(19)
(11)EP 3 779 970 A1

(12)EUROPEAN PATENT APPLICATION
published in accordance with Art. 153(4) EPC

(43)Date of publication:
17.02.2021 Bulletin 2021/07

(21)Application number: 18879061.2

(22)Date of filing:  09.11.2018
(51)International Patent Classification (IPC): 
G10L 15/26(2006.01)
(86)International application number:
PCT/CN2018/114792
(87)International publication number:
WO 2019/096068 (23.05.2019 Gazette  2019/21)
(84)Designated Contracting States:
AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR

(30)Priority: 14.11.2017 CN 201711121071

(71)Applicant: NIO (Anhui) Holding Co., Ltd.
Hefei City, Anhui Province (CN)

(72)Inventors:
  • GE, Sihan
    Shanghai (CN)
  • MA, Tianze
    Shanghai (CN)
  • LIN, Feng
    Shanghai (CN)
  • DUAN, Quansheng
    Shanghai (CN)
  • ZHAO, Haotian
    Shanghai (CN)
  • XIA, Yan
    Shanghai (CN)
  • LIANG, Wei
    Shanghai (CN)
  • ZHOU, Ji
    Shanghai (CN)
  • XU, Longsheng
    Shanghai (CN)
  • MA, Yingcai
    Shanghai (CN)
  • YIN, Lutong
    Shanghai (CN)

(74)Representative: Vossius & Partner Patentanwälte Rechtsanwälte mbB 
Siebertstrasse 3
81675 München
81675 München (DE)

  


(54)VOICE RECOGNITION AND ERROR CORRECTION METHOD AND VOICE RECOGNITION AND ERROR CORRECTION SYSTEM


(57) Provided are a method and system for error correction in speech recognition. The method comprises: extracting a place name by means of speech recognition and semantic understanding, and converting the place name into place name pinyin (S100); splitting the place name pinyin to obtain N pinyin lists, and performing a first Cartesian product operation on the N pinyin lists to generate a first pinyin candidate set, where N is a natural number (S200); filtering the pinyin in the first pinyin candidate set to eliminate irregular pinyin (S300); performing a second Cartesian product operation on the pinyin output in the filtering step to obtain a second pinyin candidate set (S400); constructing a query based on the second pinyin candidate set (S500); and querying the query in an inverted index to obtain candidate place names with high to low probabilities and display the candidate place names to a user (S600).




Description

Technical Field



[0001] The invention relates to speech recognition technologies, and particularly to a method and system for error correction in speech recognition.

Background Art



[0002] Speech recognition is the process of recognizing and converting a sound into a piece of text. Chinese speech recognition is the conversion of speech into Chinese text according to a speaker's pronunciation. However, due to the differences in personal pronunciations, non-standard Chinese pinyin may lead to a great reduction in the accuracy of speech recognition.

[0003] When speech recognition is applied to a vehicle navigation scenario, a speech dialogue system will first convert the speech into text, then perform semantic understanding, recognize a place name, query the specific information of the place name, and give navigation. However, due to great differences in speakers' pronunciations, a large number of place names, and homophones, etc., the place names obtained through speech recognition may sometimes have errors, resulting in a user being unable to find a correct place. For example, three types of recognition errors are shown in Table 1 below:
Table 1
Homophone "Navigate to Jiatinghui (

)" → "Navigate to Jiatinghui (

)"
Alveolar and velar nasals "Navigate to Xingguangjiayuan (

)" → "Navigate to Xinguangjiayuan (

)"
Flat tongue and retroflex "Navigate to Zijingyicun (

)" → "Navigate to Zhijinyicun (

)"

Summary of the Invention



[0004] In view of the above problems, the invention is intended to provide a system and method for error correction in speech recognition of a place name, which can improve the accuracy of speech recognition.

[0005] A method for error correction in speech recognition in an aspect of the invention is characterized by comprising the following steps:

a pinyin conversion step of converting speech information into pinyin after extraction by means of speech recognition and semantic understanding;

a first pinyin candidate set generation step of splitting the pinyin to obtain N pinyin lists, and performing a first Cartesian product operation on the N pinyin lists to generate a first pinyin candidate set, where N is a natural number;

a filtering step of filtering the pinyin in the first pinyin candidate set to eliminate irregular pinyin;

a second pinyin candidate set generation step of performing a second Cartesian product operation on the pinyin output in the filtering step to obtain a second pinyin candidate set;

a query generation step of constructing a query based on the second pinyin candidate set; and

a candidate place name generation step of querying the query in an inverted index to obtain candidate place names with high to low probabilities and display the candidate place names to a user.



