That would be brilliant, but it’s just me fantasizing at the moment 
3 Likes
You could be passionate from anywhere, but sadly you have to be at least a resident to push bureaucratic buttons. But I hear your point about being practical. If i figure out what I could do practically beside sharing learning resources, I’d gladly do that 
1 Like
Let’s fantasize together, I totally love the image of SSinConnacht
…Seriously though, there’s easier and cheaper way to care for proper phonetics, and that’s using dialect-specific model like on www.teanglann.ie, which was already suggested above. Still plastic, but at least it would be plastic capable of doing fada and unwritten vowels instead of doing unwholesome things to the old tongue 
1 Like
Do we have any feedback on the pronunciation accuracy from native speakers?
Also, do we know the dialect(s) the training material comes from? I do hear slight differences between the generated speakers, e.g., ceist a chur orm, but I could be making this up. 
e: I’ve been using teanglann.ie (thank you @sara-peacock-1!) to compare words and how they sound across the 3 dialects and I believe the accent of the first speaker is Connacht. The 2nd I’m unsure about.
1 Like
0003 - Eolas_faoi_na_Canúintí__1_.pdf (384.2 KB)
Haven’t heard n pronounced as r yet, which would happen in Connacht, but it’s relatively rare. My guess is, both are closer to urban Irish rather than L1 Irish. Official standard of Irish only regulates spelling, not the sound of words, but Buntús Cainte which taught a lot of people to speak was recorded with Connemara native speakers, and neutral urban Irish tends to sound like something between Connacht and English.
Lack of corrective feedback is a real issue here. In a country with more than 5 million people only about 72,000 speak Irish daily, and most of them live in Gaeltacht areas. It’s an ecosystem where it’s relatively easy to create pronunciation drift because L1 speakers are drop in the sea compared to about two million L2 speakers (which is why it’s also interesting if the person who did the initial translation into Irish was L1 or L2 speaker).
1 Like
Okay, I’ll risk to be annoying some more.
I don’t understand the target audience.
Irish has several groups that are potentially interested in speaking skills.
There are people of Irish descent, who live outside Ireland hence have about zero need for Irish for practical purposes, but they are actively interested in some Irish as part of identity and as a way of reconnecting with the culture. Need to reconnect with the culture partly explains huge success of Manchán Magan’s books in motivating people to learn Irish. They might be hugely inaccurate in etymology, but they give that feeling that you’re part of something ancient and very special. That feeling sells well. Which means in this group it’s important for example to teach gabh mo leithscéal not as “excuse me” (you can say “excuse me” in any language, English is quite sufficient for that) but as “take my half of the story” which is the literal meaning. It has flavor, it’s interesting and it’s something that you don’t say in English. But this principle clashes with the aim of making universal curriculum for all the courses.
There’s group of people who were let down by mandatory Irish at school. It’s different emotional landscape. People born elsewhere lack contact with Irish, people who went through thirteen or something years of badly taught Irish may have bad blood with the teanga and feeling it was forced upon them. But there’s also class inequality side to that, because typically working class kids had limited access to Gaeltacht so they were less successful in learning to speak. Which creates shame of not speaking ancestor’s tongue and at the same time anger of unfairness of the situation where there’s paywall between you and language which is your birthright. This group typically disliked Manchán because what they saw was a kid from a good family that could afford Gaeltach vacations, who was laughing at them for not being able to talk in Gaelic. In this group price tag might be a serious barrier.
Then there are people who are in their teens now, their childhood already had TG4 and they consider speaking Gaeilge cool (with the side bonus of parents being unable to understand them). Selling, say, to their parents might be a working strategy if it was framed as a way to teach kids speaking as if they were attending Irish-medium school. Which wouldn’t even be far from the truth, if the expected result is kids speaking Gaeilge bhriste and picking up grammar mistakes from each other. Immersion schools do that, app probably could do pretty much the same cheaper. But this, if scaled, means seriously affecting ecosystem and potentially introducing a number of grammar mistakes into the whole generation (L2 experiential learners were shown not to correct each other’s mistakes). Presuming there’s some degree of responsibility in the strategy, it would be necessary to very carefully proofread everything with somebody who speaks their L1 Irish daily.
