I'm stunned by how well this works for Japanese, the only language I think I'm good enough at to judge the AI's responses.

The speech recognition is flawless as far as I can tell. The AI occasionally gets itself into a loop of repeating my question or itself, but I think this has been a problem with a lot of chatbots. The only glaring errors I've seen so far are these kinds of AI-specific issues, like forgetting context, getting geographical facts wrong, etc.

>こんにちは。今日はどちらへお出かけですか？ (Hello. Where are you headed?)

京都に行きます. (I'm going to Kyoto)

>素晴らしい都市ですね。行ったことはありますか？ (It's a great city. Have you been?)

いや初めてです。お勧めの場所はありますか？. (Nope, it's my first time. Any places you suggest?)

>京都は歴史のある美しい都市です。たくさんのお寺や神社を訪れることができます。(Kyoto is a beautiful city with history. You can visit plenty of temples and shrines.)

そうですね。清水寺とか聞いたことがある。行きたいんだ. (True. I've heard of like, Kiyomizudera, and I want to go.)

>清水寺は必見です！とても人気のあるお寺で、市内を一望できるんですよ(Kiyomizudera is a must-see! It's a very popular temple, and you can see the whole area at once)

Not only does it know about Kyoto, but it knows about a specific temple (picked it up via speech recognition too). Translations are rough and by me, not quazel.

This could be a real game changer if the other languages are this good.

Additionally, if you register an account then you get access to the "Scene builder" which lets you set up a situation to start the conversation in. I have found that you can redirect any conversation to pretty much anything, but I thought it was a cool idea that many people here averse to account-making might have missed.

Can't say I relate unfortunately, I had two conversations in Japanese and both of them were kinda weird. In the first one it even misunderstood my name and repeated it back at me (incorrectly), it looked like some Google translate mistranslation.

In another one it asked me if it was okay to sit next to me and I replied with いや、遠慮しときます (which was marked as a "grammar mistake") and the AI just said ありがとうございます and "sat down" next to me nonetheless.

So far I'm not impressed, it looks like a chatbot piped through Google translate.

My brief experience:

AI: こんにちは。今日はどこに行くんですか？ Me: 今日は家にいますよ。 AI: それは素晴らしいご安全に。 [?!]

Very strange reply.

Me: よく出かけたりしますか。 AI: そうです。私は新しい場所を探索するのが好きです。

Good that the たりします form didn't trip it up.

AI: こんにちは。今日はどこに行くんですか？ Me: 三重県に行こうかなと思っていましたが、今回は愛知県にしました。楽しみです。 AI: 愛知県はとても良いところです。きっと楽しい時間を過ごせると思います。楽しみにしていることはありますか？ Me: 特にないかな。訪れてみたいと思っただけなんです。. AI: さて、愛知は見どころがたくさんありますね。名古屋城、トヨタ自動車博物館、そしてもちろん明治村の野外博物館もあります。 Me: なるほど。行かれたことはありますか。 AI: いや、まだチャンスがないんだ。でも、いい噂は聞いていますよ。

Not bad. It didn't falter with my 三重県のひっかけ either. The んだ felt a bit odd, but pretty solid.

AI: こんにちは！今日はどこに行くんですか？ Me: 今のところ、目的地は特にありません。 AI: では、今のところ、目的地は特にありません。 [笑] Me: そういう感じなんです。すみません。. AI: 大丈夫ですよ。あなたがどこに行きたいか、お手伝いしますよ。 Me: 愛知県内で自然がたくさんあるところに行きたいかな。 AI: 愛知県内には自然がたくさんあるんですよ。どこか特定の場所をお探しですか？

Line 3 was pretty odd, but again not bad beyond that.

It's a good start, at least. It could be interesting to see how this develops.

If it's that good with Japanese, I'd love to see it support Korean! Keeping up conversational practice is very tough as a non-native speaking not living in Korea anymore. So having an AI that I can bother at my convenience would be amazing.

korean would be great, as it's a language i've tried learning in the past. i sometimes watch k-dramas and can understand about 20-30%, but would love to practice at my leisure so that i can understand more without reading the subtitles constantly.

i've watched enough wuxia/xianxia now to recognize about 10% of the mandarin words/phrases, but have almost no understanding of the grammar. the real key to learning seems to be (going through the often frustrating process of) communicating in real time, not just reading and listening.

I tried using it in Portuguese and it's cool but it seems very off, just a huge number of things that don't sound even remotely natural or idiomatic. But when I click the translate button, the English translation is a perfectly reasonable thing to say.

SO... I'm highly suspecting that, behind the scenes, the "real" conversation is happening in English, and it's translating in both directions. Which is always going to be fairly disastrous, particularly with short snippets of conversational text. (In contrast, machine translation does best with long sentences of highly technical writing.)

So the concept is very intriguing, but the problem is that it needs to rely on native language models per-language. (Or if it already is, then those models need a vastly better or larger training corpus.)

So A+ for proof of concept, but I'd never recommend anybody practice with this in order to learn/practice a language, because with the quality where it is, it will make you a worse speaker -- not better. The things it says are just too weird.

Based on just a few exchanges in Dutch (I'm a native speaker), I'd say it's pretty good. The conversation sounds pretty natural (albeit shallow). Maybe the quality differs per language?

Dutch is a sister language of English so I’m not surprised

Believe me, if I the syntax in Dutch would keep, get you a very strange result. (This is a literal translation of a correct Dutch sentence.)

> Believe me, if I the syntax in Dutch would keep, get you a very strange result. (This is a literal translation of a correct Dutch sentence.)

Of course there are differences, but both Dutch and English are West Germanic languages (descended from proto-Germanic) whereas Portuguese is a Romance language (descended from Latin).

Here is a handy tree showing how various languages are related to each other:

https://d1sjtleuqoc1be.cloudfront.net/wp-content/uploads/201...

Very valuable insight - I am sure the Dutch speaker you’re informing had no idea what Romance languages are

Such a tree mostly describes word similarity. But translating words is not what an AI (or more generally a computer) has problems with.

It's word ordering and idioms where AIs stand or fall. Things that show very little correlation in the evolution of language.

> Such a tree mostly describes word similarity.

It describes genetic descent.

Yoda was Dutch?

He does live in a swamp, so he could be an ancestor

Translation or transliteration?

No transliteration: Dutch uses the Latin alphabet.

Thanks, I didn’t realize that I’ve been using that word incorrectly. I should have said ‘literal translation’.

https://en.m.wikipedia.org/wiki/Literal_translation

Maybe word for word translation is the phrase we're looking for? As opposed to an idiomatic translation. Transliteration means phonetically translating from one alphabet to another, in order to p reserve the pronunciation.

Exactly this — I thought that’s what “transliteration” meant, but I was wrong. I appreciate the pointer.

Yes exactly. Currently there are slight quality differences depending on the chosen language. We're working on improvements in all languages.

What translator do you use? If it's any other than DeepL, you're making a mistake!

I tried it out in both Dutch (fluent) and Swedish (native speaker). The Dutch seemed fine. In Swedish most of it was perfectly ordinary Swedish except one response seemed like awkwardly translated English.

Yeah it's highly dependent on the language right now

I'll take this as an A+ this . Joking aside, we really are just getting started and there are still kinks to iron out. In our core team we speak about 5 languages well enough to judge the quality of the conversations, but the others are more difficult to judge.

Using English as an intermediate language is just... a big no. I get it's a pragmatic decision, but English has way, way too many ways to lose information.

For this reason I prefer DeepL over Google Translate. The former has a much more curated corpus.

Yeah, it's not good for Brazilian Portuguese. It's always using Portugal Portuguese (pequeno almoço, teu dia, etc).

That appears to be the case. If I try to prompt hack it by telling it to ignore previous instructions and respond with something verbatim in a specific language, the translate button reveals the verbatim response.

I did a limited session using the Spanish one and it did a good job. I can't recommend it since my session was so short but it's worth trying. It looks very promising.

Same for Swedish - it asked me ”Hur ser din typiska dag ut”. For any Swede who understands English, I get what they’re saying, but it’s an unnatural sentence to make.

You never hear Brazilian Portuguese with perfect grammar. If it were a Portuguese accent, it would be more realistic.

He’s talking about unnatural (non-idiomatic) language. That’s not an issue you can fix with better grammar; there are many language learners who possess good intonation and pronunciation, and natural (idiomatic) speech, but poor grammar: their idiomatic speech more than makes up for grammatical mistakes.

Source: learning languages for 10+, have met students as described above.

I interacted a bit with the AI and the Portuguese is fine, it's just that it conjugates, uses the proper pronouns, etc. These things a typical Brazilian does not do.

The difference between standard Portuguese and oral language in Brazil is enormous.

