Researchers at VK presented a compact FastWave model for audio restoration

Researchers from VK and HSE University have developed FastWave — a diffusion model designed to restore high-frequency details of an audio signal (Bandwidth Extension, BWE). The solution allows increasing the sampling frequency of any audio file to studio 48 kHz.

The model's architecture is optimized for operation on devices with limited computational resources. The total number of FastWave parameters is only 1.3 million, making it one of the lightest neural networks in its class. The small weight allows integrating the algorithm directly into mobile applications and client services without significant processor load.

The BWE technology solves the problem of poor sound quality when transmitting through VoIP services or working with archival recordings. Unlike standard interpolation methods, the diffusion approach allows restoring spectral components that were lost during compression or bandwidth limitation, providing a more natural sound of voice and music.

The development is aimed at improving the quality of real-time communication. The use of compact diffusion models allows implementing complex audio processing algorithms in mass-market products, where low latency and minimal RAM consumption are critically important.