Fig. 3
Spatial speech separation with a two-channel SuDoRM-RF22 architecture. To evaluate the impact of spatial cues on speech separation performance, we trained eight instances of the SuDoRM-RF model. Instances vary in terms of CI microphone channels selected for training as well as the addition of auxiliary pre-computed spatial cues (Table 1). All models trained on two-channel speech mixture input only followed the model architecture depicted in the top row (A), while all models trained on two-channel speech mixture input augmented with pre-computed spatial cues followed the architecture depicted in the bottom row (B). The example shows the calculated IPDs (range [–180, 180] degrees). Training parameters were the same for all model instances.
