KR102217621B1 - Apparatus and method of correcting user utterance errors - Google Patents

Abstract

A method of correcting an error in a user's speech according to an embodiment includes the steps of: receiving a pronunciation sequence corresponding to the user's speech including a singer's name and a song name; Determining a plurality of candidates from the correct answer candidate dictionary based on the pronunciation sequence; Selecting one of the plurality of candidates based on the feature list corresponding to the plurality of candidates; And correcting an error in the user's speech based on the selected candidate.

Description

Method and apparatus for correcting errors in user speech {APPARATUS AND METHOD OF CORRECTING USER UTTERANCE ERRORS}

The following embodiments relate to a method and apparatus for correcting an error in user speech.

With the advancement of technology, various artificial intelligence virtual assistant products are being released. For example, a smart speaker is a kind of wireless speaker, a voice command with a built-in virtual assistant that performs user interaction with the help of one or more impressive hot words or voice commands. It can be a device. Virtual assistant products can perform functions such as listening to music and searching for information with voice commands. These virtual assistant products perform various actions based on the user's voice recognition. Therefore, errors occurring during speech recognition can be propagated to the entire system. In smart speaker and/or virtual assistant products, the accuracy of speech recognition is very important.

According to one side, a method of correcting an error in a user's speech includes receiving a pronunciation sequence corresponding to the user's speech including a singer's name and a song name; Determining a plurality of candidates from a dictionary of correct answer candidates based on the pronunciation sequence; Selecting one of the plurality of candidates based on a feature list corresponding to the plurality of candidates; And correcting an error in the user's speech based on the selected candidate.

The receiving of the pronunciation sequence may include receiving a pronunciation sequence corresponding to the user utterance when a request to play music content based on the user utterance fails.

The determining of the plurality of candidates may include calculating similarities between the pronunciation sequence and character string candidates included in the correct answer candidate dictionary; And determining the plurality of candidates based on the similarities.

The calculating of the similarity may include: an edit distance similarity (Damerau Levenshtein Edit distance) calculating the similarities based on the distance between the character string candidates and the character string candidates according to the insertion, deletion, and substitution of the pronunciation sequence; A Jaro-Winkler similarity for calculating the similarities by combining the length weight of the suffix with the replacement cost between the pronunciation sequence and the character string candidates; And calculating the similarity based on at least one of an overlap similarity that calculates the similarity by using an overlap ratio between the pronunciation sequence and the character string candidates.

The determining of the plurality of candidates may include determining whether or not a character string partially matching a pronunciation sequence corresponding to the singer name and the song name exists in the correct answer candidate dictionary based on the similarity; And determining the plurality of candidates based on the partially matched character string.

The determining of the plurality of candidates may include determining whether or not a character string perfectly matching the singer's name and the pronunciation sequence corresponding to the song name exists in the correct answer candidate dictionary based on the similarities; If a character string that completely matches the pronunciation sequence does not exist in the correct answer candidate dictionary, determining whether a character string partially matching the pronunciation sequence corresponding to the singer name and the song name exists in the correct answer candidate dictionary based on the similarity step; And determining the plurality of candidates based on the exact matched string or the partial matched string.

The correct answer candidate dictionary may be generated based on a music database including information on the singer name and song name, and synonym additional information on the singer name and the song name.

The correct answer candidate dictionary may include at least one of some correct answer candidate dictionaries and all correct answer candidate dictionaries according to a preset criterion.

The correct answer candidate dictionary may be filtered by an accumulated click log for a preset period.

The selecting of one of the plurality of candidates may include determining whether or not the plurality of candidates need to be corrected based on the feature list; And determining any one of the candidates based on whether or not the correction is required.

The feature list may include at least one of a total similarity of the singer name and the song name combined, the singer name similarity, and the song name similarity.

The method of correcting an error in the user speech may further include converting the user speech into the pronunciation sequence.

