Neural Network Hotsound Detection Optimization

The patent segments the neural network into two distinct components: a first neural network trained for keyword/keyphrase recognition and a second neural network trained for hotsound detection. This segmentation allows each network to specialize in its respective task, enabling reliable hotsound detection without interfering with keyword detection accuracy. The second network processes hotsound inputs independently while the first network continues to handle keyword recognition.

The patent introduces an intermediary optimization process that modifies the seed hotsound in the feature space generated by the first neural network. By using gradient descent algorithms to optimize the hotsound representation in the existing feature space, the system creates a modified hotsound that is detectable by the first network without triggering false positives from keyword patterns. This intermediary optimization step bridges the gap between the two detection tasks.

Instead of retraining the entire neural network, the patent segments the training process by creating a second, specialized neural network for hotsound detection. This segmentation preserves the original first network's keyword detection capabilities and training, while the second network is trained exclusively on hotsound data. The result is improved hotsound detection capability without the computational overhead of retraining the entire system.

The patent performs preliminary optimization of the hotsound in the feature space before training the second network. By using gradient descent to pre-optimize the modified hotsound representation in the feature space generated by the first network, the system prepares the input data in advance. This preliminary action reduces the training complexity and computational requirements for the second network, thereby maintaining computational efficiency.

The patent divides the detection system into two specialized networks, allowing each to be trained independently on its specific task. The first network maintains its existing keyword detection accuracy, while the second network is trained solely for hotsound detection. This segmentation eliminates the need for time-consuming full-network retraining, as only the second network requires training on hotsound data.

The patent performs preliminary optimization of the hotsound representation in the feature space using gradient descent algorithms before initiating the second network's training. This preliminary action pre-processes the hotsound data to be optimally detectable, reducing the complexity and time required for the actual training phase. The modified hotsound is already aligned with the feature space structure, accelerating the training process.