At INTERSPEECH 2025 on Kos Island, Greece, with Joonas Kalda presenting our paper, available here: isca-archive.org/interspeech_20…

☀️ We hope everyone had a great summer!!!☀️ Our 📢 NoDaLiDa/Baltic-HLT 2025 *2nd call for papers* is out! 📢 Deadline: 21 October ✍️⌚️ #NLProc #nodalida_baltichlt

Anyone in here could provide a link to that paper showing that it is beneficial to add the attenuated original mixture 🔀 to the separated sources for multi-speaker ASR? 🙏 I am at the point where I think I might have dreamed 😴 about it

wtf americans?

NoDaLiDa/Baltic-HLT is less than a month away! Did you know Talllinn is a living UNESCO treasure and also has a café culture? One such example is Maiasmokk, the oldest café in Tallinn dating back to 1864!

Deepseek fearlessly puts out its innovations in delightfully detailed papers, while other top AI companies are tight-lipped about their work. This makes Chinese AI competition look formidable and their American counterparts weak and insecure. I am not sure if this was a top-level

Last year, our lab ventured into a new domain: detecting cognitive decline from speech recordings. We competed in the IEEE 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/docum…

Tucker Carlson: Female leaders are the most warlike and bloodthirsty; they just can't wait to murdér people. That girl from Estonia [former PM Kaja Kallas], I mean, she would kill people without even flinching.

I respect Meta's work on open-source models a lot, but I must say Llama 4 doesn't look too good. It's supposed be better in multilinguality than previous open models, but on Estonian high school exams task, Llama-4-Maverick (400B MoE) is worse than LLama 3.1 70B and Gemma-3-27B.

Rumours about Llama4-related VLLM bugs seem to be true. Running Maverick again on the Estonian exams benchmark (using a cloud provider) improves Maverick's accuracy from 70.4% to 75.1%, making it comparable to Llama 3.1 70b. Gemma-3-27b-it is still 4 points ahead.

Nice article featuring two junior member in our lab: taltech.ee/en/news/where-…

Yesterday #INTERSPEECH2025 wrapped up — one of the best conferences I’ve attended! 🚀 Our research group had record presence and a record number of papers. I loved vibrant Rotterdam! 🇳🇱✨