Method, system, electronic device and storage medium for improving speech recognition accuracy

By retrieving and matching multiple voice sources in valid and invalid databases, the problem of low speech recognition accuracy when multiple command sounds in the environment are weaker than other sounds, achieving higher recognition accuracy.

CN114582321BActive Publication Date: 2024-11-26CHINA UNITED NETWORK COMM GRP CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202011389352.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-12-01
Publication Date
2024-11-26
Estimated Expiration
2040-12-01

AI Technical Summary

Technical Problem

In the prior art, when there are multiple input command sounds in the environment, especially when the command sounds of multiple inputs are weaker than other sounds in terms of height or tone characteristics, the accuracy of speech recognition is greatly reduced and cannot be accurately captured. Command sound.

Method used

By establishing valid and invalid databases, multiple voice sources currently captured are retrieved and matched in these databases in sequence, improving the recognition accuracy when multiple voice inputs exist in the environment. The specific steps include collecting voice data in real time, extracting voice feature values, sorting them, and searching and comparing them in valid, standard and invalid databases, determining the validity and standardization of voice feature values, and dynamically updating the database to improve identification accuracy.

Benefits of technology

It greatly improves the recognition accuracy when multiple voice inputs exist in the environment, can capture command sounds more accurately, and improves the accuracy of voice input recognition.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114582321B_ABST
    Figure CN114582321B_ABST
Patent Text Reader

Abstract

The present invention relates to a method for improving the accuracy of voice input recognition. The method establishes and updates a valid database and an invalid database in real time. When there are multiple voice source inputs in an environment, each source input is captured, and searched and matched in the valid database, the standard database and the invalid database in sequence to determine the valid source input. By using this method, the accuracy of voice input recognition of smart devices in noisy environments can be significantly improved. The present invention also relates to a corresponding system, an electronic device and a storage medium.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a method, a system, an electronic device and a storage medium for improving the accuracy of speech input recognition. Background Art

[0002] In recent years, with the development of artificial intelligence and the increase in computer processing power, speech recognition technology has become increasingly mature and widely used in smart terminals. Smart speakers are a type of smart device that people use more frequently. With the help of an integrated artificial intelligence module, human-computer interaction is carried out through voice input, and smart speakers can play music or stories specified by people, greatly improving the user experience. However, the devices currently on the market have problems in voice input recognition, that is, when there are multiple input command sounds in the environment, especially when the multiple input command sounds are weaker than other sounds in terms of height or timbre characteristics, the accuracy of speech recognition will be greatly reduced. Therefore, how to improve the accuracy of speech input recognition, especially the recognition accuracy when there are multiple command sounds in the environment, is a problem worth exploring.

[0003] In the prior art, for example, CN 110211609 A discloses a method for improving the accuracy of speech recognition, including real-time recording of speech data input by a user, comparing a user speech database, comparing a standard database, obtaining a final recognition result, the user judging whether the recognition result is correct, inputting speech and the correct result into the user database, refreshing the user database, and performing operations according to the correct recognition result. This method achieves the effect of improving the speech recognition accuracy of high-frequency users by establishing an exclusive speech database for the current user in addition to the standard database, adding the speech data of the user with recognition errors into the user speech database, and giving priority to model matching and retrieval of the user speech database, thereby improving the interactive use experience of the user.

[0004] CN 108428446 A discloses a speech recognition method, comprising: determining the identity information of a target user in response to a speech input signal; extracting a common phrase set including a plurality of common phrases from a stored common phrase database based on the identity information; extracting acoustic features of the speech input signal and inputting the feature into an acoustic model to obtain an acoustic model score; determining whether the content of the speech input signal is a common phrase of the target user based on the score and the acoustic model scores of the common phrases in the stored common phrase set; if so, decoding the acoustic features of the speech input signal using a language model constructed based on the common phrases to obtain a speech recognition result.

[0005] In addition, CN 108231078 A discloses a speech recognition method, including: converting the collected voice information into digital information; extracting characteristic value a and classifying and marking it, and then establishing and storing it in a speech recognition library; digitizing and extracting characteristics of the user input voice information to be recognized, and comparing and recognizing the obtained characteristic value b with the characteristic value a in the speech recognition library; if the two match, the user input sound information is unlocked; otherwise, a prompt is issued, and the background server will receive an alarm message; through the extraction and comparison of characteristic values, different timbres and commands can be distinguished, thereby improving the accuracy and security of speech recognition.

