Launch HN: Issen (YC F24) – Personal AI language tutor

306 points by mariano54 ↗ HN
Hey HN, we're Mariano and Anton from ISSEN (https://issen.com), a foreign language voice tutor app that adapts to your interests, goals, and needs.

Demo: https://www.loom.com/share/a78e713d46934857a2dc88aed1bb100d?...

We started this company after struggling to find great tools to practice speaking Japanese and French. Having a tutor can be awesome, but there are downsides: they can be expensive (since you pay by the hour), difficult to schedule, and have a high upfront cost (finding a tutor you like often forces you to cycle through a few that you don’t).

We wanted something that would talk with us — realistically, in full conversations — and actually help us improve. So we built it ourselves. The app relies on a custom voice AI pipeline combining STT (speech-to-text), TTS (text-to-speech), LLMs, long term memory, interruptions, turn-taking, etc. Getting speech-to-text to work well for learners was one of the hardest parts — especially with accents, multi-lingual sentences, and noisy environments. We now combine Gemini Flash, Whisper, Scribe, and GPT-4o-transcribe to minimize errors and keep the conversation flowing.

We didn’t want to focus too much on gamification. In our experience, that leads to users performing well in the app, achieving long streaks and so on, without actually getting fluent in the language you're wanting to learn.

With ISSEN you instantly speak and immerse yourself in the language, which, while not easy, is a much more efficient way to learn.

We combine this with a word bank and SRS flashcards for new words learned in the AI voice chats, which allows very rapid improvement in both vocabulary and speaking skills. We also create custom curriculums for each student based on goals, interests, and preferences, and fully customizable settings like speed, turn taking, formality, etc.

App: https://issen.com (works on web, iOS, Android) Pricing: 20 min free trial, $20–29/month (depending on duration and specific geography)

We’d love your feedback — on the tech, the UX, or what you’d wish from a tool like this. Thanks!

281 comments

[ 2.7 ms ] story [ 327 ms ] thread
>We didn’t want to focus too much on gamification.

Thank you so much for this. Duolingo is literally unbearable because it's so gamified. I'll try it out later. I've seen a few of these apps, can I seamlessly go between my native language and the language I'm trying to learn? If I am trying to learn Hindi, can I ask a question in English in the middle of a conversation?

Yes, we've spent a lot of time getting the STT and TTS to work seamlessly in multilingual, it works pretty well!
I said I would use it and report back. Unfortunately, the experience was not great for me. I told it multiple times I'm a beginner and not speak in full, long sentences in my target language, and it would remember that for one response maybe, then switch right back to full sentences. I had to keep telling it I couldn't understand it and that I'm a beginner. After the 4th or 5th time I gave up.
The app is optimized on the whole population, not on individual level. They even publish papers on global optimization.

These kinds of learning apps are destined to become mediocre over time.

The learning metric is so easy to capture, the learning content so easy to produce, yet no one has an individualized loop to make learning work well.

For example, I'd press "Training" on Duolingo, and would get nowhere. Same lessons all of the time. Bread and water.

Duolingo is laughably inefficient. I have an app I've made for myself to learn french, and it's amazing how little effort is required to make something 10x better. (I completed the french Duolingo tree and learned essentially nothing so I feel justified in saying that.) If you're learning french, let me know and I'll add you to the app.
The real question: the 10x, can you mainstream it ? Or is it only 10x better per your preference ?

Btw: if the app is available, I would love to take a look!

Yeah great work with this. Seems like a real opportunity given how hard Duolingo is dropping the ball.
This actually looks pretty neat. How have you been able to achieve such broad language support so quickly?

How widely have you tested your supported languages on native-speakers and learners?

The STT and LLM support many languages out of the box. For TTS we use multiple providers based on their strengths and weaknesses (for example minimax is great for Chinese)

We've done a lot of testing on Spanish, English, Italian, Japanese, and French, but much less on the others and none at all for some of the niche ones.

The language support is based on the intersection of the languages that have low word errors rates in the transcribers, as well as officially supported by LLM/TTS (like gpt4.1, eleven labs etc).

We've seen the models' quality improve consistently over the last 6 months, in all languages we tested, and now the error rates are getting really low.

Right - I think it would be appreciated by your users if you at the very least made it clear from the outset how well different languages are supported and what degree of testing you have done.

Certainly if your product were to mis-teach me important details, and I were to then find out that you had spent less time testing than I had spent learning, I would be quite angry.

