Task-Directed Residual AddUNet:Perfect-Reconstruction Routing for Full-Rate Representations
The paper frames AddUNet as a perfect-reconstruction routing system, then extends it to full-rate task-directed representations.
It argues that constrained additive U-Nets match critically sampled multirate perfect-reconstruction filter banks. The proposed residual full-rate version keeps exact reconstruction while routing nuisance or redundant structure away from the task-facing path. The claim does not depend on invertible routing operators, reconstruction loss, a learned decoder, or a matched synthesis bank. In TIMIT experiments, the front end lowers test PER from 28.60±2.09% to 25.76±0.41% under the same recognizer and training setup, while preserving exact reconstruction. ArXiv · AI/CL/LG's note
It argues that constrained additive U-Nets match critically sampled multirate perfect-reconstruction filter banks. The proposed residual full-rate version keeps exact reconstruction while routing nuisance or redundant structure away from the task-facing path. The claim does not depend on invertible routing operators, reconstruction loss, a learned decoder, or a matched synthesis bank. In TIMIT experiments, the front end lowers test PER from 28.60±2.09% to 25.76±0.41% under the same recognizer and training setup, while preserving exact reconstruction. ArXiv · AI/CL/LG's note
score 4