Xiaomi Open-Sources MiDashengLM-Spatial, the First End-to-End Audio Model That Knows Where Sounds Come From

Xiaomi's speech team has open-sourced a new audio model that does something most voice AI still cannot: locate sounds in space. MiDashengLM-Spatial, released this week with full weights and inference code under the permissive Apache 2.0 license, is described by the company as the industry's first open-source, end-to-end audio-language model that combines general audio understanding with genuine spatial awareness in a single architecture.

The model extends Xiaomi's MiDashengLM-7B, a voice understanding model already running in more than 30 scenarios across the company's smart home devices and car cockpits. The new version adds a dedicated spatial audio encoder, called Spatial-Dasheng, wired in through a hierarchical conditioning module that injects intermediate semantic representations into the spatial branch at multiple depths. The upshot is a model that can not only transcribe or describe what it hears, but also reason about where a sound originated and how sources are distributed in a room or a cabin.

Because real-world spatial training data is scarce, the team built a synthesis pipeline that renders diverse acoustic scenes complete with scene-level spatial descriptions and question-answer pairs. The approach, documented in a technical report on arXiv, lets the model learn spatial cues without giving up the general capabilities of its monaural foundation.

Benchmark numbers come from Xiaomi's own paper and should be read as company-reported rather than independently verified. On STAR-Bench, a spatial reasoning benchmark, the model scores 70.3 percent, compared with 0.0 for BAT-7B, a comparable binaural model. On the spatial subset of MMAU-Pro it reaches 55.4 percent versus 23.7 for the competitor, and it posts a 72.6 percent channel-swap accuracy, a measure of whether left-right judgments rest on real auditory cues rather than luck. On general monaural benchmarks the model remains competitive with other 8B-scale audio-language models, which is the point: spatial awareness was added without degrading what the base model could already do.

The release lands as audio becomes a differentiator in embodied and automotive AI. A cockpit that can tell which seat a passenger is speaking from, or a home hub that knows a sound came from the kitchen, needs exactly this kind of spatial reasoning. Xiaomi says the model will roll into its smart home and vehicle products as part of its human-car-home ecosystem push.

Weights and evaluation code are live on GitHub under xiaomi-research and on Hugging Face under the mispeech organization, with a project page and an arXiv paper for reference. For researchers in speech and embodied AI, the interesting question is whether a permissively licensed spatial model forces the wider ecosystem to treat sound localization as a standard capability rather than a premium feature.