Show HN: I built an AI conversation partner to practice speaking languages
apps.apple.comHi,
I built TalkBits because most language apps focus on vocabulary or exercises, but not actual conversation. The hard part of learning a language is speaking naturally under pressure.
TalkBits lets you have real-time spoken conversations with an AI that acts like a native speaker. You can choose different scenarios (travel, daily life, work, etc.), speak naturally, and the AI responds with natural speech back.
The goal is to make it feel like talking to a real person rather than doing lessons.
Techwise, it uses realtime speech input, transcription, LLM responses, and tts streaming to keep latency low so the conversation feels fluid.
I’m specially interested in feedback about: – Does it feel natural? – Where does the conversation break immersion? – What would make you use this regularly?
Happy to answer technical questions too.
Thanks
I just tried in Portuguese, and it introduces itself as ChatGPT. I was going to ask what can this do that ChatGPT can’t, but no need to answer now I guess.
Did it? It uses an openAI (gpt) model. I should refine the prompt. Thanks for the feedback!
The deeper issue is that your product is one of many chatGPT wrappers with no unique hook, not that users can find out you're one of many chatGPT wrappers with no unique hook.
This is only a problem to other sausage makers who feel the competition cheated.
Real users get won't care so long as it works.
Very curious to know where you have been in the last few years for your to regard openai.com/chat as something known only to specialists and wholesalers. To stick with the sausage-making analogy: this is like someone selling costco hotdogs in the parking lot out of the back of a van for 25% markup because they pre-squeezed the ketchup on it.
How many people in the costco parking lot do you think there are that want costco hotdogs but are unaware of the existence of costco, costco hotdogs, ketchup, and what they cost?
but users do care eventually. they just don’t yet have the vocabulary that “other sausage makers” do. dismissing these concerns implies that outcomes matter and provenance doesn’t, which may be convenient if you’re building wrappers, but it’s corrosive if you care about ecosystems, incentives, and long-term quality when the app is compared to others in the market.
that app is not the only app in the market that focuses on conversation. feedback like this is worth taking seriously rather than waving off.
Cute parable but most rely on child sweatshop labor. Users express "thoughts and prayers" level of care if pressed, but not take up a trend of sewing their shirts to spare kids they will never meet RSI.
Noting competition exists seems focused on the outcome of making money. An obligation that exists in an exploitative economy. Where's your concern for the provenance of such obligations?
Yeah. Every expression of concern is probably something like social desirability bias, appearance of concern while sticking with status quo real effort of zero change. Low effort rhetorical "care".
People express concern about global warming and drive off in their SUV. They're just parroting social script.
There's millions of people who do walk the talk. Lots of them do this in silence, without any kind of personal gains in social status.
This kind of faux-nihilistic pretension is peak irony - convincing oneself that no one actually cares to feel better about not caring.
And millions more working against your theoretical "silent majority". It will take billions. Ratios are not there.
Not saying I don't care. Am saying it will take actual change in agency of billions to make me believe the rhetoric.
Thoughts and prayers are not cutting it is what I am saying.
You go ahead, sit online repeating the same empty rhetoric and assure me your "thoughts and prayers" being cranked to 11 is meaningful effort.
Cool app. It takes a lot of courage to share here, so well done. I’d definitely highlight a bit more prominently what languages it supports though. The interface is very “Claude” so I’d be keen to hear about:
- How did you go about developing this?
- Was the entire thing vibe coded or just part of it? No shade either way, just curious.
- How long did it take?
- What were some of the harder hurdles to overcome?
- Given the use of AI, what’s your approach to security of your users data? How did you review any generated code?
Thanks! I just wanted to use it myself as it's always hard to find a native speaker to speak with when learning a language. The app is partly vibe coded. I myself am a software developer and had no problem reviewing the vibe coded code. I think it took me a little more than a month give or take, as i was not consistent with my commitment. I would say the syncing part of the conversation was the challenging part, which is still not perfect to be honest. As for user data, i don't store much of user data apart from the minutes used.