[0006] Optionally, in the first pinyin candidate set generation step, the pinyin is split into initials and finals, a list of similar initials and a list of similar finals are found according to a similarity dictionary, and a Cartesian product operation is performed on the list of initials and the list of finals to generate a candidate set for the pinyin.

[0007] Optionally, in the first pinyin candidate set generation step, the list of similar initials for the initials is:

the list of similar finals for the finals is:

then, the first Cartesian product operation is performed on

and

to generate similar pinyin for the pinyin to serve as the first pinyin candidate set, namely:



[0008] Optionally, in the second pinyin candidate set generation step, the second Cartesian product operation is performed on a candidate set for each piece of pinyin to obtain a correct pinyin candidate set: C = PC1 × PC2 ×... × PCn.

[0009] Optionally, in the filtering step, the pinyin in the first pinyin candidate set is filtered according to a pinyin dictionary.

[0010] Optionally, in the query generation step, the query is constructed based on the second pinyin candidate set according to the following formula:

where should represents an OR operation, match_phrase means that a document must contain a character string in a query statement, and weight is the weight of a query clause, the weight being calculated based on the distance between candidate pinyin and original pinyin.

[0011] Optionally, the weight is calculated as shown in the following formula:

where pi is the pinyin of the ith character in the original pinyin, qi is the pinyin of the ith character in the candidate pinyin, d(pi,qi) is the distance between pi and qi, and the value of d(pi,qi) is the number of times that initials and finals are replaced for turning pi into qi.

[0012] A system for error correction in speech recognition in an aspect of the invention is characterized by comprising:

a pinyin conversion module for acquiring speech information and converting the speech information into pinyin by means of speech recognition and semantic understanding;

a first pinyin candidate set generation module for splitting the pinyin to obtain N pinyin lists, and performing a first Cartesian product operation on the N pinyin lists to generate a first pinyin candidate set, where N is a natural number;

a filtering module for filtering the pinyin in the first pinyin candidate set to eliminate irregular pinyin;

a second pinyin candidate set generation module for performing a second Cartesian product operation on the pinyin output by the filtering module to obtain a second pinyin candidate set; and

a query generation module for constructing a query based on the second pinyin candidate set; and

a candidate generation module for querying the query in an inverted index to obtain candidate information with high to low probabilities and display the candidate information to a user.



[0013] Optionally, the first pinyin candidate set generation module is configured to split the pinyin into initials and finals, find a list of similar initials and a list of similar finals according to a similarity dictionary, and perform a Cartesian product operation on the list of initials and the list of finals to generate a candidate set for the pinyin,
wherein the list of similar initials for the initials is:

the list of similar finals for the finals is:

the first Cartesian product operation is further performed on

and

to generate similar pinyin for the pinyin to serve as the first pinyin candidate set, namely:



[0014] Optionally, the second pinyin candidate set generation module is configured to perform the second Cartesian product operation on a candidate set for each piece of pinyin to generate a correct pinyin candidate set: C = PC1 × PC2 × ... × PCn.

[0015] Optionally, the query generation module is configured to construct the query based on the second pinyin candidate set according to the following formula:

where should represents an OR operation, match_phrase means that a document must contain a character string in a query statement, and weight is the weight of a query clause, the weight being calculated based on the distance between candidate pinyin and original pinyin,
and the weight being calculated as shown in the following formula:

where pi is the pinyin of the ith character in the original pinyin, qi is the pinyin of the ith character in the candidate pinyin, d(pi,qi) is the distance between pi and qi, and the value of d(pi,qi) is the number of times that initials and finals are replaced for turning pi into qi.

[0016] A computer-readable medium of the invention, which has a computer program stored thereon, is characterized in that when the computer program is executed by a processor, a method for error correction in speech recognition as described is implemented.

[0017] A computer device of the invention, which comprises a memory, a processor, and a computer program stored on the memory and executable on the processor, is characterized in that when the processor executes the computer program, the steps of a method for error correction in speech recognition as described are implemented.

