Dunning–Kruger at its finest…
3 Likes
Just dawned on me that reader/speaker difference might run deeper than I thought. I just realized that I’m habitually trying to use data/ideas exchange as default social currency instead of signals like encouragement, and protocols do not match. I’m not sure how to switch/translate.
Definitely seems to be more complex than “add emojis”.
There are differences in logic and core ideas. Like, what’s safety, being around people or having enough data to navigate confidently. What’s contact, warmth or understanding. Language is a way to connect to people or people are a way to connect to language.
1 Like
Let’s try again, maybe. Who was the translator, what’s their level of Irish and qualification? Which provider is AI model, what was it trained on, how was it chosen? Generic expressions like “our model”, “Irish speaker” and “government official” are unverifiable and feel like flying in the dark in unsafe environment to me. I don’t need encouragement, I need to understand what exactly I am engaging with. Lack of transparency feels disturbing to me and I find it hard to communicate when I have unresolved doubts about trustworthiness. It’ll be significantly easier for me to switch to something productive if I have accurate idea of what am I looking at. I sadly don’t have several years of prior experience with nice people here, so I have to base my conclusions about proper level of trust on what I experience right now. I’d like to trust, but I can’t trust implicitly, I have to base my trust on something and personal accounts of people who differ from me significantly in cognitive style don’t really work for me in this regard. Does this make sense?
It’s absolutely fine for you not to trust us - in fact, that’s probably a sensible starting point for any new social connections - I personally choose to trust in people’s good intentions until I am given convincing evidence that I’m mistaken, but that’s just my personal preference. If you can’t trust us because we differ significantly from you in cognitive style, we will accept and respect that.
We’ve always believed in transparency, and will always answer questions as fully as possible (given time constraints).
I’m not sure we have current permission to share the name of our translator/proofreader - @Kai can maybe check?
In the current Irish course, we have a mixture of human translations from English, AI translations proofread by a human, and AI translations that haven’t been proofread yet (we’re working on an editing interface to allow people to suggest changes, but it’s non-trivial). We primarily use Anthropic’s Claude for translations and phrase generation, and either Azure or ElevenLabs for artificial voices. We are also working on an interface to allow people to contribute recordings - also non-trivial. I can’t go into any detail on what Claude was trained on because I don’t know. We chose it because it seemed in our testing to provide the best results of the current choices.
Please feel free to ask any further questions you have.
1 Like
Buíochas le Dia! That starts to feel like actually functional communication. It just occurred to me that tech specs weren’t immediately obvious signal to send in code system based on exchanging social signals mainly. Persuasive power of anecdotal evidence, beliefs, evoking authority figures and social pressure is usually lost on me completely, but I understand common sense, formal logic, hard data and tech specs. After some amount of mutual culture shock I suppose it should be possible to take best from both worlds.
Claude works pretty good on Russian, I haven’t used it much on Gaeilge, but yeah, generally it’s pretty good at translating.
Azure and ElevenLabs are not specialized for working with Irish. Irish is a known weak point for speech synthesis because of large dispersion of phonetic variants and scarce data. Abair.ie is considered state of the art for Irish speech synthesis, they don’t do any other languages except Irish and keep dialects separated. Also, that’s Trinity college project, it doesn’t get more credible than that. At this point they should have api already, it’s worth at least a try. They are working at children voices now. Still occasionally scratchy and lacking phonetic nuance compared to real McCoy, but at least their fadas are believable. (Palatalization is more of a problem, but speech synthesis is not very good at handling it generally.) For example, á here in dhá has that natural depth it’s supposed to have. That’s Connemara female model. (Fear a bhfuil dhá theanga aige, is fiú beirt é means a man that knows two languages is worth two men.)
For comparison. L in bhfuil sounds differently, that’s palatalization, it’s a big part of Irish phonetics and robots are bad at it.
AI translations generally can cope with simple Irish syntax, start to be very unreliable on complex structures (this applies to Google translate too). As SSi tends to use complex structures early, this can become a problem early. Beta-testing done by beginners would help little finding errors in complex structures (they won’t know correct variant from incorrect one), at least three L1 speakers should work through all the curriculum to be sure it’s more or less correct, and it gets much more reliable if they are also Irish language teachers or have other relevant professional skills. (Also it helps if they are in contact with each other, because if they have different ideas about complex grammar, it’s better to let them argue and wait until they agree on something.) Incidentally, how much involvement from native speakers do you have at beta-testing at this point?
1 Like
It’s really interesting reading your comments on Irish, as I come from virtually no knowledge apart from having dabbled in it a little. I’m not sure I agree with your suggestion that reading should be mandatory from the beginning though.
My first attempt at learning Irish was a few years ago now before a trip to Ireland. The only resource I had was Duolingo, and it drove me insane that it forced me to write the words. I found myself shouting at my tablet, “I’m not going to Ireland to write Irish! I just want to speak it!” and I gave up. Even with the early SSi Irish, when I wanted to learn a little, I always shut my eyes to focus on the sounds, and didn’t want to see how the words were written as it was a distraction. I wanted to get a feel for the sound of it first, and become comfortable with producing those sounds myself, before I was sidetracked by how the written words looked.
I did the same with the little bit of SSi Dutch that exists - learnt from listening and saying the phrases before I went to the Netherlands for a weekend. I saw written Dutch there and my reaction was, “Good grief! Is that how you write that word I’ve been saying?” but Dutch friends were impressed by my pronunciation.
