June 20th ABAIR Meeting
· 4 min read
Actions
- Send Ailbhe and Neasa dates of upcoming leave.
Highlights
- Reports from UK Speech held in York - possible link ups with researchers working on Manx and Sami.
- Fotheidil performance massively improved - especially where background noise is present.
- Combined Arts & Humanities funding application for €950,000 GPU machine passed first round in College.
- Anna introduced as new member - will be working as an artist for MAO.
- Got fiana (sp?) to do recordings - July. Got schedule to get equipment down. CAn do recording in her house. Can be down there most days.
- Finishing up with prompts. Got last few stories from books.
- Talked about UK Speech in York (took place earlier in the week). Mentioned PhD student from Sheffield who is developing speech synthesis for Manx. Possible tie-in in the future.
- Updates on work with Finn on Auth.
- Demonstrated MOL again, with updates on projects.
- Will be gone for ~6 weeks after today so passing over responsibilities - Joey for the Irish class, Andy for taking meeting notes, Finn for Auth, and everyone for updating MOL
- Working on Auth stuff and step 4 MAO. Developed new hook.
- Asked about deadline for MAO to judge his own deadline for the Auth system.
- Coming home in 3 weeks from Erasmus in Spain.
- UCC stuff: Will create data himself - 50 sentences - and decide which are best.
- Can train 8-billion parameter model now. Much better that 1.7 billion at current.
- Waiting for response from Elaine Ui Donnacha for use of corpora.
- Wants to see which data are best for the model for understanding Irish.
- Friend doing PhD in NLP in Edinburgh. Might come to lab Monday week to give talk.
- Asked everyone to send in dates of leave.
- Wants to talk to Muireann more about her work seperately.
- UK Speech: Cambridge doing lots of work on LLMs.
- Guy from NZ, retired, writes in Irish, is an AI data person who wants to work on LLMs in Irish with us.
- Another guy from Finland who wants to come work with us for 3 months. Works on Sami, ASR and TTS. Interested in language learning applications. Sep-Dec.
- Funding. 3 Arts and Humanities funding applications put together for 3 servers with 8 H200s. €950,000 worth of compute.
- Last Friday there was a meeting. Powers that be decided there is committee for Digital Plan. Ailbhe gave talk on funding and requirements over Zoom. People from TG4, Foras etc. on the call. TG4 want more on accessibility. Coming in on Tuesday at lunchtime.
- Neasa went and talked budgets - application needs to be in over the next few days. Hoping for more funding but should get at least the same.
- Other meeting - went there with Amanda - about disabaility and children. We were only group with something concrete developed. Helped with how Foras, COGG etc. view us. Need to make Deachtú.
- SPoke about Oskar and him making the new website.
- TG4 will love the udated Fotheidil, in addition to the accessibility stuff.
- Move towards eliciting and using spontaneous speech for children's voices.
- Need to get people to add content.
- Need to get in touch with Ackerman's (sp?) in the Netherlands about the van. Process will take a long time. Lots of peperwork to be done.
- Joanna might be coming to do a PhD.
- Introduced as artist for MAO.
- Kayla (sp?) starting next week.
- Showed Update to Fotheidil. Unsing new models e2e with 9000 hours of data. Makes them robust to noise.
- Played audio from a hurling match and a recording from 1957 with poor audio. Performance is great.
- Speaker diarisation also has come on leaps and bounds. Using SOTA. ASR & speaker diarisation running on GPU.
- Speed - 30 minute video in 70s. Accuracy much better than Kaldi.
- Will integrate capitalisation as well.
- Kayla could help get the updated models on Fotheidil.
- One issue - not enough English names in data. Maybe use Irish English data to create bilingual ASR.
- Working on backoffice. Finished step 4.
- Bulding file system for media content.
- Shared screen and showed Sai's work - allows users to modify an image and add text.