[0018] According to the method and system for error correction in speech recognition of the invention, it is possible to filter out wrong pinyin, thereby improving the accuracy of recognition. In addition, more similar pinyin can be calculated using the first pinyin candidate set generation module and the second pinyin candidate set generation module, and the candidate place names with high to low probabilities can be arranged using the query generation module and the candidate place name generation module, so that it is convenient for the user to select and the accuracy of recognition is also improved.

Brief Description of the Drawings



[0019] 

FIG. 1 is a flowchart of a method for error correction in speech recognition in an embodiment of the invention.

FIG. 2 is an example representing a similarity dictionary.

FIG. 3 is an example representing a pinyin dictionary.

FIG. 4 is a schematic diagram of the construction of a system for error correction in speech recognition in an embodiment of the invention.


Detailed Description of Embodiments



[0020] The following introduces some of multiple embodiments of the invention and is intended to provide a basic understanding of the invention. It is not intended to confirm the key or decisive elements of the invention or limit the scope of protection.

[0021] In the following description, the method and system for error correction in speech recognition of the invention will be described by taking the application thereof to a scenario for error correction in place name recognition as an example. Certainly, the method and system for error correction in speech recognition of the invention can be applicable not only to error correction in speech recognition of a place name, but also to various other scenarios, such as online search guidance, and shopping guidance.

[0022] FIG. 1 is a flowchart of a method for error correction in speech recognition in an embodiment of the invention.

[0023] As shown in FIG. 1, the method for error correction in speech recognition in an embodiment of the invention comprises the following steps:

a pinyin conversion step S100: extracting a place name by means of speech recognition and semantic understanding, and converting the place name into place name pinyin, in which case specifically, for example, "

(Shanghai Bund)" is converted into pinyin to obtain a pinyin string: "shang hai tan";

a first pinyin candidate set generation step S200: splitting the place name pinyin to obtain N pinyin lists, and performing a first Cartesian product operation on the N pinyin lists to generate a first pinyin candidate set, where N is a natural number, in which case specifically, for example, the pinyin string "shang hai tan" obtained through conversion in the pinyin conversion step S100 contains pinyin of three characters, here, a candidate set is obtained for each of the three pieces of pinyin, because "sh" and "s" are similar and "ang" and "an" are similar (obtained from a similarity dictionary), the candidates corresponding to "shang" are "sang", "shan", and "san", as such, no other candidates for "hai" can be calculated, and "tan" has a candidate "tang", such that a list of candidates for each piece of pinyin is generated;

a filtering step S300: filtering the pinyin in the first pinyin candidate set to eliminate irregular pinyin; a second pinyin candidate set generation step S400: performing a second Cartesian product operation on the pinyin output in the filtering step to obtain a second pinyin candidate set, in which case specifically, for example, the candidates for the entire pinyin string "shang hai tan" are: sang hai tan, shan hai tan, san hai tan, shang hai tang, sang hai tang, shan hai tang, san hai tang;

a query generation step S500: constructing a query (namely, "query statement") based on the second pinyin candidate set; and

a candidate place name generation step S600: querying the query in an inverted index to obtain candidate place names with high to low probabilities and display the candidate place names to a user.



[0024] Next, the above steps will be specifically described.

[0025] In the pinyin conversion step S100, the place name is extracted by means of speech recognition and semantic understanding and converted into the place name pinyin. For example, a document D = (name, pinyin) is constructed, which has two attributes: a place name and pinyin corresponding to the place name. Examples of documents are shown in Table 2 below:
Table 2
D1 : (

iia ting hui)
D2 : (

xing guang iia vuan)
D3 : (

zi iing vi cun)


[0026] An inverted index is constructed based on pinyin. This process can be performed by means of an existing resource search engine, such as: elastic search, which constructs documents for all the place names in the above manner and indexes the documents.

[0027] Next, in the first pinyin candidate set generation step S200, after the place name is extracted from a semantic understanding result of a speech dialogue system, the place name is converted into pinyin corg. Then, the place name pinyin is split to obtain a list of pinyin: P = [p1,p2, ..., pn].

[0028] Each piece of pinyin pk is split into the initials

and the finals

According to the initials and the finals, from the similarity dictionary, a list of similar initials for the initials is found:

and
a list of similar finals for the finals is found:



[0029] The similarity dictionary may be stored in a disk or a database in the form of a file, for example, and the system will read the file and load the similarity dictionary. Each record in the dictionary is a key-value pair, the key is an initial (or final), and the value is a similar initial (or final). Reference is made to FIG. 2 for an example of the similarity dictionary. There are common similar pronunciations in the dictionary, such as: "z" and "zh", "c" and "ch", etc.