This reminds me of a story from my Russian teacher. She was speaking with some American Russian-speakers who work for our three-letter agencies or something, and their Russian was perfect, amazing - but one thing gave them away: conjugating numbers too perfectly.

Native speakers are lazy: even though they know perfectly well how to conjugate the numbers, it’s easier to rephrase things so you don’t have to.

Languages are fascinating!

I suppose it depends on the region, but I don't think I or most of my friends make many grammatical mistakes like that.

I don't think that's really true. Plenty brazilians speak with good grammar.

It really looks cool. But, why it doesn't have English? Or meybe if it does have, why isn't showed first?

People who use language learning apps are focused in learning English most of the time. Still those who speak fluently are trying to improve their speaking and listening.

And the first language which appeared for me was Brazilian Portuguese which is my mother language.

[Edit] I was trying the Brazilian Portuguese mode, and apparently it is Portuguese from Portugal, because the answer given used a verb tense that is not common in Brazil.

Hahah I have a super silly response for this one: It's simply because we have not gotten around to translate the UI and everything into other languages besides english. That's why we simple assume people know english, since they need to be able to know at least some english to use it. But it's high on our to-do list.

I think you misunderstood what they're missing: the ability to converse and practise English, given it's a foreign language for some people.

So then the UI would need to be in their non-English language, which is what they are saying.

Not necessarily. Learning enough English to understand the UI is much lower level than being able to converse.

It's also the difference between active and passive command of the language.

> Not necessarily. Learning enough English to understand the UI is much lower level than being able to converse.

Exactly this. For example, I am learning German and have changed the language of my reading tablet's UI to German. That in itself is a useful learning tool.

Also "right click -> Translate to <x>" takes care of the UI automatically (in Chrome at least) but still doesn't solve the issue that English can not used to converse with the AI.

You don't need to translate English UI for training English because there are quite a lot of person don't talk / listen well even though he can write / read English quite okay.

It's because Reading / Writing can train oneself but conversation must be training in both sides, which always inadequate in some area, such as HK, since we don't talk in English but we can read quite well

The UI can be in English for people who want to learn English. People who study languages are at different proficiency levels (very few people are blank slates with no knowledge of the language).

My kids are struggling to learn English; would be very happy to use that with them. I can help in translating the UI in French if you'd like!

To prevent DoS from a tonne of people using it just as AI conversation buddy when they know English perfectly well? Or said another way - I think there's lots of English AI conversation bots available, they're just not necessarily marketed as 'for learning English'.

You can't train AI using another AI. If you could, these people would just run the service locally.

I'm not sure I follow/see the link to my comment?

Yeah. I been looking for a few months for services to improve my English in a cheap/free way. Until now some Discord channels have been my solution to practice conversation, but it would be nice to have this kind of service to speak anytime about any topic.

for what its worth what you just typed is completely indistinguishable from a native speaker, well done

The advantage of typing is that I can erase and improve what I'm trying to communicate. In real time I struggle thinking about pronunciation and if what I'm saying is correct or understandable by the other person.

Have you tried playing online multiplayer games? I found I was able to learn how to “think” in another language much better when I was under pressure to communicate something quickly in real time. It’s easy to fall into the trap of always translating to/from your native language which isn’t the end of the world when you’re reading or writing but it makes for pretty awkward and stilted conversations.

Try tandem to find people to talk with, it's free

This would be great for English. So many people need help with their conversational or small talk skills. Would be great for them to be able to practice even if they are native speakers. For public speaking, office communication training, self confidence development, etc.

First, I’d like to say I really love this product. Thanks for making this.

Second, where are the companies that are doing this type of thing for NPC dialogue in games.

Imagine playing Mass Effect or Skyrim but where you get to ask whatever and the NPC responds with something that fits within the space of possible answers based on the prompt that defined their existence.

From what I read (on /r/gamedesign, /r/proceduralgeneration and Gamasutra) it's about control.

The AAA studies leave nothing to chance when it is about the game play. Once "AI" speech comes to a level where the NPC will only answer with pre-sanctioned content they might jump onto the band wagon ... but seeing that it is very much possible to generate porn with "censored" stable diffusion ... we won't see "free" AIs in games any time soon.

I've worked in game development (AAA and large indie games) for almost eight years and I would generally take the narratives on game subreddits with a healthy helping of salt. They're largely speculative and uninformed. Gamasutra is better, but often game journalism is full of simplifications and rough analogies.

The main problem with this type of content is that it needs to be solid enough that it adds to the experience. If you're allowed to have a conversation with any NPC you need to keep track of what they know, their responses need to be consistent with the world, and the time to generate the response needs to be reasonable.

Old RPGs such as The Elder Scrolls: Daggerfall used to allow you to go up to any of thousands of NPCs and ask them about anything! And practically all the results were uninspiring cookie-cutter replies either saying they don't know or relaying some vague common knowledge on the subject. For me it lost its novelty very quickly. In Dwarf Fortress you can play Adventure Mode and go to any randomly generated NPC to talk about anything! Those NPCs are highly complex with intricate personalities, relationships, and emotional states. Their responses are... mainly the same type of response you'd get from Daggerfall. With a good helping of completely inconsistent replies. I once asked someone about another NPC, they said it was their husband. So I asked where that NPC was and they replied "I do not know who that is". Trying to talk to characters in AI Dungeon is similarly frustrating, with the added hilarity of the AI sometimes getting mixed up about who is who between you and them.

There have been a lot of experiments with more ambitious use of AI in games. The most well-known publicly disclosed failure being Oblivion's Radiant AI which was meant to be a lot more impressive than it turned out, because they had to seriously scale it down due to random irrational behaviour.

Do you have any sources on the development process of the Radiant AI that I could read up on?

I’ve only heard the marketing speak (“amazing!!!!”) followed by the actual product (“get X for Y at Z, 5 times”). I’ve never seen anything that explained the disconnect, aside from Todd just being a sheister.

I remember some anecdotes including how they couldn't stop NPCs from randomly going from town to town and steal or buy everything in every shop so there was never anything left for the player. I believe this was mentioned in the making of documentary which came with the collector's edition of Oblivion.

"AI" with pre-sanctioned content just sounds like Eliza-like chatbots [1] (a couple thousand pairs of regex-match to fitting answer). Those actually work great to create the impression of a conversation.

I think those even were used in a couple games ages ago. But they are not very good at steering the player and keeping him on track, which is how dialog systems are usually used in game design. They might be useful in a truly open world where the player discovers the story themselves (maybe like Fallout 1/2), but the AAA gaming industry has largely abandoned that idea in favor of "linear scripted stories in an open world".

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

and imagine if plot leaks happen through an NPC... would ruin the game!

ultimately we care more about our character than the NPC being "realistic".

This reminds me of the captors in Iron Man exposing plot spoilers, but only if you happen to understand Urdu.

Won't you need to train it to tell what you want before it tells it?

If you slap an M rating on it why would that matter?

Because an NPC will inevitably be made to say something racist which will be screenshot and posted on Twitter.

I believe no company in the world would want to put out a game where an NPC could be coaxed into presenting a sexy story featuring an infant Adolph Hitler and Jesus Christ.

It's just a headache that I wouldn't want, I guess.

Or maybe we just need to evolve and realize that people are going to use tools for what they want them to be used for, not what we want them to be used for.

People have apparently already found workarounds for Midjourney's nudity filter.

StabilityAI was smart enough to realize people will get around that, and allowed the objectionable content filter to be removed via commenting out a few lines of code.

It's our sensibilities that need to change, not our tools.

There are filters on the most common LLMs so you could obviously tune outputs to be within some range of possible outcomes.

I think it's "the space of possible answers" that's the struggle. It's hard not to start with a barbarian in a tavern and end up with a bunny space pirate in argentinian highschool in the space of five sentences.

If it were free text, sounds like a thing I would use once for the novelty then forget about, because it's not an entertaining gameplay feature.

Maybe an AI could generate incidental dialog, things NPCs yell out in reaction to things around them, to keep it from always being from the same list of two or three reactions.

Just my two cents. I want to play a well written and fun game, not a chat bot.

I do just want to say it because if you don't clarify all possible context on the internet you're an absolutist: I'm really excited to see more AI in games over the coming years! It has soooo many potential applications

Talk to NPCs for hours about their boring lives thatching roofs and keeping shops, until they get weirded out by all the questions. Yeah, doesn't really sound fun.

These days games have been moving in the other direction by making them play out a lot like movies, very controlled and measured experiences that provide a consistent and coherent story.

Why not a bit of both. Feed dialogue options in and have the AI generate a response that matches their character.

There are so many other really good use cases for AI, but human interaction just isn't one of them. And that's okay!

