System and method for enhancement of a degraded audio signal

1. A method for supervised training of a machine learning model, MLM, to enhance a degraded audio signal by calculating gains to be applied to frequency bands of the degraded audio signal, the method comprising the steps of:

receiving a degraded audio signal and a clean audio signal for training of the MLM;

extracting a first set of features from the received degraded audio signal, and a second set of features from the received clean audio signal, each feature corresponding to a frequency band of the respective received audio signals;

comparing each feature of the first set of features to a corresponding feature of the second set of features to derive a set of gains, each gain corresponding to a respective feature among the first set of features, and used as ground truth when training the MLM;

using the first set of features and the derived set of gains as a training set for training the MLM;

wherein the method further comprises at least one of:

a pre-processing step performed prior to deriving the set of gains, wherein the pre-processing step comprises adjusting the frequency energy distribution of the first and/or the second set of features such that the frequency energy distribution of the first set of features is more similar to the frequency energy distribution of the second set of features, and

defining a loss function of the MLM which is configured to punish a predicted gain being lower than the ground truth gain more than a predicted gain being higher than the ground truth gain.

2. A method according to claim 1 , wherein only one of the pre-processing step and the step of defining a loss function of the MLM is used.

3. A method according to claim 1 , wherein both the pre-processing step and the step of defining a loss function of the MLM are used.

4. A method according to claim 1 , wherein the loss function is further weighted according to the frequency band of the features of the training set, such that an error for a feature corresponding to a relatively higher frequency band is weighted with a relatively higher weight.

5. A method according to claim 4 , wherein an error for a feature corresponding to a frequency band exceeding 6 kHz is weighted with a higher weight compared to an error for a feature corresponding to a frequency band below 6 kHz.

6. A method according to claim 1 , wherein the first and second sets of features are extracted by converting the received degraded audio signal and clean audio signal into the frequency domain.

7. A method according to claim 6 , wherein the conversion is performed using one from the list of: a short time Fourier transform, SFTF, a modified discrete cosine transform, MDCT, and a shifted discrete frequency transform, MDXT.

8. A method according to claim 7 , wherein the first and second set of features are extracted by, for each frequency band of a plurality of frequency bands,

for frequency bins of the frequency band, combining complex features of the frequency domain representation of the respective audio signal corresponding to the frequency bins into a single feature corresponding to that frequency band.

9. A method according to claim 8 , wherein the features of the first and second set of features corresponds to Mel-frequency band powers, Bark Scale band powers, log-frequency band powers or ERB band powers.

10. A method according to claim 1 , wherein the step of pre-processing comprises balancing a frequency energy distribution of the second set of features across the entire frequency band of the received clean audio signal.

11. A method according to claim 10 , wherein the pre-processing comprises:

fitting a polynomial curve to the second set of features,

defining a filter based on a difference between the polynomial curve and a constant function,

applying the filter to the second set of features.

12. A method according to claim 10 , wherein the pre-processing comprises:

fitting a polynomial curve to the second set of features,

calculating a difference between a minimum value and a maximum value of the polynomial curve,

upon determining that the difference exceeds a threshold value:

defining a filter based on the difference between the polynomial curve and a constant function,

applying the filter to the second set of features.

13. A method according to claim 12 , wherein the threshold value corresponds to a 3 dB difference in a frequency energy distribution of the second set of features across the entire frequency band of the received clean audio signal.

14. A method according to claim 11 , wherein a value of the constant function is set to the maximum value of the polynomial curve.

15. A method according to claim 11 , wherein the polynomial curve is one from the list of: a linear curve, a quadratic curve and a cubic curve.

16. A method according to claim 1 , wherein the loss function is configured to punish a predicted gain being lower than the ground truth gain more than a predicted gain being higher than the ground truth gain by:

multiplying a distance measurement between the predicted gain and the ground truth with a weight, the weight being relatively higher when:

the predicted gain is lower than the ground truth gain, and

the predicted gain is negative,

the weight being relatively lower when:

the predicted gain is higher than or equal to the ground truth gain, or

the predicted gain is positive.

17. A method according to claim 16 , wherein the ratio between the relatively higher weight and the relatively lower weight is between 3-7.

18. A method for enhancing a degraded audio signal, comprising the steps of:

receiving a degraded audio signal;

extracting a first set of features from the received degraded audio signal;

inputting the extracted first set of features to a machine learning model, MLM, trained according to claim 1 ; and

using output gains from the MLM for enhancing the received degraded audio signal.

19. A device configured for supervised training of a machine learning model, MLM, to enhance a degraded audio signal by calculating gains to be applied to frequency bands of the degraded audio signal, the device comprising circuitry configure to:

receive a degraded audio signal and a clean audio signal for training of the MLM;

extract a first set of features from the received degraded audio signal, and a second set of features from the received clean audio signal, each feature corresponding to a frequency band of the respective received audio signals;

compare each feature of the first set of features to a corresponding feature of the second set of features to derive a set of gains, each gain corresponding to a respective feature among the first set of features, and used as ground truth when training the MLM;

use the first set of features and the derived set of gains as a training set for training the MLM;

wherein the circuitry is further configured for at least one of:

prior to deriving the set of gains, performing pre-processing comprising adjusting the frequency energy distribution of the first and/or the second set of features such that the frequency energy distribution of the first set of features is more similar to the frequency energy distribution of the second set of features, and

defining a loss function of the MLM configured to punish a predicted gain being lower than the ground truth gain more than a predicted gain being higher than the ground truth gain.

20. A device configured for enhancing a degraded audio signal, the device comprising circuitry configure to:

receive a degraded audio signal;

extract a first set of features from the received degraded audio signal;

input the extracted first set of features to a machine learning model, MLM, trained according to claim 1 ; and

use output gains from the MLM for enhancing the received degraded audio signal.

21. A computer program product comprising a non-transitory computer-readable storage medium with instructions adapted to carry out the method of claim 1 when executed by a device having processing capability.