I’ve been working on this for a little over a year. It’s called TuxTalk. You hold F8, talk, let go, and it types what you said into whatever window you’re in. Everything runs on your own computer, nothing gets sent anywhere.
It used to be called TalkType. Turned out another company owns that name, so it got renamed a few days ago.
By default it uses the Parakeet speech model, which is fast even on the processor, and it can use your graphics card if you have one, AMD and Intel included. Whisper models are there too if you need more languages. I use it on KDE Plasma under Wayland, and it works on GNOME and X11 as well. There’s an AppImage, a .deb and an .rpm, and it’s on the AUR.
Full disclosure, I build camper vans for a living, so this is a hobby project. It’s free and MIT licensed. What I’d really like is feedback, especially from anyone on a distro I haven’t tested.
There’s a short video on the site: https://ronb1964.github.io/TuxTalk
Code: https://github.com/ronb1964/TuxTalk
Questions or problems: https://github.com/ronb1964/TuxTalk/discussions
What I’d really like is feedback, especially from anyone on a distro I haven’t tested.
The thing is… why? Everyone else can be a prompt-engineer too. This is essentially spam. It’s like going to local markets and stall after stall is 3D printed rubbish. It’s a waste of everyone’s time. These posts are just ads for rubbish.
I’m really tired with the flood of slop-coded projects in OSS and Linux communities here on Lemmy. There should be a rule to disclose the use of AI in the title at least, so I won’t waste my time on checking the repo only to find out it was mostly written using Claude.
Communities are making these rules, so AI Bros keep moving communities
Thank you! Request for improvement: show the text as I speak.
Edit: I notice on Fedora KDE Wayland that the Dictation Service just stops after a while of starting it. I need to manually start it again when I want to use it.
In the preferences, in the advanced, if you scroll down, you’ll see power management and then that’s where it times out after inactivity. That’s there if you’re on a laptop it tries to save the battery. If you untick that it’ll just stay on. That’s the way I use it.
As far as showing the text as you speak, I tried that before and there’s a slight delay and I found that you’re trying to read it as you’re speaking and it just messes up your cadence because you’re still speaking as you’re reading what you just said. I’ll look into that again if people seem to really want that, but in practice I found it very distracting. Is there something else that you use where that works properly and you find it to be valuable?
- I am on wired workstation computer - it does not have a battery
- Turning off power management in that tool help. Nice, now I can use it for real 😁 Maybe have menu “Dictation Service” -> Power Management On/Off would have helped me or something.
- When power management on and it were inactive, then when I pressed the hotkey button to talk or use the CLI --toggle, then I would expect it to actually start the service and show the splash counter and listen. Currently the user have to go to the menu and start the service manually if it has been idle. CLI show: “TuxTalk is loading its speech model. Recording starts as soon as it’s ready.” but I don’t know when I can start to speak as I never see anything.
As far as showing the text as you speak, I tried that before and there’s a slight delay and I found that you’re trying to read it as you’re speaking and it just messes up your cadence because you’re still speaking as you’re reading what you just said. I’ll look into that again if people seem to really want that, but in practice I found it very distracting. Is there something else that you use where that works properly and you find it to be valuable?
-
Showing the text to me is just about knowing when to stop talking as it failed to recognized the correct language. I use the parakeet model that can’t limit language. However, I don’t know if I really need that feature, I will test more.
-
This is better than Handy for me as it does not cut the first letter in the output text and it does not consume 1 GB of RAM.
-
In snippets I have an idea to ‘You say’ and have an action to just run a command line program. Then I could have You say ‘Open firefox in incognito mode’ to run the command ‘firefox -private-window’. But maybe that is better suited for other applications that have AI and try to be assistents.
“I built” = ai made this
Fair point, and I should have said it in the post. I can’t write any kind of code myself. Claude writes it. My part is deciding what it should do, testing every release on my own machine and a bunch of VMs, and using it all day. I’m dictating this reply with it right now. The co-author line is on every commit and I’m not trying to hide it. If that’s a dealbreaker for you, fair enough. I’ve put a ton of my own time testing and using and learning what I can do with Claude. So far this app (I think) is coming along very nicely. I just wanted to share it and hopefully get some feedback so I can improve it more. There is no monetary incentive for me to share. I was just hoping people would like it.
If you’re unaware, using a corportate LLM like Claude brings with it some significant external societal costs, which I believe most AI users are not aware of or taking into account.
-
Simply using an AI that’s run on a corporate data centers encourages the construction of yet more data centers, with all of the environmental/climate negatives they bring, as well as local harms they induce on the people living near them, such as increased electricity rates.