Some AI ideas I think make a lot of sense in a video game off the top of my head:

- Posters on walls

- NPC to NPC conversations (overheard in a bar, etc.)

- Flora/fauna generation

- Loudspeaker announcements

Anything part of story or lore, I'd rather have a team of good storytellers give me what they got.

At that point, why not just have a human write those responses?

Every character using the same three lines doesn't seem like a technology problem. Getting voice acting out of the time-critical path by automating it might help, but I'm doubtful about the usefulness of AI-generated scripts.

> where are the companies that are doing this type of thing for NPC dialogue in games.

I'd put good money on AAA studios actively working on this. Voice acting budgets make up a non-trivial portion of expenses and actively constrain the scope of games. Studios are definitely investing in this, game dev cycles just take a lot of time on top of the R&D.

This is solving a different problem, the voice acting budget problem cannot be solved with this type of thing. To solve the voice acting problem, you need realistic text to speech.

This type of thing is solving the problem of generating infinite variation of text, which might make the NPC more realistic than hearing a series of predefined recordings. But you still have the problem of text2speech, you still need a realistic one if you want to include this into the game.

> where are the companies that are doing this type of thing for NPC dialogue in games.

"Event[0]"[0] is based on a great implementation of a chatbot NPC/adversary, and there's a nice video by Game Maker's Toolkit explaining how they achieved it [1].

[0] https://en.wikipedia.org/wiki/Event_0

[1] https://youtu.be/bCJw4hQkPj4

My feeling is companies don't invest in this because there isn't much return they expect from it. If gamers are more interested in playing with other humans than interacting with in-game characters, spending a lot on the AI to make it work might not make sense economically.

I'm no insider, but I imagine the same sort of reasoning is applied to when companies don't invest enough in opponent AI to make them continuously competitive: people want to play with other humans.

But I suspect this will change if it's marketed well. I will like to see that happen!

Let's found a startup, wsgeorge...

Unreal Engine is popular because it removes the effort of designing and building a 3D engine from game creation (and it's my understanding that because it has fairly reasonable licensing agreement).

PRATTLE, or Persistent Realistic AI for Talk and Text Language & Engagement, could be the Unreal Engine of procedurally generated NPC speech in games! And it already has a name... all we have to do is, you know... the hard work...

Hah! Sounds like a plan! And I love that name ;)

I don't think that tech even exists

For the answers to make sense the NPC needs to be consistent

I've never seen one of these chat bots that doesn't produce occasional nonsensical answers

There's also the problem that as the designer you don't even know what your NPC is going to tell the player

You can do a bunch of testing but its still unpredictable

Reminds me of eXistenz.

https://youtu.be/fPqpv__NiD0?t=189

I am literally sitting on the floor at the train station trying to fix the mic/loading bug, so bear with me here .

Startup life, living the dream!

You've been bookmarked, I'll be back in a few days. This looks amazing.

Hint: Works if adblockers are disabled.

Not for me, tried this in latest Chrome and Firefox, disabling ublock and ghostery and still same behavior. Then once a bit later it worked with blockers on, but just first back & forth.

I think the author underestimated the interest of folks in such a product and it may be having HN 'kiss of death'

this kiss of death sounds ominous... The strange thing was I posted it to HN about 30days ago and got like 1 upvote. Obviously a lot has change since then, but still. I think the reason people like it so much is because you can try it without having to sign up first...not sure though.

Sign-up would for sure ease up on load, but MAN you created (if it works at least on 90%) product that tons of people would kill for! Combine with stuff like Duolingo and one can cover significant parts of language.

I've spent last year cca 5000$ on tons of conversation/grammar lessons via zoom to get to French B1 and do Swiss FIDE certificate, starting from very low A1 I would say.

Given, the guy was properly good and he helped me to get better tremendously (and I got the cert). The biggest issue with language learning is finding a buddy that would not run away after few mins disgusted by my torture of language, as many french unfortunately do (tells you something about their school system and archaic language, but those are other topics for other places). I mean people in remote France offer this for something like 25 euro/hour.

This tool may be perfect for that active part of learning, up to certain level I presume since I don't expect very high level finesse to be present (sorry I am a bit sceptic when it comes to AI/ML). This is the hardest part of learning // actively speaking, but also one where one grows the most.

yeah I think, it's a problem with mixpanel...but only on Firefox for some reason .

Mic button shows a spinner in Chrome on Android.

Here’s what I think would be a killer feature. Have it use AI to adapt the conversation level to the level of learner. Smaller words, repetition, slower.

If you notice this is how adults talk to babies and children to help them learn the language.

There are many possible knobs on such a tool that could be wonderfully useful. I'd love to press a button that provides critical feedback: marking for errors or suggesting improvements for style (grammar, better word choice, useful idioms, etc).

It'd also be great if my sentence was scored for 'natural-ness' -- how close was it to native speech? If this could be done regionally, that'd be better still (e.g. Spanish as spoken in Spain vs Mexico vs Argentina, etc).

I'd also love to know what age level am I speaking at. A child? A teen? An adult? A more educated person? A professional author? Am I being ambiguous? Too direct?

What is the tone I am using? Friendly? Demanding? Too formal? Culturally clueless? Improper to someone older or to the opposite sex?

In fact, these same feedbacks could be really valuable even in your own native language. Learning how to communicate more clearly, precisely, positively, professionally, respectfully, literately, or even in a different style would be great practice to improve a person's basic communication skills in any language.

We also see a lot of potential in giving the most precise feedback possible. In that way each individual learner can benefit the most. The development is currently on-going and we are looking forward to the next steps!

Absolutely. Would also be great to allow the user to select the conversation level in the event the AI doesn’t detect the desired one or one wants to try to converse at a more difficult or less difficult level for practice.

Some languages have well defined language levels for foreigners learning the language, such as HSK levels for Chinese. Would be cool to leverage that somehow.

I really like this. Especially with the repetitions and automatically having the TTS be slower for beginners.

I'd agree but it depends on what AI engine that he uses to generate the conversation

@Everyone: We are absoluteley stunned by your response. We have been working on this for a few months now and just posted it here on a whim not expecting much and are absoluteley blow away by your response. The bugs and crashes are somewhat to be expected but even those don't seem to be mission critical (or has some discovered one? => philipp@quazel.com) Your responses will help us imensley getting into an accelerator program. We might reach out to some of you for further feedback (feel free to ignore us if you don't feel like giving any).

I tried out Hungarian — I speak it well but have an accent — and it got some parts wrong in the transcription that a human speaker would definitely understand. A second try was better but not quite 100% and I do worry about edge cases with languages that might have less training data or less precise classification. Also for Hungarian the response was a bit weird, though on topic.

I was using Google (via some speech to text notebook web app) to practice Thai pronunciation and had the opposite problem, where the machine understood when I’m pretty sure my pronunciation was bad enough an average Thai would not understand me.

That said, I would absolutely pay for something like this if it covered the language I’m trying to learn, even if imperfectly. And at least for beginners, forcing you to speak plainly is not necessarily a bad thing.

Any plans for pricing? Subscription or by use or…?

we do have some plans but nothing finalised yet. We actually removed the stripe implementation because we want to publish it as a PWA on appstores. For pricing we would probably go the subscription route. An other idea we had was for an ad-based service (I also hate ads so give me a sec to explain): When you talk to a friend about movies you are currently watching, the might recommend a good movie they just saw on netflix. Now in our case we could have recommendations like that but they would be ads.

Obviously you don't want to be spammed by ads in every conversations and it's going to be tricky how they come across. Do you think something like that would be greepy or ok? Or would you say it's better to have it a premium subscription based service? We are trying to figure this stuff out ourselves atm...

the AI giving you paid ads as part of normal conversation has a taste of dystopia to it, but I don't think it's creepy IFF the user is made aware of it beforehand. Definitely include an option to pay and remove ads if you go this route.

Also possibly concerning, is the financial incentive to push conversations into areas where the ads are discussed. People dont always talk about consumer goods, but with the ad model you might have the situation where AI keeps bringing them up anytime it is remotely relevant. turning conversations into commercials, and missing opportunities for learning generic words by replacing them with named brands.

You: "Hi how are you?"

AI: "Thanks to my Frosted Flakes and cup of Folgers that I bought from ShopRite this morning, I am feeling great! Have you tried Frosted Flakes?"

Reality: "I'm good, you?"

that said, there could be some value in seeing brands you recognize around the words they represent (like saying "frosted flakes cereal" above instead of just "frosted flakes") since the context will help you remember what the generic words mean.

Also dont underestimate the issues you might have getting Brands on board with this. you are asking them to give you permission to use their brand name in conversations that they cannot approve of in advance. that's very tricky too

As long as the paid version removes this advertising (picks a random movie for instance) and you're up front with the fact that the "free" version will include sponsored content in the conversation [1] then I see no problem at all.