The method of correcting an error in the user speech may further include outputting a result of correcting the error in the user speech.

The method of correcting an error in the user's speech may further include searching for and providing music content corresponding to a result of correcting the error in the user's speech.

1 is a view for explaining a general processing process of a smart speaker.
2 is a flow chart showing a method of correcting an error in a user's speech according to an embodiment.
3 is a block diagram of a speech correction model for correcting an error in user speech according to an embodiment.
4 is a diagram illustrating an operation of a candidate generation model according to an embodiment.
5 is a diagram for describing a process of generating a dictionary for correct answer candidates according to an embodiment.
6 is a view for explaining a method of correcting an error in user speech between a smart speaker and a server according to an embodiment.

Hereinafter, exemplary embodiments will be described in detail with reference to the accompanying drawings. The same reference numerals in each drawing indicate the same members. Various changes may be made to the embodiments described below. The embodiments described below are not intended to be limited to the embodiments, and should be understood to include all changes, equivalents, and substitutes thereto.

The terms used in the examples are used only to describe specific embodiments, and are not intended to limit the embodiments. Singular expressions include plural expressions unless the context clearly indicates otherwise. In the present specification, terms such as "comprise" or "have" are intended to designate the presence of features, numbers, steps, inputs, components, parts, or a combination of one or more other features described in the specification. It is to be understood that the presence or addition of elements or numbers, steps, inputs, components, parts, or combinations thereof does not preclude in advance.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiment belongs. Terms as defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning in the context of the related technology, and should not be interpreted as an ideal or excessively formal meaning unless explicitly defined in this application. Does not.

In addition, in the description with reference to the accompanying drawings, the same reference numerals are assigned to the same components regardless of the reference numerals, and redundant descriptions thereof will be omitted. In describing the embodiments, when it is determined that a detailed description of related known technologies may unnecessarily obscure the subject matter of the embodiments, the detailed description thereof will be omitted.

1 is a diagram for explaining a processing process of a general smart speaker. Referring to FIG. 1, the smart speaker may perform an operation corresponding to an interaction with a user through the processes of speech recognition 110, natural language understanding 120, search 130, and natural language generation 140. At this time, in the process of voice recognition 100, for example, a voice recognition environment including sound from outside such as TV, speech by family members of various ages from children to grandmothers and grandfathers, music, taxi calls, and messages Errors may occur due to various causes, such as support for multiple domains such as transmission, weather, and fortune-telling, and end point detection (EPD) when an operation is terminated while performing speech recognition. An error in the speech recognition 100 process may act as an error in the entire system, thereby reducing user satisfaction.

The above-described problem may also occur when music is played or spoken. For example, when music is reproduced and uttered, errors may frequently occur due to the names of singers and songs uttered in various forms in the process of speech recognition 100. 'Music reproduction speech' can be understood as speech related to music reproduction. Music playback is ignited, for example, may consist of a singer and a song that contains the user utterances such as "Wishing Good Day IU" or "Give me turn was the love of IKON '.

The types of errors that occur in music playback speech can be summarized as shown in [Table 1] below.

The similar pronunciation error type may occur, for example, when'seeya' is incorrectly recognized as a word having similar pronunciations such as'Jia'. The pseudo pronunciation error type may occur when the probability of the language model of the correct answer is lower than the recognition result.

The type of spacing error can occur when the correct answer and the spacing in the recognition result are different, such as'bump pum' and'pump pum'. The spacing error type may occur when the user speaks with a spacing during speech, or when the probability of a language model resulting from the spacing is high.

The English↔Hangul error type may occur when the English uttered by the user is recognized as Korean, or the Korean uttered by the user is recognized as English.

The insertion error type may occur when the utterance end point (EPD) is not accurately predicted or when the same word is repeatedly recognized, such as "run run".

The type of deletion error can mainly occur when the user utters only part of the correct answer, for example, "ABBA I have a dream" or "ABBA the dream".