[0006] However, the above three methods cannot solve the problem that when there are multiple input command sounds in the environment, especially when the multiple input command sounds are weaker than other voices in terms of pitch or timbre characteristics, the command sounds cannot be accurately captured, resulting in a significant reduction in the accuracy of speech recognition. Summary of the invention

[0007] Therefore, an object of the present invention is to provide a method for improving the accuracy of speech input recognition, as well as a corresponding system, electronic device and storage medium that can at least overcome the above-mentioned problems.

[0008] According to one aspect of the present invention, there is provided a method for improving the accuracy of speech input recognition, comprising:

[0009] Detection step, real-time collection of voice data in the current environment;

[0010] Extraction step, extracting speech feature values ​​from the collected speech data respectively;

[0011] A sorting step, sorting the speech feature values ​​according to certain rules;

[0012] A comparison step is performed to search and compare the speech feature values ​​in a database including a valid database, a standard database and an invalid database according to the sorting, and the comparison step includes:

[0013] According to the sorting, determine whether the speech feature value is in the valid database and output a valid recognition result,

[0014] If the speech feature value is in the valid database, the valid recognition result corresponding to the speech feature value is returned as the final recognition result, and the corresponding instruction is executed when the user does not modify the final recognition result.

[0015] If the speech feature value is not in the valid database, determining whether the speech feature value is in the standard database according to the sorting and outputting a standard recognition result,

[0016] If the speech feature value is in the standard database, determining whether the speech feature value is in the invalid database according to the sorting,

[0017] If the speech feature value is not in the invalid database, the standard recognition result corresponding to the speech feature value is returned as the final recognition result, and the corresponding instruction is executed when the user does not modify the final recognition result;

[0018] If the speech feature value is in the invalid database, determining that the speech feature value is invalid; and

[0019] If after searching and comparing according to the order, it is determined that all the speech feature values ​​are not in the standard database, the recognition fails and the user is prompted to re-enter.

[0020] The present invention establishes an additional valid database and an invalid database, and searches and matches the multiple voice sources currently captured in the valid database and the invalid database in sequence, thereby greatly improving the recognition accuracy when there are multiple voice inputs in the environment.

[0021] The speech feature values ​​may be sorted according to the sound height of the input speech source, or according to the intonation flatness of the input speech source.

[0022] The valid database stores: when the valid recognition result or the standard recognition result corresponding to the speech feature value is returned as the final recognition result, the speech feature value corresponding to the final recognition result is not modified by the user.

[0023] The invalid database stores: when the valid recognition result or the standard recognition result corresponding to the voice feature value is returned as the final recognition result, the voice feature value corresponding to the user's multiple modifications (which can be set to more than 3 times) to the final recognition result in the same input operation.

[0024] Dynamically update the valid database and the invalid database. When the valid recognition result or the standard recognition result corresponding to the current voice feature value is returned as the final recognition result, and the user has not modified the final recognition result, the current voice feature value is added to the valid database; when the valid recognition result or the standard recognition result corresponding to the current voice feature value is returned as the final recognition result, and the user has modified the final recognition result multiple times (which can be set to more than 3 times) in the same input operation, the current voice feature value is added to the invalid database.

[0025] According to another aspect of the present invention, a system for improving the accuracy of speech input recognition is provided, comprising:

[0026] A detection module, which detects and collects voice data in the current environment;

[0027] An extraction module, which extracts speech feature values ​​from the collected speech data respectively;

[0028] A sorting module, which sorts the speech feature values ​​according to certain rules;

[0029] A comparison module, in which the speech feature values ​​are searched and compared in a database according to the sorting, wherein the database includes a valid database, a standard database, and an invalid database:

[0030] According to the sorting, determine whether the speech feature value is in the valid database and output a valid recognition result,

[0031] If the speech feature value is in the valid database, the valid recognition result corresponding to the speech feature value is returned as the final recognition result, and the corresponding instruction is executed when the user does not modify the final recognition result.

