, and do not mistake the fluency you build there for the ability to talk to a Japanese person. Those are two different skills, and the second one still requires the one component none of this software includes, which is somebody who is actually there.
This is the fourteenth article in a series testing Japanese learning products against the language as it is actually used. If you are choosing your first tool, start with the main comparison. For why speaking is hard even when you understand, see shadowing practice. For the listener half of a Japanese conversation, see textbook Japanese versus real Japanese.