The type of cognitive error may occur, for example, when the user confuses the singer and the song and utters "Let's be mine" or "Let's be mine".

The type of symbol recognition error occurs when the pronunciation of the singer's name or the symbol included in the song name is various. Symbol recognition error types may occur when '76-71=' is uttered in various forms, such as '76 minus 71 is' and '76 minus 71 is'.

In addition, there may be various types of errors, such as errors related to numbers or errors that occur when pronouncing foreign songs.

In one embodiment, various types of errors occurring in the voice recognition process 110 are corrected through a method described below, thereby improving the accuracy of voice recognition in smart speakers and/or virtual assistant products, while also improving user satisfaction. I can.

2 is a flowchart illustrating a method of correcting an error in a user's speech according to an embodiment. Referring to FIG. 2, an apparatus for correcting an error in a user's speech (hereinafter, a'correction device') receives a pronunciation sequence corresponding to a user's speech including a singer's name and a song name (210). Here, the user utterance may be, for example, a music reproduction utterance related to music reproduction. The'pronunciation sequence' corresponds to a sequence of pronunciation symbols, and may be mapped, for example, in English characters. For example, when a request to play music content based on a user's speech fails, the calibration apparatus may receive a pronunciation sequence corresponding to the user's speech to correct an error in the user's speech. The calibration apparatus according to an embodiment may be, for example, a single server or a cloud server itself, or some modules included in the server or a smart speaker. Alternatively, the calibration device may be distributedly disposed on the server and the smart speaker.

The calibration apparatus determines a plurality of candidates from the correct answer candidate dictionary based on the pronunciation sequence received in step 210 (220). The calibration apparatus may calculate similarities between the pronunciation sequence and the character string candidates included in the correct answer candidate dictionary. The character string candidate(s) corresponds to a sequence of phonetic symbols, and may be mapped to English characters, for example, like a phonetic sequence. Unlike the pronunciation sequence corresponding to the user's utterance, it may be understood that the character string candidate(s) correspond to the text previously registered in the correct answer candidate dictionary. Since both the character string candidate(s) and the pronunciation sequence are mapped in the form of characters, similarity can be compared with each other.

For example, the correction device inserts, deletes, and replaces the phonetic sequence, and the edit distance similarity (Damerau Levenshtein Edit distance) that calculates similarities based on the distances to the character string candidates in the reverse order, and a suffix to the replacement cost between the phonetic sequence and the character string candidates. Similarities can be calculated based on at least one of the Jaro-Winkler similarity, which calculates similarities by combining the length weights of and the overlap similarity, which calculates similarities using the overlapping ratio between the pronunciation sequence and the character string candidates. I can. A method of calculating the similarities by the calibration apparatus will be described in detail with reference to FIG. 3 below.

The calibration apparatus may determine a plurality of candidates based on the similarities. The calibration apparatus may determine whether or not a character string that completely matches a pronunciation sequence corresponding to a singer's name and a song name exists in the correct answer candidate dictionary based on the similarity, for example. If a character string that completely matches the pronunciation sequence does not exist in the correct answer candidate dictionary, the correction apparatus may determine whether or not a character string partially matching the pronunciation sequence corresponding to the singer name and the song name exists in the correct answer candidate dictionary based on similarities. The calibration apparatus may determine a plurality of candidates based on a character string that completely matches the pronunciation sequence or a character string that partially matches the pronunciation sequence. A method of determining a plurality of candidates by the calibration apparatus will be described in detail with reference to FIG. 4 below.

