Signal processing using artificial intelligence (AI) has gained increasing interest because it outperforms existing solutions in many fields. A significant challenge for deep neural networks lies in meeting strict requirements regarding latency, computational load and memory, which is vital in automotive audio. This paper presents CUpGAN (Conditional Upmix GAN), a computationally efficient method for extracting upmix signals with low latency, leveraging signal separation for two upmixing concepts using a conditional generative adversarial network (CGAN). One upmix approach utilize spatial positions of direct sources within the stereo image, allowing for the distribu- tion of sources around the listener. The second approach separates direct and diffuse signals to create an ambience signal for rear surround loudspeakers. By employing phase-aware loss functions, integrating residual connections in the generator, and training with coherent input and target signals, we achieve high sound quality in the generated signals. This methodology also facilitates the computation of a cost-efficient complementary signal for both upmixing concepts through the difference between input and generated signals. The proposed technique reduces memory as 96% of the parameters can be shared between both applications, allowing seamless switching between upmixing approaches without the need for parameter loading; instead, parameters are computed by a small control network. The GAN generator is trained on synthetically generated data, enabling control over separation characteristics that surpass traditional methods. We present an evaluation using listening tests and computational metrics, demonstrating the advantages of our approach compared to classical signal processing methods.
