Launch HN: Univerbal (YC W23) – Language learning with a conversational AI tutor
We’re an AI language learning app where you talk to your AI Tutor, just like you would with a human one. Here’s a short demo of what that looks like in an actual conversation: https://www.youtube.com/watch?v=IukKGc00juY
We actually started as a Show HN over a year ago and the responses we got from the HN community led us to apply to YC and make an actual company out of a side project (original Show HN https://news.ycombinator.com/item?id=32993130). We were called Quazel which we have since realized is too hard to spell and only makes sense in German.
We started off as only offering open-ended conversations. While that is still at the core of what we do, we quickly realized that we need to solve a bigger problem for learners than simply giving them the option to hold a conversation with an AI. Tutoring is (usually) the highest quality of education you can get. The curriculum is based on your interests and needs, you move at your own pace and you get personalized feedback about your specific mistakes.
At the same time, the reality is that tutoring is prohibitively expensive for most people, let alone that it can be very time-consuming to find a tutor that you like and you have to work around their schedule. We want to fix these issues by providing an AI Language Tutor that’s available whenever you want and affordable to everyone.
Now the tricky thing about this, is that a human tutor (to no one’s surprise) actually does many things at the same time, which are very much not straightforward to approximate, even with AI. The interaction part of a tutor session is already the main interaction in Univerbal, through speaking and having the replies read out loud. The hard part intervenes when you try to replicate the progress tracking and tailoring of the curriculum.
These are things that tutors automatically do, and getting a system, even one that’s based on LLM to do something like that is very hard! Our current approach is based on “Skills” (e.g. I can introduce myself), that a user works towards. These are a measure of where the user currently stands and we use this progress in a feedback loop to come up with relevant and interesting next lessons for a learner.
I often get asked, “Ohh so is it like Duolingo?”. When I get that question, I smile, feel something inside me die and then say that, while Duolingo is a great language-learning app, we don’t really see them as our competition, rather we see online tutoring platforms (italki, verbling, cambly etc.) as the companies we’re “attacking”.
We already have a couple of success stories. One of which comes from an Australian user who’s living in Italy and has successfully prepared herself for her Italian B2 language exam by using Univerbal :). If you’d like to give it a try, you can try it without an account for free for 7 days (no credit card required). Afterwards, it’s about $10ish/month.
One thing I’d love to get HN’s opinion on is how much gamification you think we should add. On the one hand, we don’t want to become a “game” and a Tutor lesson is always a “hard” thing and kind of exhausting, but there is obviously some balance to be struck there. Thoughts?
David, Same and Philipp (me :) )
You can try it without an account at https://chat.univerbal.app
136 comments
[ 3.8 ms ] story [ 239 ms ] threadI'm not sure how much work is it for you guys to add a language, maybe it's easier for you to add languages than for other mainstream platforms thanks to the use of AI ? If it the case, it would be really nice to have one serbo-croatian language, such as croatian or bosnian.
I struggled to find ressources for these languages. Since a few months I have been trying to learn Bosnian (going on a long trek in the balkanies this summer) and I was just not able to find any "modern" service (Duolingo, Babel, etc ...) which supports these languages. I guess it makes sense since almost no-one wakes up and decide to learn Montenegrin, but maybe you can find a niche market.
Anyway, cool product ! I hope you guys make it.
Bit of feedback: the UI confused me when it asked for what language I want to learn. I didn't realize the flag was a button to choose my language and spent a little while in two different browsers thinking the space next to the microphone was a textbox to type in or the microphone was the button for letting me talk. For some reason "use the buttons above to continue" didn't make any sense to me until now. Maybe cause there was only one button?
Anyways, I don't use a lot of apps and when I do I often find the UI to be confusing, so I might be an outlier here. But I almost gave up and didn't try it.
Good luck!
It worked the first time after asking for mic permission.
But second time, I did have a problem with the mic button on iOS 16.6.1 getting stuck on but not taking audio input.
Pushing the button again (to cancel perhaps) did nothing.
Switching to tutor mode fixed it for another input.
So it showed that it was recording but it actually wasn't? Then you switched to tutor, clicked on the mic again and it worked?
I really like the concept - good execution. I picked up the ability to click on a word for translation easily.
Have struggled with learning Chinese, will give this a shot.
1. The FAQ on the landing page hides questions as answers are opened, and it is not obvious that previous answers need to be closed to see everything.
2. On the FAQ page, when opening the hamburger menu, the menu is clipped and must options are hidden.
3. The FAQ question about pricing does not answer the question, and I see no where to get an answer.
4. Re: gamification, I think it is enough to reward users with making progress, and encourage them to use all the different kinds of modes/ features.
The guided conversation is too many words at once, it’s overwhelming. I immediately see a prompt to assemble two sentences from a list of words I don’t know without translations or hints. It’s the opposite problem of Duolingo drilling the same phrase a dozen times a day.
A few comments:
For Japanese, you might want to offer a choice for the amount of kanji to use in the written transcriptions that appear on the screen. Your current transcripts use the standard orthography, including kanji; some learners might like that, while others, especially beginners, will prefer that everything is in kana or romaji.
Other languages have similar aspects that should be customizable for learners: levels of politeness and formality, dialects and accents, grammatical gender, etc. It will be hard to do that with the current LLMs, but as better models become available it should become possible.
Multimodal LLMs that were trained on audio will be necessary for the tutors to respond to users’ pronunciation and intonation, produce natural backchanneling (an important part of conversation in Japanese and some other languages), etc. Perhaps such models will be available later this year.
