Xiaomi open-sources AI that can tell where a sound comes from

Audio AI has mostly been asked one question: what is that sound? Xiaomi's new model adds a second one: where is it coming from. On 16 October, the Chinese electronics maker released MiDashengLM-Spatial, an open-source model that answers both at once.

The model is built on Xiaomi's MiDashengLM audio model and extends the company's Dasheng series from recognizing sound content to understanding the surrounding acoustic environment. It works on binaural audio, recordings made with two microphones placed like ears. Two ears hear the same sound slightly differently, and the brain reads direction from that gap; the model learns the same kind of cue from two-channel recordings. It keeps the abilities of its base model, including speech recognition, audio captioning and questions about environmental sounds and music.

The performance figures come from the project's own documentation. On the spatial subset of STAR-Bench, a benchmark of spatial audio questions, the model reaches 70.3% accuracy, against 73.7% for humans. It has about 8 billion parameters, the adjustable values that define a model. These numbers have not been independently reproduced in the material reviewed for this article. The limits are concrete: spatial answers need binaural input, while ordinary mono or non-binaural stereo audio does not meet the model's requirements, and running it requires an NVIDIA GPU supporting CUDA, the parallel computing platform that speeds up AI workloads.

For developers, the change is immediate. The weights are published on Hugging Face, the code on GitHub, and the Apache 2.0 license allows use in research and commercial products. The project's GitHub page describes it, to its authors' knowledge, as the first open-source end-to-end model that unifies general audio understanding and spatial awareness. For Xiaomi, the target is clearer than the calendar: smart-home products and vehicle cabins, where a device could turn toward a voice or react to a sound behind the driver. Xiaomi has not said when those products will arrive, and no partner or device is named.

What remains open is whether the 70.3% holds in real rooms and real cars, outside the benchmark. The documentation says the model was trained on synthetic data and still generalizes to real binaural recordings, a claim that outside testers will need to confirm.