Here, the correct answer candidate dictionary may be generated based on a music database including additional information such as information on singer names and song names, synonyms for singer names and song names, and similar words. The correct answer candidate dictionary may include some correct answer candidate dictionaries and all correct answer candidate dictionaries according to a preset criterion. Some correct answer candidate dictionaries according to a preset criterion may be, for example, a TOP 100 correct answer candidate dictionary 550 as shown in FIG. 5 below. The'TOP 100 Correct Answer Candidate Dictionary' may include, for example, information on singer names for 100 popular music contents in the latest order, information on song names, and additional information on synonyms for singer names and song names. At this time, the TOP 100 correct answer candidate dictionary may be about 100 latest music contents that are popular for each music content field, or about 100 latest music contents that are popular among all music contents. The TOP 100 correct answer candidate dictionary corresponds to an example of some correct answer candidate dictionaries according to preset criteria, and is not limited thereto. In addition, some correct answer candidate dictionaries may be determined according to various criteria.

The calibration apparatus selects one of the plurality of candidates based on the feature list corresponding to the plurality of candidates (230). The calibration apparatus may determine whether a plurality of candidates need to be calibrated based on the feature list. The feature list may include, for example, at least one of an overall similarity of a singer's name and a song name, a singer's name similarity, and a song name similarity. The calibration apparatus may determine any one candidate based on whether or not calibration is necessary.

The correction device corrects an error in user speech based on the selected candidate (240).

The calibration device may output a result of correcting an error in the user's speech in step 240 (hereinafter, a'correction result') (250). Alternatively, according to an embodiment, the calibration apparatus may search and provide music content corresponding to the calibration result.

3 is a block diagram of a speech correction model for correcting errors in user speech according to an exemplary embodiment. Referring to FIG. 3, the speech correction model 300 according to an embodiment may include a candidate generation model 310 and a correction determination model 330. The speech correction model 300 may be included in the calibration device or may be included in an external device coupled with the calibration device.

The speech correction model 300 is, for example, a speech correction model for music reproduction, and may be driven when a reproduction request for music content fails. Here, it may be understood that'the case where a request to play music content has failed' may be understood as a case in which the correction apparatus cannot search for and provide music content to be played due to, for example, a voice misrecognition and a speaker's recognition error. The speech correction model 300 may be driven when it is impossible to search for music content to be played in response to the user speech after analyzing the singer name and the song name through music domain classification and natural language understanding for the user speech.

The candidate generation model 310 may convert the user's speech into a pronunciation sequence, and determine a plurality of candidates from a correct answer candidate dictionary based on the pronunciation sequence. The candidate generation model 310 may calculate similarities between the pronunciation sequence and character string candidates included in the correct answer candidate dictionary, and determine a preset number (eg, three) of candidates in the order of the highest similarity.

According to an embodiment, the candidate generation model 310 may determine a preset number (eg, 5) of candidates exceeding the threshold value by comparing the similarity values with the threshold values for each similarity. The candidate generation model 310 may calculate similarities by, for example, an edit distance similarity, a magnetic path-winkler similarity, and an overlap similarity calculation method.

The edit distance similarity may be calculated based on a distance between various character string candidates and a pronunciation sequence configured according to insertion, deletion, substitution, and reverse order, for example. The edit distance similarity may be calculated for each of a song and an artist as shown in Equation 1 below.

는 발음열과 다양한 문자열 후보(들) 간의 거리에 해당하고,

는 가중치에 해당할 수 있다. Here, x denotes a pronunciation sequence, and y denotes a character string candidate(s) included in the correct answer candidate dictionary.

Is the distance between the pronunciation sequence and the various character string candidate(s),

May correspond to the weight.

The ruler-winkler similarity is a method of calculating the similarity by putting a weight on the prefix of the pronunciation sequence. The ruler-winkler similarity can be calculated by combining the weight of the length of the suffix with the replacement cost between the pronunciation sequence and the character string candidate. For example, the magnetic-winkler similarity may be calculated for each of the song name and singer name, for example, as shown in Equation 2 below.

는 발음열과 문자열 후보 간의 치환 비용에 해당하고,

는 접미사의 길이 가중치에 해당할 수 있다. here,

Corresponds to the replacement cost between the pronunciation sequence and the string candidate,

May correspond to the weight of the length of the suffix.

