Deep learning based method and system for processing sound quality characteristics

1. A method comprising:

extracting features from audio data associated with a user;

determining, based on the extracted features, data characteristics of the audio data;

determining that the data characteristics are associated with artificial damage;

based on a determination that the data characteristics are associated with the artificial damage, training a baseline model by:

artificially processing original audio data samples of audio databases to obtain lossy audio data samples;

extracting features of the original audio data samples and the lossy audio data samples respectively to obtain lossy audio data characteristics and original audio data characteristics;

using the lossy audio data characteristics as input layer data and using the original audio data characteristics as output layer data; and

based on the data characteristics and the baseline model, generating a sound quality processing result of the audio data associated with the user,

wherein the baseline model comprises a neural network model.

2. The method of claim 1 , wherein the audio data associated with the user comprises data indicating a user's preferences for sound quality characteristics.

3. The method of claim 1 , wherein the audio data associated with the user comprises at least one of user behavior data, user audio list data, audio data in playback, ambient noise data, or user device data.

4. The method of claim 1 , wherein the data characteristics comprise one or more of time domain or frequency domain characteristics.

5. The method of claim 1 , further comprising:

determining the neural network model by performing an unsupervised deep learning method on the audio data associated with the user.

6. The method of claim 1 , further comprising:

based on a determination that the data characteristics are associated with empirical fusion, training the baseline model by:

extracting category tag features and data characteristics from audio data associated with a plurality of users; and

training the baseline model by using the data characteristics of the audio data associated with the plurality of users as input layer data and using the category tag features as output layer data.

7. The method of claim 1 , further comprising:

based on a determination that the data characteristics are associated with sound quality feedback, training the baseline model by:

extracting features from environmental data of audio data associated with a plurality of users and behavioral data corresponding to the environmental data respectively to obtain environmental data characteristics of the environmental data and sound quality setting characteristics of the behavioral data; and

using the environmental data characteristics as input layer data and using the sound quality feedback as output layer data.

8. The method of claim 1 , further comprising:

based on a determination that the data characteristics are associated with empirical formula adjustment, training the baseline model by:

extracting features from audio data associated with a plurality of users to obtain original data features;

adjusting sound quality of the audio data associated with the plurality of users and extracting features from the adjusted sound quality of the audio data to obtain adjustment characteristics;

establishing a network model that maps the features from the adjusted sound quality of the audio data to subjective evaluation through subjective evaluation experiments to obtain optimized sound quality adjustment parameters; and

using the original data features as input layer data and using the optimized sound quality adjustment parameters as output layer data.

9. The method of claim 1 , further comprising:

based on a determination that the data characteristics are associated with signal domain calculation, training the baseline model by:

extracting sound quality characteristics from audio data associated with a plurality of users;

clustering the sound quality characteristics to obtain a clustering model;

extracting sound quality characteristics from user preferred audio data;

inputting the extracted sound quality characteristics from the user preferred audio data into the clustering model to generate a vector of sound quality characteristic distributions; and

training the baseline model based on the vector of sound quality characteristic distributions.

10. The method of claim 1 , further comprising:

based on a determination that the data characteristics are associated with noise, training the baseline model by:

extracting data features and corresponding category tag features from audio data associated with a plurality of users, wherein the audio data associated with the plurality of users comprises noise or normal sound category tag features; and

training the baseline model based on the data features and the corresponding category tag features.

11. The method of claim 10 , further comprising:

determining a noise identification result based on the baseline model;

based on a determination that the noise identification result indicates that noise is filterable, determining to use a trained noise reduction baseline model to obtain noise-reduced audio data; and

based on a determination that the noise identification result indicates that the noise is non-filterable, sending a notice to a user, wherein the notice indicates deleting the audio data associated with the plurality of users.

12. The method of claim 1 , further comprising:

based on a determination that the data characteristics are associated with other data characteristics, training the baseline model by:

obtaining audio data, associated with a plurality of users, from an audio database, wherein the audio data is associated with the other data characteristics;

extracting features of the other sound quality characteristics from the audio data associated with the plurality of users; and

using the features of the other sound quality characteristics as input layer data, and using (a) user preferred features of the other sound quality characteristics or (b) subjective perception evaluation obtained by mapping the other sound quality characteristics to a network model of subjective evaluation as output layer data.

13. A method, comprising:

extracting preference data of a user in real time;

determining data characteristics of the preference data;

determining that the data characteristics are associated with artificial damage;

based on a determination that the data characteristics are associated with the artificial damage, training a baseline model by:

artificially processing original audio data samples of audio databases to obtain lossy audio data samples;

extracting features of the original audio data samples and the lossy audio data samples respectively to obtain lossy audio data characteristics and original audio data characteristics;

using the lossy audio data characteristics as input layer data and using the original audio data characteristics as output layer data; and

generating a personalized sound quality characteristic processing model by learning, based on the baseline model, the preference data of the user; and

generating a personalized sound quality characteristic processing result based on the personalized sound quality characteristic processing model,

wherein the baseline model comprises a neural network model.

14. The method of claim 13 , wherein the learning the preference data of the user comprises:

outputting sound quality characteristics processing results based on the baseline model, and updating the baseline model to obtain the personalized sound quality characteristics processing model based on feedback data from the user on the sound quality characteristics processing results; or

based on the user's feedback data on the personalized sound quality characteristics processing result, updating the personalized sound characteristic processing model.

15. The method of claim 14 , wherein the learning the preference data of the user comprises: if a number of individual user feedback behaviors in user behavior data exceeds a predetermined value, relearning the baseline model or the personalized audio sound quality characteristics processing model based on the preference data of the user.

16. A system, comprising:

an audio playback terminal configured to store user preference data;

a cloud platform configured to:

determine that data characteristics of the user preference data are associated with artificial damage;

based on a determination that the data characteristics are associated with the artificial damage, train a baseline model by:

artificially processing original audio data samples of audio databases to obtain lossy audio data samples;

extracting features of the original audio data samples and the lossy audio data samples respectively to obtain lossy audio data characteristics and original audio data characteristics;

using the lossy audio data characteristics as input layer data and using the original audio data characteristics as output layer data,

wherein the baseline model comprises a neural network model; and

an online learning device configured to receive the user preference data and learn the baseline model online to generate a personalized audio sound quality model.

17. The method of claim 13 , wherein the data characteristics comprise one or more of time domain or frequency domain characteristics.

18. The system of claim 16 , wherein the data characteristics comprise one or more of time domain or frequency domain characteristics.

19. The system of claim 16 , further comprising:

processing, based on the personalized audio sound quality model, audio data.

20. The system of claim 16 , wherein the cloud platform is configured to train the baseline model by performing an unsupervised deep learning method on the user preference data.