How the SSI-263 learned to speak again: the whole story, day by day, from the first ROM read to this release, and who found what along the way. read it at:
https://github.com/tgeczy/ssi263-speech/blob/main/docs/history.md
Open questions: what we still don't know, and where the answer would come from. This is where I need the community to help. If you are a chip expert or knows someone who is in the SSI263, I need to get in touch with that person or team up.
https://github.com/tgeczy/ssi263-speech/blob/main/docs/open-questions.md
Remote
Tamas G
@Tamasg@mindly.social
mastodon 4.5.18Originally from Hungary, but living and working in the US now. Accessibility Web engineer at Spotify full-time, , avid philosopher. Fun, random, optimistic. Friend to many, with an open mind. Passionate about accessibility and usability. Tinkering with Raspberry Pi, meditation, ham radio (K7HUN,) more. eng posts do not reflect my employer's opinions or views
460 Followers
318 Following
50 Posts
Joined January 20, 2023
You know what would suck on a birthday though? a badly written pull request from someone, full of errors and vibe coded crap that pulls in dependencies the project never needed in the first place and just muck it all up.
RE: https://mastodon.social/@cary_rowen/117343878605032126
lol. I guess my Masto followers can decide that, I'll say getting 10 stars on the repo in 24 hours felt good.
Quoting
@Tamasg@mindly.social Although I'm reluctant to believe it, this is very likely garbage produced by someone who doesn't know how to use AI.
Open quoted postI think @serrebi@serrebiradio.com now that the SSI263 add-ons are live on Git your store should be able to crawl and pick them up soon.
As I use all this retro tech around me, I'm reminded of something deeper. The peace wasn't discipline to only use one or two apps at a time, it was the hardware enforcing it for you. You didn't monotask because you were wiser; you monotasked because opening a second heavy thing would've brought the machine to its knees. The scarcity was the feature. And the quietly funny part is that we've now spent twenty years building software with things like focus modes, tab suspenders, "do not disturb," app blockers, to manufacture the exact constraint the old machines gave us for free. We're paying to re-impose the limit we celebrated escaping. And I find a huge irony in this. Back then, you went online to download files, check the news and read your favorite forums. You then closed the browser, and ditn' think of it until next evening. That's how it worked. There was no "let's check Mastodon because Claude's grinding away on a work task" thought.
And the latency did something we lost too: it gave every act a beginning and an end. Downloading an album was a thing you did because you committed, you waited, and at the end you had it. It was yours, finite, sitting on a disk you owned. Streaming is infinite and borrowed and never finishes, so nothing is ever a completed act, it's just ambient flow you dip in and out of forever. A finite collection you've curated has a shape. Infinite access has none. That's a real loss to me.
Clicks Communicator brings 4.03-inch AMOLED display and physical keyboard to Android 17.
For 12 gigs of RAM and a fairly mid-range but OK chipset, these specs are not half bad to see on a keyboard phone.
https://linuxgizmos.com/clicks-communicator-brings-4-03-inch-amoled-display-and-physical-keyboard-to-android-17/
Ok, I get that on one hand, we need to get used to vibe coding as it's here, but also like, forcing that view on those skeptical or very rational about the limits of vibe coding vs where human logic and creativity, ingenuity come in will not get you anywhere but toward a place of more hatred and condemnation. Remember that just because "vibe coding is here to stay" doesn't mean everyone has to accept or be OK with it either in their projects or their lives, and that's OK. We humans should be able to respect that view and not have to force infinite vibe-coded projects on them out of rath or because you can. Especially if 3-quarters of them don't work as expected. Not directed at anyone by the way, just seeing this from many sides in the community as of late.
And yes, it does go the other way too: No big meteorite is coming for Earth just to wipe out AI. That's doom-talking nonsense. So while AI vibe coding is here to stay, the choice to actively not use or subscribe to things that pump out more Vibe coded things is also well justified and I think a human right to choose.
Replying to
@prism@infosec.exchange haha OFC not surprised :D I see this pattern I'd say at least once a month and sadly the gap's just gotten worse with AI, not better. Again we can help build MCP servers that extract knowledge info into nice queries for people, but it is by choice a developer invokes it or installs and connects for auth, so, the dice is rolled on whether it gets majority use. I think for us the only real blessing AI has given as a team is that we have good dashboards we can spin up for leadership that can tell the accessibility story with better numbers reflected across all the products used internally. Now for that concentration of work, I'm all for.