[0030] Then, the first Cartesian product operation is performed on

and

to generate similar pinyin for the pinyin, that is, generating a candidate set:





[0031] Since not all the pinyin in the generated candidate set is regular pinyin, the irregular pinyin needs to be filtered out. Therefore, in the filtering step S300, the pinyin in the generated candidate set is filtered to eliminate the irregular pinyin. This filtering step can be performed through a pinyin dictionary. The pinyin dictionary may be stored in a disk or a database in the form of a file, for example, and the system will read the file and load the pinyin dictionary. The dictionary contains all the regular Chinese pinyin, and each record in the dictionary is one piece of regular Chinese pinyin. Reference is made to FIG. 3 for an example of the pinyin dictionary.

[0032] Then, in the second pinyin candidate set generation step S400, the second Cartesian product operation is performed on a candidate set for each piece of pinyin to obtain a correct place name pinyin candidate set: C = PC1 × PC2 × ... × PCn.

[0033] Finally, in the query generation step S500, for the place name pinyin candidate set calculated in the above step, the query is constructed according to the candidate set to query a corresponding document in the inverted index. During the construction of the query, an OR operation is performed on each piece of candidate pinyin, and each piece of candidate pinyin has a different weight. The candidate pinyin that is closer to the original place name pinyin is more likely to be the correct place name pinyin, and thus has a higher weight. Here, an example of constructing a query in this embodiment is as follows:

where should represents an OR operation, and match_phrase means that a document must contain a character string in a query statement, and here, means that a pinyin field in the document must contain a pinyin field in a query clause. weight is the weight of the query clause, and the weight is calculated based on the distance between candidate pinyin and original place name pinyin. The distance is calculated using the following formula:

where pi is the pinyin of the ith character in the original pinyin, qi is the pinyin of the ith character in the candidate pinyin, d(pi,qi) is the distance between pi and qi, and the value of d(pi,qi) is the number of times that initials and finals are replaced for turning pi into qi.

[0034] An example of the process of implementing the above formula using computer codes will be distance(corg,c) described next. The specific code implementation process of the algorithm is as follows:

Input: corg - the original place name pinyin, c - the pinyin in a candidate set

Output: the distance between the original pinyin and the pinyin in the candidate set

  1. 1: Split corg to pinyin list P = [p1,p2, ..., pn]
  2. 2: Split c to pinyin list Q = [p1,q2,.... qn]
  3. 3: For i in range (0, n-1)
  4. 4: t → distance between pi and qi
  5. 5: distance → distance + t
  6. 6: End for
  7. 7: Return distance.



[0035] Finally, in the candidate place name generation step S600, the query can be queried in the inverted index to obtain a list of candidate place names with high to low probabilities. After the list of candidate place names are displayed, the user can select the correct place name so that error correction is implemented.

[0036] The method for error correction in speech recognition of a place name in an embodiment of the invention has been specifically described above. Next, a system for error correction in speech recognition of a place name in an embodiment of the invention will be described.

[0037] FIG. 4 is a schematic diagram of the construction of a system for error correction in speech recognition of a place name in an embodiment of the invention.

[0038] As shown in FIG. 4, the system for error correction in speech recognition of a place name in an embodiment of the invention has:

a pinyin conversion module 100 for acquiring speech information and extracting a place name from the speech information by means of speech recognition and semantic understanding, and converting the place name into place name pinyin, in which case specifically, for example,

is converted into pinyin to obtain a pinyin string: "shang hai tan";

a first pinyin candidate set generation module 200 for splitting the place name pinyin to obtain N pinyin lists, and performing a first Cartesian product operation on the N pinyin lists to generate a first pinyin candidate set, where N is a natural number, in which case specifically, for example, the pinyin string "shang hai tan" obtained through conversion in the pinyin conversion step S100 contains pinyin of three characters, here, a candidate set is obtained for each of the three pieces of pinyin, because "sh" and "s" are similar and "ang" and "an" are similar (obtained from a similarity dictionary), the candidates corresponding to "shang" are "sang", "shan", and "san", as such, no other candidates for "hai" can be calculated, and "tan" has a candidate "tang", such that a list of candidates for each piece of pinyin is generated;

