Show HN: I made an open source and local translation app
github.comA few years ago, right after high school, I decided to try to make a simultaneous translation app for Android as a side project, it took longer than expected (about 2 years) and I had to make a lot of compromises (I had to use Google's API and therefore make users use a developer key because at the time there were no free solutions for speech recognition and translation that had good quality). At the end of university, I decided to pick it up again and finally, using OpenAi's Whisper for speech recognition and Meta's NLLB for translation (with both running locally on the phone), I managed to make it free and totally open-source (as it was meant to be from the beginning). The app is still in beta, so I would love your feedback.
If you are doing translation locally on the device, why does your Privacy Policy say you are sending voice and transcriptions as well as personal location data to Google?
> RTranslator in any case collects and processes data that will then be sent to Google, such as: audio, the transcription of which is transmitted at a later time via bluetooth to the phone with which you are communicating, and the transcription of the audio received by the other user, to carry out the translation.
— https://github.com/niedev/RTranslator/blob/v2.00/privacy/Pri...
The position sent to Google is an error that I didn't see (I will correct it in 5 minutes), the rest, as I say in the readme, is due to the fact that the privacy policy used for RTranslator 2.0 for now is the same as RTranslator 1.0 (that used Google's Api), I will make a new privacy policy contract for RTranslator 2.0 with a lawyer soon.
Nice one! Thanks for sharing this.
If possible, it would be good to include a list of devices that you know performance is good. We can all understand that most flagship mobiles will run it smoothly, but what about the average user?
"...to be able to use the app without the risk of crashing you need a phone with at least 6GB of RAM, and to have a good enough execution time you need a phone with a fast enough CPU."
Thank you! The app can be used on any device that has more than 6GB of RAM, with a phone with the right amount of RAM but a slower CPU the only effect is that the app takes longer to perform the translation and voice recognition (at least from my tests it shouldn't lag, so it's always usable), but yes, maybe in the future when I have the opportunity to test it with enough devices I will make a list with the performances for various Socs.
The app looks very cool, but you might have a licensing issue; Meta's NLLB models are CC-BY-NC 4.0 (non-commercial use only). I could recommend checking out the OPUS-NLP models, which are truly open.
In that case, it is not open source. Part 6 of the open source definition by OSI (the only commonly accepted definition) includes this requirement:
> 6. No Discrimination Against Fields of Endeavor
>
> The license must not restrict anyone from making use of the program in a specific field of endeavor. For example, it may not restrict the program from being used in a business, or from being used for genetic research.
This app looks really cool and I'm glad OP made it. I'm just asking that the inaccurate "open source" descriptor be removed. Don't let my comment discourage you.
Meta defines NLLB as open source, but assuming it isn't (I know Meta It's not a company that has a problem with lying), my code is open source, so how should I define my app? I personally make a distinction between open source and completely open source (or 100% open source), because otherwise there is no intermediate definition, according to the OSI definition my app is not open source, but it is not closed source either, OSI does it have an intermediate definition?
Source-available[1] is likely the closest commonly used term for it
[1] https://en.wikipedia.org/wiki/Source-available_software
Ok, I think I found the perfect solution, In the readme I added near open-source "(almost)" with a link to the libraries and models section when I explained clearer what components are open-source, closed-source, and cc-by-nc.
it's the closest, but it's a very broad and not so well known term, maybe defining my app open source is not technically a perfect term, but it's the clearest in describing it (considering that in the libraries and models section I specified which are open-source and which are not, Ml-Kit, used to recognize the language, for example is closed source), if a developer doesn't see the open source writing anywhere he will only get confused in understanding what the license of my app is. However, based on your feedback I added the specification (in the "libraries and models" section) that NLLB has a non-commercial license.
Instead of "almost", maybe "Open Source Code (and free for non-commercial use due to model licencing)" since your contributions are free as in freedom,and that is amazing!
Thank you for building this, I have been using a web interface connected to a local server for inference but the latency was about 1 second, too much for my taste!
I think it is too long for the preface, but I will better specify that my code is open source in the libraries and models section, thank you for the suggestion and the appreciation!