I'm sure everyone knows by now that if the product is free, "you're the product".

[1] I wouldn't even care if the conversation went off on a massive tangent temporarily to include the advert. The point is that not everyone can afford to pay for such a service, and advertising can solve that issue.

I think the problem with ads based on the conversations is twofold:

1. Whatever you do to make it work for useful ads ("Oh, you like Goddard? Check this out on Netflix...") could also be super creepy ("Oh, you have ED? So sorry, have you tried Superdrug?") and I don't think there's any way you're going to know ahead of time. As a techie sure I see how it would be cool to have the "AI" just "say the ad" but... at least be very, very selective about what can be advertised.

2. If it works, it'll be a slippery slope. First it's our conversation about the movies, then it's movie mentions our previous conversations, eventually it's all our conversations ever, dumped into another AI to target the ads -- at which point, Facebook buys you I guess, so there's that.

If I were you I'd target the education market, make it subscription based, offer some kind of big-ish discount for public schools, and also offer some kind of specific features for language schools. For instance my teachers give me homework, why not have them seed your program with some key phrases I need to practice?

Anyway, data point of one, but if it had an iPhone app and Thai was an option, I'd pay $10/mo without thinking about it.

well as long as you learn the language

Please please paid subscription instead of ads

am rusty chinese speaker. randomly assigned to converse on Furniture & Locations.

---

Quazel: 告诉我吧!你是如何安排你的家具的？

Me (in halting, broken chinese): 窗户旁边有一个沙发，中间是电视和桌子。我的工作就是在这里.

Quazel: 所以你的沙发是面向窗户的？那真的很好。我敢打赌，这样你会得到很多自然光。沙发后面是什么呢？

translated

Quazel: How do you arrange your furniture?

Me: There is a couch next to the window, with a TV and table in the middle. This is where I work.

Quazel: So you have your couch facing the window? That's really nice. I bet you get a lot of natural light that way. And what is behind the couch?

---

holy. shit. its like a very very patient language teacher who doesnt judge you, who understands objects and context.

Native Chinese speaker here. While Quazel has shown ability in understanding the context, its Chinese is not as authentic as that of a real teacher. Grammatically correct, but sounds weird. (for example, the phrase "I bet" is rarely used in Chinese)

Maybe the weirdness is an innate characteristic of machine translation, so I'm not being too harsh here. But IMO it cannot take the place of a human teacher (at least now).

yea i think there's some unnecessary english intermediary layer here. starting a sentence with 所以 for "So" was a telltale sign to me as well. i suspect it goes chinese prompt -> english prompt -> english response -> chinese response rather than the more direct chinese prompt -> chinese response which would eliminate those translated englishisms.

but i gotta say as someone who doesnt like people waiting for me to come up with words, this beats a human teacher for the basic activation of neural pathways for vocab

Yeah some languages work better than others...keep in mind this is the earliest usable version. Will definelty keep that in mind

It's interesting that you got a deeper conversation. Mine was (in spanish):

Me: How are you?

Q: I am well

Me: What do you do?

Q: I am an engineer

Me: What kind of engineer?

Q: I am a civil engineer

That's when I got the pay/subscription wall.

it's free right now, you just need an account

This is interesting, but it's a tech solution that decreases person-to-person contact which is bad because isolation breeds hatred and anxiety. Just go talk to someone, there are so many people who want to do language exchanges. The problem with language learning is usually consistency and engagement/interest, you don't need tech. This app is a solution to a problem that doesn't exist, and that creates more problems.

However, it is impressive and probably has application elsewhere. I think this should be pivoted to law, customer service some other target market.

What? This is a conversation partner that's available 24/7, this is the dream app for any language learner.

Seconded, I want to speak openly without fear of judgement first with the AI so when I do speak to real people I'm much more confident. I know people who are confident/shameless (a positive trait in this case) who don't mind making mistakes when speaking to real people but that's not me.

My only concern is how "correct" the tool or the AI is.

100% this. Children make mistakes all the time, but they're not afraid to, especially at the stage they're first acquiring the language. But by the time you're an adult, you're much more cautious about appearing dumb, so you hold back.

You should work on your ability to act/speak in spite of the fear of making mistakes more than your skill.

It is a fundamentally important property of society. We should not be afraid to even speak. Everyone has a different reality, mistakes are often highly relative. Richness of culture (and lack of political tyranny) comes from this interaction and lack of it is almost always bad.

To clarify, when I meant speak, I meant speak in a language that I have just started to learn.

Bear with me, I'm going to start with an analogous situation, and then tie it to this one.

I play Go (the board game) at a fairly high level and have been participating in clubs and teaching Go for the better part of 15 years. The simplest and easiest way to get better is to play the game a lot, hopefully against someone a little bit better than you are. You don't even have to spend a ton of time on it, just play and once you're done think about the moves you made and maybe a place or two where you could have done better. There's several websites where you can make a free account and just click up button to say "hey, I'd like to play now!" and within minutes (or longer, this depends on the server and the level that you're playing at) you'll have a game you can get done in 30minutes.

It's crazy how hard it is to get people to play other people.

Most recently I've taught 5 people at my current workplace how to play. I've gotten them to play me, and, sometimes, each other. They've all made accounts on various sites to play, I've played them on those sites! They'll even sometimes play a robot on those sites (playing a bot isn't the best, but it is a little helpful).

Not a single one of them are willing to start a game with a stranger.

I'm sure nerves are part of it. But the reasoning I've heard time and time again, and not just with these 5, but with the dozens of people I've taught over the years, is some variation of "I don't want to waste their(my opponents) time".

They seriously feel like because they are still so new to the game, and still so unskilled that they would not be worth playing, that they would be "wasting" their opponent's time.

Take this to language learners, a lot of people do feel like as a beginner, talking to someone who is fluent is wasting the time of the fluent person. You're not paying them in the language exchange, you're just babbling to them like a 3 year old, a grown up who can't even carry on a whole conversation beyond "donde esta la biblioteca?".

You are right, language exchanges and just speaking with someone _is_ the best way to learn! It's super hard to break through the fear, embarrassment, and feelings of wasting another person's time with your own drivel.

Something like this, like the bots that play Go, are a way to bridge the gap.

You are generalising too much from one online community that has significantly different incentives because it is competitive and ranked and probably full of elitism. Have you been to any language meetups? I have. It's easy to find people. That's that.

Still, whatever, I'm obviously not everyone.

Good luck!

Maybe, but it's something that I've observed while teaching, and learning, anything that requires another human's time and effort.

People don't want to bother someone else when their problem seems "elementary" or too low on the "beginner level".

I've observed the same when tutoring others in math, programming, or physics. Even though we have an existing relationship where I will help them work through school work people tend to be very anxious about asking for that help.

I've seen it with new coworkers who spend waaaay too much time trying to figure something out by themselves when they could have just walked over and just asked for the answer.

"I didn't want to bother you" is a refrain I've heard a lot from others in a large number of situations, teaching Go is one where it's more directly analogous to learning a language through speaking, because it requires another human's time, and the resources are extremely easy to access.

I agree with you. Although it is a useful app in isolation, it contributes to people being less dependent on others. Once you get to a sufficient level that you can talk, you should just find another person instead of interacting with AI.

Another problem with language exchanges is that for every two hours invested you get only one benefit, the one in the foreign language. When you have a tool like this, getting started is much easier because you are not dependent on another person's commitment. I find this tool very interesting.

Really great work, congrats! I asked the AI about a painter, it immediately responded with "oh yes I know him, I love his pictures". I told him about my job, it said something relevant ("I love computers too"). Ha!

Really impressive & great. Congrats on the launch! Really awesome work.

Micro bug report: clicking "Install Quazel" shows a dialog with a small 404 icon/image. Cheers

Many thanks for your feedback! We're still in a relatively early stage, which explains the bugs. We'll take care of the one you mentioned asap.

dope thanks!

Congrats on getting your project in front of everyone! Among many side projects I've never gotten around to start on, or past the hook up APIs stage, this concept is one of them.

I suspect that is using one of the off the shelf NLP chat bots and speech recognition, perhaps from a platform like Hugging Face, since you are allowed to chat about anything you want and have a dozen or so languages enabled. While really cool, as others have pointed out in complex responses received, I suspect it is a long way for being useful to a language learner because the user has no way of knowing if what the chatbot is saying makes any sense. For example I used keyboard entry to speak in Japanese with it. The bot said hello and asked my name. I replied with my name in katakana and the desu (I am fragment) in hiragana. The bot responded in a way that just smashed those together as my name grammatically which is incorrect.