Replying to
@prism@infosec.exchange yeah, if you can tie data to the actual corporate mission statements, goals, and internal bets, and drive a narritive of a story behind it that is glanceable, you got it down. But that is way way harder than what I'm saying in a sentence, LOL. It requires knowing your data sources well, and just being resourceful internally in the first place, which sadly I mean just as at any workplace, I don't see in everyone. There are allies of course in whom I do, but it takes time and years of experience to connect with just the right folks too. We had some data engs help out with pulling data for instance, but it was through a random meeting we were able to surface our ideas and that person was in the room at the time.
Replying to
@prism@infosec.exchange yeahI mean for me it is specifically the passion and knowing the things I say carry weight to engineers into improving something in the A11Y space, but my work can get torn down anytime with a new refactor too and that's a toil of a risk I'm always optimizing to minimize. Ultimately if they were to move me to say, localization or another team tomorrow I'd have to say bye to the job because my heart wouldn't be in it even if that position change came with perks.
Got in touch with the developer of FlexVoice MindMaker.
Also, today the maker of PC Robot graciously offered to mail me (yes, physical mail) the original ISA card with the PCF8200 chip inside.
Also, the author of Brailab himself was happy for my Internet Archive upload.
Mind=blown all day. Just, blown. The fact that I'm able to talk to these people who are well into their 80s and get a chance to tell them, "hey, your work lives on. Don't worry, time won't eat it." feels so good.
Replying to
FlexVoice 3.01 shipped in 2002–2003 into a company whose lead speech-industry investor had just been destroyed in one of Europe's largest accounting frauds — taking the entire Flanders Language Valley funding ecosystem with it, and thousands of small Belgian investors besides. Meanwhile everything L&H had swallowed, Dragon and Dictaphone included, consolidated into Nuance. A small elegant engine from San Jose had nowhere left to sell. Truly sad, really. The product didn't fail — its ecosystem did, and the timing was almost cruel: it was finished right as the ground went out.
The rest fits József's account exactly. Look who's left standing on that investor list: Creative Technology, GIC (Singapore's sovereign wealth fund), InnoMedia, Institute for Infocomm Research — all Singapore. When Mindmaker wound down, the technology went to the surviving investors.
FlexVoice wasn't abandoned because it wasn't good enough. It was orphaned by a fraud in Belgium — and twenty-three years later its author is still improving it, by email, because he cared about the technology. One of the stranger tales of text to speech I've heard if there ever was one.
TikTok is telling people to put their phones into the fridge.... Sigh, just why. Bet it was some AI that gave the idea to someone, just like when they told us eating a rock a day is a good thing. If you want a phone with internals covered in water droplets as condensation takes its toll it's absolutely the best idea though.
I guess what gets me is: We're out here in the first world complaining about how RAM prices are destroying the tech industry, meanwhile, all it takes is one single Earthquake to take all that away.
My heart really aches for Venezuela - don't care how much they are behind economically, each of those 900+ who passed were part of a network of families. A brother, a son, a daughter, a wife, a husband, a cousin, an uncle, an Aunt. Whatever the case may be. And no matter how many 128 gigabyte-RAM Macbook Pros you build up, one tragic event if you're in near any of the hundreds and thousands of fault lines, can take it all away. At this point I'd rather put any desires for money spending into a relief org like the World Central Kitchen and others who promptly showed up over there.
Replying to
Why: before first unlock, Android's user storage is still encrypted (Direct Boot). Third-party TTS engines live there, so they can't load. TalkBack itself IS direct-boot aware. Google TTS is too. But TalkBack won't fall back to it — it just stays silent. This has been known since 2021: GrapheneOS filed a request for direct-boot support in RHVoice, and eSpeak NG actually added it at their request. https://github.com/RHVoice/RHVoice/issues/271
Plot twist: my new Galaxy Watch 9 gets this RIGHT. Samsung's TalkBack fork boots speaking with Samsung TTS (the direct-boot-aware system engine), then hands off to my chosen engine after unlock. I deliberately rebooted into my lock screen tonight to test. It just… spoke. This is software, not silicon — nothing stops Google's TalkBack from doing the same on any Pixel Watch.
