Granted Patent B2

US 12,236,975 · App. 17/526,810 · Granted Feb 25, 2025

Inventors: Trung Bui (San Jose, CA); Subhadeep Dey (Martigny, CH); Seunghyun Yoon (Seoul, KR)

Assignee: Adobe Inc.

G10L25/63G06F17/16G06F17/18G06N3/047G10L25/30

Abstract

Claims (47)

1. A system comprising:

one or more memory devices comprising:

an audio bi-directional recurrent encoder that generates an audio feature vector for one or more words in an acoustic sequence;

a textual bi-directional recurrent encoder that generates a textual feature vector for the one or more words in a textual sequence corresponding to the acoustic sequence;

a multi-hop neural attention model that generates an attention output at each hop that alternates from utilizing the textual feature vector and the audio feature vector as context; and

a hidden feature vector generator that generates a hidden feature vector based on the attention output and one or more of the audio feature vector and the textual feature vector; and

one or more processors configured to cause the system to determine an emotion of the acoustic sequence based on the hidden feature vector.

2. The system of claim 1 , wherein the multi-hop neural attention model further comprises a first-hop attention model that generates a first attention output based on the audio feature vector at a final state of the acoustic sequence and the textual feature vector at each state of the textual sequence.

3. The system of claim 2 , wherein the hidden feature vector generator generates the hidden feature vector as a first-hop hidden feature vector based on the textual feature vector at each state of the textual sequence and the first attention output.

4. The system of claim 1 , wherein the multi-hop neural attention model further comprises a second-hop attention model that generates a second attention output based on the audio feature vector at each state of the acoustic sequence.

5. The system of claim 4 , wherein the second-hop attention model generates the second attention output based on the hidden feature vector.

6. The system of claim 4 , wherein the hidden feature vector generator generates an additional hidden feature vector as a second-hop hidden feature vector based on the second attention output and the audio feature vector at each state of the acoustic sequence.

7. The system of claim 6 , wherein the multi-hop neural attention model further comprises a third-hop attention model that generates a third attention output based on the textual feature vector at each state of the textual sequence and the second-hop hidden feature vector.

8. The system of claim 7 , wherein the hidden feature vector generator generates another hidden feature vector as a third-hop hidden feature vector based on the third attention output and the textual feature vector at each state of the textual sequence.

9. The system of claim 1 , wherein the one or more processors are configured to cause the system to determine the emotion of the acoustic sequence based on two or more of a first-hop hidden feature vector, a second-hop hidden feature vector, or a third-hop feature vector.

10. A system comprising:

one or more memory devices comprising:

an audio encoder that generates an audio feature vector for one or more words in an acoustic sequence;

a textual encoder that generates a textual feature vector for the one or more words in a textual sequence corresponding to the acoustic sequence;

a first neural attention model that generates a first attention output by applying attention to the textual feature vector using the audio feature vector as context;

a first hidden feature vector generator that generates a first hidden feature vector based on the first attention output;

a second neural attention model that generates a second attention output by applying attention to the audio feature vector using the first hidden feature vector as context; and

a second hidden feature vector generator that generates a second hidden feature vector based on the second attention output and the audio feature vector; and

one or more processors configured to cause the system to determine an emotion of the acoustic sequence based on the first hidden feature vector and the second hidden feature vector.

11. The system of claim 10 , wherein:

the first neural attention model generates the first attention output based on the audio feature vector at a final state of the acoustic sequence and the textual feature vector at each state of the textual sequence; and

the first hidden feature vector generator generates the first hidden feature vector based on first attention output and the textual feature vector at each state of the textual sequence.

12. The system of claim 10 , wherein:

the second neural attention model generates the second attention output based on the audio feature vector at each state of the acoustic sequence; and

the second hidden feature vector generator generates the second hidden feature vector based on the audio feature vector at each state of the acoustic sequence.

13. The system of claim 10 , wherein the one or more memory devices further comprise:

a third neural attention model that generates a third attention output based on the textual feature vector at each state of the textual sequence and the second hidden feature vector; and

a third hidden feature vector generator that generates a third hidden feature vector based on the third attention output and the textual feature vector at each state of the textual sequence.

14. The system of claim 13 , wherein the one or more processors are configured to cause the system to determine the emotion of the acoustic sequence based on the third hidden feature vector.

15. The system of claim 10 , wherein:

the audio encoder comprises a bi-directional recurrent encoder; and

the textual encoder comprises a bi-directional recurrent encoder.

16. A non-transitory computer-readable medium storing instructions that, when executed by at least one processor, cause a computing device to:

generate, utilizing an audio bi-directional recurrent encoder, an audio feature vector for one or more words in an acoustic sequence;

generate, utilizing a textual bi-directional recurrent encoder, a textual feature vector for the one or more words in a textual sequence corresponding to the acoustic sequence;

generate, utilizing a neural attention model, an attention output by applying attention to the audio feature vector using the textual feature vector as a context vector;

generate, utilizing a hidden feature vector generator, a hidden feature vector based on the attention output and the audio feature vector; and

determine an emotion of the acoustic sequence based on the hidden feature vector.

17. The non-transitory computer-readable medium of claim 16 , further comprising instructions that, when executed by the at least one processor, cause the computing device to generate, utilizing the neural attention model, an additional attention output by applying attention to the textual feature vector using the audio feature vector as a context vector.

18. The non-transitory computer-readable medium of claim 17 , further comprising instructions that, when executed by the at least one processor, cause the computing device to generate, utilizing the hidden feature vector generator, an additional hidden feature vector based on the additional attention output and the textual feature vector.

19. The non-transitory computer-readable medium of claim 18 , further comprising instructions that, when executed by the at least one processor, cause the computing device to determine the emotion of the acoustic sequence based on the additional hidden feature vector, the emotion corresponding to an emotion category of sad, happy, angry, or neutral.

20. The non-transitory computer-readable medium of claim 16 , further comprising instructions that, when executed by the at least one processor, cause the computing device to generate a transcription of the acoustic sequence including the one or more words in the textual sequence.

Assignments (1)

Continuity (2)

References Cited (37)