Speech LLMs don’t fail on benchmarks.
They fail in real conversations.
That’s why we launched the 2nd MLC-SLM Challenge 2026 — to push Speech LLM research closer to real-world multilingual conversational speech.
Registration Links: https://t.co/6X1VlnRAzV
This year’s challenge brings a bigger and more diverse dataset:
✅ ~2,100 hours of two-speaker conversations
✅ 14 languages
✅ Designed for multilingual speech segmentation, recognition, and spoken dialogue understanding
Why does this matter?
Because real conversational AI is far more than ASR. It involves:
• speaker overlap
• turn-taking
• acoustic variability
• semantic understanding in multi-turn dialogue
And that’s exactly where today’s Speech LLMs still struggle.
We’re also excited that the MLC-SLM 2025 Summary paper has been accepted by ICASSP. This month, members of the paper team will present their work, and the session will be published across our media channels.
If you’re working on multilingual conversational AI, this is a challenge worth watching — or joining.
Explore more here: https://t.co/Llhvrs94bd
#SpeechLLM #MLCSLM #SpeechAI #MultilingualAI #ConversationalAI #ASR