Since I don’t own an iPhone, I can only give you feedback on the landing page itself:
- The font in your “Stop studying. Start speaking.” screenshots is both hard on the eyes and strangely blurry.
- Your ad copy needs an overhaul - it feels clipped and rushed.
> I built TalkBits because most language apps focus on vocabulary or exercises, but not actual conversation
There are MANY language apps which focus on actual conversation. You are in a SUPER competitive space. You need to call-out what makes your app different. In just the last few months alone (just on HN) I've seen many Foreign Language Chat Apps:
SpeakLanguageOnline – Voice-only AI language tutor
https://news.ycombinator.com/item?id=46779716
Malan Chat - Full immersion language learning app for 62 languages
https://news.ycombinator.com/item?id=46768430
EnglishCall - AI that calls you and practices spoken English with you
https://news.ycombinator.com/item?id=46714857
TongueFu - Gamified voice-first app for communication
https://news.ycombinator.com/item?id=46553017
Orratio - Practice spoken English by discussing news articles
https://news.ycombinator.com/item?id=46510414
I really appreciate this comment. So much of marketing is "puffery" or based on what would be best for the app creator for the audience to believe. Sometimes that lines up with reality, sometimes it doesn't.
In this case the marketing copy seems more like creator wishcasting or living in the past. I've fallen into this trap before too, of creating things as if AI was a secret in my basement and not something of intense focus by a plurality of 20 million+ software developers. Rarely does a technology landscape change so fast.
How you're framing it is helpful as well. Instead of saying "this already exists (so you shouldn't do it)", I think it's valuable to highlight what a competitive space this is. The creator will need to think hard about what they can do to differentiate their offering in 2026, both in marketing and functionality.
Thank you very much for your feedback. I will work on improving the app store screenshots and their descriptions.
Issen https://news.ycombinator.com/item?id=44387828
Univerbal - Language learning with a conversational AI tutor
https://news.ycombinator.com/item?id=39188760
I'm curious how the experience is different from using live voice mode in ChatGPT or Gemini?
I haven’t tried this app yet, but my attempts at speaking to ChatGPT in Spanish were pretty dull. Almost every sentence the model replied with “Claro!” and tended to repeat itself often
To be honest i didn't try them. My purpose was just to provide a nice UI/UX to the AI voice model.
You didn't try them!? If you have the ChaptGPT app it's super simple and worth a try. I talk to it in French a lot. It gets tedious though, because after almost every question it tries to end the conversation with "let me know if I can do anything else to help". I really want it to pretend to be curious and continue the conversation. I've thought that this could be fixed with some better prompting, so I'm excited to try out this app.
I always had trouble getting AIs such as ChatGPT to get out of their assistant mode and give natural and conversational replies. No matter what I put in the system instructions it always responds with the same old walls of text and same old it's x not y etc.
Does this one have a prompt that actually takes care of this problem? Does anyone have some nice prompts that make the ai actually useful as a practice partner?
Crazy that when I search for this in the App Store on my phone, it first shows me results for Talbots, and then when I click "show results for Talkbits" it shows me a bunch of other apps. I scrolled through a bunch but then gave up and copied/pasted the URL. Ridiculous!
I'm surprised that an app like this can be rated are 4+, since any open web browser has to be adult-only. I would think that an app like this, unless it has very secure guardrails.
How does Apple deal with rating apps that tie into LLMs?
Congrats on shipping! The focus on actual conversation practice is spot on - there's a huge gap between knowing vocabulary and being able to use it under real-time pressure.
Re: your questions - one thing that often breaks immersion in conversational AI is when the responses feel too 'helpful' or educational mid-conversation. Real native speakers don't stop to explain grammar when you make mistakes. They might just naturally rephrase, or not even notice.
Have you considered adding a 'correction mode' toggle? Sometimes you want the AI to be forgiving and keep the flow (like talking to a patient friend), other times you want explicit feedback after the conversation ends.
