🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

✨In January we explored synergies between 2 🇪🇺 projects (UM Department of Advanced Computing Sciences VOXReality & AI4LT @KIT Meetween). 🌍 Galvanizing to 👀 all this #MT work w/ Jan Niehues Yusuf Can Semerci Abderrahmane Issam Tu Anh Dinh Paweł Mąka + more! 🤝 Looking forward to closer collab 🔗 voxreality.eu/fostering-inno…

🔖 Weekly pick from the #MeetweenScientificWatch: "Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research" - A massive, open dataset for advancing LM research. 🌍📚 #AI #ML arxiv.org/abs/2402.00159

🔖 Weekly pick from the #MeetweenScientificWatch: "ZeroST: Zero-Shot Speech Translation" - A novel framework leveraging multilingual models to perform speech-to-text translation without direct supervision. 🗣️🌐 #AI #NLP #SpeechTranslation merl.com/publications/d…

Weekly pick from the #MeetweenScientificWatch: "A Suite for Acoustic Language Model Evaluation" - SALMon: A novel benchmark suite for evaluating models on noise, emotion, speaker identity, and more. 🎶📊 #AI #SpeechProcessing arxiv.org/abs/2409.07437

Few days left for registration. The SpeechTek Lab will contribute with 3 papers, which are results of our efforts under ELOQUENCE AI , Meetween ,FondazioneFAIR

Weekly pick from the #MeetweenScientificWatch: "SlowFast-LLaVA: A Strong Training-Free Baseline for Video LLMs" - A two-stream design captures spatial detail & temporal context, outperforming many fine-tuned models! 🎥📚 arxiv.org/abs/2407.15841

Weekly pick from the #MeetweenScientificWatch: "Vcoder: Versatile Vision Encoders for Multimodal LLMs" - A novel encoder boosts object perception in MLLMs, outperforming GPT-4V in visual reasoning! 🌆👀 arxiv.org/abs/2312.14233

Weekly pick from the #MeetweenScientificWatch: "SpeechAlign: Aligning Speech Generation to Human Preferences" - A novel approach to improving speech language models with human feedback! 🎤🔗 arxiv.org/abs/2404.05600

Weekly pick from the #MeetweenScientificWatch: "Learning Source Disentanglement in Neural Audio Codec" - Introducing SD-Codec for domain-aware audio compression & source separation! 🎵🔊 arxiv.org/abs/2409.11228

Great result from a paper funded by the #Meetween project!

Weekly pick from the #MeetweenScientificWatch: "An image is worth 1/2 tokens after layer 2" - FastV slashes LVLM computational costs by up to 45% while maintaining top-notch performance! ⚡📷 arxiv.org/abs/2403.06764

Weekly pick from the #MeetweenScientificWatch: "ST-LLM: Large Language Models are effective temporal learners" – ST-LLM excels in video-based dialogue with spatial-temporal modeling, setting new SOTA! 🎥🤖 arxiv.org/abs/2404.00308

Weekly pick from the #MeetweenScientificWatch: “LongVU: Spatiotemporal adaptive compression for long video-language understanding” – Efficiently processes hour-long videos with minimal detail loss. 🎥🤖 arxiv.org/abs/2410.17434

Weekly pick from the #MeetweenScientificWatch: “Generating Data with Text-to-Speech and Large-Language Models for Conversational Speech Recognition” – Boost ASR with synthetic multi-speaker data. 🎙️🤖 arxiv.org/abs/2408.09215

Weekly pick from the #MeetweenScientificWatch: “Video-SALMONN: Speech-enhanced audio-visual large language models” – Redefining video comprehension with speech-aware AV-LLMs and groundbreaking QA accuracy. 🎥🎤🤖 arxiv.org/abs/2406.15704

Weekly pick from the #MeetweenScientificWatch: “PaliGemma 2: A family of versatile VLMs for transfer” – An upgraded VLM family excelling in OCR, captioning, and radiography, setting new state-of-the-art benchmarks! 📖🖼️📊 arxiv.org/abs/2412.03555

Treasure Hunt at the Meetween project meeting in Shmoopy, Karlsruhe last week! 🏴‍☠️🔍 Any guesses on what's inside the treasure? 👀💡 #Meetween #TreasureHunt #Karlsruhe

Weekly pick from the #MeetweenScientificWatch: “The Ultra-Scale Playbook: Training LLMs on GPU Clusters” – A deep dive into scaling LLM training from 1 to 1000+ GPUs, demystifying parallelism techniques with practical examples. 🚀🖥️📖 huggingface.co/spaces/nanotro…

📢 The evaluation period of the Instruction Following task at IWSLT 2025 just started! 🖥️ Consider submitting your speech-to-text system! The outputs can be easily uploaded on the SPEECHM platform developed in the Meetween project! ➡️ iwslt2025.speechm.cloud.cyfronet.pl

Weekly pick from the #MeetweenScientificWatch: Overtrained language models are harder to fine-tune — This study reveals how longer pretraining can harm downstream adaptability, introducing "catastrophic overtraining" as a real risk. 🧠📉 arxiv.org/abs/2503.19206