Data driven radio enhancement

1. A system comprising:

a processing apparatus that is configured to:

access a demodulated audio signal that is based on a demodulated analog radio signal;

determine a window of audio samples based on the demodulated audio signal;

apply an audio enhancement network to the window of audio samples to obtain an enhanced audio segment, in which the audio enhancement network includes a machine learning network that has been trained using demodulated audio signals derived from analog radio signals to take the window of audio samples, in a pulse-code modulated format, as input and output the enhanced audio segment, in a pulse-code modulated format, such that power of an electromagnetic atmospheric noise present in the window of audio samples is reduced in the enhanced audio segment; and

store, play, or transmit an enhanced audio signal based on the enhanced audio segment.

2. The system of claim 1 , in which the processing apparatus is configured to:

access a training demodulated audio signal that is based on a demodulated radio signal;

receive a streamed audio signal corresponding to the radio signal;

determine a training window of audio samples based on the training demodulated audio signal;

determine, based on the streamed audio signal, a ground truth window of audio samples corresponding to the training window of audio samples; and

train the audio enhancement network using the training window of audio samples as a training input with feedback based on a loss function comparing a resulting enhanced audio segment to the ground truth window of audio samples.

3. The system of claim 2 , in which determining, based on the streamed audio signal, the ground truth window of audio samples corresponding to the training window of audio samples comprises:

determining cross-correlations between the training demodulated audio signal and the streamed audio signal; and

resampling, based on the cross-correlations, the streamed audio signal to align samples of the ground truth window of audio samples with the training window of audio samples.

4. The system of claim 2 , in which determining the training window of audio samples based on the training demodulated audio signal, and determining, based on the streamed audio signal, the ground truth window of audio samples corresponding to the training window of audio samples comprises:

determining cross-correlations between the training demodulated audio signal and the streamed audio signal;

based on the cross-correlations, identifying portions of the training demodulated audio signal and the streamed audio signal that are mismatched portions; and

omitting samples based on the mismatched portions of the training demodulated audio signal and the streamed audio signal from the training window of audio samples and the ground truth window of audio samples respectively.

5. The system of claim 2 , in which determining, based on the streamed audio signal, the ground truth window of audio samples corresponding to the training window of audio samples comprises:

determining cross-correlations between the training demodulated audio signal and the streamed audio signal;

based on the cross-correlations, determining a time shift between the training demodulated audio signal and the streamed audio signal; and

applying the time shift to the streamed audio signal to align samples of the ground truth window of audio samples with the training window of audio samples.

6. The system of claim 1 , in which the processing apparatus is configured to:

input the window of audio samples to a classifier to obtain a classification, in which the classifier includes a neural network and the classification takes a value from a set of multiple classes of audio; and

select, based on the classification, the audio enhancement network from a set of multiple audio enhancement networks, in which the audio enhancement network includes a neural network that has been trained using demodulated audio signals of a type associated with the classification.

7. The system of claim 6 , in which the set of multiple classes of audio includes a speech class and a music class, and the set of multiple audio enhancement networks includes a speech enhancement network and a music enhancement network.

8. The system of claim 1 , in which the processing apparatus is configured to:

determine a sequence of partially overlapping windows of audio samples, including the window of audio samples, based on the demodulated audio signal;

determine a sequence of respective enhanced audio segments, including the enhanced audio segment, based on the sequence of partially overlapping windows of audio samples using the audio enhancement network; and

determine an enhanced audio signal by blending the sequence of respective enhanced audio segments, including combining corresponding samples from multiple segments in overlapping portions of the enhanced audio segments.

9. The system of claim 1 , in which the audio enhancement network includes a convolutional neural network.

10. The system of claim 1 , comprising:

a speaker configured to play the enhanced audio signal.

11. A method comprising:

demodulating an analog radio signal to obtain a demodulated audio signal;

determining a window of audio samples based on the demodulated audio signal;

applying an audio enhancement network to the window of audio samples to obtain an enhanced audio segment, in which the audio enhancement network includes a machine learning network that has been trained using demodulated audio signals derived from analog radio signals to take the window of audio samples, in a pulse-code modulated format, as input and output the enhanced audio segment, in a pulse-code modulated format, such that power of an electromagnetic atmospheric noise present in the window of audio samples is reduced in the enhanced audio segment; and

storing, playing, or transmitting an enhanced audio signal based on the enhanced audio segment.