There are also language geeks who’ll try everything and then say quite a lot about what they think. At this point it takes five minutes of reading reddit to figure out there’s consensus on Duolingo Irish among people who have good Irish, and it’s not favorable. Duolingo Irish is not advised for beginners precisely because it has low quality AI audio which trains for wrong pronunciation patterns. SaySomething is not that visible there yet, but it will be with some more advertising, and at this point low quality AI-generated content can destroy credibility very, very quickly. If the target audience are absolute beginners, getting this kind of visible feedback is not exactly good for business.
I see a lot of interesting possibilities at this point, but they all require doing something differently.
2 Likes
teanglann.ie is excellent, I’d say is one of the most best pronunciation sources on the internet for Irish, along with fuaimeanna na gaeilge
2 Likes
…Also critical insights like the necessity of calibrating speech to native speakers probably should be somewhere in the app itself, not on forum. Forum seems barely alive. If, say, a group of about five people had this discussion and agreed it’s good to check with teanglann.ie and listen to Gaeltacht radio, that’s good for them (and presumably their pronunciation), but that’s five of how many people?
And some more annoying staff. I totally get cozy family&friends local culture, but it’s sadly not sustainable when business gets scaled. Please forgive me for pressing the matter. Quality control will become an issue sooner or later, better start worrying about that at the stage of open beta than later when it will cost more in reputational risks.
At this point, how many of these questions can be answered in a way that does not make the company look shady?
— How many people on the team are Irish speakers? What training/education in Irish did they receive aside from playing SaySomething? Do they have any documents to confirm their language proficiency?
— Was the person that did the translation L1 or L2 Irish speaker? How often do they speak Irish themselves? What training/education did they receive? Do they have any documents to confirm their language proficiency?
— Was there any proofreading on the translation? If there was, who did it (same questions as above)?
— If any grammar/lexical/pronunciation mistakes are identified in the course by learners, what’s the procedure for correcting them? How much time would it typically take and whose responsibility is that?
— Which AI model was used on sound files, what data set was used for training (L1 speakers, L2 speakers, what level of proficiency, which dialects?), was there any evaluation of the model for pronunciation accuracy by L1 speakers? How was it chosen and which alternatives were considered?
— If learners have questions concerning grammar, is there any designated person on team tasked with and qualified to answer the questions?
— Was there any independent formal evaluation of skills acquired by playing SaySomething in Irish (specifically, for grammar accuracy)?
— What procedures are currently implemented for quality control?
1 Like
I can answer one of your questions. I remember reading on here that the translations were all checked by a native speaker.
2 Likes
What do we know about their education/profession? If i task, say, random L1 Russian speaker with editing/proofreading, it might be poor editing/proofreading (there’s a reason why those are both professions for which people get trained).
Possible missed grammar mistakes are important point here, because SaySomething curriculum is grammatically ahead of typical A1 stuff, which means even if learner does not plan on staying illiterate and reads grammar concurrently, when they get to the point where they know enough grammar to see that something is off, it might be already cemented by generous amount of spaced repetition. Environmental corrective feedback for reasons explained above is not something to count on. So it’s critical that the curriculum is checked thoroughly.
I don’t think all errors are equally worrisome in that regard. There are errors that frequently fossilise and errors that don’t (and you couldn’t have reached as high a level in English as you have without knowing that as well as I do).
My impression from the Italian and Japanese courses is that the AI is good at avoiding those kinds of errors, perhaps because AIs learn languages in a very different way from people and tend to produce different kinds of errors.
3 Likes
It is relatively easy to reach high level in English because the environment is rich with it and it’s constantly needed for practical purposes.