Regarding gamification, how about offering choices to your users? Some learners will like gamification and benefit from it, while others don’t need it and will find it annoying.
It’s not clear from the free demo whether the characters one converses with are persistent or not. Especially for intermediate and advanced learners, it will be very valuable if your customers can chat with the same character repeatedly and that the character remembers the content of previous conversations and adapts accordingly. LLM context windows are getting longer, so that should be feasible. (Conversely, you will lose customers if they find that they are having the same conversations again and again or being encouraged to talk about things they aren't interested in.)
Also, you might consider setting up three-way conversations: two bots and the learner. One-on-one conversation practice can be tiring to the learner, and the learner doesn’t get a chance to observe how fluent speakers talk to each other. If the bots sometimes interacted with each other, the learner would both get a break and have a chance to learn by listening to the bots converse.
I have worked in language education for many years, and it seems that I was the first person to post a video to YouTube about using ChatGPT for language learning, on December 5, 2022 [1]. If you might find it useful to discuss ideas with me, feel free to get in touch. The URL of my website is in my profile.
[1] https://youtu.be/NVPHY3fYfmc
I would absolutely lean into the game aspect as long as it doesn't impact the quality of the tool. People love games and it lets you tie everything into a story, which people also love. For instance, people love stardew valley and it has inspired a bunch of youngsters to become farmers, but what if it actually taught farming?
I have a feature request: if I don't have a pinyin IME installed, it's very hard to use - it would be nice to have an in-browser one!
[0]: https://github.com/statusfailed/gptlingo
A couple more bits of feedback:
(1) The "suggestion" / "I'm unsure" etc. feedback is fantastic
(2) Word segmentation doesn't seem to be working properly, and so the context lookup doesn't work right. Example:
中国 should be parsed as a single word ("china"), but it's parsed as individual characters ("middle", "kingdom").
This means I have to tab out to a dictionary to look up words, and it's a bit annoying to select the right text.
Not sure if you saw it, but we already have pinyin in there. If you open up the settings and tick "show pronunciations" they will appear above the chat messages.
Or find all substrings that are listed in a dictionary (≈everyone uses cc-cedict https://www.mdbg.net/chinese/dictionary?page=cc-cedict ) and give translations for all of them. That way, the user won't be limited to any particular chunking granularity, which is always a finicky aspect of word segmenters to fine-tune.
The "show pronounciations" setting just turns on pinyin above characters - what I want is to type pinyin and enter chinese characters. Actually showing the pinyin above characters is quite distracting!
[0]: https://pypi.org/project/jieba/
You definitely fall into camp 1
Not sure what else to say beyond this: if you have a project that you've been wanting to start, believe in yourself and build it!
How do you measure linguistic progress?
>* The hard part intervenes when you try to replicate the progress tracking and tailoring of the curriculum.*
Perhaps you guys just aren't thinking about it in the right way ;) Progress tracking and measurement is significantly easier from a conceptual standpoint than you might think. I'm not just bullshitting here - I've actually done it myself so I'm speaking from experience.
Disclaimer: I've learned several languages over the course of my life; also co-founded a language learning company that used AI to help track and measure improvement. Oddly enough, someone from HN hit me up yesterday to ask about it. The company shut down in 2022 and I'm in a completely different industry now (non tech) so am fully able to talk about what we learned, where we made mistakes and where we innovated.
Also, how do you teach things like slang, jargon, body language, hand movements, etc.. that are essential to effective social or professional communication in a given culture? Memorizing words, grammar and conjugations, etc... is great but is only a small part of the language learning journey.
You're overthinking it, my friend. The solution is much simpler than you realize. Email me.
Before being forwarded into the app did you open a website that had some text like "Your email is now confirmed"?
I had an account previously (before the rename) and I'm using it to log in. Maybe it's something to do with that?
I wonder if it's the delay or just the bleak conversation with an AI. Might get to usable level for me once LLMs get faster and more fun. (Twitter model maybe? :D)
When you start the intro, and it says "Hi, what language do you want to learn" the button should be labeled "Select..." not a default language. It was unclear I had to click it to select a different language.
The languages in the app, and on your FAQ about which languages are available, aren't in sync.
The "Select..." makes a lot of sense
- The errors in my english sentences, mainly word order and missing articles, are involuntary and the corrections are reasonably coherent with deepl and deepl writer (that also uses AI).
- the french words were correctly identified as french words, and the words were correctly translated, except maybe for one: "essentiellement" that I used in the meaning of "mainly" and was translated as "essentially" which I'm not sure it's the good choice
in any case I add it to my bookmarks
Love the rebrand, by the way - I saw your posts about Quazle, and I never wanted to try it then for some reason. Can’t really explain it, but I guess that’s the power of branding right?
One suggestion I would have is that it takes a while when I click on a word to get a definition. Is there a way you could preload definitions on device to make this faster? Sometimes I want to click like 7 words in a row, and I don’t do that now because it’s so slow.
I noticed in this app, trying a language I don’t know, I immediately get blocked on lesson 1 step 2 - my answer is wrong, but I know nothing and there are zero mechanisms to learn without prior context to guess and extrapolate upon.
Seconding this. I got noticeably better at Chinese after using duolingo every day. I feel like I hit a ceiling now and it's not helping too much, but it definitely worked.
Re:loading You can click on the message itself and you'll get the translation of the entire message. Preloading the translation of every word would be a bit costly, but we could probably squeeze a bit more performance out of it.
My core task loops are phrase translation and audio transcription, and I've always been motivated by gamification and eventually data driven task selection