A frozen general-audio codec can locate sounds too: AURA beats the strongest dedicated system by 10.31 points in dual-source direction accuracy

Spatial question answering no longer has to depend on a separately trained spatial encoder: the frozen features of a general-audio codec never exposed to spatial supervision suffice to identify and locate sources in multi-source scenes. Earlier spatial audio-language models drew spatial information from a dedicated encoder, splitting spatial and semantic processing into separate feature spaces, while general-purpose audio-language models fed monaural input sat at or below chance on spatial questions, with Qwen2-Audio reaching 8.85% single-source direction accuracy against 12.57% for a random predictor. On dual-source SpatialSoundQA, AURA's direction accuracy exceeds the strongest dedicated system, DSpAST-BAT†, by 10.31 points, with 64.63M trainable parameters, about 40% of the roughly 161M in BAT and DSpAST-BAT†.

The frozen codec already holds spatial information, but it is distributed across layers and the lateral cue lives in the difference between the two ear channels, which the model must receive explicitly to use it. A controlled layer-wise analysis shows shallow residual vector quantization layers are shaped more by source content, while deeper layers respond increasingly to interaural differences, room acoustics and reverberation; lateral information is only partly recoverable from either channel alone and becomes reliably decodable through the difference between the two. Linear probes on single-source SpatialSoundQA raise left/right accuracy from 74.0% with monaural features to 93.0% with binaural features, while front/back and up/down move only from 71.7% and 72.1% to 72.9% and 73.4%, concentrating the gain on the lateral axis.

A conventional interface trained with answer-only supervision collapses on exactly the left/right axis, and even where front/back and up/down exceed chance it reads scene-level attributes without binding them to the queried source. On dual-source training the four-slot interface matches the majority baseline on left/right while exceeding it on front/back and up/down; stratified by whether the two sources lie on the same or opposite side of the queried axis, same-side accuracy is high and opposite-side accuracy falls to chance, with three-axis predictions matching the queried and the other source about equally often. A controlled experiment crosses four- and six-slot representations with answer-only and scene supervision, sharing the frozen codec, Qwen3-8B backbone, projector, LoRA configuration and training samples; either intervention alone lifts left/right above chance, while opposite-side binding improves clearly only when the two are combined.

AURA aggregates shallow and deep quantization layers into a content-oriented and an acoustic-oriented band, adds the left-right channel difference for each band to form a six-slot representation, and aligns text through scene and geometric supervision. The four-slot interface hands the two channels to the language model as separate tokens and leaves cross-channel comparison to the model, whereas scene supervision requires describing each audible source's category, direction and distance before answering. The codec and the Qwen3-8B backbone stay frozen throughout, and only a projector and LoRA modules totaling 64.63M parameters are trained; on SpatialSoundQA, single-source direction accuracy is 73.58% and dual-source 48.21%, with a dual-source distance error rate of 64.72%, the lowest among the compared systems.

The demonstrated scope is binaural static-source scenes in SpatialSoundQA, with training and evaluation using room impulse responses simulated by SoundSpaces 2.0 in Matterport3D buildings; the authors interpret the results as evidence that a frozen general-audio codec is sufficient for competitive spatial question answering on that dataset, not as evidence that it outperforms a dedicated encoder.

The authors note that AURA and the baselines differ in several components, so the comparison is system-level and the role of each component is established separately by the controlled experiment; trainable binaural spatial question-answering corpora remain scarce, and BiDepth shares sources, simulator and buildings with SpatialSoundQA, offering no independent test of generalization.