Personalized bandwidth extension

A computer-implemented method for training a bandwidth extension model for personalized bandwidth extension, wherein the method comprises:

• obtaining an audio dataset comprising one or more first audio signals with a first bandwidth;
• obtaining a hearing dataset comprising a hearing profile comprising one or more user profiles including a first user profile for a user of an audio device;
• applying a bandwidth extension model to the one or more first audio signals to generate one or more bandwidth extended audio signals with a second bandwidth,
• determining one or more perceptual losses associated with the one or more bandwidth extended audio signals based on the first user profile;
• determining a masking threshold and a personalized bandwidth based on the hearing dataset;
• determining an audible range and audible levels based on the masking threshold and the personalized bandwidth;
• and
• training, based on the one or more perceptual losses, the bandwidth extension model,
• wherein the audible range is a personalized bandwidth for the user of the audio device,
• wherein the audible levels are used to determine masked and unmasked components of the generated bandwidth extending audio signals,
• wherein each of the one or more perceptual losses is a perceptual loss function,
• wherein the training of the bandwidth extension model includes training the bandwidth extension model to generate spectral content that is audible within the personalized bandwidth for the user of the audio device.