The overlap similarity can be calculated as in Equation 3 below by using the overlapping ratio (or overlap ratio) between the pronunciation sequence combining the singer's name and the song name and the character string candidate.

Depending on the embodiment, the candidate generation model 310 equally combines various similarity calculation methods between the phonetic sequence and the character string candidates included in the correct answer candidate dictionary, such as edit distance similarity, ruler-winkler similarity, and/or overlap similarity. Thus, the final similarities may be calculated as shown in Equation 4 below.

The operation of the candidate generation model 310 will be described in more detail with reference to FIG. 4 below.

The calibration discrimination model 330 may determine whether to correct candidates output from the candidate generation model 310. The calibration discrimination model 330 may determine whether or not to be corrected by, for example, a binary classification of a probability that a correction is required and a probability that a correction is not required. For the candidates output from the candidate generation model 310, the calibration discrimination model 330 calculates 16 feature information as described in the feature list in [Table 2] below, and uses the machine-learned classifier features. Can be used.

The feature list may be divided into, for example, the overall similarity of the singer name and the song name combined, the singer name similarity, and the song name similarity. Features obtained by applying normalization and reverse to the editing distance similarity and the ruler-winkler similarity may be added to the singer name similarity and the song name similarity. Normalization may correspond to a state in which symbols and spaces are removed from a character string or phonetic sequence, and the reverse order may correspond to a state in which the character string or phonetic sequence is reversed and applied.

The calibration discrimination model 330 may include, for example, a random forest (RF) classifier. The random forest classifier is one of an ensemble model that improves performance by combining several classification models, and may use a decision tree as an individual model.

The correction discrimination model 330 may correct an error in the user's speech by a candidate having the highest probability of not requiring correction according to a result of determining whether or not the candidates output from the candidate generation model 310 are corrected.

4 is a diagram for describing an operation of a candidate generation model according to an exemplary embodiment. Referring to FIG. 4, the structure and operation of a candidate generation model 310 including a pronunciation conversion module 410, an exact match search module 430, and a similar candidate search module 450 are illustrated.

The candidate generation model 310 converts the user's speech including the singer's name and the song name into a pronunciation string by the pronunciation conversion module 410, and performs an exact match search for the pronunciation string through the perfect match search module 430, and is similar. The candidate search module 450 may perform a partial match search for the pronunciation sequence. In this case, the complete match search and the partial match search may be performed by sequentially searching the TOP100 correct answer candidate dictionary and all correct answer candidate dictionaries corresponding to some correct answer candidate dictionaries, respectively.

The exact match search module 430 may determine whether or not a character string identical to the pronunciation sequence converted from the user's speech, that is, exactly matching the pronunciation sequence, exists in the correct answer candidate dictionary. The exact match search module 430 may determine whether or not a character string that completely matches the pronunciation sequence exists in the TOP 100 candidate dictionary 431 and then determine whether it exists in the entire candidate dictionary 433. For example, the exact match search module 430 does not have an error in recognizing the user's speech, but is temporarily misrecognized due to a network problem or other various causes, and is prepared in case the problem is resolved in the future. Can be.

The similar candidate search module 450 may operate when a character string that completely matches the pronunciation sequence does not exist in the correct answer candidate dictionary. The similarity candidate search module 450 may determine whether a character string (or a similar character string) partially matching the pronunciation sequence exists in the correct answer candidate dictionary. The similar candidate search module 450 may determine whether a character string partially matching the pronunciation sequence, that is, whether a similar candidate exists in the TOP 100 candidate dictionary 451, and then determine whether it exists in the entire candidate dictionary 453.

Both the exact match search module 430 and the similar candidate search module 450 may calculate a degree of similarity through Equation 4 described above, and output as many candidates as a preset maximum number (max N).

