Method and apparatus for determining output token

1. A processor-implemented method of determining an output token, comprising:

predicting a first probability of each of candidate output tokens provided from a first model;

predicting a second probability of each of the candidate output tokens provided from a second model different from and interworking with the first model;

adjusting the second probability of each of the candidate output tokens based on the first probability; and

performing an inference operation including a speech recognition by determining the output token among the candidate output tokens based on the first probability and the adjusted second probability.

2. The method of claim 1 , wherein the adjusting comprises:

determining rankings of the candidate output tokens of the first model based on the first probability; and

adjusting the second probability based on the rankings.

3. The method of claim 2 , wherein the adjusting the second probability based on the rankings comprises:

adjusting the second probability using a function determined based on the rankings.

4. The method of claim 2 , wherein the adjusting the second probability based on the rankings comprises:

adjusting the second probability based on one or both of a scale coefficient and a threshold coefficient that are determined based on the rankings.

5. The method of claim 1 , wherein the adjusting comprises:

determining rankings of the candidate output tokens of the first model based on the first probability;

extracting target candidate output tokens for which the adjusting is to be performed from among the candidate output tokens based on the rankings; and

adjusting the second probability corresponding to the extracted target candidate output tokens.

6. The method of claim 5 , wherein the adjusting the second probability corresponding to the extracted target candidate output tokens comprises:

adjusting the second probability corresponding to the extracted target candidate output tokens using a function determined based on the rankings.

7. The method of claim 1 , further comprising:

normalizing the adjusted second probability.

8. The method of claim 7 , wherein the normalizing comprises:

normalizing the second probability such that a total sum of the adjusted second probability is 1.

9. The method of claim 1 , further comprising:

performing a beam search on a preset number of candidate output tokens among the candidate output tokens; and

normalizing the adjusted second probability while performing the beam search.

10. The method of claim 1 , wherein the determining the output token comprises:

calculating a weighted sum of the first probability and the adjusted second probability; and

determining, to be the output token, a candidate output token having a greatest weighted sum among the candidate output tokens.

11. The method of claim 7 , wherein a weight applied to the first probability is greater than a weight applied to the second probability.

12. The method of claim 1 , wherein the predicting the first probability comprises:

predicting the first probability based on an output token previously determined by the first model and the second model,

wherein the predicting the second probability comprises:

predicting the second probability based on the output token previously determined by the first model.

13. A non-transitory computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform the method of claim 1 .

14. An apparatus for determining an output token, comprising:

one or more processors configured to:

predict a first probability of each of candidate output tokens provided from a first model;

predict a second probability of each of the candidate output tokens provided from a second model different from and interworking with the first model;

adjust the second probability of each of the candidate output tokens based on the first probability; and

perform an inference operation including a speech recognition by determining the output token among the candidate output tokens based on the first probability and the adjusted second probability.

15. The apparatus of claim 14 , wherein the one or more processors are configured to:

determine rankings of the candidate output tokens of the first model based on the first probability; and

adjust the second probability based on the rankings.

16. The apparatus of claim 15 , wherein the one or more processors are configured to:

adjust the second probability using a function determined based on the rankings.

17. The apparatus of claim 15 , wherein the one or more processors are configured to:

adjust the second probability based on one or both of a scale coefficient and a threshold coefficient that are determined based on the rankings.

18. The apparatus of claim 14 , wherein the one or more processors are configured to:

determine rankings of the candidate output tokens of the first model based on the first probability;

extract target candidate output tokens for which the adjusting is to be performed from among the candidate output tokens based on the rankings; and

adjust the second probability corresponding to the extracted target candidate output tokens.

19. The apparatus of claim 18 , wherein the one or more processors are configured to:

adjust the second probability corresponding to the extracted target candidate output tokens using a function determined based on the rankings.

20. The apparatus of claim 14 , wherein the one or more processors are configured to:

normalize the adjusted second probability.

21. The apparatus of claim 20 , wherein the one or more processors are configured to:

normalize the adjusted second probability such that a total sum of the adjusted second probability is 1.

22. The apparatus of claim 14 , wherein the one or more processors are configured to:

perform a beam search to search the candidate output tokens for a preset number of candidate output tokens; and

normalize the adjusted second probability while performing the beam search.

23. The apparatus of claim 14 , wherein the one or more processors are configured to:

calculate a weighted sum of the first probability and the adjusted second probability; and

determine, to be the output token, a candidate output token having a greatest weighted sum among the candidate output tokens.

24. The apparatus of claim 14 , wherein a weight applied to the first probability is greater than a weight applied to the second probability.

25. The apparatus of claim 14 , wherein the one or more processors are configured to:

predict the first probability based on an output token previously determined based on the first model and the second model; and

predict the second probability based on the output token previously determined based on the first model.

26. An apparatus comprising:

one or more processors configured to:

rank candidate tokens based on first probabilities of the candidate tokens;

determine a weight for each of the candidate tokens based on a ranking of each of the candidate tokens;

adjust second probabilities of the candidate tokens by weighting each of the second probabilities by the respective weight;

calculate a weighted sum for each of the candidate tokens based on the first probabilities and the adjusted second probabilities; and

perform an inference operation including a speech recognition by outputting a candidate token having a largest weighted sum.

27. The apparatus of claim 26 , wherein the one or more processors are configured to:

rank the candidate tokens by grouping the candidate tokens into two or more ranked groups based on the first probabilities of the candidate tokens; and

determine a different weight for each of the ranked groups.

28. The apparatus of claim 26 , wherein the one or more processors are configured to determine the weight for each of the candidate tokens such that a first candidate token, which has a higher first probability than a second candidate token, has a larger weight than the second candidate token.

29. The apparatus of claim 26 , wherein the one or more processors are configured to determine the weight for each of the candidate tokens to be larger than one, one, or between zero and one based on the ranking of each of the candidate tokens.

30. The apparatus of claim 26 , wherein the one or more processors are configured to rank the candidate tokens is descending from a candidate token having a highest first probability to a candidate token having a lowest first probability.