# What two-channel audio adds to Xiaomi's MiDashengLM-Spatial > Xiaomi researchers released an audio-language model for spatial questions. Its preprint reports stronger directional benchmark scores with two-channel input, while complex scenes remain difficult. By BIG CHANGE Editorial Published: 2026-10-16T09:43:13.526Z Updated: 2026-10-16T09:43:13.526Z Canonical: https://bigchange.ai/blog/xiaomi-midashenglm-spatial-two-channel-audio ![Conceptual charcoal illustration of a seated listener viewed from behind, with a handbell to the left and an alarm clock to the right; curved orange marks suggest sound from each side.](https://bigchange.ai/api/media/file/midashenglm-spatial-listening-concept-v1.png) AI-generated conceptual illustration by BIG CHANGE; no Xiaomi product interface, test setup or measured result is depicted. Xiaomi researchers released MiDashengLM-Spatial, an audio-language model that accepts two-channel sound and answers questions about where sounds are coming from as well as what they contain. The October 8 paper reports substantially higher scores than the tested monaural models on two spatial-audio benchmark subsets. Its public code and checkpoint make the method inspectable, although BIG CHANGE has not run the model or repeated the evaluations. The distinction matters when a recording contains more than one source. A conventional audio question might ask what can be heard. A spatial question also asks whether a source is left or right of the listener, whether it moves, and which sound an answer refers to. Mixing two channels into one removes the differences between them that can carry directional information. ## The big change - **What changed:** The researchers added a spatial audio branch to MiDashengLM and released a paper, code and model checkpoint. The model takes binaural audio for spatial questions and retains a path for general audio understanding. - **Why it matters:** A text answer about a sound's location needs both event recognition and location cues. The dual-branch design lets the spatial branch use intermediate semantic features while leaving the semantic branch's forward path structurally unchanged. - **What to watch:** The authors' channel-swap benchmark tests whether left-right answers follow the audio channels. More independent tests, including varied recording setups and crowded scenes, are needed before treating the reported results as dependable in applications. ## How the two branches divide the task The [arXiv v1 paper](https://arxiv.org/abs/2610.11156v1) describes a 32-layer semantic encoder inherited from MiDashengLM and a 12-layer Spatial-Dasheng encoder for multichannel audio. The semantic path extracts what is in the recording. Spatial-Dasheng is trained to detect active sound events frame by frame and estimate their direction and distance, including when sources overlap. Intermediate semantic representations enter the spatial branch at four depths through adapters. The final outputs are fused into audio tokens for a language model to generate text. That arrangement addresses a specific association problem. If two sources overlap, a location estimate alone does not say which sound belongs there. Conditioning spatial features on semantic features is the authors' way of joining those clues without feeding new spatial representations back through the original semantic path. The paper reports that its training uses roughly 1 million synthesized binaural scenes, about 13,000 hours of audio, and 6 million question-answer pairs, followed by a mix of monaural and binaural data. Those are the authors' dataset figures, not independently audited counts. The [pinned repository README](https://github.com/xiaomi-research/midashenglm-spatial/tree/f314b3455e7fd94c60ad078655a39d69bcd05232) documents a Transformers example that places audio before a text instruction and accepts real binaural audio as a `(2, T)` array. It says mono input still runs but carries no direction. That statement does not establish compatibility with a particular microphone, device or acoustic setting. The code path calls `from_pretrained` with `trust_remote_code=True`, so users should inspect the repository and checkpoint code before executing it. The [checkpoint tree](https://huggingface.co/mispeech/midashenglm-spatial/tree/160ae51b51becade390ace4294feaa5b078d5665) lists about 33.5 GB of files and an Apache-2.0 license; file size alone does not specify working memory or inference cost. ## What the authors measured On the Spatial Reasoning subset of STAR-Bench, the paper reports 70.3% spatial accuracy for MiDashengLM-Spatial, compared with 44.3% for its monaural MiDashengLM baseline. On the spatial subset of MMAU-Pro, it reports 55.4% versus 18.2%. These figures come from the authors' evaluation of benchmark recordings, not a BIG CHANGE test or a measured deployment. The paper also checks whether answers depend on the two channels. For questions whose correct left-right answer reverses when the audio channels are swapped, its all-correct rate requires both the original and swapped answers to be right. The authors report 72.6% on STAR-Bench for their model and 0% for the monaural MiDashengLM baseline. That paired check is more informative about directional grounding than a single accuracy score, though it remains a benchmark result under the paper's conditions. The paper gives reasons for caution. In its own synthetic scene tests, exact matches for motion trajectories remain below 20%, and the model performs worse on complex scenes with multiple overlapping sources. For the stereo STARSS23 localization evaluation, the authors restrict the task to left-right azimuth because stereo recordings give limited cues for resolving up-down and front-back ambiguity. General audio results vary across tasks rather than improving uniformly. These limits matter for anyone considering the checkpoint for research: spatial benchmark gains do not establish robust localization across recording environments or a safe basis for navigation. ## Sources & further reading - [Hu and colleagues, “MiDashengLM-Spatial: Unifying General Audio Understanding and Spatial Awareness,” arXiv v1 (October 8, 2026)](https://arxiv.org/abs/2610.11156v1). Primary preprint for model components, training data, author-reported evaluations and limitations; no independent replication is claimed here. - [Xiaomi Research, MiDashengLM-Spatial repository, pinned commit `f314b3455e7fd94c60ad078655a39d69bcd05232` (checked October 16, 2026)](https://github.com/xiaomi-research/midashenglm-spatial/tree/f314b3455e7fd94c60ad078655a39d69bcd05232). README, example and evaluation code document inputs, setup and reported tables; BIG CHANGE did not execute them. - [Xiaomi Speech Team, MiDashengLM-Spatial checkpoint tree, pinned revision `160ae51b51becade390ace4294feaa5b078d5665` (checked October 16, 2026)](https://huggingface.co/mispeech/midashenglm-spatial/tree/160ae51b51becade390ace4294feaa5b078d5665). Public files, size and license listing; no runtime measurement is inferred. ## Sources - [Hu et al., MiDashengLM-Spatial, arXiv v1 (October 8, 2026)](https://arxiv.org/abs/2610.11156v1) — Primary preprint for architecture, training, author-reported benchmarks and limits. BIG CHANGE did not reproduce the results. - [Xiaomi Research MiDashengLM-Spatial repository, pinned commit](https://github.com/xiaomi-research/midashenglm-spatial/tree/f314b3455e7fd94c60ad078655a39d69bcd05232) — README and code document two-channel input, setup and evaluation recipes. BIG CHANGE did not run them. - [Xiaomi Speech Team MiDashengLM-Spatial checkpoint tree, pinned revision](https://huggingface.co/mispeech/midashenglm-spatial/tree/160ae51b51becade390ace4294feaa5b078d5665) — Lists model files, size and license. File size does not establish runtime memory.