The first run experience isn't great, you'll need to download all the files from the model, start the app, and then go to settings and configure the model directory. It runs well on Mac and Windows; I haven't tested it on Linux in a couple of months since my Linux install is out of commission currently.
Nice - I'm definitely going to take a look at this. I've built my own cross-platform (Mac/Win/Linux) live captioning app on top of Nemotron, and it works well but dealing with ONNX is kind of annoying. With this having Rust support (I built it on Rust/Tauri) it should be a pretty solid candidate; I'll have to see if I can find a Silero VAD implementation that doesn't depend on ONNX, or maybe I'll see if the clankers can migrate it for me.
Wasn't mentioned on mine, either (Ohio, United States). My subscription is through in-app purchase, so I'm assuming that'll go through Apple's usual "your subscription price is increasing" flow.
I have a 2023 Crosstrek, my wife has a '21 Ascent. I have the same habit you do - edging away from large trucks slightly - and both of them do the same thing you described to me.
It's essentially that Subaru's lane system actually has two levels: it has lane keeping where it's just trying to keep you inside the lines, and then on top of that it also has lane centering which is pretty much what it says.
Just a note for you or anyone reading who has a recent Subaru and doesn't know already: if you find the centering really bothersome, you should be able to be able to go into the settings on the instrument cluster display (up/down arrows at the lower left behind the wheel, toggle it until you get to the "hold for settings" option), find the Eyesight settings, and turn off lane centering. It will still try to keep you inside the lane markers but won't try to park you right in the center of the lane. In that mode, it's more like the Honda Sensing system I had on my 2016 Civic.
I go back and forth a bit on it but mostly keep it in lane centering mode now - I've gotten used to how it positions the car in the lane, and it lets me focus more on what's going on around me than micromanaging lane position and such.
I'm deaf, so I test a lot of speech to text and transcription apps from an accessibility point of view.
My answer to "why have a monthly subscription" would be that you need capabilities that Whisper doesn't handle well, like real-time transcription in noisy environments.
That's not the niche you're targeting here, though. :)
My experience is that Whisper - not being built for real time speech to text - isn't as good at it as other tools are. You can hack something together by stacking together progressively more audio frames to feed to Whisper to give it context, but IME, you're going to get better results from a model that's designed for real-time STT in the first place, or by using a service like Azure Speech to Text which has excellent noise resilience... but which is also an ongoing cost which would justify a subscription. Real-time Whisper also devours your battery quickly.
That said - while I've had very good experiences with Parakeet in MacWhisper, I'm curious if you evaluated Apple's SpeechAnalyzer APIs at all. It's unfortunately limited macOS/iOS/iPadOS 26+ since it's a new API, but it's on device, has comparable quality of results to Whisper Large v3 Turbo and Parakeet, and seems to be better on battery usage.
Yep, I'm also deaf (since age 6), went through a lot of speech therapy, and have a very pronounced deaf accent. I live in the midwestern US (specifically, Ohio) and at least once a year I get asked where I'm from - England being the most common guess, but I've also had folks ask if I'm Scottish or Australian.
AI struggles massively with my accent. I've gotten the best results out of Whisper Large v2 and even that is only perhaps 60% accurate. It's been on my todo list to experiment with using LLMs to try to clean it up further - mostly so I can do things like dictate blog post outlines to my phone on long car rides - but I haven't had as much time as I'd like to mess around with it.
Yeah, I've been deaf for over 40 years now and captioning glasses are something that I've wanted ever since I was a kid. I'm not a particularly big fan of Meta and I have some serious reservations around privacy that need to be satisfied, but at the same time it's really exciting to see this going from "pie in the sky thing I dreamed about having when I was ten" to "actual existing product."
There's a few other companies/startups working on this too, but a lot of the glasses they're producing are very ugly. There's a couple that didn't look bad, but from what I'm seeing Meta's are a combination of the best-looking ones and best display so far, and I'll be very curious to see the reviews.
One of my weird hobbies is radar chasing storms, and all of that stuff is completely normal. NEXRAD is very sensitive, especially when it's in clear air mode (it has different modes depending on if it's raining in the area) and can pick up things like dust, birds, bats, and insects. There's also ground clutter from things like buildings, wind farms, and even cars.
> In the last 10 years has technology actually made my life better?
In my case? Yes, absolutely. Automatic speech to text is now cheap or free, ubiquitous across most platforms (even Linux!), and generally very effective. Total game changer to my ability to participate in meetings at work and in society generally.
I would say it's cool in the sense that building anything is cool, but I find myself mostly in agreement with your take, although with a caveat.
I can't find the quote now, but someone (I think simonw?) said that they feel a bit of an obligation to spend at least as much time working on writing something as it would take to read it, and I agree with that... if you want me to spend time reading your post, I'd like to know you actually made an effort on it.
For me, writing is thinking, and helps me refine my thinking, so I don't use AI to assist writing process. I agree with the comments that AI writing tends to have a specific voice, and I don't care for that voice and don't want my writing to come across that way.
