Last year, our lab ventured into a new domain: detecting cognitive decline from speech recordings. We competed in the
ICASSP PROCESS Challenge and secured 2nd place in the MMSE prediction task (out of ~30 teams)! 🏆 Our paper is now published: ieeexplore.ieee.org/abstract/doc...
Coat of arms of Ukraine.
I just donated to u24.gov.ua
The University of Helsinki is looking for a lecturer in Finno-Ugric linguistics: jobs.helsinki.fi/job/Helsinki...
"And I'll see the day that anyone gives us #1 without being forced to do so ..."
There are many LLM projects that are open about training and evaluation data, such as AllenAI OLMo, several EU projects (EuroGPT, HPLT), and several Huggingface projects. I don't think anybody forced them to do so.
@nature.com asked me to write a short comment piece about the SeamlessM4T paper from @metaai.bsky.social
(nature.com/articles/s41...), here it is: nature.com/articles/d41.... I think SeamlessM4T is still the best publicly available multilingual ASR/speech-translation model.
Great challenge but very little time...
What is the maximum length of a test utterance (important considering limited GPU RAM on the test server)?
Is ASR CER case sensitive? Are spaces taken into account when computing CER?
What is the maximum length of a test utterance (important considering limited GPU RAM on the test server)?
Is ASR CER case sensitive? Are spaces taken into account when computing CER?
Very interesting challenge! Unfortunately there is very little time, considering that participants would have to prepare some kind of container that decodes the test data on the Dynabench server.
Some questions follow...