I tried the Web Version. Started, then tried to create an account, but it kept looping, informing me that my email address does not exist in your system. Well, the “Create New Account” got kicked off and gets me in a loop of “Do not Exist”. I just went through the whole process again, and I'm back to the beginning.

I’m going to assume this works better on the App.

Unclear what issue you hit, we'll look into it. Thanks for sharing.
Played around > tried creating an account > says email not in system > can neither create nor sign in. So, I don't have an account even if I wanted to create one.
I appreciate your comment about gamification. I’ve kept a streak alive on other apps for no other reason than keeping a streak alive. Not learning a thing.
Yeah, this is the biggest gripe we hear about much of the existing language learning landscape. That they're effectively gaming apps masked as language learning apps.
You should still gamify it. Gamification is orthogonal to whether the tool actually works and positively correlated with whether the user actually uses it.
I would argue that games are a great analogue to language learning as well. Contrary to our ideals, people do like to enjoy themselves and are more likely to pick an activity they enjoy than one they don't. Games and puzzles are able to present frustration as enjoyment (provided there is appropriate reward and perceived growth) making them great tools for learning.

However, gamification can only do so much and I'm afraid language learning is a lot like learning to code: many people want to want it but few actually want it. In that case, presenting as a "want it" when you are a "want to want it" is social proof and largely unrelated to whether you are actually learning (as long as the pretense is kept up) — hence the success of Duolingo despite the relatively poor real-world outcomes. In Duolingo's case the streaks are even explicitly considered to be social proof.

Looks great. I have been looking for something like this
Luis von Ahn spoke in the early 2010s—probably around 2014—at The LAB in Wynwood, Miami. He recounted how his fascination with crowd-sourcing led first to reCAPTCHA and then to his latest venture, Duolingo. He made it clear that his real passion wasn’t language per se, but building a crowd-sourced human translation service as a business model. At that point, Duolingo had roughly 24 employees—and, much to his surprise, only two were focused on the crowd-sourcing engine. He explained how they’d enlisted some of the world’s leading language-education researchers as consultants. Their very first question: “Which part of speech should learners tackle first?” The experts confessed they didn’t know, so the team gathered the data and used A/B testing coupled with statistical analysis to pinpoint the answer.

Today, it’s not only easier than ever to launch a platform to challenge Duolingo, but its core product—its crowd-sourced human translation service—has been distrupted.

This morning, I found myself thinking about how all those decade-old learning platforms—like Coursera, as reflected in its ever-falling stock price—are being distrupted.

Your product looks awesome and I hope you distrupt all the language learning platforms. Thank you for sharing.

(I had ChatGPT fix my grammatical errors and now this comment doesn't sound like me, sorry.)

> (I had ChatGPT fix my grammatical errors and now this comment doesn't sound like me, sorry.)

And it didn't correct "distrupted" to disrupted?

> https://imgur.com/a/m3svUky

That's an image of text. Is it supposed to provide more evidentiary value than the word "Nope" above it would by itself?

I'll bet you I can show you a screenshot where "ChatGPT" says whatever your heart desires.

I started using o3 yesterday. Turns out it has sycophancy behavior. That is good to know. Regardless, I'm beginning to enjoy my errors because it reminds me of my humanity.
Coursera is failing because its platforms are infested with Big tech cert slop.

And mid 2010s view was MOOCs were supposed to disrupt University education!

Add it to the pile.

Do you store conversations? And what's the general privacy philosophy behind the app?
We store the messages, but not the audio. We also store session summaries and a "user facts" summary that gets regenerated after every session, based on all session summaries, everything in our AWS DB.

You can delete your account at any time to fully wipe all your data, but there is no way to delete sessions ATM.

Those FAQ boxes on the main page don't expand?
I can't wait to try this! I studied a few languages in school and have lost any semblance of proficiency -- mainly because I never have a real occasion to use anything other than English. I've been waiting for someone to build something like this
Why not use the Gemini flash voice-api directly instead? Cost? I ask because from the demo, the tutor's voice seems mechanical. I've played with the gemini voice api and it's quite impressive for conversation with low latency, I'd say perfect for your use case. It even switches languages if I say "Okay, let's talk in $foo language".

The vocabulary tooling looks neat and well thought out.

Multiple reasons (which also apply to openAIs realtime API): - it's less intelligent than the non voice apis - intelligence degrades even further with lots of context - more expensive - latency is not a free lunch, it comes at the cost of more interruptions from the tutor, which is a really bad UX. We prefer to interrupt less and have higher latency

Also, we prefer the eleven labs voices, but there is definitely varying quality. I'm guessing later this year or next, the voice to voice models will become good enough, and we will switch over.