[0032] If the speech feature value is not in the valid database, determining whether the speech feature value is in the standard database according to the sorting and outputting a standard recognition result,

[0033] If the speech feature value is in the standard database, determining whether the speech feature value is in the invalid database according to the sorting,

[0034] If the speech feature value is not in the invalid database, the standard recognition result corresponding to the speech feature value is returned as the final recognition result, and the corresponding instruction is executed when the user does not modify the final recognition result;

[0035] If the speech feature value is in the invalid database, determining that the speech feature value is invalid; and

[0036] If after searching and comparing according to the order, it is determined that all the speech feature values ​​are not in the standard database, the recognition fails and the user is prompted to re-enter; and

[0037] An output module is used to output the final speech recognition result.

[0038] According to another aspect of the present invention, there is provided an electronic device, comprising:

[0039] Processor; and

[0040] a memory storing a computer program,

[0041] in,

[0042] When the computer program is executed by the processor, the processor performs the method according to the present invention.

[0043] According to another aspect of the present invention, a storage medium is provided, on which computer-readable instructions are stored, wherein when the computer-readable instructions are executed by a processor, the processor performs the method according to the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flow chart of a method for improving the accuracy of speech input recognition according to an embodiment of the present invention; and

[0045] Figure 2 is a schematic diagram of a system for improving speech input recognition accuracy according to an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The present invention establishes a valid database and an invalid database, and searches and matches the multiple voice sources currently captured in the valid database and the invalid database in sequence, so as to solve the problem that when there are multiple command voice sources in the environment, especially when the height or timbre characteristics of the multiple command voice sources are weaker than other voices, the command voice cannot be accurately captured, thereby greatly reducing the accuracy of voice recognition.

[0047] Please refer to the following Figure 1 and Figure 2 , specifically describes a method and system for improving the accuracy of speech input recognition according to an embodiment of the present invention. Please note that the embodiments shown herein are only exemplary and are not intended to limit the scope of the present invention, which is defined by the appended claims.

[0048] Figure 1 is a flowchart of a method for improving the accuracy of speech input recognition according to an embodiment of the present invention, and Figure 2 is a schematic diagram of a system for improving speech input recognition accuracy according to an embodiment of the present invention.

[0049] like Figure 1 As shown, the method for improving the accuracy of speech input recognition according to an embodiment of the present invention includes:

[0050] S100: collecting voice data in the current environment in real time and extracting voice feature values ​​from the collected voice data;

[0051] S200: sorting the speech feature values ​​according to certain rules;

[0052] S300: Search and compare the speech feature values ​​in the valid database according to the order to determine whether the speech feature values ​​are in the valid database, and output a valid recognition result:

[0053] S310: If the speech feature value is in the valid database, the valid recognition result corresponding to the speech feature value is returned as the final recognition result, and the corresponding instruction is executed when the user does not modify the final recognition result;

[0054] S320: If after the search and comparison according to the sorting is completed, it is determined that all the remaining speech feature values ​​are not in the valid database, all the speech feature values ​​that are not in the valid database are sorted according to the rules of S200; then

[0055] S330: Search and compare the speech feature values ​​in the standard database according to the order to determine whether the speech feature values ​​are in the standard database, and output the standard recognition result:

[0056] S331: If the speech feature value is in the standard database, the speech feature value is searched and compared in the invalid database to determine whether the speech feature value is in the invalid database:

[0057] S331-1: If the speech feature value is not in the invalid database, the standard recognition result corresponding to the speech feature value is returned as the final recognition result, and the corresponding instruction is executed if the user does not modify the final recognition result; otherwise

[0058] S331-2: If the speech feature value is in the invalid database, the speech feature value is invalid, and the process continues to step S330; and

[0059] S322: If after searching and comparing according to the order, it is determined that all speech feature values ​​are not in the standard database, the recognition fails and the user is prompted to re-enter.

[0060] In another embodiment, the subsequent steps can also be performed before all speech feature values ​​are searched and compared in the valid database, that is, if a part of the speech feature values ​​is determined to be not in the valid database, the part of the speech feature values ​​has been searched and compared in the standard database and the possible invalid database, as long as the search and comparison in the valid database, the standard database and the invalid database are all performed on the speech feature values ​​according to the ranking. For example, the speech feature value ranked as 1 can be searched and compared in the valid database and the valid recognition result can be output. If the valid recognition result is no, the speech feature value ranked as 1 can be searched and compared in the standard speech database and the standard recognition result can be output. If the standard recognition result is yes, the speech feature value can be searched and compared in the invalid database. If the speech feature value is in the invalid database, it is determined that the speech feature value is invalid, and the above steps are continued to be executed cyclically for the speech feature value ranked as 2.