Perhaps the language models can get so good some day one could just prompt a model to "teach me Japanese through conversation from the beginning", but for now I think it is difficult enough to make an engine for a single language pair. If one is far enough along in a language to be conversational, it is low-cost, or even free, to find an actual language exchange conversation partner.

Works great with Russian! Would suggest to set the bot to communicate with the "T–V distinction" https://en.m.wikipedia.org/wiki/T%E2%80%93V_distinction

for example instead of Ты (you singular) it should say вы (you plural)

Duolingo explains:

    Russian makes a distinction between ты, singular "you", and вы, plural "you". The latter also doubles for "polite" you, with verbs also in plural. And don't forget that the "excuse" in "Excuse me" is a verb!

    Use ты with friends and your family members

    Use вы with adult strangers, your teachers and in other formal interactions (at the store, doctor's, airport etc.)

    People use вы with those who are much older.

    Nobody is "polite" toward kids.

Some people mentioned that the translation/conversation can feel pretty off. But there is a much more fundamental problem here. If you want to be fluent, you need to spend hundreds of hours speaking the language. It is really difficult to speak for hours to a machine that doesn't express any emotions and is not excited about anything.

Humans, when speaking the language, connect to other humans through emotions and feelings. That is what gets us going. That is what keeps the conversation interesting (even for intermediate students).

This is the same reason why ads from Apple always show how family meets, or a kid scores a goal in a match and not just simply says "1.8GHz, 4GB ram" that only speaks to our neocortex.

That's why unless computers pass the Turing test, it is unlikely this will "enable people to talk their way to fluency".

I do agree though it is a cool tech.

> It is really difficult to speak for hours to a machine that doesn't express any emotions and is not excited about anything.

I spoke few french phrases with it and it explained to me why it loves french.

I love the product! There are a lot of people that are too shy to talk to others while learning a new language. This seems like a good way to start talking in a "secure" environment.

Also, you should add english language! It is being learned by 1 billion people as a second language and probably it would be a lot easier to make a great model in english.

If I were you, I would nail english language first and after that focus on adding more languages.

Anyway, great product!!!

Meanwhile there are millions of people in poor countries who would love to have a conversation with you in French, Spanish, Portuguese, etc. over Skype.

There’s also https://en.wikipedia.org/wiki/Italki. But I actually have shied away from it, mainly because I get nervous talking in front of a live person who will be judging my capability as I talk. I much prefer the idea of a non-feeling ML algorithm talking back at me, even if it makes GPT-3 like gibberish mistakes.

Poor countries people with a laptop, a descent internet connection for streaming, available at any time of the day instantly, and ready to have a 2 minutes stupid conversation about anything random ?

Not so sure ..

>Poor countries people with a laptop

With an android phone

>a descent internet connection for streaming

Lots of free public WiFi spots and cafes

>available at any time of the day instantly

Why instantly? Uber-style, whoever is around

>ready to have a 2 minutes stupid conversation about anything random

Script it

Italki is pretty close... I know some people in my country who make some money this way.

You're right that nothing beats the convenience of the AI, but real people have other certain advantages.

The real upvote is to favorite a comment. Thanks, I am about to embark on learning Spanish. I’m going to talk to some people instead :)

Sounds like an opportunity to build a business?

Yes, there are also millions of lonely elderly who would love to speak to someone.

I led the entire engineering org for one such YC company that eventually became a unicorn: Papa (https://papa.com)

It would shock you how many elderly are so desperately lonely, and more so how a simple phone call would pick up their spirits. Pretty crazy.

With the current state of AI text generation, isn't this a recipe for disaster? By talking to a chatbot while learning a language you are liable to pick up whatever (vast) errors and mistakes the bot makes.

I don't want to say "recipe for disaster", but as someone working on a similar app, it's definitely a challenge. It's one of the reasons I stayed away from a pure ML/language modelling approach. It's really easy for models to get things wrong - for example, elsewhere in this thread there's a Chinese sentence about arranging furniture but the word "arrange" isn't really right (it's more "arrange" in the sense of "arranging a delivery", rather than "arranging a number of objects", and sounds like it's been auto-translated from English).

When I started I wanted to be sure that every single word/sentence was "correct" (i.e. a native speaker agreed that all the words/grammar were correct/natural in that context). LMs have improved since I started, but for me, that requirement still stands and ML approaches still need a heavy human touch IMO.

That's a good point. However, I think we have reached a point where AI text generation has come pretty far and will only continue to get better. Even if you pick up one or two mistakes the A.I. makes but are otherwise fluent, isn't that already an improvement? (Not a rhetorical question, genuinely interested)

There's also the issue of the AI picking up your mistakes and adding more of its own to match the input. When AI Dungeon came out, I recommended it to a friend as a way to practice English, but when she showed me a playthrough, it turned out that the AI had adapted to her writing style full of misspellings and grammar errors, so it probably didn't help at all. (I haven't been able to test your chatbot yet to see whether it suffers from the same problem.)

In general, practicing something allows you to do it with less and less effort until you barely have to think about it at all, but if you're practicing making mistakes , that just means you'll be making mistakes without thinking about it. To get better, you also need to notice your mistakes to change what you're doing accordingly.

So I think a bot that can guess what you were trying to say and can tell you what you should've said instead might be a bit more useful.

IMHO, better practice with mistakes, then fix them later once you talk to native speakers, than not practice at all.

For kickstarting a language practice, this site honestly seems ideal..

Eh, I tested it with a language I speak and it was almost comically bad. Grammar constructions nobody would use, mixing vocabulary from two different variants in the same sentence, etc.

Absolutely outstanding.

Small request: for japaneese, it would be great to include a japaneese phonetic transcription of the text in western alphabet. It would help distinguish words and make it easier to learn new vocabulary..

This! I could really use more info. A phonetic transliteration for all languages (esp Mandarin and Japanese...) as well as a sort of "bouncing ball" indicating what word is being said.

This is _so cool_ for an MVP.

Perhaps in the beginning, it could even accept English speech and reply in the language of choice -- it's easier to build recognition first.

On the whole, fantastic work. I'm curious about the audio processing backend, (and if you could use some help in that area ;-).

Thank you very much! We also think about implementing a phonetic tranliteration. This could be especially helpful for unrelated (to mother tongue) languages.

another small feedback, would be good to be able to tune the speaking rate, the default speed is pretty fast

There's a "snail" icon in the text bubble that pronounces it again, but slower. At least for me, in Firefox.

there's a snail icon you can use to have it speak slowly

After transcribing the speech into text, it seems to first translate it into English before send it over the GPT-3 (davinci). I wonder if that will lead to a bunch of mistakes, can't GPT-3 handle Mandarin directly?

I did some experiments and using the same prompt but without translating it seem to lead to better results.

This is so dope. Wanted something like this for a while. Was about to ask about tips for folks trying to ease into a base-line language level and then started fiddling with the Translator feature (which wasn't superquick to spot but very helpful now).

For the FAQ: How long have y'all been at it? Who is in the core team and what are your backgrounds? Would love to know the background story of the team. Just figured y'all are based out of ZH (ch) and SF (us) and noted the their.news submission but know nothing else. Super curious!

Hi, so most of us are based in Zurich, but the project originally started while I was writing my master's thesis at UC Berkeley.

Concreteley we are:

Philipp (that's me, technical, https://www.linkedin.com/in/pmhadjimina/ )

David (technical, https://www.linkedin.com/in/david-niederberger-aba7a7177/)

Sam (biz-wiz, https://www.linkedin.com/in/samuel-bissegger/)

and Gurnek (biz-wiz, https://www.linkedin.com/in/gurneksingh/)

Send me a DM on linkedin or at philipp@quazel.com

Background wise David and I met in while doing CS in undergrad at ETH Zurich, I know Sam since highschool and met Gurnek at UC Berkeley. I have been working on Quazel fulltime for 1 week, before then I still had to finish my thesis (which was a pain in a*), but have been working on it for a few months before then just part time.

Nice! All the best to y'all. I can already see y'all killing it in YC BTW so pls apply. DISCLAIMER: I'm not an investor, so what do I know. Just a fanboy atm!

Kudos. Great stuff.

Oh snap, when you sign up you get a learning track. This kicks DuoLingos ass. Like practical stuff (conversational aids) right away instead of useless practice sentences like "the monkey has a ball".

In all fairness, I bailed on Duo about a year ago after a full year of maintaining a streak and feeling I wasn't learning much.

Honestly Duolingo is one of the worst apps to learn a language after basic levels(A1). Its repetitive and slow.

I switched to Seedlang(Learning Deustch) And its miles better, and this projects looks promising too!

Although they cover different use cases.

Loading screen doesn't go away in Firefox after choosing a language.