Sidenote: I don't think ML model weights are even copyrightable
I know that, but since I make money only from donations without giving anything in return of the money I am considered non commercial. For now I decided to use NLLB instead of Opus because it has good quality and is a single model for all languages (so I didn't have to implement the download management logic for the various languages and I was able to make the app faster), but it is a temporary solution, I will certainly replace NLLB in the future, either with OpusMt Big or with other models with a less restrictive license.
Great we are more than a century ahead of schedule for Star Trek universal translator [1].
Joking aside, this will be very beneficial software if it can work seamlessly in real-time for countries like Japan who can't speak English, and France who won't speak English.
[1] Star Trek’s Translator Technology, Explained:
https://gamerant.com/star-trek-translator-technology-explain...
Hi, I made a non-local version that supports translation to other languages.
Android: https://play.google.com/store/apps/details?id=org.evergreenl... iOS: https://apps.apple.com/app/3po/id6503194251
p.s. I contemplated calling it Sato, ended up using 3PO as the name. :-)
this is very cool the first time i use listing mode on ios the app crash but now its working fine, im in ios 18
thanks. Looking into crash log. I am working on better support for different dialects (e.g. different flavours of Indian and Chinese, etc.) Version 1.1 should come out soon.
On the input language nothing appears
is a option to delete old conversations?
That would be the end goal (very far :D). But yeah, joking aside, I aim to make the translation better and better as technology advances, the current level is a usable start but maybe one day it will be truly seamless.
Suggestion
> RTranslator is an open-source (almost), free, and offline real-time translation app for Android.
change to:
> RTranslator is an (almost) open-source, free, and offline real-time translation app for Android.
Reading it the first way, it's possible to misunderstand and think it's almost (but not completely) free, rather than almost (but not completely) open-source
Thank you, you are right, I will do that.
As said, a video demo would be great, especially since it is not on iOS yet, but it looks great!
Thank you. I'll probably make one these days.
I like the idea, it's awesome! Too bad that NLLB, despite being advertised as specialized translation LLM, doesn't work with Turkish - it generates rubbish like Google Translate, which doesn't make much sense. I live in TR, but don't know Turkish :( I tried all major translators, for now the only option which performs well is ChatGpt. I created a custom GPT there, instructed it to be my translator, and that works so well, even recognizes poorly written messages without punctuation. Anyway, this is just my observations. Best of luck with the app!
Thank you! Just tested with Turkish and yes the quality doesn't seem to be the best, however in the future I plan to add translation models with higher quality, so I hope that those will do better with this language.
So, there's a hope, cool
Just for the reference, what I have to face with: "Sizinde Sular kokuyormu çok kötü koku var baneda mutfakta"
NLLB gives: "You smell like water. It smells bad in the kitchen."
The right translation is "Do your waters also smell? There is a very bad odor in my kitchen too."
"Bu bina elektronik bir sistem var dış kapı -otomatik- uydu -asansör elektrik yoksa hiç biri çalışmaz yanlış sigorta kapatılıp açılırsa böyle sıkıntı olur"
NLLB: "This building has an electronic system. An external door. An automatic satellite. An electric elevator. If none of them are working, it's a problem."
Should look like "This building has an electronic system with an automatic exterior door, satellite, and elevator. If there is no electricity, none of them work. If the wrong fuse is turned off and on, such problems occur."
They have regular punctuation marks like dots, commas, questions, and exclamations, but, for some reason, they don't use these in chats. I don't know why, but it hardens translations a lot.
I bookmarked/starred the repo to track the updates :) Just to clarify - no pressure, obligations, expectations, etc. Just my feedback with couple of samples, hope that helps.
Feedback are always helpful, thank you.
The UI is nice and clean, but I don't have a secondary device to test out the conversation mode. A video demo of how this works would be nice.
The first translation I tried was incorrect:
"Bonjour" (french) -> "- How are you ?" (english)
I suppose that's the limitation of local models.
Great job getting it to a functional beta though. I think the idea is cool, it just needs some polishing.
Yeah, for short inputs the models are not that precise, but for long and complicated ones they work quite good, probably NLLB has been trained with very few short texts.
This app is very cool! In walkie talkie mode, you may want to tweak the UI to make it clearer when the microphone is listening and when it's not. I saw the microphone icon changes color, but a stronger visual hue may help. Google translate may be a good point of reference - in its conversation mode the shape of the microphone icon changes when the mic is active. I've also noticed that my message is sometimes cut short when translated, it only translates the first half.