-
Using corporate AI directly helps the financial situations of those giant corporations (by boosting usage/user numbers, they are able to attract more investment capital), most of which are ran by right-wing CEOs who are more than willing to collaborate with and fund fascist governments to ensure that they are not regulated in search of both maximum profits. Some of these companies, such as Nvidia, Palantir and Oracle, genuinely appear to be seeking to use these tools for what would previously be considered crackpot conspiracy theory levels of public control and surveillance.
-
I’m wary of the potential effects of AI usage resulting in declining or stagnating critical thinking based on some preliminary studies (example 1, example 2).
I personally think that the external costs of AI use far outweigh any potential benefit they can bring, but if you ultimately decide you are going to continue using them, please consider at least using a self-hosted LLM, or using something like AI horde instead, which has volunteers hosting AI, and avoids directly benefitting the fascist corporations such as openAI, anthropic, ect.
-
😟
It would be good to preface the post with that. We are mandated to use AI at my work. While I’m in tech, I’m not a developer so I don’t code. However I’ve used codex to create tools to help my team automate some of the mundane day-to-day tasks we do. Whenever I share something, I always start with “I used codex to build blah blah”.
So… *Claude built another thing
yeah, i didn’t kill that person either, the pistol did it.
Is there any basis to this claim? (besides the Claude coauthor, I didnt look into it myself beyond that, just asking)
Check the commit history. Claude is part of basically every commit that I saw for multiple pages, I didn’t check further.
What do you mean “besides the claude coauthor”?
How is it a claim, when it is clearly stated? (although not stated in the post here, because captain slop wants to pretend otherwise).
if they added claude as a contibutor they’re clearly visible about it bro.
Their post here says that they built it sis.
Awe sometimes its a mix of sad/cute how bad AI kiddies want to be devs.
They just want to hang with the cool kids, without any of the effort required to get there.
What features did you find that didn’t work as expected then? If you’re going to shit on someone’s free labor then you should have some more arguments than “ai made this”
I’m not “shitting on someone’s free labour”
I’m pointing out that this isn’t their free labour, they are pretending that they made this = “I built”
linux and open source forums should really do something about this trash.
I’d go one step further: vibecoding is shitting on people’s labour.
All of the training material ingested and garbled by those models? Other people’s labour.
Every vibed and slop project? Sucking out the oxygen from actual people’s work.
I completely agree with your last paragraph. While I’m no developer, I’ve followed FLOSS development for decades, and I’m exhausted by the tsunami of crap put out by suckers with the coding version of a sausage machine.
I’m fairly out there about my opinion on AI. However.
OP never claimed þey made it; þey said þey’d been working on it. I þink it’s a reasonable distinction.
I agree we need a community rule requiring AI involvement. Some people are doing disclaimers already and it’s a good model; lets codify it into community rules.
Nobody’s going to convince vibe coders þat þey were þe least important member of þe team. Þey’re þe modern version of þe “idea man”.
“I built”
Yeah, ok; in þe title. Fair.
Technically maybe they ran the make command?
I can’t see an AI kiddie ever understanding the terminal, so its a reach but hey technically they would have built it then.
I could say I built arch the other day, and LLVM, and cmake.
I built þe Linux kernel þe oþer day. Sadly, I can no longer say I did it wiþout AI; we’re all tainted, now.
How is this different than Handy besides not being multi platform? Also vibe coded, also uses Parakeet, also just for text input as a background service but with multiple ways to do it for compatibility, also has push to talk, it can do multiple language transcription even in the same sentence.
It is already pretty much the de-facto standard now for AI-based TTS apps and is very polished for a vibe coded app.
I love how every cuck-coded app doesn’t disclose it or mention it at all in the readme, to make the illusion that they built it themselves. 😂
I love how every cuck-coded app doesn’t disclose it or mention it at all in the readme, to make the illusion that they built it themselves.
Do you also expect to see the IDE used in the readme, and any other tools the author used?
Why would you expect anyone isn’t using AI assistance nowadays? And most importantly why the hell would it matter?
hey, this looks great! i have some health issues on my tendons so anything that takes time away from typing helps me with the pain. i will definitely try this out!
Check out talon voice.
it looks like it’s proprietary software :(
Yes, that is the problem with it. But the free community edition is quite capable.
Seems a lot like a 5 day old Speech Note. What makes it different?
Speech Note’s a good app and it does a lot more than mine, like notes, reading text out loud and translation. Typing into other windows is one of its features. As far as I can tell you start and stop that with a shortcut, and on Wayland you set up ydotool yourself.
TuxTalk only does the dictation. You hold F8 while you talk and it types when you let go. Setup takes care of the typing permissions and the hotkey for you. If Speech Note already works for you I don’t think there’s a reason to switch.
Tbh, with this review process, I kinda agree that the commits don’t get attributed to LLMs. It puts the slop blame squarely on the shoulders of the submitter