Also curious - how do you handle the awkward pauses when users are thinking? That's where a lot of language learning apps feel unnatural. Real conversations have natural silence.
Really interesting approach to the conversation practice problem. The latency optimization you mentioned is crucial - I've seen so many AI voice apps where the delay completely kills the natural flow.
One thing I'm curious about: how do you handle pronunciation feedback? In my experience, the hardest part of speaking practice isn't the vocabulary but getting feedback on how you're saying things. Does TalkBits correct pronunciation errors, or is the focus purely on fluency and keeping the conversation going?
Also, do you adapt difficulty based on the user's level, or is it more freeform?
Congratulations on shipping!
I’ve been working on a similar idea for similar reasons, but only for personal use.
Initially, I thought that if I had a better platform to chat with (better than say, Duolingo) then this would unlock a big growth area for me where I could get more realistic conversational experience. But, as I’ve been building and experimenting, I realize that during conversations I still fall back mostly on the expressions that I’m comfortable with. So, I’ve been experimenting with different modes that will push me to use more advanced forms of grammar and focus in different areas, and so on. Also, I allow myself to select a level of proficiency and dialect (e.g., B1 Mexican Spanish) so that I can get corrections and suggestions that are more specific to my goals.
I’m curious to know if, as a user of your application, you feel like it’s pushing you into awkward situations that will force you to grow your skills?
The 'speaking under pressure' insight is spot on. Most apps drill vocabulary but people freeze when it's time to actually produce speech in real time.
Question: do you give any feedback on pronunciation, or is it purely about conversational fluency? I've found that 'am I saying this right?' anxiety is often bigger than grammar anxiety.
Also curious about the latency you've achieved - realtime conversation needs to feel instant or the brain switches back to translation mode. What's your target round-trip time?
The latency optimization is key here - what's your approach to detecting end-of-turn vs mid-thought pauses? That's one of the hardest parts of real-time conversation AI. In natural speech, people pause to think all the time, but if the AI jumps in too early it breaks immersion completely. Curious if you're using VAD with some confidence threshold or something more sophisticated like prosodic analysis.
I really love this concept for an app. Downloading and getting going was a breeze.
That said, I’m finding some latency issues when I respond. I’ll say something in Spanish, hit the red button, and none of the text I sent appears.
Then I’ll hit the red button again to start talking but before I do the assistant responds to my response and the words I spoke show up
Some of the back and forth is showing up out of order.
Also, error messages are in Spanish, and my Spanish isn’t good enough to read them so I’m not sure what to do
Again, love this concept and would love to have an ai assistant I can have daily conversations with to start sharp
Would possibly be nice to prime the ai before the convo saying “I’d like to talk about x” in English
Keep going! I’ll keep playing with it
Interesting approach focusing on conversation rather than vocab drills. The latency point is crucial - I've tried other voice-based language tools where even 500ms delays completely break the natural flow.
What's your typical round-trip time from when someone stops speaking to when the AI starts responding? Also curious if you're handling interruptions (when the user starts talking while the AI is still speaking) - that's a big part of natural conversation.
The speaking practice angle is interesting - have you found that users actually stick with it, or does engagement drop off like most language apps? I've noticed the hard part isn't building the AI conversation, it's creating enough novelty/progression to keep people coming back when they could just chat with native speakers online for free.
> chat with native speakers online for free
Where can you do that?
I pay for LinguaTalk and I also took a close look at TalkPal… how does your feature set compare to theirs?
Thanks for sharing your app. I've been trying a couple of conversations, but the AI is hearing its own voice and then responding to itself, without letting me speak. I hope I'm not doing something wrong, but if I am, this could be a problem faced by many others.
Yes you are right. This happens when you are using outer speaker. I am trying to fix it. Thanks!
so, from a PT-PT perspective, I guess I can understand, the clanker begins in PT-BR. After I instruct it to switch to PT-PT, it starts mixing formal and informal phrase constructions (think like the German Du and Sie forms) which is odd but then becomes really jarring with the accent/tone which is sounds mechanical and robotic unlike the uncanny valley of gemini or openAI models. Anyways, I like the app and can imagine interesting use cases while abroad!