Thank you for the feedback! In a little while I plan to redo the entire graphic interface, and yes a method to make it clearer when the microphone is listening is already in the plans. As for the cutting of sentences you can probably solve it by increasing the microphone's sensitivity from the app settings (from there you can also change other settings regarding the activation of the microphone, but most likely it is a sensitivity problem).
This looks neat. Surprised to find that it only got 40 upvotes on Producthunt.
Just bookmarked on Github. Any chance to launch it for iOS as well?
Thank you! I've done the launch on ProductHunt years ago, when the app still used Google APIs and required the developer key, I'll make another one soon. Regarding the release of the iOS version for now it's not in my plans, but if the app is successful and I make enough money with donations in the future I might consider it (since I would have to buy a Mac and at least 2 iPhones to test the Conversation mode), and always if the App store's policies will not be too restrictive.
As an iOS developer, that could definitely use this, I completely support you, in keeping it Android native.
I’m a believer in native software (I write native software). I think it results in much higher Quality user experience, than hybrid development.
There’s a reason that it’s not super-popular, though. It’s a fair bit more expensive than hybrid approaches, and generally requires a higher skill level, as well. Most companies find that hybrid approaches are “good enough,” and they make much better margins.
That said, you could probably test it, using the Xcode Simulator. It’s gotten pretty good with I/O, lately. Bluetooth still requires a physical device, but most of the stuff could probably be tested fine, if you used a modular approach (my preferred methodology).
Yes I agree with you, moreover for now I am much better as a Java or C developer than as a web developer, so for me the native approach is also simpler. However, yes, I practically never use the emulator even for Android precisely because I have to test Bluetooth, which at least in my experience on Android, is the most difficult part.
I wrote up this series, a few years ago[0]. It was actually #1 on HN for a day, but is probably a bit "dated," now.
I did do an actual video tutorial on it for try!Swift[1].
[0] https://littlegreenviper.com/series/bluetooth/
[1] https://github.com/ChrisMarshallNY/ITCB-master
Really cool tutorial! When/if I start developing for iOS this will be really useful, thanks!
Congratulations to the translator, very nice and clean. Will you release a version for iOS or Progressive Web App in the future?
Thank you, for now it's not in my plans, but if the app is successful and I make enough money with donations in the future I might consider it (since I would have to buy a Mac and at least 2 iPhones to test the Conversation mode), and always if the App store's policies will not be too restrictive.
FYI, After model download, I get an error:
Using LineageOS+Microg and eSpeak for TTS (equivalent android version = 13). This setup works fine for TTS with the commercial map application, Here WeGo Maps.Unfortunately you are not the only one having this problem (https://github.com/niedev/RTranslator/issues/8), these days I will try to solve this bug, for now it has top priority.
Thanks for pointing me to that bug report. And, thanks for working on it :).
Your initial fix of not disabling the entire app's functionality if it can't get TTS working sounds like a good one.
After downloading library I got error: There was an error with the tts initialising. Retry won't do anything.
For now I have tested the app with Samsung's and Google's TTS, and with these 2 it should work, I am investigating the problem and if possible I will add support for other TTS engines.
Samsung with LineageOS 19 without gaaps to be precise.
Samsung TTS? That's strange
If I was to guess, it might be RHVoice from FDroid as they are using a custom rom.
this looks really exciting! sadly running into an issue with TTS initialization on CalyxOS with the eSpeak TTS engine.
I tested the app with the Google and Samsung TTS, in these days I will also test with other TTS including the 2 mentioned here and I will try to solve it, I will update you in the issue you opened.
I have same issue on LOS 21 and RHvoice.
Any change on an fdroid release?
I was considering adding it in the future but I don't know if it meets all the necessary conditions given that uses Ml-kit (closed source) for language recognition in the WalkieTalkie mode.
What does the latency look like with this?
With a Snapdragon 8 Plus Gen 1, I measured about 2 seconds for speech recognition of 30s of audio (although for shorter durations it takes not much less than 2s), while for translation about 2 seconds for a text of 300 characters (in this case, however, the performance scales linearly based on the length of the text).
[dead]