Researchers from HSE University and VK have developed FastWave, a lightweight diffusion model designed for audio bandwidth extension (BWE). The model restores high-frequency details in audio, supporting any input sampling rate up to 48 kHz.
FastWave features 1.3 million parameters and a computational complexity of approximately 50 GFLOPS. By implementing depthwise separable convolutions and Global Response Normalization, the team reduced the parameter count by 30% compared to the NU-Wave 2 baseline.
The model utilizes the EDM training methodology, which includes continuous noise scheduling and weighted L2 loss. These optimizations allow the model to achieve high-quality results in 4 inference steps, doubling the speed compared to the 8 steps required by NU-Wave 2.
Training FastWave on a single V100 GPU takes up to 30 hours, whereas the baseline model required two weeks on two A100 GPUs. The model achieves a real-time factor (RTF) that processes audio approximately 6 times faster than real-time, making it suitable for streaming applications.
The project results were evaluated on the VCTK dataset, which includes 110 speakers and 44 hours of audio. FastWave demonstrates performance comparable to SOTA models like FlowHigh and NU-Wave 2, while maintaining a significantly smaller footprint.