So, Google: TalkBack should fall back to a direct-boot-aware engine before first unlock. Your own TTS already supports it. Samsung already ships exactly this. A blind person locked out of their own device because their screen reader has no voice isn't an edge case — it's the lock screen.
"Hell" is considered a mild swear word in English because of Protestant cultural history. While many languages view the literal concept of hell objectively, English speakers historically grouped the word into blasphemy and "liturgical swearing". Calling upon or wishing someone to a place of eternal damnation was viewed as violating the sacred.
Imagine, then, me as a little 8th grader from Hungary. I had told my teacher that a test can go to hell. Sent to principles office. Very confused. Why, when in Hungary, hell was never a swear or cuss?
However, by contrast, in Hungary, to tell someone to go "francba" (to France!") is an extremely bad swear kids are discouraged to say. Instead, you use "Fenékbe" (into your butt) - so like, "go put yourself in your own ass!" - The French thing comes from the numerous Venereal Disease cases the French (supposedly) gave Hungarians.
This prime day stuff kind of makes me sad, so I'm choosing to not browse it. Basically you buy a new computer now at the price that it should have been from the beginning. Then we see how the Steam Machine suffered in the pricing from the shortage, also coming with single-channel RAM, and yeah, it's enough of a clusterfuck for me to not want to look at electronic aisles. Same with the Rog Ally X. It's a great time to buy one, as Best Buy also discounted it to $899, but that was the price it was at before they hiked it up a little. The new MSI Claw is double that.
The 4 TB Samsung SSD in my desktop is now $1099 instead of the $335 I paid. Oh wait, but for Prime day, you can snag it as low as $889. Just, what a joke.
Also if Google Pixel Watch 5 is seriously going to use just an updated Wear Gen 2 accelerated chip, maybe with slightly higher clocks or GPU cores but not quite as good as the Single A78-cluster plus 3X a55 cluster of the SW6100 (the Wear Elite in Galaxy Watch 9) I'm seriously going to be disappointed. Like, seriously. That will seal any tinge of sorrow I may have even felt not going with them, despite their perfect-circular shape. While The lack of a crown on the Watch 9 bothers me a little, the circular motion around the edge of the screen behaves the same for scrolling in the first place and I was a bit surprised at how easy it was to get used to it.
Replying to
and if anyone's considering a Pixel Watch versus a Galaxy Watch, go with Galaxy until Google gets their shit right on this matter. It's sad to not be able to recommend Google's own Pixel Watches because of it. Having access to your pin entry on first boot is non-negotiable.
Well well, the US government strikes again, all access to Fable 5 models is disabled for all customers, just wow.
Trying to pick it as the Claude model just redirects here: https://www.anthropic.com/news/fable-mythos-access
Replying to
@Jage@mas.to @TheQuinbox@dragonscave.space and, windows plus minus types an endash, – looks like this wasn't added until 25H2 itself, so it's one that got a quiet rollout.
Google Gemini is for sure back for me both at home and at work, so I'm thinking the outages are finally solved for them. Took out most of the day though, had to focus on work that was more doable with Claude Code and the web and not built on using Gemini models like image description stuff that use my work API key. Alas, so goes the fragility of the AI world and our dependence on it (and often not by choice for situations such as work, where staying ahead matters.)
I see Anthropic removed the ARIA from the "open sidebar" button, making it no longer be a toggle-button but a standard one for which the label changes, presumably still via an Aria-label (although, who knows. Could be a visually-hidden Span but I won't raise my bar that high.) Still, one less ARIA attribute usage and a native button keeps it more consistent across platforms in how it announces, so a small win there.
so iOS 27 so far isn't horrible. The weird comma-inflection for Eloquence on home-screen icons will take some to get used to, along with how the status bar has no pauses between the item and role. Swiping feels a tad faster if anything, and all my apps like banking still work, so I haven't found a reason to go back just yet. The only new VoiceOver change (aside from the stuff in VoiceOver recognition, which now has a toggle for only using local models) is that the Braille "Item overview" mode can now be customized with the number of letters shown of each item across the display. So, yes. It does feel like a performance-focused update, and of course even for me there's a waitlist for the new Siri, so have not played around with it yet. The ability to ask about an image though with VoiceOver (in addition to exploring like prior through the rotor actions) is presumably powered by this same system, though.
