Method for separating target sound source from mixed sound source and electronic device thereof

Patent metadata

Field	Value
Publication number	US-12424241-B2
Application number	US-202318235664-A
Country	US
Kind code	B2
Filing date	Aug 18, 2023
Priority date	Aug 18, 2022
Publication date	Sep 23, 2025
Grant date	Sep 23, 2025

Abstract

A method for separating a target sound source, includes: obtaining a mixed sound source including at least one sound source; obtaining, based on the mixed sound source, scene information related to the mixed sound source; converting, based on the scene information, a first embedding vector corresponding to a designated sound source group into a second embedding vector; and separating, based on the mixed sound source and the second embedding vector, the target sound source from the mixed sound source.

First claim

What is claimed:

1. A method for separating a target sound source, the method comprising:
obtaining a mixed sound source including at least one sound source;
obtaining, based on the mixed sound source, scene information related to the mixed sound source;
converting, based on the scene information, a first embedding vector corresponding to a designated sound source group into a second embedding vector; and
separating, based on the mixed sound source and the second embedding vector, the target sound source from the mixed sound source.

2. The method of claim 1 , further comprising:
performing a first pre-training process to learn a scene information vector based on obtaining an input sound source,
wherein the obtaining the scene information comprises outputting the scene information vector based on the mixed sound source.

3. The method of claim 2 , wherein the performing the first pre-training process comprises outputting the scene information vector based on the obtaining of the input sound source and learning to classify a designated scene based on the output scene information vector.

4. The method of claim 1 , wherein the obtaining the scene information comprises generating an instance vector corresponding to the scene information based on the mixed sound source.

5. The method of claim 1 , wherein the first embedding vector corresponds to an entirety of the designated sound source group, and
wherein the converting the first embedding vector into the second embedding vector comprises converting the first embedding vector into the second embedding vector corresponding to at least a portion of the designated sound source group based on the obtained scene information.

6. The method of claim 1 , further comprising:
performing a second pre-training process to learn partial embedding vectors, each of the partial embedding vectors corresponding to a respective sound source included in the designated sound source group,
wherein the first embedding vector is a sum vector of the partial embedding vectors.

7. The method of claim 6 , wherein the converting the first embedding vector into the second embedding vector comprises:
identifying the target sound source included in the designated sound source group corresponding to the scene information; and
converting the first embedding vector into the second embedding vector to correspond to the target sound source.

8. The method of claim 1 , wherein the designated sound source group includes at least one target sound source designated to correspond to the scene information.

9. The method of claim 1 , further comprising performing a third pre-training process, based on an embedding vector between specific scene information and a specific target sound source corresponding to the specific scene information, to learn the conversion from the first embedding vector into the second embedding vector.

10. The method of claim 1 , wherein the separating the target sound source comprises generating the target sound source having a vector form with the same size as the mixed sound source by applying the second embedding vector to the mixed sound source.

11. An electronic device comprising:
an input interface;
a memory storing at least one instruction; and
at least one processor operatively connected with the input interface and the memory,
wherein the at least one processor is configured to execute the at least one instruction to:
obtain, from the input interface, a mixed sound source including at least one sound source,
obtain, based on the mixed sound source, scene information related to the mixed sound source
convert, based on the scene information, a first embedding vector corresponding to a designated sound source group into a second embedding vector, and
separate, based on the mixed sound source and the second embedding vector, the target sound source from the mixed sound source.

12. The electronic device of claim 11 , wherein the at least one processor is further configured to execute the at least one instruction to:
perform a first pre-training process to learn a scene information vector based on obtaining an input sound source, and
as at least part of the obtaining the scene information output the scene information vector based on the mixed sound source.

13. The electronic device of claim 12 , wherein the at least one processor is further configured to: as at least part of the performing the first pre-training process, output the scene information vector based on the obtaining of the input sound source and learn to classify a designated scene based on the output scene information vector.

14. The electronic device of claim 11 , wherein the at least one processor is further configured to execute the at least one instruction to: as at least part of the obtaining the scene information, generate an instance vector corresponding to the scene information based on the mixed sound source.

15. The electronic device of claim 11 , wherein the first embedding vector corresponds to an entirety of the designated sound source group, and
wherein the at least one processor is further configured to: as at least part of the converting the first embedding vector into the second embedding vector, convert the first embedding vector into the second embedding vector corresponding to at least a portion of the designated sound source group based on the scene information.

16. The electronic device of claim 11 , wherein the at least one processor is further configured to execute the at least one instruction to perform a second pre-training process to learn partial embedding vectors, each of the partial embedding vectors corresponding to a respective sound source included in the designated sound source group, and
wherein the first embedding vector is a sum vector of the partial embedding vectors.

17. The electronic device of claim 16 , wherein the at least one processor is further configured to execute the at least one instruction to: as at least part of the converting the first embedding vector into the second embedding vector, identify the target sound source included in the designated sound source group corresponding to the scene information and convert the first embedding vector into the second embedding vector to correspond to the target sound source.

18. The electronic device of claim 11 , wherein the designated sound source group includes at least one target sound source designated to correspond to the scene information.

19. The electronic device of claim 11 , wherein the at least one processor is further configured to execute the at least one instruction to perform a third pre-training process, based on an embedding vector between specific scene information and a specific target sound source corresponding to the specific scene information, to learn the conversion from the first embedding vector into the second embedding vector.

20. The electronic device of claim 11 , wherein the at least one processor is further configured to execute the at least one instruction to: as at least part of the separating the target sound source, generate the target sound source having a vector form with the same size as the mixed sound source by applying the second embedding vector to the mixed sound source.

Assignees

• Samsung Electronics Co Ltd
• Univ Hanyang Ind Univ Coop Found

Inventors

• Yang Jaemo
• Chang Joonhyuk
• Kim Geeyeun
• Moon Hangil
• Bang Kyoungho
• Kim Dail
• Kim Yungyeo
• Baek Minsang
• Choi Wongook
• Choi Jeonghwan

Classifications

• G10L25/30

  using neural networks · CPC title

• G06V10/82

  using neural networks · CPC title

• G06N3/08

  Learning methods · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N20/00

  Machine learning · CPC title