Where I do find it useful in writing, however, is as an editing pass in an advisory role. I don't ask it to rewrite anything for me, but I will ask it to double-check for excessive passive voice, tone, does it raise unanswered points, etc. I typically write my draft posts in Zed, and use Zed's AI chat panel to throw a request at Claude. The big thing though is not blindly accepting every suggestion the AI makes - I read them, think about it, and sometimes adjust the post based on that feedback. It's a useful sanity checking step and while a real human editor would be preferable, I can't justify the cost to hire an editor for my little blog that probably gets zero hits most days. :)
Yup. I grew up in Hancock Co and used to ride occasionally with the bike club there, had a couple of days when the ride out was brutal because of persistent, endless wind, but then the ride back was awesome for the same reason. :)
The tools for in-person are getting better, but aren't frictionless to set up and sometimes require you to spend time futzing with getting your iPad or iPhone to actually see an external microphone. I don't know if Android is better about this or not, unfortunately. I would _hope_ that interviewers would extend people a bit of grace about this, but who knows.
As an aside - I saw your post on Apple Live Captions, and completely agree with you. I've been slowly adding to a collection of reviews of various captioning tools, and was _very_ critical of some of the choices Apple made there.
I have the OG 13" MBP M1, and it's been great; I only have two real reasons I'm considering jumping to the 14" MBP M4 Pro finally:
- More RAM, primarily for local LLM usage through Ollama (a bit more overhead for bigger models would be nice)
- A bit niche, but I often run multiple external displays. DisplayLink works fine for this, but I also use live captions heavily and Apple's live captions don't work when any form of screen sharing/recording is enabled... which is how Displaylink works. :(
Not quite sold yet, but definitely thinking about it.
Best retro I ever had, on a small team of seniors: we all sat down, looked at each other, agreed that a sprint happened and we couldn't think of anything that was good or bad about it. Then we called in our manager, who also acted as scrum master, so we could do planning for our next sprint. I thought this was reasonable enough - ostensibly we'd do retro and then planning back to back, and none of us minded the chance to take a few minutes and reflect if we had anything we should discuss.
By contrast, I've worked with scrum masters who were strict about the process and insisted _every_ retro needed to have at least one improvmenet or action item out of it, preferably more. I found this pointless and I've rarely seen them actually followed up on.
Yup - I've done double-ended safety razors, I've tried the Gillette and Harry's disposable ones, but in 30+ years of shaving what I keep coming back to is Braun electric razors. So far, they're the only option I've found that doesn't leave me with razor burn.
Currently using a ten-year-old Braun Series 7. I can get away with infrequent shaves (~3x a week) since my beard doesn't grow very quickly, so I replace the foil and cutter heads every 2-3 years. It's not as cheap as a double-ended, but for me, it's a better shave, worth the cost, and cheaper/less waste than disposables.
If I upgrade anytime soon, it will probably be to something that's designed to be used in the shower - I do miss a nice shaving cream sometimes, and unless I have a bottle of 'Lectric Shave handy, I don't usually like to shave right after showering.
Yes, I was on an internal project recently that wanted to use LLMs in a way that was appropriate to evaluate if changes between two versions of a text were semantically meaningful, and limited to that scope, it would've been a really valuable tool.
We had a directive from management to, for political reasons, use AI in the tool as much as possible to show how innovative and forward-thinking the company is. This led to a bunch of poorly-thought-out choices and while the project is in production and has internal users... I don't think it was particularly successful.
Not all of that is due to the "use AI" directive; there were also poor technology and deployment stack choices that made things overly complicated and cost us a bunch of time.
I did something similar recently with my daughter's school calendar and ChatGPT with gpt-4o; her school has a ton of closures/teacher work days, I fed the PDF of the calendar in and asked it for all the dates that impacted 2nd grade, then asked it to create an icalendar file for them.
Oddly, it didn't want to create the actual file, but gave me a Python script for doing so; you just need to be sure to tell it what time zone you are working in and that you want it to be an all-day event, or you'll get suboptimal results.
Speaking as someone who's deaf and uses these services a lot: for speech to text, the AI stuff is getting rather good.
I'm not saying it's perfect for every situation, but I have a very high success rate using InnoCaption[0] for captioned phone calls, including to places like restaurants with a lot of noise going on in the background. InnoCaption does both live person and AI-based captioning; since they started offering the AI-based option I've left that on, and I've never had to switch to human operators to continue a conversation.
That said - I'm not deaf from birth (lost my hearing in elementary school), so I voice for myself and that does simplify the process. I have used the old school text-only relay services and that was always such a miserable experience for me that I would crawl over broken glass to avoid making phone calls, especially going through phone trees. That's one area that relay operators still have a major advantage on. IIRC, Google's Pixel phones are supposed to be able to navigate phone trees for you, but since I use iOS I have no personal experience there.
I've been running Consent-o-matic [1] in both Chrome and Firefox for quite a while now, which automates a lot of them. You can set your preferences for what categories of cookies you want to allow.
Right now it only supports languages supported by parakeet-rs and Nemotron (so... English only as far as I'm aware) and you'll need the ONNX version of Nemotron: https://huggingface.co/altunenes/parakeet-rs/tree/main/nemot...
The first run experience isn't great, you'll need to download all the files from the model, start the app, and then go to settings and configure the model directory. It runs well on Mac and Windows; I haven't tested it on Linux in a couple of months since my Linux install is out of commission currently.