Granted Patent B2

US 12,354,599 · App. 18/434,665 · Granted Jul 8, 2025

Inventors: Yiheng Huang (Shenzhen, CN); Xiaozheng Jian (Shenzhen, CN); Liqiang He (Shenzhen, CN)

Assignee: TENCENT TECHNOLOGY (SHENZHEN) COMPANY LIMITED

G10L15/083G10L15/063G10L19/008G10L2015/085G10L2015/088

Abstract

Claims (63)

1. A method for speech decoding performed by a computer device, the method comprising:

obtaining audio data corresponding to a speech, the audio data including a first audio frame and a second audio frame;

decoding the first audio frame using a first decoding network corresponding to a low-order language model and a second decoding network corresponding to a differential language model to obtain a plurality of first tokens, each first token having a corresponding decoding score according to the first and second decoding network;

obtaining a decoding score corresponding to a target token of the plurality of first tokens having a smallest decoding score;

determining a first pruning parameter according to a preset value and the decoding score corresponding to the target token;

traversing non-empty input edges in the first decoding network by using a first state of the target token as a start state to obtain a second pruning parameter;

rescoring a first decoding path formed by the non-empty input edges in the second decoding network by using a second state of the target token as a start state to obtain a third pruning parameter; and

decoding the second audio frame using the first decoding network and the second decoding network according to the first token list and the first pruning parameter, the second parameter, and the third parameter.

2. The method according to claim 1 , wherein the traversing non-empty input edges in the first decoding network by using a first state of the target token as a start state to obtain a second pruning parameter comprises:

obtaining a first acoustic score of the first decoding path under an acoustic model; and

obtaining the second pruning parameter according to the first acoustic score, the preset value, and the decoding score corresponding to the target token.

3. The method according to claim 1 , wherein the rescoring a first decoding path formed by the non-empty input edges in the second decoding network by using a second state of the target token as a start state to obtain a third pruning parameter comprises:

obtaining a first acoustic score of the first decoding path under an acoustic model;

obtaining a first path score of the first decoding path in the first decoding network;

obtaining a first rescoring score of the first decoding path in the second decoding network; and

obtaining the third pruning parameter according to the first acoustic score, the first path score, the first rescoring score, the preset value, and the decoding score corresponding to the target token.

4. The method according to claim 1 , wherein the decoding the second audio frame using the first decoding network and the second decoding network according to the first token list and the pruning parameters comprises:

for each first token in the first token list, traversing non-empty input edges in the first decoding network by using the first state of the first token as a start state, and determining whether to skip the first token according to the second pruning parameter and a second decoding path formed by each non-empty input edge;

rescoring the second decoding path formed by the non-empty input edge in the second decoding network by using the second state of the first token as a start state, and determining whether to skip the first token according to a rescoring result and the third pruning parameter;

obtaining a second token by performing a state jump on the first token, the second token comprising an updated state pair and a decoding score;

combining second tokens corresponding to all the first tokens into a second token list; and

determining a second token with a smallest decoding score in the second token list as a decoding result of the second audio frame.

5. The method according to claim 1 , wherein each first token includes a pair of states for characterizing a correspondence between a first state of the first token in the first decoding network and a second state of the first token in the second decoding network.

6. A computer device, comprising a processor and a memory, the memory storing computer-readable instructions that, when executed by the processor, cause the computer device to perform a method for speech decoding including:

obtaining audio data corresponding to a speech, the audio data including a first audio frame and a second audio frame;

decoding the first audio frame using a first decoding network corresponding to a low-order language model and a second decoding network corresponding to a differential language model to obtain a plurality of first tokens, each first having a state pair that includes a first state and a second state having a correspondence with the first state;

obtaining a decoding score corresponding to a target token of the plurality of first tokens having a smallest decoding score:

determining a first pruning parameter according to a preset value and the decoding score corresponding to the target token;

traversing non-empty input edges in the first decoding network by using a first state of the target token as a start state to obtain a second pruning parameter:

rescoring a first decoding path formed by the non-empty input edges in the second decoding network by using a second state of the target token as a start state to obtain a third pruning parameter; and

decoding the second audio frame using the first decoding network and the second decoding network according to the first token list and the first pruning parameter, the second parameter, and the third parameter.