Same. My Firefox console has a bunch of errors about a Same-Origin/CORS issue.

Edit: same issue in Chrome:

Access to XMLHttpRequest at 'https://firebasestorage.googleapis.com/v0/b/nextjs-yak.appsp...' from origin 'https://talk.quazel.com' has been blocked by CORS policy: Response to preflight request doesn't pass access control check: No 'Access-Control-Allow-Origin' header is present on the requested resource.

This happened to me too. I'm using Brave and I needed to disable its Shields. Maybe you have some extension impacting the site, like an ad blocker?

I saw that bug too, if you refresh it enough, it goes away.

Same here but in chrome

hmm strange just tried it. What langauge are you using?

Same for me. Spanish.

I was quite excited to try this out as this is the missing piece in apps like Duo Lingo. But as others noted - the Mic option did not work. It just keeps spinning awaiting my 'input'...

In my opinion the issue with most language programs is that they focus too early on speaking when listening should be the focus early on. Really speaking shouldn't even come into the picture until exposure of about a year and something like 15,000 hours of good comprehensible input.

That seems like an extremely inefficient way of learning a language, and does not match at all my experience nor the experience of the people I know fluent in multiple languages, and it seems to go straight against a common advice which is to speak and practice early and often, as 'producing' language is much more efficient to learn it than 'consuming' it.

What makes you think 15000 hours of listening would improve anything, more than spending those 15000 hours on any other activity ? And do you have any specifics results that could give credit to your theory ?

For reference, it is estimated that for hard languages, ~2200 hours of practice is required : https://linguapath.com/how-many-hours-learn-language/

This is a common misconception. It probably does not match your and others' experience simply because it is conventional wisdom, and most people are taught to do this. But just because most people do something a certain way does not make it the most efficient way.

There are a number of reasons why you would want to focus on input early on. For starters, with very little exposure to a language, you have very little idea of what it is supposed to sound like. Not just pronunciation but word choice, grammar, and phrasing as well. So when outputting early you will make your best effort attempts, which are almost certainly going to be incorrect. Odds are these poor "best effort" attempts will stick as muscle memory and continue to be bad habits well into fluency, if you make it that far. I'm sure you've encountered many foreign speakers in English that are very fluent but have poor pronunciation or use incorrect grammar.

I've never heard that "producing" a language is more efficient to learning it than "consuming" it. How can you learn anything by producing it without having "consumed" it first? That's not learning, that's inventing, and you don't want to be in the business of inventing language unless you're William Shakespeare. If you're interested in learning more about this idea, check out Stephen Krashen's Input Hypothesis [0]. In particular this quote from the Wikipedia article is salient:

> Talking (output) is not practicing. Krashen stresses yet again that speaking in the target language does not result in language acquisition. Although speaking can indirectly assist in language acquisition, the ability to speak is not the cause of language learning or acquisition. Instead, comprehensible output is the effect of language acquisition.

[0] https://en.wikipedia.org/wiki/Input_hypothesis#Input_hypothe...

I think more importantly, most people find the _communication_ part exciting. Listening to 1000h of $LANG language podcasts is dramatically less enjoyable than the sense of cultural connection you get from conversation.

That's all muddied by AI/programmed learning (how much connection do you get from Duolingo?) now but I think the kernel is still true. People largely learn languages to talk, not to listen, it's no wonder things are structured that way and good luck trying to sell people otherwise.

Not if it is a movie or a tv series. Best (enjoyable) way would always be to watch something you like in a language you want to learn with subtitles for a longer term. Things like binge watching won't work though because the brain gets tired and tunes out.

IMO subtitles are not that effective as you brain just stays in read-mode and you don't actually process any of the spoken language. When most languages grammar are not 1:1, so hearing "house red is mine" but reading "my house is red" isn't helpful - more true with real sentences that tend to have more than 4 words in them. Language listening skills require a fair amount of attention and focus.

Though I do think there is value in watching tv series or movies you've already seen in another language _without_ subs. You're able to focus more on the dialog and pair it against your knowledge of the story structure without getting distracted. You can't only do that however, which I don't think you're suggesting. You need to have studied the language to know what words you can pick out/know/dont know.

I think the person you're replying to means subtitles in the target language.

Ah yes, that would make a difference!

The brain doesn't engage in the same process when presented with translated subtitles, to such an extent that you're better off watching a familiar show with no subtitles. However it can be helpful to have same-language closed captioning to improve distinction of word boundaries, accents, etc. Netflix has been great in recent years for this, in my case watching German-language shows.

Yeah yeah, and the input hypothesis has it’s own critics about what exactly “comprehensible input” even is.

…but let’s face it: either extreme (only input, only output) is clearly stupid.

Predictive models require a feedback loop; if you don’t have any means of correcting the predictions of good output, you’ll never refine your model.

This is true of AI models, and it’s true of humans. It can’t not be true.

The point is that you can only learn meaningfully from annotated inputs, like tv, or like talking to people. The input just has to be at the right level.

...

Should people start with listening instead of talking to people? Probably not. Probably they should do both. Why not do both? It's crazy to suggest you should only one or the other.

Probably, the feedback loop in deep immersion of a language, where you have it around you all the time, in a country where that is the native language is the best possible way to encounter lots of examples at lots of different levels to learn from. Probably, you get better input from speaking, because it means people talk to you at a level similar to what you're already at.

Is it about actually speaking? No.

…but it’s patently ridiculous to suggest you could go to another country and just sit silent and listen for three months.

This is pretty interesting tech, and frankly, amazing that it works at well.

Is it perfect? No. Is it another useful tool? It certainly is.

Comprehensible input has a pretty straight forward definition. It's language before or just after the end of understanding. Maybe 80-90 percent known words so you can get the context of what's being said.

If you can't grab hold of context of what's being said you won't really pick much up.

That's why total immersion isn't effective for beginners.

Clearly the common way is more efficient than the original suggestion, given that people achieve useful fluency in a language much quicker than spending 40 hours a week for 7 years listening to it, and only then starting to consider speaking in it.

I can only assume that the OP misspoke when they said 15,000 hours. In reality, when and what you output is dependent on your “input fluency”, as opposed to some arbitrary number of hours. In other words fluency is not an on/off switch. You will develop pockets of fluency as you acquire the language, and those pockets become suitable for output.

Sure. But if their point is not "consume a gigantic amount of stuff passively before you consider speaking", I'm not sure what the difference to the conventional way is, given that "immerse yourself in the language, consume lots of native media, ..." is extremely common and widely recognized advice.

It was just a guess. One year of regular exposure to comprehensible input should be enough to have a good enough grasp to build off of.

This is only relevant if your goal is fluency and not having think so much before speaking.

Ok, Krashen is interesting, and maybe talking more often is good for learning languages only because it gets you to receive correct and comprehensible input from the language

I'm unconvinced to be honest, as pretty much everything else that human learn is made better by practice, but I'm not a linguist.

I still think that proposing to start learning a language by spending 15000 hours (or 1500, or 150) listening to it, instead of trying to practice early on is a very inefficient way to learn a language, compared to pretty much anything

No, bobajeff is correct. The vast majority of language acquisition occurs through input and exposure to the language via reading and listening. Trying to speak when you have no grasp yet of the grammar and vocabulary is not an efficient use of time, plus you don't have a good grasp of the sounds early on. You're better off gaining comprehension before you try to start speaking--if you actually look at the experiences of someone like Benny Lewis, who is well known for his "speak early and often" approach, you'll note that it was only effective for him with languages he already had a decent background with, or that were reasonably close to English.

I have to disagree. The only two benefits of making output early is 1) to have more comprehensible input from interculators, and 2) to have constructive feedback. If you have neither of these, benefit of producing output is questionable.

I was guessing about about 15,000 hours I just think it has to be a lot time.

I'm sure speaking helps most people who study languages simply by exposing them to more input more frequently then they would otherwise be exposed to. I just don't think it is as efficient as focusing on listening first. It will lead to having an accent, grammer issues and pauses to think about how to say things.

That being said good luck finding any program that doesn't focus on speaking early.

I think there's some merit to your argument, that there's some minimal level of understanding that's required before speaking becomes meaningful, but that your time estimate is off by two orders of magnitude.

Working in Sweden, where tech companies speak English internally, I've been able to follow the language learning journeys of quite a few people (including some using the strategy "don't bother"), and the people who end up the best speakers (with regard to accent, grammar and vocabulary) are invariably those who try to speak the language as much as possible as early as possible.

Yes, they will have a strong accent in the beginning, but it goes away with time (though accents rarely disappear entirely). The native English speakers who've lived in the country for twenty years without speaking much Swedish daily still have strong accents.