12. The method of claim 11 , comprising:

demodulating a radio signal to obtain a training demodulated audio signal;

receiving a streamed audio signal corresponding to the radio signal;

determining a training window of audio samples based on the training demodulated audio signal;

determining, based on the streamed audio signal, a ground truth window of audio samples corresponding to the training window of audio samples; and

training the audio enhancement network using the training window of audio samples as a training input with feedback based on a loss function comparing a resulting enhanced audio segment to the ground truth window of audio samples.

13. The method of claim 12 , in which determining, based on the streamed audio signal, the ground truth window of audio samples corresponding to the training window of audio samples comprises:

determining cross-correlations between the training demodulated audio signal and the streamed audio signal; and

resampling, based on the cross-correlations, the streamed audio signal to align samples of the ground truth window of audio samples with the training window of audio samples.

14. The method of claim 12 , in which determining the training window of audio samples based on the training demodulated audio signal, and determining, based on the streamed audio signal, the ground truth window of audio samples corresponding to the training window of audio samples comprises:

determining cross-correlations between the training demodulated audio signal and the streamed audio signal;

based on the cross-correlations, identifying portions of the training demodulated audio signal and the streamed audio signal that are mismatched portions; and

omitting samples based on the mismatched portions of the training demodulated audio signal and the streamed audio signal from the training window of audio samples and the ground truth window of audio samples respectively.

15. The method of claim 12 , in which determining, based on the streamed audio signal, the ground truth window of audio samples corresponding to the training window of audio samples comprises:

determining cross-correlations between the training demodulated audio signal and the streamed audio signal;

based on the cross-correlations, determining a time shift between the training demodulated audio signal and the streamed audio signal; and

applying the time shift to the streamed audio signal to align samples of the ground truth window of audio samples with the training window of audio samples.

16. The method of claim 11 , comprising:

inputting the window of audio samples to a classifier to obtain a classification, in which the classifier includes a neural network and the classification takes a value from a set of multiple classes of audio; and

selecting, based on the classification, the audio enhancement network from a set of multiple audio enhancement networks, in which the audio enhancement network includes a neural network that has been trained using demodulated audio signals of a type associated with the classification.

17. The method of claim 16 , in which the set of multiple classes of audio includes a speech class and a music class, and the set of multiple audio enhancement networks includes a speech enhancement network and a music enhancement network.

18. The method of claim 11 , comprising:

determining a sequence of partially overlapping windows of audio samples, including the window of audio samples, based on the demodulated audio signal;

determining a sequence of respective enhanced audio segments, including the enhanced audio segment, based on the sequence of partially overlapping windows of audio samples using the audio enhancement network; and

determining an enhanced audio signal by blending the sequence of respective enhanced audio segments, including combining corresponding samples from multiple segments in overlapping portions of the enhanced audio segments.

19. The method of claim 11 , in which the audio enhancement network includes a convolutional neural network.

20. A non-transitory computer-readable storage medium comprising executable instructions that, when executed by a processor, facilitate performance of operations comprising:

accessing a demodulated audio signal that is based on a demodulated analog radio signal;

determining a window of audio samples based on the demodulated audio signal;

applying an audio enhancement network to the window of audio samples to obtain an enhanced audio segment, in which the audio enhancement network includes a machine learning network that has been trained using demodulated audio signals derived from analog radio signals to take the window of audio samples, in a pulse-code modulated format, as input and output the enhanced audio segment, in a pulse-code modulated format, such that power of an electromagnetic atmospheric noise present in the window of audio samples is reduced in the enhanced audio segment; and

storing, playing, or transmitting an enhanced audio signal based on the enhanced audio segment.

21. The non-transitory computer-readable storage medium of claim 20 , comprising executable instructions that, when executed by a processor, facilitate performance of operations comprising:

accessing a training demodulated audio signal that is based on a demodulated radio signal;

receiving a streamed audio signal corresponding to the radio signal;

determining a training window of audio samples based on the training demodulated audio signal;

determining, based on the streamed audio signal, a ground truth window of audio samples corresponding to the training window of audio samples; and

training the audio enhancement network using the training window of audio samples as a training input with feedback based on a loss function comparing a resulting enhanced audio segment to the ground truth window of audio samples.

22. The non-transitory computer-readable storage medium of claim 20 , comprising executable instructions that, when executed by a processor, facilitate performance of operations comprising:

inputting the window of audio samples to a classifier to obtain a classification, in which the classifier includes a neural network and the classification takes a value from a set of multiple classes of audio; and

selecting, based on the classification, the audio enhancement network from a set of multiple audio enhancement networks, in which the audio enhancement network includes a neural network that has been trained using demodulated audio signals of a type associated with the classification.