IT Home, October 16: Xiaomi announced yesterday the introduction of MiDashengLM-Spatial, a significant new member of the Dasheng series, fully releasing the model weights and inference code under the Apache 2.0 license.
Over the past two years, Xiaomi has successively released and open-sourced the foundational audio model Xiaomi Dasheng and the large audio understanding model MiDashengLM-7B. These technologies have been deployed in over 30 applications across Xiaomi’s smart home and automotive cockpit scenarios, enabling unified perception and understanding of speech, environmental sounds, and music.
According to the introduction, the new model is built on the MiDashengLM-7B foundation and further equips the large audio understanding model with “a pair of ears,” enabling it not only to comprehend the content of sounds but also to perceive their origin and spatial distribution.
Xiaomi stated that MiDashengLM-Spatial is the industry's first open-source end-to-end spatial audio understanding large model, which retains foundational audio understanding capabilities while adding spatial awareness and reasoning, advancing audio large models from "understanding sounds" to "comprehending the spaces in which sounds occur."
This model will be gradually applied to scenarios such as Xiaomi's smart home and automotive cockpits, enhancing the interaction experience of the "person-car-home full ecosystem" through spatial audio capabilities. Relevant test data is as follows:
IT之家 related links:
- Project homepage: https://jinbo-hu.github.io/midashenglm-spatial-demo/
- Technical report: https://arxiv.org/abs/2610.11156
- Evaluate code (GitHub): https://github.com/xiaomi-research/midashenglm-spatial
- Model architecture and weights (Hugging Face): https://huggingface.co/mispeech/midashenglm-spatial