Neural Network Upmixing for Low-Complexity Multichannel Audio

The encoder pre-calculates and transmits control data values that guide the decoder's neural network processing. By performing complex analysis at the encoder side and providing simplified control signals to the decoder, the system maintains audio quality while reducing real-time computational burden at the decoder.

The patent replaces traditional signal processing operations with a trained neural network at the decoder. The neural network learns optimal upmixing strategies during training, substituting complex real-time calculations with pre-learned patterns that require fewer computational resources while maintaining or improving audio quality.

The control data values serve as an intermediary between the encoder and decoder. These compact parameters carry essential guidance information from the encoder to the decoder, enabling sophisticated encoder-side control without requiring complex real-time communication or synchronization mechanisms.

The system uses a set of control data values that parameterize the decoder's neural network operation. By transmitting these parameters from the encoder, the system enables dynamic adaptation of the decoding process to different audio content and encoding conditions without hardwiring complex control logic.

The neural network is trained offline in advance on large datasets of audio content. This preliminary training phase allows the network to learn optimal reconstruction patterns that can be applied efficiently during actual decoding without requiring heavy computational resources at runtime.

Instead of implementing complex upmixing algorithms from scratch at the decoder, the system uses a copied neural network model that has been trained on comprehensive audio data. This pre-trained model encapsulates learned patterns that would be extremely resource-intensive to compute without the benefit of prior training.