Speech bandwidth extension (BWE) aims to recover missing spectral content from band-limited audio, but most generative methods are designed for a fixed input bandwidth and rely on iterative sampling. The new BDR-BWE framework addresses both limitations with a three-stage generative approach that works across multiple bandwidths in a single inference step.

The method first uses a cutoff-aware conditional flow bridge in the compressed complex STFT domain, focusing generation on the missing high-frequency band while preserving observed spectral information. Progressive consistency distillation then compresses the multi-step flow transport into a one-step generator, and adversarial refinement further improves high-frequency detail and harmonic coherence.

On the VCTK benchmark, BDR-BWE achieves state-of-the-art log-spectral distance across all evaluated bandwidths with one-step inference, while maintaining competitive perceptual quality. The authors also report good generalization to the EARS dataset, and audio samples are available online.