[0061] The above-mentioned effective database stores: when the effective recognition result or standard recognition result corresponding to the speech feature value is returned as the final recognition result, the speech feature value corresponding to the final recognition result is not modified by the user.

[0062] The invalid database stores: when the valid recognition result or standard recognition result corresponding to the voice feature value is returned as the final recognition result, the voice feature value corresponding to the user's continuous (referring to the same input operation) multiple modifications (can be set to more than 3 times) to the final recognition result.

[0063] The valid database and invalid database are updated dynamically. When the valid recognition result or standard recognition result corresponding to the current voice feature value is returned as the final recognition result, if the user has not modified the final recognition result, this voice feature value will be added to the valid database; when the valid recognition result or standard recognition result corresponding to the current voice feature value is returned as the final recognition result, if the user has modified the final recognition result multiple times (in the same input operation) continuously (can be set to more than 3 times), this voice feature value will be added to the invalid database.

[0064] like Figure 2 As shown, a system 1000 for improving the accuracy of speech input recognition according to an embodiment of the present invention includes:

[0065] Detection module 1001: responsible for detecting and collecting voice data in the current environment

[0066] Extraction module 1002: responsible for extracting the speech feature values ​​of all collected speech data

[0067] Sorting module 1003: responsible for sorting the extracted speech feature values ​​according to certain rules

[0068] Comparison module 1004: responsible for searching and comparing in the database

[0069] Output module 1005: responsible for outputting the final speech recognition results

[0070] The following is an explanation through specific examples.

[0071] Examples:

[0072] A smart speaker device 1000 includes at least five modules: a detection module 1001, an extraction module 1002, a sorting module 1003, a comparison module 1004, and an output module 1005. Each module is mainly responsible for:

[0073] Detection module 1001: responsible for detecting and collecting voice data in the current environment

[0074] Extraction module 1002: responsible for extracting the speech feature values ​​of all collected speech data

[0075] Sorting module 1003: responsible for sorting the extracted speech feature values ​​according to certain rules

[0076] Comparison module 1004: responsible for searching and comparing in the database

[0077] Output module 1005: responsible for outputting the final speech recognition results

[0078] Specifically, the collaborative workflow between modules is as follows:

[0079] Step 1, detection module 1001: real-time collection of voice data in the current environment

[0080] Step 2, extraction module 1002: extracting speech feature values ​​from the output results of the detection module 1000;

[0081] Step 3, sorting module 1003: sorting the output values ​​of the extraction module 1001, which can be sorted according to the sound height of the input speech source, or according to the smoothness of the tone of the input speech source;

[0082] Step 4, comparison module 1004: search and compare the speech feature value ranked as 1 in the valid database and output the valid recognition result:

[0083] The output result represents whether the speech feature value is in the valid database. If so, the output result is yes; if not, the output result is no;

[0084] If the comparison module 1004 outputs a yes result, the output module 1005 returns the valid recognition result corresponding to the speech feature value ranked as 1 as the final recognition result, and executes the corresponding instruction when the user does not modify the final recognition result;

[0085] If the comparison module 1004 outputs a negative result, the comparison module 1004: searches and compares the speech feature value ranked as 1 in the standard database and outputs a standard recognition result;

[0086] If the speech feature value is in the standard database: the speech feature value is searched and compared in the invalid database; if the speech feature value is not in the invalid database, the output module 1005: the standard recognition result corresponding to the speech feature value is returned as the final recognition result, and the corresponding instruction is executed when the user does not modify the final recognition result; if the speech feature value is in the invalid database, the speech feature value is invalid, and the speech feature value ranked as 2 is continuously looped to execute the above step 4;

[0087] If the speech feature value is not in the standard database, then continue to loop through step 4 for the speech feature value ranked 2; and