Can't wait to try it; my kids need to learn French in school and I've been trying to keep up with them with Duolingo; but something is missing there.

For me a key feature will be a family plan; Duolingo is great in that regard.

Speaking of translation with LLMs I've been looking for a solution to quickly open a bi-directional translation context without having to prompt ChatGPT or any other LLM every time. iOS lets you set the action button to use the default translation app quickly, but the translation it provides is vastly inferior to LLMs.

Even some basic app that can pre-load the prompt doesn't seem to exist?

Thanks for working on this! Language learning really needs a breakthrough.

Now, I tried the web app and chose to learn Greek as a beginner. And while I had better experience with your app than with ChatGPT or Gemini voice modes, I still got lost 5 minutes in because the AI tutor doesn't seem to have a plan for me, nor does it "see" my struggles. For example, after asking me about a hobby, it gives me a long sentence in Greek about how how it is nice to hike in mountains. Being absolute noob I cannot reply to it, nor even repeat it. And I don't even know what it is expected from me at the moment. A human tutor here would probably repeat a part of the sentence with a translation and ask me to repeat, or would explain something. The AI just sits there waiting for me to make a sound, and when I make it, it goes on on a tangental subject of beach vacations. :)

Again, this is still relatively not bad, and I'm going to give it another try.

I had a similar feeling with Swedish just now. It isn't really much different than conversing with ChatGPT in advanced voice mode - it's up to me to drive the conversation and it all feels quite arbitrary (and I find myself instinctively falling back on topics I know how to talk about, which quite defeats the purpose). I was hoping for a more structured learning plan that strategically expands my comfort zone and skills in a guided way.
Thanks for the feedback. Yeah we need to improve the beginner experience, it's more tailored towards intermediate/advanced students at the moment.
Do you mean that the experience is meant to have more structure if you pick the intermediate or advanced level? (fwiw I did pick intermediate for my Swedish level in the app).

My thinking is - I can have unstructured conversations with Advanced Voice Mode or in real life here in Sweden. What I'd really appreciate is a guided learning experience taking me up from intermediate/slightly above intermediate to fluent in the most efficient possible way (as opposed to just having us 'ramble' about random topics of my own choosing).

There is a structured curriculum that gets generated after the intro lesson (if you responded yes to the curriculum question).

This is available for all proficiencies. It's just much harder to talk for hours in a new language as a beginner. It's usable but requires more effort.

I see it now, thank you! This looks like what I was hoping for. I wonder if there's some way to communicate very clearly that you'll need to talk for around n minutes to get a structured curriculum prepared - and maybe even show a progress bar of some sort so the user can have an idea of when they're going to get to "the good stuff"?

Another thing is that the trial period seems incongruent. To me the structured curriculum is what I really want to _try_. I want to see what the planned lessons are like, how guided they are, etc. But the trial runs out and tries to make me pay right after the unstructured more all-over-the-place feeling introductory conversation, and I'm not prepared to pay at that point since I feel like I haven't gotten to evaluate the main part of the product at all. I would suggest leaving the trial unlocked to maybe the first three structured lessons of the learning plan. Let the user really experience what they'll be paying for.

So basically if you are starting a new language from zero, then this is not for you?
Why wouldn't intermediate/advanced students just talk directly to ChatGPT? From what I see, I thought your value prop was for the beginners.
ISSEN is designed from the ground up for this use case.

* curriculum, completely customizable, with grammar, roleplay, topics, speaking speech, transcript, dictionary, corrections, etc

* prompting and AI models all chosen to be a better fit for multilingual, easy to understand, etc.

* the tutor actively tries to teach you, it's not an assistant

* integrated flashcards that go hand in hand with the speaking immersion

I'm an advanced learner but I stopped after a few moments because it's boring. It's asking me questions that you'd ask a beginner (although a beginner wouldn't understand the questions). It just asked what food I like to eat, where I like to travel, whether I like the weather, etc. I have a language tutor IRL and I have found that we run out of things to talk about too. So we often find ourselves just discussing the latest events from the news. I think you should feed fresh conversation topics daily from a data source like the news, localized to the user. There are global news APIs you can subscribe to.
I'm not sure that would solve the problem. Ultimately this (and speaking with LLMs in general) feels a lot like filling in an adaptive form rather than talking to a human being. The LLM, for example, is never going to go on and tell you some anecdote from its life (and if it were prompted to do so it would come across as quite insincere). It's not going to say, "Oh that neighborhood is not safe," when you tell it about where you'll live when studying abroad. It's not going to recommend supplementary material (or worse, it will...with broken links that never existed).

