Bridge, Distill, Refine: High-Fidelity One-Step Generation for Multi-Bandwidth Speech Extension
Compare the same VCTK utterance across input bandwidths and generation methods. All audio players use 48 kHz output.
Listen and compare
Two selected utterances per input sampling rate. Lower LSD is better.
About these samples
The number on each tab is the input sampling rate: 4, 8, 12, 16, or 24 kHz. The corresponding frequency cutoffs are approximately 2, 4, 6, 8, and 12 kHz. All results target 48 kHz.
The 8/12/16/24 kHz comparisons include NU-Wave2, Fre-Painter, AP-BWE, FlowHigh, and our NFE=1 and NFE=2 GAN models. The 4 kHz comparison omits AP-BWE because no official 4 kHz checkpoint is available.
Cases were selected from a shared VCTK test subset (seed 3407) for clear examples. They are illustrative, not an estimate of average performance. LSD is measured against the full 48 kHz reference on the displayed waveform, with a 2,048-point STFT, 512-sample hop, and power floor of 10⁻⁸.
Source speech: CSTR VCTK Corpus v0.92 by Yamagishi, Veaux, and MacDonald, licensed under CC BY 4.0. The input and model audio are derived from those recordings.
Paper abstract
Speech bandwidth extension (BWE) aims to recover missing spectral content from band-limited speech. However, most generative BWE methods are designed for fixed input bandwidths and rely on iterative sampling, limiting their flexibility and efficiency. We propose BDR-BWE, a three-stage generative framework for multi-bandwidth speech extension. First, to enable conditional generation from the observed band-limited spectrum, we design 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. Second, progressive consistency distillation compresses the multi-step flow transport into a one-step generator. Third, adversarial refinement further improves high-frequency detail and harmonic coherence. BDR-BWE achieves state-of-the-art log-spectral distance (LSD) on VCTK across all evaluated bandwidths with one-step inference, while maintaining competitive perceptual quality and generalizing well to EARS.