Duolingo does this, but only for a few languages and it needs some improvement.
Thanks for your feedback. Yes it definitely needs improvement.
Without using headphones, the mic picks the AI speech, meaning it ends up in a discussion loop with itself
Yes that's what i am trying to fix right now. Thanks though!
Looks pretty good, the speech pipeline feels noticeably faster than the general-purpose apps I've used for lang conversations. Dead air kills immersion after all
Thanks for the feedback!
Did you vibe code the whole thing? Talk thru the stack and your experience/findings
Not the whole. I vibe coded part of it. I am a software developer and previously worked as a mobile developer. My stack is just react-native and a simple nodejs backend hosted in a vps.
OK:) congrats, was just prompting you to do a technical writeup
Tangential, but on the topic of learning/practicing language, I was thinking that in India, with such high population density, if I want to practice speaking any language, I can find quite a few people within a short walking distance. Personally, I can speak three languages (default for most Indians), but it is very common to find people who speak 5+ languages. I can also understand many other languages that share similar sounds and commonalities to Hindi, so I can be around Gujarati, Haryanavi, Bengali, Marathi, etc., without getting totally lost. Unfortunately, despite living in the South and making many attempts, I could never pick up any Southern language beyond a few words to get by.
Yesterday, while on a walk with a friend discussing SAP, he stopped to greet someone and spoke in Oriya. When I asked, he said he can speak 5 languages fluently and can get by in another 5 or so.
My daughter needs help with her French; we have a neighbor for that (not an App). I’m at three words—Oui, Bonjour, and Bonsoir.
What about Japanese? :)
The other comments about calligraphy, copy, competitors, etc are not useful for getting users. This app is already over-engineered.
Go on TikTok and Instagram Reels, scroll for a week 15 mins a day in the language and travel niches. Don't post until you've done that!
Then post funny, scroll stopping videos. In the comments, and in your bio, mention your app.
It's funny, making an app these days is the easier part. Getting people to use it, any substantial quantity of them, is where it's difficult.
You don't need to warm for a full week
What do you need to do instead?
I would be a potential customer at some point in the near future. So I would like to know:
- what languages can it handle perfectly from A1 to C2?
- Pricing
- Any daily caps? I am guessing I can't talk for 24 hours, can I?
- How long can a conversation be until a new one with a new context starts?
The levels would be great. I’ve tried prompting general purpose LLMs to only use words at particular levels, but they pretty quickly diverge from the prompting guardrails. One might just need to building a custom model trained only on a limited vocabulary.
I am on Android. Could you make the app on this platform too?
How about a webapp that runs in desktop/mobile browser?
Does the AI store & train on my voice?
Android is my next step. It doesn't train on your voice.
The best part of learning a language is being able to build connections with real people. It's about being vulnerable — learning to take chance — to embrace being uncomfortable. Somehow the uncertainty and vulnerability with other people ends up being one of the most enjoyable parts of learning a language. Taking that chance and getting to know what it's like to communicate with another person across the unknown.
What makes learning a language so wonderful is being, "Lost in Translation."
My observations:
The calligraphic font is antithetical to the theme of your app. The apps colors and ux suggest playful. The font suggests school marm, which is it?
The app itself doesn’t differentiate itself enough to stand out on first use as unique. What does this provide over other similar platforms? How is this different?
The space you’ve chosen is highly competitive. Most of the big players bear a unique signature from ux down to the syllabus they teach.
I applaud you for sharing with us. Sharing here takes grit. Good fortune with your endeavors mate.
What's different between this and all other 10s of different ones doing the same?
This question maybe formulated too harsh, but it is valuable. There are quite a few similar applications (I think I tried 2 or 3 of them). Some are around for couple of years. What is new / unique in your approach?
[dead]
[dead]
[dead]
[dead]
AI slop commenter account above