OMG. Toy Story 5. Totally 100% recommend. Even if Forky's role was a bit more minor than I would have liked, but at least you do se a lot of classics reappear like Slinky And Rex. And I love how they weave the starting point of the story throughout it, there was serious creativity placed into this film. Swift singing the credits was also amusing. All in all, high recommend.
oof. Notepad++ creator issues warning to users about an "unauthorized" macOS fork
https://www.neowin.net/news/notepad-creator-issues-warning-to-users-about-an-unauthorized-macos-fork/
Inside interoception: The hidden sense of how you feel inside
https://www.technologyreview.com/2026/06/12/1138833/inside-interoception-brain-body/
Been throwing up all second half of the day, not sure if it's the BBQ chicken from the freezer (was only about 2-3 weeks old, but chicken, so...) or my heart-burn getting so up to my chest that I slept only like 1-2 hours the night prior. Regardless, the nausea has finally come to pass.
also, y'all, Apple really, really destroyed the Weather app with VoiceOver in iOS 27. It's a huge mess of repeated controls, items not grouped together properly to their parent relational element, so, so bad. Swiping right does an infinite-scroll thing where you see the next city's weather if you're not careful. Explore by touch brings you some sanity, and the audio graphs are still there, so like, clearly it's a mix of the old and new Weather app UI but boy it's a mess. It being beta 1, I'm not superbly concerned, so I have not filed feedback on it yet, figured I'd give their devs time on this until beta 3 at least or so
Replying to
@robini71@dragonscave.space @fireborn@dragonscave.space I'm just laughing at the fact that they called it "talk forward" and consider it even a serious project with that kind of naming. It's like, you know the thing was made out of spite with AI.
Replying to
@zersiax@cupoftea.social @fireborn@dragonscave.space @FreakyFwoof@universeodon.com @MostlyBlindGamer@dragonscave.space hm, senior dev. Then at least they should have some skills to dogfood their own code, surely that's one you learn well and to spin up various emulators or test devices with what you're vibe coding. But who am I to judge, maybe he really does so and just has no clue how to use screen readers, in which case having dedicated groups in closed testing that can provide you critical show-stopping feedback is still a skill to have learned by senior dev days too.
Replying to
@danestange@caneandable.social ha. I saw that news story! It made me laugh so so much! Yeah, I genuinely should give GPT-5.5 class a try sometime though for a sesh, I hear they're falling behind on some benchmarks in Opus-4.7 land with OpenAI releasing that. For Chinese models, I mean, technically I have qwen3:30b-a3b here locally and I have my Claude call into Ollama through MCP connectors to manage the server and then call it for extra review of its thinking. It's helped a lot. Even though it's Alibaba it's very competitive with Gemma4, and so tripple-checking something between Claude's brain with a Qwen model plus Gemma has been a nicer workflow in some things than just blindly trusting Claude alone.
NVDA Store: I don't know if it's a good idea for Panthera and Outspoken. Claude's suggesting it, but for one, I have zero idea how difficult submission is, and for another, maybe they would disallow it for being an Apple engine. Who knows. I have not considered it tonight until it was direct-asked again. So yeah, it's up in the air for now.
And yes I know of the repo. I've been poking at it. https://github.com/nvaccess/addon-datastore
"Anthropic says Claude AI hacked three organisations during cyber tests" Watch OpenAI come out with, "wait, we actually found ours hacked four!" Only in today's world is it completely normalized to have a dick slinging contest about who hacked the most of what and how they did so.
Replying to
@socheat@dragonscave.space hahahhaaha was google TTS rejected, really wow. Good to know considering it's not even bundling anything by itself as an add-on, the repo took the correct path on that.
Replying to
@prism@infosec.exchange now that definitely sounds like a great example of overengineering for overengineering's sake (just because you can) maybe because AI or the visual nature of what you're coding turned someone into that type of engineer at the developer's studios. Would be my guess. But boy is it annoying to come up against over-engineered shit like that
"To display this content from X (Twitter), you must enable advertisement tracking and audience measurement." ope, nope, never, fuck you X for limiting your content like that, but that ain't gonna make me give in.
