Using a predictive model to automatically enhance audio having various audio quality issues

A system for enhancing audio, the system comprising:

• a training subsystem configured to train a prediction model to enhance source audios, the training subsystem configured to:
• access training data comprising a plurality of tuples, the plurality of tuples comprising source audios and target audios, each tuple comprising a source audio and a corresponding target audio, wherein the corresponding target audio is a target version of the source audio;
• cause the prediction model to generate predicted audios based on the source audios;
• apply a loss function to the predicted audios and the target audios, the loss function comprising a combination of (i) a spectrogram loss based on spectrogram versions of the predicted audios and the target audios and (ii) an adversarial loss based on feedback from a discriminator of a generative adversarial network (GAN);
• and
• update the prediction model to optimize the loss function comprising the combination of the spectrogram loss and the adversarial loss.

A non-transitory computer-readable medium embodying program code for enhancing audio, the program code comprising instructions that, when executed by a processor, cause the processor to perform operations comprising:

• providing a generative adversarial network (GAN) comprising a prediction model and a discriminator to be jointly trained;
• obtaining training data comprising a plurality of tuples, the plurality of tuples comprising source audios and target audios, each tuple comprising a source audio and a corresponding target audio, wherein the corresponding target audio is a target version of the source audio;
• training the GAN based on the training data, wherein the training comprises:
• causing the prediction model to generate predicted audios based on the source audios;
• applying a generator loss function to the predicted audios and the target audios, the generator loss function comprising a combination of a spectrogram loss and an adversarial loss;
• and
• updating the prediction model to optimize the generator loss function comprising the combination of the spectrogram loss and the adversarial loss;
• wherein, after the training, the prediction model is configured to generate new predicted audios based on new source audios.