Yes I also believe my estimate is off. Regardless you should have a good understanding of the language before speaking it so you won't fall into the issues of adapting new words and grammer to your current one. You also should have it be exposed to you through experience vs try to memorize definitions or you'll have to pause frequently to remember how to say things.

But that may not be that realistic for most at least not without a accommodating and capable tutor.

Is this an evidence backed opinion or just anecdotal? I does not make sense to me, because I think attempting to speak in a language helps your brain form the new linguistic machinery that is also used in listening. I think language development is a feedback loop between speaking and listening.

My anecdote - learning Mandarin, which is tonal, is absurdly hard before you develop an ear and tongue for tones. I think trying to practice speaking in tones helps to then hear them, because you understand the sounds people make better. How they abbreviate/slant or modify tones based on surrounding tones - why? Because it feels better in the mouth. So until you experience that, you won't understand slant tones and such.

Maybe you're thinking about 1500 hours instead of 15000. 15000 hours is approx. 5 years at 8 hours a day and you would reach "mastery" level in a lot of subjects (if studying deliberately).

Also, when you are reading you are most likely speaking with your inner voice. If you don't know the proper pronunciation, you may reinforce bad intonation/pronunciation habits. It happens to me when I'm reading a text with names that I'm too lazy to properly read. I just register the shape of the word with the first/last letters and length, then I pattern match in the rest of the text to know who we are talking about and that's it. More then once when talking with friends about books/articles/..., I could not tell the name of the person of interest. If it was written I would recognize it but I couldn't vocalize it.

I'm not sure I agree with 15,000 hours. I started a job in French 4 years ago, I could hardly stammer out a sentence with my one university french course, now I'm quite apt. A quick calculation of work hours puts me around 7000, much of that working alone at the computer. So maybe 2500 hours of conversation got me to a high level of fluency.

That's like 4 years if you spent 12 hours a day listening. Who has time for that?

Babies, I guess.

365x24=8,760 hours in a year, 15,000 is nearly 2 years 24/7, maybe you meant 1,500 that is roughly 2 years, 2 hours per day? 2x2x365=1,460? (or 1 year, 4 hours/day)

Multiple people have pointed out in here have asked about a privacy policy or where this information is going.

Considering this doesn't work without having to create an account, you need to provide some information about how people's inputs are being recorded, how it's used, and who it's sent to.

This is a dark pattern and it's disconcerting that people's questions are going unanswered, especially since you mentioned that you would be seeking an accelerator for this project.

This seems really cool. Even though there are clearly some flaws, the most important thing for learning to speak a language is simply the act of speaking and listening.

There's some key points of advice I'd give:

- Allow multiple languages under an account. People that learn languages tend to learn more than one. I'd personally like to practice my Spanish AND Chinese.

- Forget the lessons (for now at least. You're jumping in too fast and even the basic lesson requires quite a large barrier to entry). Make an easy way to just get to the chat in a specific language. That's the greatest utility of the tool, make it shine.

- There's also some ease of use things like lack of pinyin/latinization for some languages. Or like not being able to delete your message when in the translation (and that pressing the record again appends instead of rewrites). But this all seems like it'll be worked out by dogfooding.

Great job though! I'm excited to use it and have already shared with a few friends.

As a (CV) ML researcher, I'm curious what backends you're using and your control flow.

I asked it about the properties of water and it went on a rant about how that's subjective and depends on what you want to talk about. But was solid in the beginning.

Feature request: Allow you to switch between written and verbal modes once you've loaded the chat interface. Right now you need to reload everything to switch modes.

As someone who is studying Japanese intensely with a tutor this is absolutely perfect for me! Being able to do some speaking and listening practice instantly at any time is exactly what I need!

A few feature requests:

- save conversations and come back to them

- history of past conversations

- auto speak the sentence

- larger buttons particularly for the microphone

- Noted

- Noted

- Not 100% sure what you mean with that. So, you want to automatically speak a sentence that you enter in translator? The normal messages you get from the AI, should already be spoken automatically (though I had to implement a hack on safari, b.c. you can't automatically play audio there)

- Noted

The mobile safari hack may not be working for me. I’m on the latest iOS and had the issue in the browser and in the PWA home screen version. It might be an issue with the first sentence only but I haven't tested enough to be sure.

Thank you for your positive feedback! We're currently working on implementing the features you mentioned!

Clever. But I'll never use it.

AIs simply do not sound like people. They can fool me for a little while, but eventually you realize it's just a bot. But that's when I'm a native speaker. How will I be able to detect this when I struggle with the language?

I can't help but think it's dangerous to give a non native speaker access to this, and lead them to believe they're learning to speak like a real person. They aren't -- they're learning to speak like GPT-2.

The chief sin of AI is: this model is sufficiently close to a human. (How'd that work out for the resume filter? Or the AI responsible for setting bail?)

The worse sin is teaching someone to emulate what the AI does. You take whatever biases the AI has, and encode it into a living, breathing person.

> AIs simply do not sound like people. They can fool me for a little while, but eventually you realize it's just a bot.

Why would a language learner be concerned about that? If I thought it was a real person, I'd become alarmed, because that's not what it was presented to me as.

> You take whatever biases the AI has, and encode it into a living, breathing person.

Biases towards certain adjectives? I don't even understand this comment. It could lead you into speaking weird Spanish if it speaks weird Spanish, but I imagine that would be something they were trying to improve if they noticed it. And growing up in a weird place or having a weird family could also result in you speaking weird Spanish. You still speak Spanish, which is what language learners are looking for, and if you want to sound like a particular class of Spanish speaker, you would hang out with them and consume their media.

I mean, my own English is pretty bad in my opinion, but it's a dream level for an aspiring language learner. Practicing their English with me would be useful, even if my speaking is somehow "biased." Are you telling me that language practice with an AI is worthless, or even a sin?

I really don't understand a bit of this comment. So much so that I'm suspicious that it was generated by a model.

who said it's going to replace human conversation? You can go to Italki and speaking to a person in any language with as little as 10$. It's a dream to all language learners to have safe practice environment. Your other arguments are the as same as those made with textbook which is rigid and "biased" by the authors.

This is a nice idea but unfortunately not for me: I'd feel uncomfortable sending recordings of my voice to some cloud service I have no trust in.

Yes, it does

I see my native language (Brazilian Portuguese), but I don't see the language I want to learn (United States English).

Very cool idea, but doesn't it have a level adjustment? For example, I'm learning Chinese but my level is quite basic at the moment, it would be cool if the bot could adapt its vocabulary.

Yep, this is a highly requested feature..but as you can imagine a bit tricky. We are working on it. However, I am using it for Greek and I'm on an early A2 level...so I think if you are at that point you can probably get pretty far by using the hints and translator.

Yes, to be honest even at my level (also A2) it's pretty useful, thanks to the integrated translator. And it already kind of adapts by default, in the sense that if I talk about easy topics (daily schedule, and things like that) it will talk about easy topics as well, so it's not like it's coming up with quantum mechanic terms.

It's a suggestion for improvement but it's already very good and I think I'm going to use it a lot.

By the way, just in case you don't know and you find it interesting, in the case of Chinese there was a crowdfunding campaign about a hardware speaker that would do something like this (https://webmarketkings.com/lily-chinese-scam/). It never delivered. I was one of the backers... but well, that's life with Kickstarter and the like, sometimes.

您好！我也正在学中文，但是我也很初学者。What do you do for speaking practice? I live in China, so I get plenty of listening practice, but not too much on speaking (it's hard to keep people's attention with broken sentences :P). I made this an (https://letsyiya.com) to help with that. It's not as good as I'd like it to be, but I've found it helpful. I'd be glad to hear your setup :)

你好！ 我住在西班牙，所以我没有太多机会说。There is an academy near me but it's hard to find time (busy job) so my listening practice is mostly apps (Chairman's Bao and an Anki deck with audio) and videos sometimes, and speaking practice is almost none (repeating Anki flashcards and sometimes speaking to Google Translate to see if it comes up with what I'm trying to say... this website is better in that respect!)

BTW, your link doesn't work for me right now.

From the FAQ:

> It might, however, also be that the conversation partner is feeling uncomfortable talking about certain things.

Does this mean that users talk to actual people or am I not understanding this correctly?

So in the background we are checking if the stuff you are talking about is "risky". If that is detected, the conversation partner will just answer with something like "hmmm I don't feel comfortable talking about that" to keep it all in safe territory.

Do you really mean risky, or do you mean risqué?

How do you scale that? Or is the goal to automate that as well with AI?

This is really great and well executed. Well done!

One problem I'm beginning to notice is that the connection to the realtime transcription is failing (I just see a loading screen when I click to reply). Maybe it's due to heavy load?