Replying to
@prism@infosec.exchange yep and even if there's a main corporate design system, which a lot of companies have (Spotify's Eng blog has some of my early work with Encore detailed publically,) a million smaller local design systems can pop up. And all of those could be forking a different version of your main one. Then a step worse are teams who just outright say why bother when an AI can build me components for cheap without context beyond design tokens to the main one to use. Sadly there's just not one easy solution to that corporate environment whatsoever. Beyond raising awareness through education, documentation, skills, yada yada yada, and hope + pray a team at work is diligent enough to read them. So, TLDR, corporate environments will always be a shitshow for A11Y to some degree, and whether at Microsoft, Spotify, Shopify, Atlassian, whatever your poison, they all come with it without exceptions. The politics, the siloed nature of work within only a main studio or org, yup.
loool should I feel a bit glad that Mist World hasn't made it to AgNow yet? Probably. Imagine people playing from there for like 12 hours at a time. ha.
Replying to
@mcourcel@allovertheplace.ca yeah, sadly the core configuration hasn't really changed since Pixel Watch 2nd generation, when they got the quad-core A53. By today's standards those are almost 10-year-old cores, so even the A55 is a bit newer and does better per clock cycle.
Replying to
@gklein88@dragonscave.space not surprised! Good luck out there. I feel the same way when I have to dig deep into the domain of what data engineers or backend peeps do with their pipelines. So much configuring, figuring out Docker local run guides, bla. But, I do it anyway to be able to better assert a test for a team or something, the sacrifice's worth it. Of my sanity. xD
Replying to
@JamminJerry@mastodon.stickbear.me @mcourcel@allovertheplace.ca oh yeah for the ultra one, but also like, for me, a 47MM screen is huge :D I suppose still not as big as the Apple Watch Ultra at 49MM but still! xD the larger 44MM watch has a 445MAH battery which is still doable.
Replying to
@Jage@mas.to @TheQuinbox@dragonscave.space from a website (not AI):
Em dashes are longest. They act like commas or parentheses and create a break in a sentence.
En dashes, shorter than Em dashes but longer than hyphens, express a range of values or connect related items.
Hyphens, the shortest of the three, connect compound words and help avoid ambiguity.
This is very nuanced and I suspect AIs often use the emdash when in reality an endash would suffice, but they don't get the distinctions
Replying to
@JamminJerry@mastodon.stickbear.me @mcourcel@allovertheplace.ca I mean, definitely not. Think their's tops out around 600 MAH on the Ultra 3 for sure. I bet Samsung is using a Silicon Carbon battery in these watches. Apple's 46MM one is also only using a 327MAH battery as opposed to the nearly 450 in the Galaxy one, so my guess is they had to do something to keep the weight of the watch this good but boost the capacity a little. Watch 8 had around 430 MAH so it's still about a 10 to 20% battery increase.
Replying to
@simon@procrastodon.net @pitermach@dragonscave.space @sparksexist@dragonscave.space you're definitely making me wonder with that last point. Does contacting some of these hard-to-reach companies (or ones super dismissive of their A11Y practices being important) from a corporately-owned or nonprofit e-mail change their tune of importance? It might be why also Atlassian and Slack get back to me in 24-48 hours with clearly read responses to the feedback, though I'd like to think they both also have stronger practices internally which is more likely. However, you never know, I guess.
Happy Friday! If it is your birthday today, or you know of someone who has theirs, maybe it's their lucky day. you want to hear an old Hungarian formant synthesizer sing like it did in 1991? Listen no further.
I think like, 90% of my followers (and people I speak with day to day) are kind people who wake up every day here to better the world. Even if the percentage of "kind and friendly" people may be lower in the actual world, what matters y'all is who you surround yourself with.
If it's a lot of those who are passive aggressive or easily develop an attitude over your smallest of questions, or become argumentative, then surely you will think the world is made up of high strung not nice people, too.
Replying to
@serrebi@serrebiradio.com I mean, I vibe code myself :) There's just a lot of difference between human-tested vibe coding, like, programmer actually sees thoughts and code written during the process and actively steers the AI towards other choices or decisions out of expertise. Versus just having an AI go at it for hours with minimal prompting and agreeing to what it should do next based on what it's proposing always. So quality certainly matters, but yeah, I've also just seen a lot of people wave it in skeptic's faces like "want proof it's here to stay? Here are 3 other vibe-coded projects that are good!" which then puts everyone on defensive ground and does nobody any good.