[0088] If the speech feature values ​​of all captured speech sources are not in the valid database and the standard database after the sequential comparison is completed, the recognition fails and the output module 1005 reminds the user to re-input.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present disclosure, rather than to limit them. Although the present disclosure has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A method for improving the accuracy of speech input recognition, comprising: Detection step, real-time collection of voice data in the current environment; Extraction step, extracting speech feature values ​​from the collected speech data respectively; A sorting step, sorting the speech feature values ​​according to certain rules; A comparison step is performed to search and compare the speech feature values ​​in a database including a valid database, a standard database and an invalid database according to the sorting, and the comparison step includes: According to the sorting, determine whether the speech feature value is in the valid database and output a valid recognition result, If the speech feature value is in the valid database, the valid recognition result corresponding to the speech feature value is returned as the final recognition result, and the corresponding instruction is executed when the user does not modify the final recognition result. If the speech feature value is not in the valid database, determining whether the speech feature value is in the standard database according to the sorting and outputting a standard recognition result, If the speech feature value is in the standard database, determining whether the speech feature value is in the invalid database according to the sorting, If the speech feature value is not in the invalid database, the standard recognition result corresponding to the speech feature value is returned as the final recognition result, and the corresponding instruction is executed when the user does not modify the final recognition result; If the speech feature value is in the invalid database, determining that the speech feature value is invalid; and If after searching and comparing according to the order, it is determined that all the speech feature values ​​are not in the standard database, the recognition fails and the user is prompted to re-enter.

2. The method according to claim 1, characterized in that The speech feature values ​​are sorted according to the sound height of the input speech source.

3. The method according to claim 1, characterized in that The speech feature values ​​are sorted according to the intonation smoothness of the input speech source.

4. The method according to claim 1, characterized in that The valid database stores: when the valid recognition result or the standard recognition result corresponding to the speech feature value is returned as the final recognition result, the speech feature value corresponding to the final recognition result is not modified by the user.

5. The method according to claim 1, characterized in that The invalid database stores: when the valid recognition result or the standard recognition result corresponding to the speech feature value is returned as the final recognition result, the speech feature value corresponding to the multiple modifications made by the user to the final recognition result in the same input operation.

6. The method according to claim 1, characterized in that dynamically updating the valid database and the invalid database, When the valid recognition result or the standard recognition result corresponding to the current speech feature value is returned as the final recognition result, and the user does not modify the final recognition result, the current speech feature value is added to the valid database; When the valid recognition result or the standard recognition result corresponding to the current speech feature value is returned as the final recognition result, and the user modifies the final recognition result multiple times in the same input operation, the current speech feature value is added to the invalid database.

7. The method according to claim 5 or 6, characterized in that: The multiple modifications are more than 3 times.

8. A system for improving the accuracy of speech input recognition, comprising: A detection module, which detects and collects voice data in the current environment; An extraction module, which extracts speech feature values ​​from the collected speech data respectively; A sorting module, which sorts the speech feature values ​​according to certain rules; A comparison module, in which the speech feature values ​​are searched and compared in a database according to the sorting, wherein the database includes a valid database, a standard database, and an invalid database: According to the sorting, determine whether the speech feature value is in the valid database and output a valid recognition result, If the speech feature value is in the valid database, the valid recognition result corresponding to the speech feature value is returned as the final recognition result, and the corresponding instruction is executed when the user does not modify the final recognition result. If the speech feature value is not in the valid database, determining whether the speech feature value is in the standard database according to the sorting and outputting a standard recognition result, If the speech feature value is in the standard database, determining whether the speech feature value is in the invalid database according to the sorting, If the speech feature value is not in the invalid database, the standard recognition result corresponding to the speech feature value is returned as the final recognition result, and the corresponding instruction is executed when the user does not modify the final recognition result; If the speech feature value is in the invalid database, determining that the speech feature value is invalid; as well as If after searching and comparing according to the order, it is determined that all the speech feature values ​​are not in the standard database, the recognition fails and the user is prompted to re-enter; as well as An output module is used to output the final speech recognition result.

9. An electronic device, comprising: processor; and a memory storing a computer program, It is characterized in that When the computer program is executed by the processor, the processor performs the method according to any one of claims 1 to 7.

10. A storage medium having computer-readable instructions stored thereon, characterized in that: When the computer readable instructions are executed by a processor, the processor performs the method according to any one of claims 1 to 7.