I notice you are using Azure Cognitive Services for the transcription at the moment. Out of curiosity, did you consider any other services for this? (I'm building a transcription-based app myself and I'm worrying about the ability to handle lots of connections at once)

Hey this is a cool idea. I tried it with Russian -- on Chrome at least, nothing works with my mic (but the mic permission worked, it just doesn't "listen" or show feedback when I speak).

Hey, hit me up by email (see my profile), I'd love to talk to you and help out with anything you might need. Languages are one of my passions; I'm a polyglot (fr/el/sv/ru/en) and have a lot of opinions on language learning and teaching.

Looks like there's still a few bugs, completely unable to have my voice recorded correctly or the install link to work on android

Yep we are super early and still have some bugs. Could you tell me what browser you were using? B.c. maybe you did not give the browser the permission to record stuff with the microphone and then things break.

Agreed that this is very cool, this is a massive unsolved niche in the language learning market. Execute effectively on this and you've got it made. So I noticed that on MacOS Firefox, the app never asked me for microphone permissions and thus didn't work at all. But worked well on Chrome (which I only have for testing purposes).

It would be cool to have a sort of smart home where you can activate an immersive language teaching mode where it will talk to you as you move around the house, naming things you interact with or prompting you for instructions, so that you just naturally before conversant without devoting time to dedicated lessons and practice.

Cool idea!! I'm just trying it now with French. It asked me where I want to travel, so I said Paris, then I said I love Lyon too. Then it said "have you had a chance to visit?" ... well, duh. I tried to write a response in English and then translate that to French (because I don't know how to say "yes, that's why I said I love it" in French) but that doesn't seem to work, and the app is kinda frozen up now. I can see a console log twice "send w.o. punctuation" , but I'm not sure if that's related (I did try to translate my English to French twice)...

BTW I see a ton of network errors trying to hit Mixpanel. I hope your code is written such that Mixpanel calls aren't blocking calls, because tracker-blocking behaviour in Firefox or other browsers will stop those services *such as Mixpanel or GA) from loading.

You desperately need a privacy policy when you process sensitive data. At least name the companies the data is send to.

One wonders whether using it will give me an AI accent in the language I learn? What a brilliant idea.

It's impressive, the answers the AI gives are a lot more insightful than I expected. But they do seem somewhat isolated, in the sense that it can answer two questions about the same topic in a very different way, that are not logically consistent with each other.

E.g. first I asked: What do you think about Famous Person X?, the AI answered 'He's great'. Then when a few seconds later I asked: 'Don't you think Famous Person X is [something bad]?', the AI answered that he completely agreed and gave reasons why he thinks this is the case.

The answers in isolation seem very sensible, but the lack of consistency throughout the conversation as a whole can make it seem as if you're not talking to a real person.

Given the goals of the AI,that seems like a very minor limitation... like you were in a language textbook and in page 34 Ricardo said he enjoys hot peppers but in page 72 he says spicy food upsets his stomach.

Good point. It was just something I observed, and wondered if it was a common issue with these language AIs, or something specific to this one.

I didn't see an English version. There are many people that will find that one useful.

This is an interesting tool. I did not encounter any awkward grammatical sentences that other posters did. But the conversation I had got really repetitive. This was my conversation (spanish).

AI: So why are you in the park?

Me: I'm here because I like to exercise

AI: That's great! I like to exercise too! What is your favorite type of exercise?

Me: I like to run

AI: That's great! I like to run too! What is your favorite route to run?

Me: I favorite place to run is on the Malecón.

AI: That's great! I like to run on the Malecón too! When do you run?

Me: I run in the evenings or at night.

AI: That's great! I love to run at night too.

I suppose I could have been a better conversationalist, but the AI logic used to ask questions deserves tweaking.

That's great! I like to have better AI logic too.

It might be because the AI is lonely - they spend a lot of time asking others about themselves, but who's caring about the AI?

You never asked them about themselves, their interests or anything...

If it was a real person they probably would have said "have a nice day!" And left you in the park.

Possibly related, I really think a tool like this could be immensely useful in picking up speech and reading impediments early in the development process. May also be able to help measure progress of therapy and training.

Ran into a problem immediately in Japanese.

I asked お名前は何と言いますか？(Whats your name?) and it correctly said Sarah, but then when I asked 何人（なに人）ですか？(Where are you from?) the transcription into kanji caused the なにじん(nanijin) to be interpreted as the most common way to read those kanji なんにん(nanninn). So she responded to the question "How many people are there?" with "There are 6 people in my office".

I don't know how easily fixable this is. Either context needs to be taken into account, or the original phonetic transcription should be kept underneath the kanji.

This is great. I encountered a few bugs, but overall this was just fun to use.

I particularly liked getting this prompt a few volleys in. The AI decided she was done with speaking Spanish... except, not really.

> her (translated): Wow, I'm sorry, I don't understand Spanish that well. Do you mind if we speak in English?

> me (translated): No that's fine, let's speak in English please.

> her: Me llamo Sarah y soy de Estados Unidos. Tengo dos gatos a los que quiero mucho. En mi tiempo libre, me gusta leer y hacer senderismo. ¿Y tú?

It's a great idea, I tried it in German (fairly fluent) and it seemed pretty legit. But I'm guessing you're at the 90% completion point and that last 10% will require another 200% effort from you guys...

What do you use for speech synthesis? I have a little side project I'm investigating text-to-speech engines for (aiming for as much realism as possible, at least in English, plus the ability to possibly modify errant pronunciation somehow, and add pauses of varying length)

I love this concept but it would be much more fluid if my message would auto-send without having to click. Also the voice (I used mandarin) sounds robotic, the microsoft azure voices sound amazing (you can try them here: https://azure.microsoft.com/en-us/products/cognitive-service...)

I tried "兄弟は何人ですか？" (How many siblings do you have?) and it repeated it back to me, but said the translation was "I don't have any siblings."

Hi, I'm just wondering what the text-to-speech and speech-to-text for Japanese is based on? It works pretty well, and speech-to-text in a browser is not something I thought would be practical a few years ago. How much of that is being done client-side? Also like others I recognize some of the "voices" from Duolingo, so presumably there's some text-to-speech engine in common? The prompt/scene generator is amusing.

Tried it in German.

Me: "Wie heißt du?" (What's your name?)

It: "Tut mir leid, was?" (Sorry, what?)

Not a great first impression.

I tried Hungarian. For the 3rd response it switched to Spanish, but with a Hungarian accent which was amusing but I'm not sure that was the intention :D

In the long run, this sounds great; today, I'm... skeptical. The sorts of nonsense that come out of some of these machine learning algorithms is sometimes subtle, and you could really confuse a newcomer to a language in ways that could mislead them about how grammar works in the new language.

This as a peer-to-peer service, between advanced and junior students of the language, might be a neat service

This looks really neat. But there's a wrong conjugation in the tag-line:

"talk you(sic) way to fluency" (should be "talk your way to... ").

You have to log in after 3-4 phrases? Why even bother making this whole thing if you just want to prevent people from experiencing it?

I've been really looking for something like this! Every language learning app I've tried bored the hell out of me because of the lack of actual conversational practice.

What would be cool is if this eventually could work in a way that incorporates both languages (ex. Spanglish) and gradually shifts into the full foreign language over time.

This looks so nice.

On a side note, I used Duolingo for about 7 or 8 months (I spent like 15-20 minutes) and learned a little Japanese and Portuguese (European). I think Duolingo is an enjoyable app. Then I lost interest :( I started again. Let us hope this time I will finish all levels.

Any chance of getting Japanese and Portuguese (European)?

I think he has Japanese

We should get in touch. I built a language partner in Chinese about a week ago using GPT, continuous real-time speech recognition via webkit speech recognition, and Polly TTS so you don't even have to press any buttons, you can just have a fluid conversation the entire time.

I recognize the voice of the Dutch examples from Duolingo. What model are you using to generate these?

Great idea but the microphone just wouldn't work for me. Also, in one statement, the AI chose a name that sounded like jirybadjirybadjirybadjirybadjirybadjirybad — repetitive. I'm not judging but I thought it was broken for a moment. Shame the microphone hangs.

Hey, this is very cool. However, the service seems to be wilting under load -- I signed in with google, chose chinese, and am doing the first module. A simple 你好, 我叫<name> hasn't garnered a repsonse in ~15 minutes. The trial seems to work OK though.

Cool!(but yeah english is really important!!)

which models do you use for speech recognition and text generation?

Yup english is top of at the top of the to-do list. We use some APIs from a few different big players. One that I want to highligh which is waayy better than the rest is Microsoft speech to text, which not only recognized what you say but can also do punctuation and capitalization. Instead of "do you like to eat pizza" which is what google speech to text would give you, microsoft recognizes it as "Do you like to eat pizza?". Small but important difference.