a filtering module 300 for filtering the pinyin in the first pinyin candidate set to eliminate irregular pinyin;

a second pinyin candidate set generation module 400 for performing a second Cartesian product operation on the pinyin output by the filtering module to obtain a second pinyin candidate set, in which case specifically, for example, the candidates for the entire pinyin string "shang hai tan" are: sang hai tan, shan hai tan, san hai tan, shang hai tang, sang hai tang, shan hai tang, san hai tang; and

a query generation module 500 for constructing a query based on the second pinyin candidate set; and

a candidate place name generation module 600 for querying the query in an inverted index to obtain candidate place names with high to low probabilities and display the candidate place names to a user.



[0039] The first pinyin candidate set generation module 200 is configured to split the place name pinyin into initials and finals, find a list of similar initials and a list of similar finals according to a similarity dictionary, and perform a Cartesian product operation on the list of initials and the list of finals to generate a candidate set for the pinyin,
wherein the list of similar initials for the initials is:

the list of similar finals for the finals is:

the first Cartesian product operation is performed on

and

to generate similar pinyin for the pinyin to serve as the first pinyin candidate set, namely:



[0040] Further, the second pinyin candidate set generation module 400 is configured to perform the second Cartesian operation on a candidate set for each piece of pinyin to generate a correct place name pinyin candidate set: C = PC1 × PC2 × ... × PCn.

[0041] Furthermore, the query generation module 500 is configured to construct the query based on the second pinyin candidate set according to the following formula:

where should represents an OR operation, match_phrase means that a document must contain a character string in a query statement, and weight is the weight of a query clause, the weight being calculated based on the distance between candidate pinyin and original place name pinyin,
and the weight being calculated as shown in the following formula:

where pi is the pinyin of the ith character in the original pinyin, qi is the pinyin of the ith character in the candidate pinyin, d(pi,qi) is the distance between pi and qi, and the value of d(pi,qi) is the number of times that initials and finals are replaced for turning pi into qi.

[0042] The system for error correction in speech recognition of a place name of the invention can filter out wrong pinyin using the filtering module 300, thereby improving the accuracy of recognition. In addition, more similar pinyin can be calculated using the first pinyin candidate set generation module 200 and the second pinyin candidate set generation module 400, and further, the candidate place names with high to low probabilities can be arranged using the query generation module 500 and the candidate place name generation module 600, so that it is convenient for the user to select and the accuracy of recognition is also improved.

[0043] Furthermore, the invention provides a computer-readable medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the above method for error correction in speech recognition of a place name is implemented.

[0044] Furthermore, the invention provides a computer device, which comprises a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the above computer program, the above method for error correction in speech recognition of a place name is implemented.

[0045] For the computer-readable medium, there is a magnetic recording device, an optical disc, a magneto-optical recording medium, a semiconductor memory, etc. For the magnetic recording device, there is an HDD, an FD, a magnetic tape, etc. For the optical disc, there is a DVD (Digital Versatile Disc), a DVD-RAM, a CD-ROM, a CD-R (Recordable)/RW (Rewritable), etc. For the magneto-optical recording device, there is an MO (Magneto Optical Disk), etc.

[0046] The above examples mainly illustrate the system and method for error correction in speech recognition of a place name of the present invention. Although only some specific implementations of the invention are described, a person of ordinary skill in the art should understand that the invention may be implemented in multiple other forms without departing from the essence and scope of the invention. Accordingly, the present examples and embodiments are considered to be illustrative rather than restrictive, and the invention may encompass various modifications and alterations without departing from the spirit and scope of the invention that are defined by the appended claims.


Claims

1. A method for error correction in speech recognition, characterized by comprising the following steps:

a pinyin conversion step of converting speech information into pinyin;

a first pinyin candidate set generation step of splitting the pinyin to obtain N pinyin lists, and performing a first Cartesian product operation on the N pinyin lists to generate a first pinyin candidate set, where N is a natural number;

a filtering step of filtering the pinyin in the first pinyin candidate set to eliminate irregular pinyin;

a second pinyin candidate set generation step of performing a second Cartesian product operation on the pinyin output in the filtering step to obtain a second pinyin candidate set;

a query generation step of constructing a query based on the second pinyin candidate set; and

a candidate information generation step of querying the query in an inverted index to obtain candidate information with high to low probabilities and display the candidate information to a user.


