Comparing Pro Audio Source Separation

We covered the announcement that Soundcore (aka GPU Audio) have released a Source Separation module powered by the GPU. In this video they compare the speed of non-realtime platforms, then look at the leading Real-time Source Separation software - compared to their new model. They also cover some use cases which is not limited to just the application shown in the video. The model is based on HS-TASNET and this video shows just how quick the model is compared to other Pro Audio platforms.

The GPU Audio SDK is free to download, and available from https://gpu.audio/sdk

Overview of the Technology
Start by reviewing the motivation and core principles behind our technology here: https://github.com/gpuaudio/gpuaudio-sdk?tab=readme-ov-file#motivation-and-introduction

1. SDK Installation
Download and install the SDK by following the instructions here: https://github.com/gpuaudio/gpuaudio-sdk/blob/main/installation/main.md

2. Build Initial Example Projects
We recommend beginning with simple examples such as the Gain and FIR processors to verify your setup: https://github.com/gpuaudio/gpuaudio-sdk?tab=readme-ov-file#example-projects

3. Evaluate a Target Module (Realtime Source Separation – HSTasNet)
For example, you can test realtime source separation based on the HSTasNet neural network:
https://github.com/gpuaudio/gpuaudio-sdk?tab=readme-ov-file#real-time-sound-source-separation-rt3s

Please make sure to test this module within your DAW environment by building it as a plugin (details here): https://github.com/gpuaudio/gpuaudio-sdk/blob/main/installation/main.md#use-plugins-with-gpu-acceleration-in-the-daw-1

For a detailed technical explanation of how we re-implemented HSTasNet using the GPU Audio Platform, please refer to: https://github.com/gpuaudio/gpuaudiosdk/blob/main/Guides/RealtimeSourceSeparation.pdf

Performance Results - Using the following configuration:
STFT size: 1024
Overlap: 512 samples
Sample rate: 44.1 kHz
Buffer size: 512 samples.

We observed the following realtime ratios (RTR - how many times faster than real time the system processes one audio buffer):

CPU version (Eigen): 0.18× RTR = ~5x times slower than realtime
Meta PyTorch (GPU): 0.58× RTR = can't process one buffer on time still
GPU Audio SDK: 6.95× RTR 

This represents approximately:
~12× performance improvement over standard GPU (PyTorch)
~36× performance improvement compared to CPU implementation 

The GPU Audio SDK is free to download here: https://gpu.audio/sdk

• ttiimmee: Turn Signals into Grain Clouds 10-Nov-26
• Mefjus Studio Insights 10-Nov-26
• Free ReBirth RB-338 Clone Released 09-Nov-26
• Mr. Scruff & His Most Iconic Track 09-Nov-26
• Arca on Music, Sound Design and Workflow 06-Nov-26