As far as I know speech gen says what was written, so there’s issue of questionable pronunciation (relatively easily solved by switching to better model) and separate issue of grammar mistakes that can contain human errors if the translation and proofreading were done by humans. Now, if the only humans with any significant level of language proficiency in Irish were those who’ve done translation and proofreading, and they have not been involved since, anything they missed will stay indefinitely (possibly bleeding into wider population, because Irish does not have the mass of native speakers to absorb the impact of scaled faulty teaching) and we’re getting the same issue as Duolingo learners who’ll insist that something they learned on Duo is the right way, because they never considered there could be mistakes in the curriculum. I’m working through TY in Irish, and even my modest A1.5 is enough to notice there are occasional spelling mistakes in every unit. That’s okay, I worked in book publishing, and I know that the only book supposed to be without mistakes is hand-copied Torah. But I can notice the mistakes in my textbook because everything I see is more or less on my level and I understand the grammar. With Say Something the grammar is ahead of what I know analytically so if it’s off, I won’t know. With Welsh I can safely trust nice Jones couple to know what they are doing with their own language, with Irish I’m not so sure. If everyone involved is flying blind pretty much the same way I do, it does not look trustworthy.
Theoretically, with target audience consisting not only of novices and more active forum it would be possible to use collective attention to catch those. I’d say that’s something to think about.
Different mistakes with different size of effect is an interesting idea though. The way I experience world, I see a dense system of causal links and influences, I can see the how stuff fits together and what can cause what, but I don’t feel weights and probabilities, so I can see how something can happen, but I have little idea how likely or how big will be the impact. Just that it is possible.
1 Like
Were there, really? 
With the Welsh course, that I started with basically 0 prior knowledge of the language, I remember the effort of listening carefully to the sentences and wondering if that sound really had changed or even disappeared from that other previous sentence. 
I don’t remember any comments about “mutations” but I know my fried brain took a sort of break whenever the explanations and intros came so I may have forgotten them entirely.
However I know that thanks to this it was much easier for me to start speaking compared to many Welsh people who had maybe studied it in traditional classes for years and years, so I think there’s no real need to worry either way.
p.s. planning to come back and re-try the Irish (which I had attempted for the first time when I was in school and so. much into Ireland and everything Irish but never moved forward beyond a handful of words).
1 Like
I’m not sure about the app, classical challenges definitely had those remarks about consonants softening. I don’t remember if it was specifically called mutation though. But I appreciated that!
2 Likes
I can remember a comment along the lines of “you might have noticed that the sound softened there - this is a special feature of Welsh, and we advise you not to worry too much about it for now, but notice where it happens” or somesuch. This was on the original “old course”, so I don’t know about the new app.
3 Likes
I don’t know about the app either. I did the challenges, Southern version.
But I really don’t remember much of the comments (despite being without any doubt a very auditory learner, as I was never able to remember anything written unless i could “play” it in my head or at least turn it into images).
I guess part of the reason could be I did Level 1 and 2 in just a bit more than a month - so there was probably no room for unessential words in English (my second language anyway) in my brain - although maybe I absorbed the idea of not worrying about whatever subconsciously. 
3 Likes
gee, I want this ability to stop worrying! I only stopped being bothered when I got a formal explanation of treiglad meddal and saw there’s method to the madness. But at least I was bothered less
I’m on Northern version. Maybe they differ.
1 Like
Interesting how seeing logic is calming. Dia duit can sound in so many different ways, duit can be dits, or it can be ghich. I was bothered until I learned ts and ch are both normal ways for slender t to sound, and gh is how dh sounds in broad position (because it was spelled dhuit before language reform, it’s not written like that anymore, but speech retained the variant). Stuff that’s irregular but does not fit any known pattern evokes stress response. Tuning into wrong phonetics is a sudden flood of reacting to every single sound that is off. That’s a thing with getting feedback from Gaeltacht folk, people often complain that when they tried to communicate in the Gaeltacht they are told about so many things that are wrong with their Irish at once, they are afraid to try again. Probably the same thing, nervous reaction to the flood of too many wrong signals.
Musicians typically start sounding good somewhere second year in the training, but people who play violin are expected to produce horrifying sounds for the first three years. Maybe it’s the same with Irish, you just have to stoically accept that you’ll sound badly and annoy other people with sounds you produce for quite a long time before you’ll sound properly melodically.