 
2. The method for error correction in speech recognition of claim 1, characterized in that
in the first pinyin candidate set generation step, the pinyin is split into initials and finals, a list of similar initials and a list of similar finals are found according to a similarity dictionary, and a Cartesian product operation is performed on the list of initials and the list of finals to generate a candidate set for the pinyin.
 
3. The method for error correction in speech recognition of claim 2, characterized in that
in the first pinyin candidate set generation step, the list of similar initials for the initials is:

the list of similar finals for the finals is:

then, the first Cartesian product operation is performed on

and

to generate similar pinyin for the pinyin to serve as the first pinyin candidate set, namely:


 
4. The method for error correction in speech recognition of claim 3, characterized in that
in the second pinyin candidate set generation step, the second Cartesian product operation is performed on a candidate set for each piece of pinyin to obtain a correct pinyin candidate set: C = PC1 × PC2 × ... × PCn.
 
5. The method for error correction in speech recognition of claim 3, characterized in that
in the filtering step, the pinyin in the first pinyin candidate set is filtered according to a pinyin dictionary.
 
6. The method for error correction in speech recognition of claim 5, characterized in that
in the query generation step, the query is constructed based on the second pinyin candidate set according to the following formula:

where should represents an OR operation, match_phrase means that a document must contain a character string in a query statement, and weight is the weight of a query clause, the weight being calculated based on the distance between candidate pinyin and original pinyin.
 
7. The method for error correction in speech recognition of claim 6, characterized in that
the weight is calculated as shown in the following formula:

where pi is the pinyin of the ith character in the original pinyin, qi is the pinyin of the ith character in the candidate pinyin, d(pi,qi) is the distance between pi and qi, and the value of d(pi,qi) is the number of times that initials and finals are replaced for turning pi into qi.
 
8. A system for error correction in speech recognition, characterized by comprising:

a pinyin conversion module for acquiring speech information and converting the speech information into pinyin;

a first pinyin candidate set generation module for splitting the pinyin to obtain N pinyin lists, and performing a first Cartesian product operation on the N pinyin lists to generate a first pinyin candidate set, where N is a natural number;

a filtering module for filtering the pinyin in the first pinyin candidate set to eliminate irregular pinyin;

a second pinyin candidate set generation module for performing a second Cartesian product operation on the pinyin output by the filtering module to obtain a second pinyin candidate set; and

a query generation module for constructing a query based on the second pinyin candidate set; and

a candidate information generation module for querying the query in an inverted index to obtain candidate information with high to low probabilities and display the candidate information to a user.


 
9. The system for error correction in speech recognition of claim 8, characterized in that
the first pinyin candidate set generation module is configured to split the pinyin into initials and finals, find a list of similar initials and a list of similar finals according to a similarity dictionary, and perform a Cartesian product operation on the list of initials and the list of finals to generate a candidate set for the pinyin,
wherein the list of similar initials for the initials is:

the list of similar finals for the finals is:

the first Cartesian product operation is further performed on

and

to generate similar pinyin for the pinyin to serve as the first pinyin candidate set, namely:


 
10. The system for error correction in speech recognition of claim 9, characterized in that
the second pinyin candidate set generation module is configured to perform the second Cartesian product operation on a candidate set for each piece of pinyin to generate a correct pinyin candidate set: C = PC1 × PC2 × ... × PCn.
 
11. The system for error correction in speech recognition of claim 10, characterized in that
the query generation module is configured to construct the query based on the second pinyin candidate set according to the following formula:

where should represents an OR operation, match_phrase means that a document must contain a character string in a query statement, and weight is the weight of a query clause, the weight being calculated based on the distance between candidate pinyin and original pinyin,
and the weight being calculated as shown in the following formula:

where pi is the pinyin of the ith character in the original pinyin, qi is the pinyin of the ith character in the candidate pinyin, d(pi,qi) is the distance between pi and qi, and the value of d(pi,qi) is the number of times that initials and finals are replaced for turning pi into qi.
 
12. A computer-readable medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, a method for error correction in speech recognition of any one of claims 1 to 7 is implemented.
 
13. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that when the processor executes the computer program, the steps of a method for error correction in speech recognition of any one of claims 1 to 7 are implemented.
 




Drawing
















Search report