An Saol Ó Dheas is all you need. Munster Irish, outside of Kerry, is very weak.

As for Abair, I dunno. I know they're working on Speech-to-Text models too, but I also know they opened the recording challenge to everyone with no filter for quality of Irish. It immediately makes me skeptical to the quality of their models, as they let people self-identify as natives and what dialect they speak, which won't cause any issues, I'm sure. I can name many who self-identify as natives and as having a dialect when they clearly don't. Now, supposedly that's only for their speech-to-text models, but still I want some assurance of quality assurance of their actual text-to-speech models being trained solely on competent, Gaeltacht-raised natives.

I agree. An Saol Ó Dheas is the first thing that came to my mind too.



If that's only for their speech-to-text models, then it's the right move to include all manner of speakers. Such models need to be able to correctly interpret even the pronunciation of non-native speakers. Siri or Alexa wouldn't be considered very good if they couldn't understand the accents of non-native English speakers, and it's arguably non-native speakers who could benefit most from speach-to-text systems. It would allow them to ask a digital assistants "what's the definition of X?" or "how do I use Y in a sentence?"

From a technical perspective, too, training AI models on large amounts data, though some of it may not be good quality, is better than training on a small amount of high quality data. Repeated studies have shown the benefit of transfer learning, i.e. even training on data from different languages can improve models for less resourced languages for which a sufficient quantity of training data does not exist.

Of course, ideally many native speakers would take part also, but it's entirely up to individual native speakers whether or not they want to be involved. There's certainly not much point complaining (not that I'm accusing anyone here of complaining) that systems like this don't understand native speech if native speakers aren't interested or inclined to help in their development.

See, I tend to think the opposite way. If their Irish pronunciation is so bad it can't be understood correctly by things trained on native speech, maybe they'd realise it's not just their 'dialect' and that 'My English accent came from Irish' and realise that pronunciation is supposed to be entirely different.



The issue is when the non-native speech vastly outweighs the native as it likely does with Irish. That's a huge issue. It's also one of my issues with Corpas Naisiúnta na Gaeilge, which only compiles Irish from 2000 onwards (and Foclóir's, where they give non-natives the same weight; perhaps even worse, as it's what they use for dictionaries! But at least the corpus for them can be filtered by natives).

A hard line to take. Reminds me of this.

Still, I think they probably just wouldn't use it in that case.



Well, we don't know if the non-native speech used for training a speech-to-text model does vastly outweigh native speech. If it does, we really need to assess why so few native speakers would want to be involved in a project like this. But I'm inclined to doubt it's very likely that non-native speakers are so over-represented that it could negatively impact an AI model.

In any case, I don't think you can compare a dictionary or text corpus to an AI speech-to-text model. It's a big issue if a non-native term is presented matter-of-factly in a dictionary or text as if it were something a native speaker would use. If a speech-to-text model learns to recognise a mispronunciation which is common among non-native speakers, that just makes the model more accessible to those speakers. It doesn't outweigh the correct pronunciation, or make the model less capable of recognising the correct pronunciation.

Thank you for all the considered, interesting, comments. It is obviously a very difficult complicated area. In the meantime I will definitely give Saol Ó Dheas a listen and delve into the huge amount of resources gathered on Cork Irish.

The elevator sketch is very funny. It had me laughing out loud. Thank you for sharing.

A supplemental question if I may: Do Fuaimeanna.ie and Teanglann.ie have the same or separate challenges when it comes to reliability?

Fuaimeanna is better in general as Teanglann sometimes has weaker speakers for some words. But both use fully recorded audio, not AI.

That is helpful, thank you.

Yes and Dara Ó Cinnéide's pronunciation on fuaimeanna is really the gold standard for Munster...