Method and apparatus for recognizing speech, electronic device and storage medium

1. A method for recognizing a speech, performed by a speech recognition device including a search engine, an acoustic model, a language model and a decoder, the method comprising:

obtaining, through the search engine, a speech to be recognized;

obtaining an acoustic feature of the speech to be recognized by inputting the speech to be recognized into the acoustic model and obtaining a language feature of the speech to be recognized by inputting the speech to be recognized into the language model, wherein the acoustic model is a model trained by Gaussian Mixed Model (GMM)-Hidden Markov Model (HMI) or Deep Neural Network (DNN)-HMM, and the language model is a model trained by an N-Gram (which is a statistic-based language model) or a Neural Network Language Model (NNLM);

inputting the speech to be recognized to a pronunciation difference statistics to generate a differential pronunciation pair corresponding to the speech to be recognized; and

generating, through the decoder, text information of the speech to be recognized based on the differential pronunciation pair, the acoustic feature, and the language feature;

wherein the pronunciation difference statistics is trained by:

obtaining a target sample text under a target scene;

generating a sample recognition result by recognizing the target sample text;

obtaining a first audio corresponding to the target sample text and obtaining a second audio corresponding to the sample recognition result;

obtaining a differential pronunciation pair sample corresponding to a pronunciation difference between the first audio and the second audio being greater than a preset threshold; and

training the pronunciation difference statistics based on the differential pronunciation pair sample.

2. The method of claim 1 , wherein obtaining the target sample text under the target scene comprises:

obtaining a sample text;

determining whether the sample text belongs to a target scene by inputting the sample text to a target scene text classifier;

determining the sample text as the target sample text based on the sample text belonging to the target scene; and

discarding the sample text based on the sample text not belonging to the target scene.

3. The method of claim 2 , wherein, the target scene text classifier is trained by:

obtaining a target scene sample and a non-target scene sample;

obtaining a first word vector representation of the target scene sample and a second word vector representation of the non-target scene sample;

inputting the first word vector representation as a positive sample and the second word vector representation as a negative sample to an initial target scene text classifier to train the initial target scene text classifier.

4. The method of claim 1 , wherein generating the text information of the speech to be recognized based on the differential pronunciation pair sample, the acoustic feature, and the language feature comprises:

inputting the differential pronunciation pair, the acoustic feature, and the language feature to a decoder to generate the text information of the speech to be recognized.

5. An electronic device, comprising:

a search engine, an acoustic model, a language model, and a decoder;

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein, the memory stores instructions executable by the at least one processor, wherein the instructions are executed by the at least one processor, so that

the search engine obtains a speech to be recognized;

the processor obtains an acoustic feature of the speech to be recognized by inputting the speech to be recognized into the acoustic model and obtains a language feature of the speech to be recognized by inputting the speech to be recognized into the language model, wherein the acoustic model is a model trained by Gaussian Mixed Model (GMM)-Hidden Markov Model (HMM) or Deep Neural Network (DNN)-HMM, and the language model is a model trained by an N-Gram (which is a statistic-based language model) or a Neural Network Language Model (NNLM);

the processor inputs the speech to be recognized to a pronunciation difference statistics to generate a differential pronunciation pair corresponding to the speech to be recognized; and

the decoder generates text information of the speech to be recognized based on the differential pronunciation pair, the acoustic feature, and the language feature;

wherein the pronunciation difference statistics is trained by:

obtaining a target sample text under a target scene;

generating a sample recognition result by recognizing the target sample text;

obtaining a first audio corresponding to the target sample text and obtaining a second audio corresponding to the sample recognition result;

obtaining a differential pronunciation pair sample corresponding to a pronunciation difference between the first audio and the second audio being greater than a preset threshold; and

training the pronunciation difference statistics based on the differential pronunciation pair sample.

6. The electronic device of claim 5 , wherein the processor is further configured to:

obtain a sample text;

determine whether the sample text belongs to a target scene by inputting the sample text to a target scene text classifier;

determine the sample text as the target sample text based on the sample text belonging to the target scene; and

discard the sample text based on the sample text not belonging to the target scene.

7. The electronic device of claim 6 , wherein the target scene text classifier is trained by:

obtaining a target scene sample and a non-target scene sample;

obtaining a first word vector representation of the target scene sample and a second word vector representation of the non-target scene sample;

inputting the first word vector representation as a positive sample and the second word vector representation as a negative sample to an initial target scene text classifier to train the initial target scene text classifier.

8. The electronic device of claim 5 , wherein the processor is further configured to:

input the differential pronunciation pair, the acoustic feature, and the language feature to a decoder to generate the text information of the speech to be recognized.

9. A non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are configured to cause a computer to execute a method for recognizing a speech, wherein the computer includes a search engine, an acoustic mode, a language model and a decoder, the method comprising:

obtaining, through the search engine, a speech to be recognized;

obtaining an acoustic feature of the speech to be recognized by inputting the speech to be recognized into the acoustic model and obtaining a language feature of the speech to be recognized by inputting the speech to be recognized into the language model, wherein the acoustic model is a model trained by Gaussian Mixed Model (GMM)-Hidden Markov Model (HMI) or Deep Neural Network (DNN)-HMM, and the language model is a model trained by an N-Gram (which is a statistic-based language model) or a Neural Network Language Model (NNLM);

inputting the speech to be recognized to a pronunciation difference statistics to generate a differential pronunciation pair corresponding to the speech to be recognized; and

generating, through the decoder, text information of the speech to be recognized based on the differential pronunciation pair, the acoustic feature, and the language feature;

wherein the pronunciation difference statistics is trained by:

obtaining a target sample text under a target scene;

generating a sample recognition result by recognizing the target sample text;

obtaining a first audio corresponding to the target sample text and obtaining a second audio corresponding to the sample recognition result;

obtaining a differential pronunciation pair sample corresponding to a pronunciation difference between the first audio and the second audio being greater than a preset threshold; and

training the pronunciation difference statistics based on the differential pronunciation pair sample.

10. The non-transitory computer-readable storage medium of claim 9 , wherein obtaining the target sample text under the target scene comprises:

obtaining a sample text;

determining whether the sample text belongs to a target scene by inputting the sample text to a target scene text classifier;

determining the sample text as the target sample text based on the sample text belonging to the target scene; and

discarding the sample text based on the sample text not belonging to the target scene.

11. The non-transitory computer-readable storage medium of claim 10 , wherein the target scene text classifier is trained by:

obtaining a target scene sample and a non-target scene sample;

obtaining a first word vector representation of the target scene sample and a second word vector representation of the non-target scene sample;

inputting the first word vector representation as a positive sample and the second word vector representation as a negative sample to an initial target scene text classifier to train the initial target scene text classifier.

12. The non-transitory computer-readable storage medium of claim 9 , wherein generating the text information of the speech to be recognized based on the differential pronunciation pair sample, the acoustic feature, and the language feature comprises:

inputting the differential pronunciation pair, the acoustic feature, and the language feature to a decoder to generate the text information of the speech to be recognized.