Submitted to ICASSP 2027

SPATEX

Array-Flexible Multichannel-to-Multichannel
Target Speech Extraction Using Direction Clue

Wen Wen^*, Changda Chen^*, Yu Xi, Haoyu Li, Qiang Zhou, Guanyu Chen,
Xiaoyu Gu, Bohan Li, Shuai Wang, Kai Yu^†

^* Equal contribution ^† Corresponding author

Audio examples

One model · epoch 85

A target direction in, a target speech image at every microphone out.
Choose a microphone to compare the same excerpt. Playback stays aligned within each row.

Loading audio examples…

6-second excerpts · 8 kHz · Shared gain across signals and microphones in each scene.

The idea

Reproduce SPATEX

SPATEX uses microphone coordinates and a target direction to extract speech across different arrays. Geometry-aware channel exchange and shared temporal attention connect microphone streams, while IPD and coherence losses preserve their spatial relationships. The same model reconstructs a separate target image for every microphone.

SPATEX architecture: shared encoding, geometry-aware channel exchange, direction-conditioned temporal co-attention, and microphone-wise reconstruction. — Shared temporal attention, microphone-specific reconstruction.

About these examples

All three examples use the released variable-array epoch-85 checkpoint, trained with 2–8 microphones. Fixed and matched examples use linear and planar layouts; the unmatched example uses a random layout with 1 cm microphone-position perturbations. Speech comes from LibriSpeech test-clean with simulated room responses.

Examples are the first saved scenes matching each geometry and microphone count, without selection by output score. Playback uses one common attenuation gain per scene. Waveforms and spectrograms share scales across signals and microphones. Downloads retain the original floating-point multichannel output.

Metrics in the expanded rows describe individual six-second examples, rather than the paper’s two-second benchmark averages. Generation details · Sample provenance

Citation

@unpublished{wen2026spatex,
  title  = {{SPATEX}: Array-Flexible Multichannel-to-Multichannel
            Target Speech Extraction Using Direction Clue},
  author = {Wen, Wen and Chen, Changda and Xi, Yu and Li, Haoyu
            and Zhou, Qiang and Chen, Guanyu and Gu, Xiaoyu
            and Li, Bohan and Wang, Shuai and Yu, Kai},
  year   = {2026},
  note   = {Submitted to ICASSP 2027},
  url    = {https://github.com/Wenanzhi/SPATEX}
}