Does whisper support Irish?

[liad-inon]

I'm getting an error when trying to use the Irish language code ga, and it seems that it does not appear in the tokenizer.py LANGUAGES dictionary. However, I couldn't find any source that explicitly states whether Irish is supported or not. Whisper has almost all European languages, so I find it quite odd.

This is a really basic question, but I couldn't find any official full language support list outside of the code in tokenizer.py. And even the language codes in the tokenizer are quite a mess. For example it has Norwegian Bokmal on the general Norwegian code no.

[ryanheise]

The source code (which you consulted) is definitive. You could try searching for anyone who has finetuned the model for Irish.

It's worth mentioning that Irish is also a low resource language so you might not be happy with the results compared to other languages where more training data is available.

The issue with the Norwegian language code has been discussed here: #1250

[liad-inon]

Thanks!!

[sproocht]

No, Whisper does not support Irish Gaelic language (ga). However, if you have some data, you may fine-tune the model for Gaelic, using a supported proxy language like English or Welsh (you need to do some testing here). You may even add Irish Gaelic (ga) to the tokenizer in more advanced settings.
Have a look at some discussion here #13

[aindriu80]

I’ve also been wondering about Irish (Gaeilge) support. From what I understand, Whisper can be nudged into producing something for Irish with tweaks, but the quality isn’t close to what native speakers would expect.

One thing worth flagging is that there is already excellent Irish TTS work happening, for example at https://abair.ie/ (University of Dublin / ABAIR project). Their voices sound natural and are used in education and accessibility.

Rather than reinventing the wheel, it might be interesting to explore whether:

• existing Irish speech resources (like ABAIR, GaelTech projects, etc.) could help with datasets,
• Whisper/whisper-transcribe could be adapted or fine-tuned on proper Irish recordings, and
• people working on Irish speech technology might be open to collaborating.

Irish is a minority language with an active technical community, so having a solid pathway for Gaeilge here would be really valuable.

[ryanheise]

Using Irish TTS actually sounds like an interesting avenue. Training is more effective when exposed to a wide diversity of different speakers, and so you would want to use different TTS engines to get as wide a diversity of recordings as you can. But of course there is also a limited amount of written Irish text to be synthesised into speech (which is also why LLMs are currently also exceedingly poor in Irish).

Since you are in Ireland, have you thought about organising a project/effort to build up a larger collection of authentic text and speech? Common Voice has 14 verified hours of audio. Much of that is probably non-native speakers, or actual Irish people but speaking Irish with an English accent (which is still OK for diversity, actually). But Common Voice also has a very easy-to-use web page where contributors can record themselves speaking. It just needs more people to get involved in both recording speech and suggesting new sentences to be read out (all public domain). I'd like to see someone champion that.

Or thinking outside the box, with more effort someone could build a voice chat app where you chat with random people, AND people sign up with the express intent and knowledge of having every conversation recorded into the public domain for the preservation of the language. A low friction app that those who are interested in preserving the language would hopefully willingly use..

[aindriu80]

I really appreciate the thoughtful reply, lots to think about here.

You’re right that the main bottleneck is high-quality, authentic Irish speech. Common Voice is great, but 14 hours is tiny, and accents vary a lot (which is interesting, but also tricky).

There are organizations here who care about this, for example Conradh na Gaeilge and the ABAIR team, but I’m not sure how much of their material can legally be released for open datasets. I might reach out and at least ask the question, or see if there is already a project underway that people here could help with.

I also really like the idea of building new corpora through community tools (like Common Voice or even a dedicated chat/recording app). If something concrete starts to form, I’d be happy to help test, share, or coordinate with Irish-language communities.

Go raibh maith agat, this is encouraging!