5 is a diagram illustrating a process of generating a dictionary for correct answer candidates according to an embodiment. Referring to FIG. 5, a process of generating a correct answer candidate dictionary 570 including a TOP 100 correct answer candidate dictionary 550 and an all correct answer candidate dictionary 560 is illustrated.

The correct answer candidate dictionary 570 may be generated by processing data included in the music DB 510. The music DB 510 may include, for example, singer information, song information, and additional information such as synonyms and similar words corresponding to singer information and song information.

The correction apparatus according to an embodiment may use the TOP 100 correct answer candidate dictionary 550 and the entire correct answer candidate dictionary 560 to compare the similarity between the pronunciation sequence corresponding to the user's speech and the character string candidates included in the correct answer candidate dictionary.

For example, in order to reduce the search range of the correct answer candidate dictionary 570, the correction device filters some data by filtering the music DB 510 with an accumulated click log for a preset period (6 months). Can be extracted (520).

The calibration apparatus may separately construct a TOP 100 correct answer candidate dictionary 550 to respond to a new song due to the nature of the music domain, for example. The calibration apparatus may select, for example, music content that is released within 3 months and falls within the 100th ranking of the overall popularity as TOP 100 by using additional information on some data extracted in 520 (530). The calibration apparatus may acquire character string candidates corresponding to the TOP 100 by inputting the TOP 100 to the pronunciation conversion module 540. The calibration apparatus may construct a TOP 100 correct answer candidate dictionary 550 using character string candidates corresponding to TOP 100. \

The correction apparatus may construct the entire correct answer candidate dictionary 560 by inputting some data extracted in 520 to the pronunciation conversion module 540.

The pronunciation conversion module 540 may be included in the calibration device or may be external to the calibration device.

The correct answer candidate dictionary 570 may be stored in a maximum heap structure in order to store as many candidates as a tri shape and a maximum number of candidates in which the edit distance is considered for searching for similar candidates.

6 is a diagram illustrating a method of correcting an error in a user's speech between a smart speaker and a server according to an exemplary embodiment. 6, a smart speaker 610 and a server 630 are shown.

For example, let's Smart Speaker (610) that receives a user utterance including a singer and song like "IU Give me a good turn." The smart speaker 610 may convert the user's speech into a corresponding pronunciation string, and then transmit the pronunciation string to the server 630. The server 630 may determine a plurality of candidates from the correct answer candidate dictionary based on the pronunciation sequence as described above, and select any one of the plurality of candidates based on a feature list corresponding to the plurality of candidates. The server 630 may correct an error in the user's speech based on the selected candidate and may directly output the correction result, or may search for music content corresponding to the correction result and provide it to the smart speaker 610. Smart Speaker 610 may provide the music content received from the server (630), with comments such as "I'll play IU good day" to the user.

The operations described with reference to FIG. 6 may be performed by the smart speaker 610 or the server 630, and as described through FIG. 6, some operations are performed by the smart speaker 610, and some other operations. May be performed in the server 630.

Although not shown in the drawings, the calibration apparatus according to an embodiment includes a communication interface and a processor. The calibration device may further include a memory, a speaker, and a display. The communication interface, processor, memory, speaker, and display may be connected to each other through a communication bus. The calibration device may be, for example, a smart phone, a wearable device, a smart speaker, or a user device or a server that performs the same or similar function.

The processor may perform at least one method or an algorithm corresponding to at least one method described above through FIGS. 1 to 6. The processor can run programs and control the calibration device. Program codes executed by the processor may be stored in a memory.

The embodiments described above may be implemented as a hardware component, a software component, and/or a combination of a hardware component and a software component. The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The above-described hardware display device may be configured to operate as one or more software modules to perform an input of an embodiment, and vice versa.

As described above, although the embodiments have been described by the limited drawings, various modifications and variations are possible from the above description by those of ordinary skill in the art. For example, the described techniques are performed in an order different from the described method, and/or components such as the described system, structure, display device, circuit, etc. are combined or combined in a form different from the described method, or a different configuration Appropriate results can be achieved even if substituted or substituted by elements or equivalents. Therefore, other implementations, other embodiments, and claims and equivalents fall within the scope of the claims to be described later.