For me, I think it’s important that the reading aspect is optional. Personally I would get a grasp on speaking some Irish first, then turn the text on and surprise myself with what the words actually look like … and until I started learning with SSi, I’ve always considered myself very much a visual learner! That’s how I learnt languages in school, and when I went to a class that focused just on speaking and we didn’t write anything down, I couldn’t remember a thing from one week to the next. Of course that didn’t have the constant spaced repetition of SSi, which is what makes a huge difference.
But it will be interesting to see what other people think once they have that opportunity to choose.
3 Likes
I suspect we might be getting into learning preferences again, but I’m with you on the Irish Deborah. I have mostly been doing the Irish course with the phone in my pocket. Now and again I’ll look to see how something is spelled, but that’s mainly to check my assumptions about grammar. I have also been doing Duolingo on the side (and getting frustrated about spelling) as well as dipping into some grammar books.
I’ve booked to go to a residential course in Galway at the end of the month. It will be very interesting to me to test out (a) how my pronunciation so far stands up, (b) how much of an issue spelling is on that course, and (c) what their “traditional” teaching methods are like (and how they compare to the Dysgu Cymraeg curriculum).
I’m at the beginning of brown belt now - should be on black by the time I go. The course is a 5-day one for beginners (so A1, I think) and I’ll miss the first two days of it because of work. I figured that I should have enough Irish to be able to cope with missing the first two days, and that if the Irish they are teaching is far more basic than the level I have (entirely possible) I will at least be in a Gaeltacht village with people who are willing to speak to learners…
3 Likes
Hi @elizaveta! I’ll ask if we can share the name, but I can tell you that she is a native speaker (as in, grew up speaking Irish at home with native speaker parents, and uses Irish at work most of the time!), and we worked together, face-to-face for two days in Baile Átha Cliath 
For the methodology to work we always need to stretch the language a little bit every now and then. A small amount of our time was spent actually translating (I would get claude to give it a go in very small batches, our translator would fix them and give feedback, then Claude would try again, rinse and repeat. Over time Claude got better at it, but she still checked every single phrase), then most of our time was spent discussing the bits that didn’t map very well, so that we could make it fit the methodology without compromising too much on correctness/naturalness. So I take responsibility for any weird bits - they’re all there because we had to try to follow the methodology and she agreed it would probably not be too weird to do it that way.
Then, we spent some time going through the way the seeds were broken down into the bits that get introduced, putting more emphasis on the start of the course to save time, but still reading through them all. And then, when I generated a bunch of practice phrases for the course after getting back home, she did a run through proofreading them and deleting all the bad ones. After that I had to generate some more (giving claude the context of everything we deleted and kept, so it might do better), and those are still in the course and still need proofreading (our translator had to start another project by that point)
Hope that helps clarify things!
4 Likes
Very interesting! It’s almost like watching the beginning of 2 distinct languages, as must have happened millions of times throughout history.
1 Like
IMHO, that’s not what will happen here, though, Deborah. To me, it seems obvious that the more beautiful of the two languages and the one with the better claim to continuity with Ireland’s history and the history of its people will die unless the full weight of conservation efforts are thrown behind that variety.
1 Like
That could be the case. I’m watching with interest to see what happens here in the Basque Country, with the unified “Batua” version being taught in schools, especially in areas like Araba province, where the language was virtually lost and is being brought back.
In other areas there are dialects, which I hope won’t be lost. A teacher explained to me that they teach Batua to the children as the “official” language, but they use the local dialect on a casual basis. Some older first language speakers seem to be quite anti Batua, but I’m not sure how much that means they don’t want to speak to them though.
1 Like
Thanks for the abair.ie tip, Liza, we’ll definitely have a look at that when we get some time, and see if it would be possible for us to integrate it with our current set of building tools.
We will definitely be looking to make it possible for people to contribute human voices to the project - we’re working on a layer for non-LLM building at the moment - and if we get the growth to make it possible, we’ll be particularly likely to fund (or co-fund) voices for the Celtic languages, as our core mission.
2 Likes
I’m almost half-way through Brown/WhiteStripe now - hoping to get to Black before I go off to Galway!
There are a few niggles along the way (missing lenition, additional words, etc) but I have to confess to being a selfish user and not stopping to take screenshots (as I’m usually walking or driving and the phone’s in my pocket). Sorry. Doesn’t really bother me though - I just plough on.
But there’s an odd glitch that I did want to report - around the last quarter of Brown and the first quarter of Brown/WhiteStripe. The algorithm seems to get fixated on variations of the sentence “she could bring her brother on Monday” to the exclusion of everything else. When a new element is introduced there are a handful of chunks to practice it, but then it’s back to this blessed woman and her brother.
I thought getting on to white stripe might dislodge it, but it kept on with the same pattern, so I skipped over a lot. Now I’m over a quarter of the way through it’s started bringing other chunks back in (which is both a relief and a challenge, because I’d forgotten quite a few of them!)
Not complaining at all - it’s still working a dream for me Just in case it’s helpful feedback.
3 Likes
You’re doing so well, @sara-peacock-1! I can’t wait to hear how you get on when you go to Galway.
That’s very useful feedback, and I have no idea why the spaced repetition algorithm would keep giving you variations on that woman and her brother (he must be a very special brother 
) but I’ll look into it and pass information to the Tech Team.
Go raibh maith agat!
2 Likes
My word you’re doing brilliantly! I absolutely can’t wait to hear how your visit goes! Oh, also, I strongly recommend that you try to get one conversation practice under your belt before going, ideallly for two hours - that should really fire it all up for you. I know some Irish speakers based in/near Cardiff who would probably be up for it, if that would be helpful
1 Like
Thought this might be of interest here