Speech to text, Voice typing is cool.
F-droid lists opensource apps that support Whisper and Vosk models.
Futo voice is not opensource, but is source available.
It’s good for English. For my mother tongue, Malayalam, the support seems to be underdeveloped currently.
Handy seems to be good for Windows, but have not used it much.
Then for Translations:
https://f-droid.org/packages/dev.davidv.translator/
Using Mozilla’s translation Models. A good enough offline and Open source alternative to GTranslate.
Are there other cool uses of offline ML/AI that you know?
You must log in or register to comment.
Occasionally I’ll spin it up on my local machine and have a chat with it about technical stuff.
Sometimes it’s fun just running scenarios with it because it’s novel. Or I use duck.ai kinda like an instruction manual you can ask things to. Never ever on anything with actual stakes.
I do find it’s good at deciphering Linux stuff, and I do end up actually learning rather than just blindly trusting whatever it spits out.
I don’t want chatbots to replace friends asking each other things or discussing technical topics on forums though. That would be sad. But sometimes it helps with those simple “been asked a thousand times” issues where I’d get bogged down reading 15 tabs and still not be certain what the best move could be.
So I guess I find it good at distilling existing information.
Using it to “create” anything though? As an artist and game dev, I find the thought genuinely disgusting. I suck at coding, but I’m not selling out my coder brethren and sistren to a machine, just like I’d be disappointed in them doing the same to me as an artist.
mostly translation and making audio books
Thank you. Which all apps/models do you use?
no need for any of that stuff for me, so i don’t use it
Object detection and recognition for my self hosted security cameras.
They use a ML model to recognize which objects are in the picture.
I use it to generate little bash and python scripts for mundane tasks, random ideas I don’t have a good starting point for, or boilerplate code I don’t have the energy to write manually. Or also to decipher cryptic error messages and logs. Nothing that I publish though, just all testing and personal scripts.
I do this with models running locally on my gpu. 14B parameters is surprisingly decent and fast enough for a 16 GB slot-powered workstation model.
For fun, since the model is local, I also toss in some excerpts of my writing, see what it can glean from my style and vocabulary, something I’d never do with an LLM hosted by someone else for obvious reasons.
My spam filter on my email server. Autocorrect on Heliboard. I think that’s it. Luckily I picked up English at a young age so I don’t need it for translation.
Also for people having a kneejerk reaction, “AI” and “ML” as terms long predate the current slopwave. The ML I am talking about in the previous paragraph is not genAI, it is the kind of ML to classify things (emails into spam/ham, words into correct/incorrect), that you can feed training data to improve its accuracy rate.
Thank you.
There seem to be offline models for many languages other than English too.I agree. People need to focus on the AI ownership aspect more. Should it not come under public control or oversight?
It’s by no means new, but I’ve found OCR to be useful.
Thank you.
Translator has an OCR option. I currently use it for OCR.
Before that: https://f-droid.org/packages/io.github.subhamtyagi.ocr/
I use it for translation help, speech-to-text, text-to speech, routefinding, and of course controlling mobs in videogames.
Route finding as in OsmAnd-like apps?
Or some custom implementation?yes, maps apps is what mean
I just don’t, and wont
Machine Learning (and even Large Language Models) do have genuine use cases they excel at and are probably more efficient at.
One good example is protein folding. ML transformed biomedical research by basically solving the protein folding problem in minutes on a GPU instead of essentially trying to brute force a solution that was millions to tens of millions times more complex. I’m glossing over details, but it’s mind blowing how much better (especially when you consider resource usage) how much better ML is in this case. Think $1 and minutes vs $100000+ and years… for a single protein.
The problem is using LLMs and Generative AI for things they seem good at but actually aren’t.
Why? Opensource AI is good if it is the corporate ownership and data safety aspect, right?
Why use Google translate when you can use offline opensource models to get translation? Why send your data to Google or some other company?
Also, you don’t need internet connectivity for offline models.
I just like talking to Gemini once in a while to sort out all the thoughts in my head and keep me from indulging in lingering depression spikes.
Use duck.ai. at least your mental state records will not be used against you later.
I don’t. Nor will I.
Why?
It is an unethical technology. I refuse it.
Technology cannot be unethical, right?
It’s use/application, safety regulation aspects, ownership etc. would be the focus, right?The corporations who supports copyrights are stealing data(from others and other corporations) for training AI.
By calling the technology unethical, we are letting them off the hook by making it vague and on the technology, rather that their decision to not provide credit or remuneration for the original data creators, right?You use it every day without realizing it, it’s been there for decades. The problem is not the technology but the people in control of it and their new bs hype bubble
Third party open weight models through OpenRouter for code review on personal projects
