Back to results
US 12,738,263 B2Granted
Method and system for constructing speech recognition model and speech processing
AssigneeHuawei Technologies Co., Ltd.
InventorsShi Yin, Tao Huang, Yibin Zhai, Zhaohui Wu
Life of this patent
Estimating the term…
Abstract
A method for constructing a speech recognition model includes obtaining a target keyword; determining a synonym group semantically associated with the target keyword; training a language model based on the target keyword and the synonym group; to obtain a target language model; generating a first decoding graph based on the target language model, where the first decoding graph indicates a plurality of decoding paths that satisfy a syntax constraint rule that is based on the target keyword and the synonym group; and determining the speech recognition model based on the first decoding graph.
Claims20
A method for constructing a speech recognition model, wherein the method comprises:
- obtaining a target keyword, wherein the target keyword comprises a first keyword and a second keyword;
- obtaining a synonym group semantically associated with the target keyword, wherein the synonym group comprises a first synonym group semantically associated with the first keyword and a second synonym group semantically associated with the second keyword;
- training, based on the target keyword and the synonym group, a language model to obtain a target language model;
- generating, based on the target language model, a first decoding graph indicating a plurality of decoding paths that satisfy a syntax constraint rule that is based on the target keyword and the synonym group;
- and
- determining, based on the first decoding graph, the speech recognition model, wherein determining the speech recognition model comprises:
- obtaining, from the first decoding graph, a first group of decoding paths and a second group of decoding paths, wherein the first group comprises first decoding paths corresponding to the first keyword and the first synonym group, and wherein the second group comprises second decoding paths corresponding to the second keyword and the second synonym group;
- generating, based on the first group, a first subgraph;
- generating, based on the second group, a second subgraph;
- and
- determining, based on the first subgraph and the second subgraph, the speech recognition model.
8 more elements
A method for speech processing comprising:
- receiving a speech input;
- obtaining a target keyword, wherein the target keyword comprises a first keyword and a second keyword;
- obtaining a synonym group semantically associated with the target keyword, wherein the synonym group comprises a first synonym group semantically associated with the first keyword and a second synonym group semantically associated with the second keyword;
- training, based on the target keyword and the synonym group, a language model to obtain a target language model;
- generating, based on the target language model, a first decoding graph indicating a plurality of decoding paths that satisfy a syntax constraint rule that is based on the target keyword and the synonym group;
- determining, based on the first decoding graph, a speech recognition model, wherein determining the speech recognition model comprises:
- obtaining, from the first decoding graph, a first group of decoding paths and a second group of decoding paths, wherein the first group comprises first decoding paths corresponding to the first keyword and the first synonym group, and wherein the second group comprises second decoding paths corresponding to the second keyword and the second synonym group;
- generating, based on the first group, a first subgraph;
- generating, based on the second group, a second subgraph;
- and
- determining, based on the first subgraph and the second subgraph, the speech recognition model;
- and
- determining, using the speech recognition model, text representation associated with the speech input.
10 more elements
An electronic device, comprising:
- a memory configured to store instructions;
- and
- a processor coupled to the memory and configured to execute the instructions to cause the electronic device to:
- obtain a target keyword, wherein the target keyword comprises a first keyword and a second keyword;
- obtain a synonym group semantically associated with the target keyword, wherein the synonym group comprises a first synonym group semantically associated with the first keyword and a second synonym group semantically associated with the second keyword, and wherein obtaining the synonym group comprises:
- determining first semantics of the target keyword;
- and
- determining the synonym group based on the first semantics, wherein a first difference between second semantics of each synonym in the synonym group and the first semantics is less than a difference threshold, and wherein determining the synonym group comprises determining the synonym group further based on a first length of the target keyword, and wherein a second difference between a second length of each synonym in the synonym group and the first length is less than a length threshold;
- train, based on the target keyword and the synonym group, a language model to obtain a target language model;
- generate, based on the target language model, a first decoding graph indicating a plurality of decoding paths that satisfy a syntax constraint rule that is based on the target keyword and the synonym group;
- and
- determine, based on the first decoding graph, a speech recognition model, wherein determining the speech recognition model comprises:
- obtaining, from the first decoding graph, a first group of decoding paths and a second group of decoding paths, wherein the first group comprises first decoding paths corresponding to the first keyword and the first synonym group, and wherein the second group comprises second decoding paths corresponding to the second keyword and the second synonym group;
- generating, based on the first group, a first subgraph;
- generating, based on the second group, a second subgraph;
- and
- determining, based on the first subgraph and the second subgraph, the speech recognition model.
14 more elements
Dependencies are inferred from the claim text.