7. The computer device according to claim 6 , wherein the traversing non-empty input edges in the first decoding network by using a first state of the target token as a start state to obtain a second pruning parameter comprises:

obtaining a first acoustic score of the first decoding path under an acoustic model; and

obtaining the second pruning parameter according to the first acoustic score, the preset value, and the decoding score corresponding to the target token.

8. The computer device according to claim 6 , wherein the rescoring a first decoding path formed by the non-empty input edges in the second decoding network by using a second state of the target token as a start state to obtain a third pruning parameter comprises:

obtaining a first acoustic score of the first decoding path under an acoustic model;

obtaining a first path score of the first decoding path in the first decoding network;

obtaining a first rescoring score of the first decoding path in the second decoding network; and

obtaining the third pruning parameter according to the first acoustic score, the first path score, the first rescoring score, the preset value, and the decoding score corresponding to the target token.

9. The computer device according to claim 6 , wherein the decoding the second audio frame using the first decoding network and the second decoding network according to the first token list and the pruning parameters comprises:

for each first token in the first token list, traversing non-empty input edges in the first decoding network by using the first state of the first token as a start state, and determining whether to skip the first token according to the second pruning parameter and a second decoding path formed by each non-empty input edge;

rescoring the second decoding path formed by the non-empty input edge in the second decoding network by using the second state of the first token as a start state, and determining whether to skip the first token according to a rescoring result and the third pruning parameter;

obtaining a second token by performing a state jump on the first token, the second token comprising an updated state pair and a decoding score;

combining second tokens corresponding to all the first tokens into a second token list; and

determining a second token with a smallest decoding score in the second token list as a decoding result of the second audio frame.

10. The computer device according to claim 6 , wherein each first token includes a pair of states for characterizing a correspondence between a first state of the first token in the first decoding network and a second state of the first token in the second decoding network.

11. A non-transitory computer-readable storage medium, storing computer-readable instructions that, when executed by one or more processors of a computer device, cause the computer device to perform a method for speech decoding including:

obtaining audio data corresponding to a speech, the audio data including a first audio frame and a second audio frame;

decoding the first audio frame using a first decoding network corresponding to a low-order language model and a second decoding network corresponding to a differential language model to obtain a plurality of first tokens, each first token having a corresponding decoding score according to the first and second decoding network;

obtaining a decoding score corresponding to a target token of the plurality of first tokens having a smallest decoding score;

determining a first pruning parameter according to a preset value and the decoding score corresponding to the target token;

traversing non-empty input edges in the first decoding network by using a first state of the target token as a start state to obtain a second pruning parameter;

rescoring a first decoding path formed by the non-empty input edges in the second decoding network by using a second state of the target token as a start state to obtain a third pruning parameter; and

decoding the second audio frame using the first decoding network and the second decoding network according to the first token list and the first pruning parameter, the second parameter, and the third parameter.

12. The non-transitory computer-readable storage medium according to claim 11 , wherein the traversing non-empty input edges in the first decoding network by using a first state of the target token as a start state to obtain a second pruning parameter comprises:

obtaining a first acoustic score of the first decoding path under an acoustic model; and

obtaining the second pruning parameter according to the first acoustic score, the preset value, and the decoding score corresponding to the target token.

13. The non-transitory computer-readable storage medium according to claim 11 , wherein the rescoring a first decoding path formed by the non-empty input edges in the second decoding network by using a second state of the target token as a start state to obtain a third pruning parameter comprises:

obtaining a first acoustic score of the first decoding path under an acoustic model;

obtaining a first path score of the first decoding path in the first decoding network;

obtaining a first rescoring score of the first decoding path in the second decoding network; and

obtaining the third pruning parameter according to the first acoustic score, the first path score, the first rescoring score, the preset value, and the decoding score corresponding to the target token.

14. The non-transitory computer-readable storage medium according to claim 11 , wherein each first token includes a pair of states for characterizing a correspondence between a first state of the first token in the first decoding network and a second state of the first token in the second decoding network.

Priority Claims (1)

Continuity (3)

References Cited (30)