Back to results
US 2026/0313304 A1Published application
Method for training a neural network
AssigneeOticon A/S
InventorsJan M. DE HAAN, Poul HOANG, Michael Noes GÄTKE, Alexander W. KING, Katarzyna ORLINSKA, Philippe GONZALEZ
Life of this patent
Estimating the term…
Abstract
Disclosed herein are embodiments of a method for training a neural network. The method can include receiving training input and training targets from a sound database. The method can further include providing a high-resolution training sub-band input based on the training input. The neural network can be configured to determine neural network parameters and sub-band filters based on high-resolution training sub-band input.
Claims12
A method for training a neural network comprising:
- receiving, by a sound database, a training input, and wherein the training input comprises a speech component and a noise component;
- receiving, by the sound database, a training target comprising the speech component;
- providing, by a high-resolution analysis filter bank, a high-resolution training sub-band input based on the training input,
- determining, by the neural network comprising a plurality of neural network parameters, one or more sub-band filters based on the high-resolution training sub-band input;
- providing, by a low-resolution analysis filter bank, a low-resolution training sub-band input based on the training input;
- providing, by filtering the low-resolution training sub-band input using the one or more sub-band filters, a training sub-band output;
- providing, by a low-resolution synthesis filter bank, a training output;
- and
- training the neural network by updating, by a network optimizer, the plurality of neural network parameters based on reducing a neural network loss between the training output and the training target.
6 more elements
Dependencies are inferred from the claim text.