I'm not even sure this helps with speaking practice since it's just a test of whether what you said can be transcribed by Whisper, which is not at all a test of correct pronunciation. I just tried it with the most horrid, butchered accent I could muster, and it still worked...if I practiced for months on end like that, I'd end up in a very difficult place as a language learner.

As an intermediate German speaker I thought it was great!
I think this is a pretty big limitation of the architecture (STT->LLM->TTS) they've chosen. The intonation around struggling to speak or difficulty with certain phrases is totally lost when the text is transcribed.
I think the point here is for you to practice (i.e. develop "muscle memory" for speaking), not to learn.
As far as using language goes, those aren't different things.
I paid for Memrise to polish up French. The scripted lessons alwere great but it dropped me into an AI conversation assistant that did exactly the same. It forgot the vocab and grammar level that the scripted lessons had taught, and often broke into idiom. I haven't picked it up since.
I'm a Memrise beta member w/ lifetime premium access for my contributions to the site in its early days. I cannot recommend anyone use Memrise for anything nowadays it has been so heavily enshittified. In fact, I recommend against using it in favor of Anki (Memrise's biggest strength over Anki in the early days was the community mnemonics and courses (Anki equivalent "community decks") - none of which really exist in any way today).

I tried following the modern Japanese track on Memrise and was appalled at how bad it is nowadays.

Funnily enough I said my native language is Greek but then it responded with an error and reset my onboarding guide. Then, I lied that my native language is English, which worked. But now it calls me Anton, rather than the name I said I have!
I also got Anton. Looks like something's hard coded - or maybe a caching issue?
Language Transfer[0] will continue to be a better resource than any AI course. It’s very hard to beat a human that has put in time crafting a logical way to teach a language with the appropriate ramp ups. The Greek course on there is fantastic. And it’s free with zero ads. Best language learning tool I’ve ever used period.

[0]: https://www.languagetransfer.org/

What's the best way to listen to this on your mobile in a way that will remember your location? SoundCloud app?
They have their own app. It’s pretty minimal but it does save your spot.
Agreed, languagetransfer is fantastic. Much better than any "AI tutor".
Agree, it was a game changer for me with Spanish. Learning a second language is just plain hard and LT is the closest to “a breakthrough” I think we will get, but it’s still hard and people don’t like that.
This looks great, congrats! As someone that has gone through Assimil courses and done lots of comprehensible input for various languages, language production is typically the weak point that isn't covered well. I've done plenty of lessons on iTalki, but I've been wanting something more structured and this seems like it could cover it. Definitely going to give it a shot!

The feature request I make for all language course makers: please consider Bengali support in the future! It's wild to me that the 7th most spoken language in the world, with a deep culture around literature and poetry [1], gets zero attention from language course makers. I can buy an Assimil course on Breton, spoken by 200k people, and not Bangla, spoken by 284 million.

[1] https://en.wikipedia.org/wiki/Charyapada

I had a go and it failed at the first hurdle I’m afraid. It was hallucinating my responses and inserting phrases that I wasn’t saying.

The teacher kept switching into an American accent when I was trying to learn French and the responses were getting very slow bitty.

Hopefully this is just an initial load of issues because the concept is great.

Yes, this is weird. I said something like "I want to learn Spanisch to go shopping" and it just added "That's great" to my sentence.

Also for the flash cards the audio that says the word starts at the same time with the audio for the example sentence.

I'm trying to learn vietnamese, but the lessons are really really rough and borderline bad advice.

---

AI: Anh mệt is good if bạn are a man speaking about yourself. You can also say, “Em mệt” if you’re a woman.

this isn't correct. If you are of "older brother" age and are male, you say Anh. Em is for if you are "younger person" (does not matter the gender). Women tend to prefer being called "em" (even if they are older), because women prefer to be identified as younger than their true age... But that doesn't mean you can't call younger men em.

A good tutor would know your age relative to theirs and explain this context.

---

It would say english phrases with a vietnamese accent.

---

It also would give me really complex vietnamese phrases that I am not ready for. when I prompt for an explaination or translation, it would get off track from the original thing we were learning.

---

Way more people in Vietnam (and the globe) speak southern Vietnamese, but the tutors seem to be from north Vietnam.

---

The STT also was very forgiving if I pronounced things incorrectly. Or it would confuse english and vietnamese. I would say, "Phai", but it heard "bye"

---

I was ready to pull out my credit card, but I can't trust it to teach me the right information. I pay $160/mo for Vietnamese tutoring ($20 per class). This would be way cheaper and I don't have to schedule my classes.

