Contribution
Data Generation for Training of Neural Networks Dealing with Underwater Speech Transmission
* Presenting author
Abstract:
Neural networks for speech quality enhancement have proven to be a great success. In acoustic communication over the underwater channel, the degradation of speech in terms of reverberation and noise is a severe problem. Communication via the underwater channel must ensure that all information is clearly understood. To set up a neural network that can compensate for the deterioration of the speech signal due to the transmission via the underwater channel, a large amount of training data is required. However, the amount of available data for training the neural network models is limited. By artificially imposing noise and reverberation, training data can be generated and used to train the models. This contribution presents the generation of training data, which includes both the communication protocol used for transmission and the effects and disturbances imposed by the underwater channel by realistic channel modelling.