October 6th ABAIR Meeting
· 3 min read
- Asked Ciadhla if tranztool working ok.
- Putting together a few things for the Oireachtas. Plan in action.
- Her dad said sometimes first words not transcribed.
- 11:15 on Wed - Udaras coming in - Thomas O Shuiochain (in person).
- Storm Zoom at 3pm on Wed.
- Need to come up with a plan for how we present. Liam 10 mins Fotheidil etc., John, S2S, GDPR, Cybersecurity, scalability 10 mins, Joey LLMs 10mins, Andy Ciadhla 10mins.
- LLM - just need to show the benchmarks, no need for an extensive demo.
- Explained about matched funding from the north being a problem when the north power sharing breaks down.
- Suggested maximum chunk size for chunks in Fotheidil to avoid errors Liam mentioned.
- Asked about skipped recordings on Tranztool. Ciadhla said some files with code-switching.
- Going to use pretrained model for new voice.
- Evaluation for Finn's project will be key.
- Decide on issues in Google doc - do we substitute words or not. People might be angry if we choose incorrectly. Can adjust if we get compaints.
- First, don't touch the spelling, be conservative, then adapt.
- Thinks LTS is more or less there, just needs to correct things - finish this week.
- Question on whether replacing all words. Rith said as 'ruth'.
- Makes point that this is a new model - a subdialect of a dialect - maybe more specified rules required.
- Bug in Fotheidil - Voice Activity Detection separated into long chunks, which broke the capitalisation and punctuation system. Fixed this issue. Possibly run cap-punct on GPU machine.
- 4.5k hours of GPU European supercomputer.
- Asked if new phonemes on new voice - Christoph said yes.
- Exlained Fotheidil to Muireann.
- Found bilingual training script for PIPER.
- Point of dialect specific voices is they are dialect specific.
- Asked Christophe how long until the LTS are ready.
- Asked about accomodation for Oireachtas- may need to wait to get paid to book - may be quite late.
- Ailbhe asked about doing a demo. Not sure what to do as LLM is not robust enough.
- Tranztool work is all done. Need to get stuff up on Fotheidil. Liam to make account.
- Will introduce agenda idea next week.
- Final year project - reading paper on evaluating serious games, to help understand user engagement on MAO. And another paper on Duolingo - interesting grouping of users.
- What's the craic with the Oireachtas?
- Looking into technologies for the API - Grafana & Prometheous.