Xiaomi Open-Sources MiDashengLM-Spatial: First End-to-End Large Model for Spatial Audio Understanding and Localization

Over the past two years, Xiaomi has released and open-sourced its audio foundation model Xiaomi Dasheng and the large-scale audio understanding model MiDashengLM-7B. These technologies have been deployed in over 30 real-world applications across scenarios such as Xiaomi’s smart home ecosystem and automotive cockpits, enabling unified perception and understanding of speech, environmental sounds, and music.

According to reports, the new model builds upon MiDashengLM-7B, effectively granting large-scale audio understanding models a "pair of ears." This enhancement allows the model not only to comprehend the content of sounds but also to perceive their spatial origin and distribution within an environment.

Xiaomi states that MiDashengLM-Spatial is the industry's first open-source end-to-end large model for spatial audio understanding. While retaining fundamental audio comprehension capabilities, it introduces spatial perception and reasoning abilities, advancing audio LLMs from merely "understanding sound" to "understanding the space where sound resides."

The model will be progressively integrated into Xiaomi’s smart home and automotive cockpit scenarios, leveraging spatial auditory capabilities to enhance the interactive experience of the "Human-Car-Home Full Ecosystem." Relevant test data is shown below:

IT Home provides the following related links:

• Technical Report: https://arxiv.org/abs/2610.11156
• Evaluation Code (GitHub): https://github.com/xiaomi-research/midashenglm-spatial
• Model Architecture and Weights (HuggingFace): https://huggingface.co/mispeech/midashenglm-spatial