| 摘要 |
A neural network for noise reduction for speech recognition in a noisy environment uses an algorithm for automatic network generation which automatically selects a suitable signal representation. Nodes may be added to the input layer of the neural network successively, with a new node being trained by calculating and minimizing a mapping error. A squared mapping error may be formed and the mapping error may be assigned a weight dependent on the importance of the vectors. In addition, a neural network that performs neural noise reduction by reducing, in a training phase, a mapping error between noise-free vectors at an output of the neural network and noise-reduced vectors at the output of the neural network using an iterative process, has the mapping error further reduced by additional information which is selected from a suitable signal representation at the input of the neural network.
|