This sounds very much like the kinds of mistakes that LLMs typically make. It's a pity, I would love a good language learning platform.
A fundamental problem with language learning built around an LLM is that the one thing you can guarantee is that no two people will have a consistent experience, nor is there ever going to be a 100% freedom-from-error. That makes it very hard to predict and therefore market what or how people will learn.

I think this company will end up pivoting into a B2B context before long. Hopefully they will still stick to the mission, but who knows (and I wouldn't fault them if they don't – survival comes first).

The trend I've seen in these AI tech companies is they launch their MVP using base models (or in this case fine tuning gpt4). This gives them enough traction for a seed round, but 2+ years later, they don't have the talent to actually improve the product beyond this.

If OpenAI puts resources to language learning, they could build a great product. But 3rd party devs relying on someone's tech hasn't proven to be a good strategy.

> nor is there ever going to be a 100% freedom-from-error

That is not a problem. Language is messy, you don't need 100% accuracy to learn. The problem is that LLM errors are fundamentally different from human errors, and you won't even know how to recognize them.

Your interlocutors can work around human errors, because they tend to follow the same patterns in same language. But they will freak out with LLM errors.

We are really sorry for the subpar experience. We did not test Vietnamese, and it seems like the quality is not sufficient. Thanks for pointing out the issues.
Hey, it’s great to see other people learn Vietnamese! And your feedback is on point. I’m still at the beginning and just built a tool to help me learn basic phrases. I will very much appreciate the feedback! https://envn.app
I haven't figured out what works for me yet when learning Vietnamese, so I'm not really sure yet was advice to give.

Trying out your tool, I'd really like to know if the sound is north or southern Vietnamese. I think your tool is southern vietnamese, but idk.. I personally prefer learning southern, but all the AI TTS tools use the north dialect. Ideally, I'd like a 'pure' southern accent and not a hybrid.

For your tool, You might want to get into the way to address people (Anh, em, ba, co, etc). You seem to just use toi (which I hear vietnamese people using with each other too...) but my understanding is the (Anh/Em/Ba/etc) are more 'intimate' whereas toi is more formal/business like?

One idea I haven't tried too much of yet is making flash cards that teach me a sentence structure, but introduce new vocabulary. Learning a diaspora of phrases works for short 2-3 word ones, but when I try to learn more complex sentences, my brain isn't able to draw the patterns as nothing is connected.

For example, trying to learn "bạn tên là gì" and "nhà vệ sinh ở đâu" (from your website) is harder than learning "Bạn tên là gì?", "Bạn nghề là gì?", "Bạn số điện thoại là gì?"

The other huge challenge I have is feeling like I am making progress. I'm definitely getting better, but its pretty disheartening to study for 40+ hours and still can't pronounce words like Can Tho properly, despite knowing how to read and write.

---

My email is in my profile. Feel free to reach out to me if you have more updates or want to bounce ideas.

how're you handling latency on turn overlaps : buffered stream with early intent cutoff or full duplex with partial decoding?
We transcribe after 400ms of silence in 200ms chunks. 3 voice chunks (VAD) automatically interrupts, unless it's a back channel like "yeah" or "right" or something like that.

Whisper can transcribe in <100ms. We then wait for the turn detection model, LLM, and tts to trigger a streamed response back to eh client.

Hopefully people that use these AI language tutors don't end up being clowned on by native speakers because they sound like robots.
I'll try it, but that seems pricy compared to a Duolingo subscription. And while I understand that they are different, will your average lead know that?
This s super interesting! i have been wanting to learn other languages, but it i have been unsatisfied with most mainstream solutions. From what i have seen and for the price, i could see myself giving this a shot!
Edit: Never mind it seems to be an issue on my device.

The faq wont expand on tap for me on android firefox. Dm me if you need more info.

Looks like a great app and I can't wait to try it for Japanese!

Can the cards be exported to anki?

Exporting to Anki is something we're currently discussing, as various users have already requested this feature!
Congrats on the launch!

I tried the Japanese track. I'm a total beginner and the first lesson wasn't helpful at all. The AI asked about maybe mixing up Japanese<>English, but it didn't actually follow through. It either spoke fully in Japanese or fully in English. Maybe this is a standard practice for language lessons? I remember going to the first day of French class in a community college, and the teacher only spoke French, which was extremely overwhelming. Perhaps it's the standard way of teaching? Even if it is, I'm not sure if it works when compressed down to the shorter times I see myself opening the app.