Speechmatics - API for Voice AI Agents
Mini toolsTinyMU
By Xiquan Li
TinyMU is a compact Music-Language Model built for music understanding and reasoning tasks such as describing and answering questions about audio. The project and paper describe it as a 229M-parameter model that achieves about 82% of state-of-the-art LALM performance on the MuChoMusic benchmark while being 35x smaller. It uses the MATPAC++ self-supervised audio encoder with a lightweight linear projector, and is trained with MusicSkills-3.5M, a 3.5 million-sample music-grounded QA dataset spanning multiple-choice, binary, and open-ended supervision.
New
Multimodal
Gen 3
Overview
TinyMU is a compact 229M audio-language model for music understanding and reasoning. It is designed to answer music-related questions efficiently, using a small architecture that targets strong performance under tight compute budgets instead of scaling to large audio-language model size.
Tools using TinyMU
No tools found for this model yet.
Last updated: April 27, 2026
