Real Time Source Separation on GPU

GPU Audio (now Soundcore for Pro Audio) have announced the release of their new SDK Module, for Real Time Source Separation. Premiered at the Audio Developer Conference in 2025, this new module allows for low-latency source separation for audio applications in live shows, music production, and more. The model used is HSTASNET, as investigated by L-Acoustics, taking both time and frequency domain data into account for ultra high resolution, artifact free processing.

The SDK is free to download, but a licesne must be obtained for commercial use. https://gpu.audio/sdk

More here: 
 

Showcased at 2025's Audio Developer Conference event in Bristol, we've delivered genuine real-time source separation utilizing deep learning for music demixing. Based on the HS TasNet, which has been researched and investigated by L-Acoustics, this new GPU Audio SDK module provides a neural network adapted for real-time low-latency applications. Opening huge potential for remixing and noise-cancelling audio streams in real time - for live shows and many other applications. 

The HS TasNet (Hybrid Spectrogram Time-domain Audio Separation Network) utilizes a combination of spectral and waveform domains which L-Acoustics state: "demonstrate the potential of efficient demixing for real-time low-latency music applications". In their white paper, a latency of 23 ms is suggested, which we have developed to be just 512 samples with the GPU, even in its early stages. 

Some stats:

- CPU version (with Eigen) gives you a 0.18x of RTR
- Meta's PyTorch (GPU) gives you only 0.58x of RTR
- GPU Audio gives you a 6.95 of RTR

This is around a 12x performance boost with our Technology

These numbers are scalable on both single and multiple devices, with just a few milliseconds latency; no historical data (compared with Meta Demucs' v4), and the opportunity to rebuild using our SDK and achieve almost a 7x RT factor. Find out more about HS TasNet in the L-Acoustics paper entitled: "Real-time Low-latency Music Source Separation Using Hybrid Spectrogram-TASNET". 

The updated SDK is available NOW for Windows and Mac on our Github page.

Try the updated SDK for FREE by visiting https://gpu.audio/sdk

• Artists Favourite Calibre Tracks 21-Oct-26
• Mulch Box 21-Oct-26
• Look Mum No Computer Interview 20-Oct-26
• KORG MS2000R Hardware-Level Emulator 20-Oct-26
• STR-500 Gestural Granular Synthesizer 16-Oct-26