NDSI-BWE uses chaos theory to restore audio with 8x fewer parameters

A team of researchers (Tamiti, Das, Mamun, Barua) has published a new paper introducing NDSI-BWE, an adversarial speech bandwidth extension (BWE) framework that leverages chaos theory to recover high-frequency audio components lost during bandwidth compression. The system employs seven distinct discriminators, each designed to capture different temporal behaviors from nonlinear dynamics: a Multi-Resolution Lyapunov Discriminator for sensitivity to initial conditions, a Multi-Scale Recurrence Discriminator for self-similar patterns, a Multi-Scale Detrended Fractal Analysis Discriminator for long-range scale invariance, a Multi-Resolution Poincaré Plot Discriminator for hidden latent relationships, a Multi-Period Discriminator for cycles, plus Multi-Resolution Amplitude and Phase Discriminators for amplitude-phase statistics. These discriminators guide a complex-valued ConformerNeXt generator with a dual-stream Lattice-Net that simultaneously refines magnitude and phase components.

By using depthwise convolution at the core of each discriminator's convolutional blocks, NDSI-BWE achieves an eightfold reduction in parameters while maintaining superior performance. The generator leverages transformer-based conformer blocks for global dependency modeling and ConvNeXt blocks for local temporal modeling. Evaluated across six objective metrics and subjective listening tests with five human judges, NDSI-BWE establishes a new state-of-the-art in bandwidth extension. The paper (arXiv:2507.15970) is currently available as a preprint, with applications ranging from telecommunications to high-fidelity audio on resource-constrained devices.