Claims (15)

Receiving a pronunciation sequence corresponding to a user utterance including a singer's name and a song name;
Determining a plurality of candidates from a dictionary of correct answer candidates based on the pronunciation sequence;
Determining whether the plurality of candidates need to be corrected based on a feature list corresponding to the plurality of candidates, the feature list including feature information of the plurality of candidates calculated by a machine-learned classifier;
Selecting one of the plurality of candidates with the highest probability that the correction is not required; And
Correcting an error in the user's speech based on the selected candidate
Including, a method for correcting an error in user speech. The method of claim 1,
The step of receiving the pronunciation sequence
When a request to play music content based on the user utterance fails, receiving a pronunciation sequence corresponding to the user utterance
Including, a method for correcting an error in user speech. The method of claim 1,
The step of determining the plurality of candidates
Calculating similarities between the pronunciation sequence and character string candidates included in the correct answer candidate dictionary; And
Determining the plurality of candidates based on the similarities
Including, a method for correcting an error in user speech. The method of claim 3,
The step of calculating the similarities is
An edit distance similarity (Damerau Levenshtein Edit distance) for calculating the similarities based on the distance between the character string candidates and the character string candidates according to insertion, deletion, substitution, and reciprocal of the pronunciation sequence;
A Jaro-Winkler similarity for calculating the similarities by combining the length weight of the suffix with the replacement cost between the pronunciation sequence and the character string candidates; And
Overlap similarity for calculating the similarities by using an overlap ratio between the pronunciation sequence and the character string candidates
Calculating the similarities based on at least one of
Including, a method for correcting an error in user speech. The method of claim 3,
The step of determining the plurality of candidates
Determining whether a character string partially matching the singer's name and the pronunciation sequence corresponding to the song name exists in the correct answer candidate dictionary based on the similarities; And
Determining the plurality of candidates based on the partially matched character string
Including, a method for correcting an error in user speech. The method of claim 3,
The step of determining the plurality of candidates
Determining whether or not a character string that completely matches the singer's name and the pronunciation sequence corresponding to the song name exists in the correct answer candidate dictionary based on the similarities;
When a character string that completely matches the pronunciation sequence does not exist in the correct answer candidate dictionary, determining whether a character string partially matching the pronunciation sequence corresponding to the singer name and the song name exists in the correct answer candidate dictionary based on the similarity step; And
Determining the plurality of candidates based on the exact matched string or the partial matched string
Including, a method for correcting an error in user speech. The method of claim 1,
The candidate dictionary for the correct answer above is
A method for correcting an error in a user's speech, which is generated based on a music database including information on the singer name and song name, and synonym additional information on the singer name and the song name. The method of claim 1,
The candidate dictionary for the correct answer above is
A method for correcting an error in a user's speech, including at least one of some correct answer candidate dictionaries and all correct answer candidate dictionaries according to a preset criterion. The method of claim 1,
The candidate dictionary for the correct answer above is
A method for correcting errors in user speech, which is filtered by a cumulative click log for a preset period of time. delete The method of claim 1,
The list of qualities above is
A method for correcting an error in a user's speech, comprising at least one of a total similarity of the singer name and the song name, a singer name similarity, and a song name similarity. The method of claim 1,
Converting the user speech into the pronunciation sequence
A method for correcting an error in user speech further comprising a. The method of claim 1,
Outputting a result of correcting the error of the user's speech
A method for correcting an error in user speech further comprising a. The method of claim 1,
Searching for and providing music content corresponding to the result of correcting the error in the user's speech
A method for correcting an error in user speech further comprising a. A computer program stored in a computer-readable recording medium for executing the method of any one of claims 1 to 9 and 11 to 14 combined with hardware.

Images