Artificial intelligence apparatus and method for recognizing speech by correcting misrecognized word

1. An artificial intelligence apparatus for recognizing speech by correcting misrecognized word, comprising:

a microphone; and

a processor configured to:

obtain, via the microphone, speech data including speech of a user;

convert the speech data into text by using an acoustic model and a language model;

determine whether an uncertain recognition exists in an acoustic recognition result according to the acoustic model;

based on a determination that the uncertain recognition exists in the acoustic recognition result, determine whether the converted text is a normal sentence by:

generating intention information from the converted text by using a natural language processing model,

generating a dropout intention information set from the converted text by applying a dropout technique to the natural language processing model, and

determining a ratio of dropout intention information that is the same as the generated intention information among pieces of dropout intention information included in the dropout intention information set, wherein the converted text is determined to be a normal sentence based on the determined ratio,

determine a sentence most similar to the converted text among sentences pre-trained by using the language model based on a determination that the converted text is not a normal sentence, and

replace the converted text with the determined most similar sentence based on the determination that the converted text is not the normal sentence; and

generate a speech recognition result corresponding to the speech data by using the converted text based on at least one of a determination that the uncertain recognition does not exist in the acoustic recognition result, a determination that the converted text is a normal sentence, or a determination that the determined ratio is greater than a second reference value.

2. The artificial intelligence apparatus according to claim 1 , wherein the processor is further configured to:

determine probabilities corresponding to each phoneme for each predetermined window unit with respect to the speech data by using the acoustic model;

determine a word recognition reliability for each word included in the speech data by using at least one of a largest probability value (p1) among the determined probabilities, a difference (p1−p2) between the largest probability value (p1) and a second largest probability value (p2) among the determined probabilities, or an entropy corresponding to the determined probabilities; and

determine whether an uncertain recognition exists in the acoustic recognition result based on the determined word recognition reliability.

3. The artificial intelligence apparatus according to claim 2 , wherein the processor is further configured to:

determine an average of p1 or an average of p1−p2 corresponding to phonemes included in a word for each word; and

determine the determined average as the word recognition reliability of the corresponding word.

4. The artificial intelligence apparatus according to claim 2 , wherein the processor is further configured to distinguish words included in the speech data from each other based on a blank or a silence.

5. The artificial intelligence apparatus according to claim 2 , wherein the processor is further configured to determine, as an uncertainly recognized word, a word whose determined word recognition reliability is smaller than a first reference value among the words.

6. The artificial intelligence apparatus according to claim 1 , wherein the processor is further configured to determine that the converted text is a normal sentence based on the determined ratio being greater than the second reference value.

7. The artificial intelligence apparatus according to claim 1 , wherein the processor is further configured to:

divide the converted text into a normal recognition region and a misrecognition region based on a word recognition reliability determined for each word based on a determination that the converted text is not a normal sentence;

select sentences similar to the normal recognition region among sentences pre-trained by using the language model;

determine a sentence most similar to the misrecognition region among the selected sentences; and

replace the converted text with the determined sentence.

8. The artificial intelligence apparatus according to claim 7 , wherein the processor is further configured to determine, as a word included in the misrecognition region, words whose determined word recognition reliability is smaller than a third reference value among the words.

9. The artificial intelligence apparatus according to claim 7 , wherein the processor is further configured to use the language model to determine a sentence most similar to the converted text among sentences related to a function supported by the artificial intelligence apparatus or sentences included in a speech log based on a determination that the converted text is not a normal sentence.

10. The artificial intelligence apparatus according to claim 1 , wherein at least one of the acoustic model, the language model, or the natural language processing model is configured to include an artificial neural network and is trained using a machine learning algorithm or a deep learning algorithm.

11. A method for recognizing speech by correcting misrecognized word, comprising:

obtaining, via a microphone, speech data including speech of a user;

converting the speech data into text by using an acoustic model and a language model;

determining whether an uncertain recognition exists in an acoustic recognition result according to the acoustic model;

based on a determination that the uncertain recognition exists in the acoustic recognition result, determining whether the converted text is a normal sentence by:

generating intention information from the converted text by using a natural language processing model,

generating a dropout intention information set from the converted text by applying a dropout technique to the natural language processing model, and

determining a ratio of dropout intention information that is the same as the generated intention information among pieces of dropout intention information included in the dropout intention information set, wherein the converted text is determined to be a normal sentence based on the determined ratio,

determining a sentence most similar to the converted text among sentences pre-trained by using the language model based on a determination that the converted text is not a normal sentence, and

replacing the converted text with the determined most similar sentence based on the determination that the converted text is not the normal sentence; and

generating a speech recognition result corresponding to the speech data by using the converted text based on at least one of a determination that the uncertain recognition does not exist in the acoustic recognition result, a determination that the converted text is a normal sentence, or a determination that the determined ratio is greater than a second reference value.

12. A non-transitory recording medium storing a program for a processor to perform a method for recognizing speech by correcting misrecognized word, the method comprising:

obtaining, via a microphone, speech data including speech of a user;

converting the speech data into text by using an acoustic model and a language model;

determining whether an uncertain recognition exists in an acoustic recognition result according to the acoustic model;

based on a determination that the uncertain recognition exists in the acoustic recognition result, determining whether the converted text is a normal sentence by:

generating intention information from the converted text by using a natural language processing model,

generating a dropout intention information set from the converted text by applying a dropout technique to the natural language processing model, and

determining a ratio of dropout intention information that is the same as the generated intention information among pieces of dropout intention information included in the dropout intention information set, wherein the converted text is determined to be a normal sentence based on the determined ratio,

determining a sentence most similar to the converted text among sentences pre-trained by using the language model based on a determination that the converted text is not a normal sentence, and

replacing the converted text with the determined most similar sentence based on the determination that the converted text is not the normal sentence; and

generating a speech recognition result corresponding to the speech data by using the converted text based on at least one of a determination that the uncertain recognition does not exist in the acoustic recognition result, a determination that the converted text is a normal sentence, or a determination that the determined ratio is greater than a second reference value.