Method, device and computer equipment for detecting misidentification of noise

Abstract

The present application provides a method, device and computer equipment for detecting misidentification of noise. The speech recognition system analyzes the degree of confusion of the decoding path by sorting the decoding results of the speech data and the probability values ​​of the corresponding candidate words. If the current speech recognition result is the result of misidentification of noise, its decoding result will be relatively scattered, that is, the degree of confusion of the decoding path will be greater than the threshold. This method for detecting misidentification of noise is simple in calculation and does not require the introduction of additional discriminant models. It can greatly reduce the probability of misidentification of noise and greatly improve the user experience and the accuracy of speech recognition.

Description

Technical Field

[0001] The present application relates to the field of speech recognition technology, and in particular to a method, device and computer equipment for detecting misidentification of noise. Background Art

[0002] When the wake-up word and command word recognition technology is applied, it can monitor the user's voice in real time and give corresponding feedback when specific words are detected (for example, waking up the device when the wake-up word is detected). In actual applications, misidentification and false wake-up often occur. For example, the voice received by the device does not contain the command word or is completely noise, but it is judged as a command word. Such misidentification and false wake-up will bring a very bad user experience to users. Summary of the invention

[0003] The main purpose of the present application is to provide a method, device and computer equipment for detecting misidentification of noise, aiming to solve the problem of misidentification of noise in the existing command word recognition technology.

[0004] To achieve the above objectives, the present application provides a method for detecting misidentification of noise, comprising:

[0005] Collecting voice data, and performing recognition and decoding on the voice data to generate a lattice decoding graph, wherein the lattice decoding graph includes a plurality of decoding results;

[0006] According to the arrangement order of each of the decoding results in the lattice decoding diagram, weights are assigned in turn according to a preset decreasing rule to obtain weights corresponding to each of the decoding results;

[0007] According to the weights corresponding to the decoding results and the candidate words corresponding to the decoding results, the probability values ​​corresponding to the candidate words are calculated;

[0008] According to the probability values ​​corresponding to the candidate words, determining whether the degree of confusion of the decoding path of the lattice decoding graph is greater than a threshold;

[0009] If the degree of confusion of the decoding path of the lattice decoding graph is greater than a threshold, it is determined that the speech recognition result of the current speech data is a misrecognition of noise.

[0010] The present application also provides a device for detecting misidentification of noise, comprising:

[0011] A recognition module, used to collect voice data, recognize and decode the voice data, and generate a lattice decoding graph, wherein the lattice decoding graph includes a plurality of decoding results;

[0012] An assignment module, used to assign weights according to a preset decreasing rule in accordance with the arrangement order of each decoding result in the lattice decoding diagram, to obtain weights corresponding to each decoding result;

[0013] A calculation module, used for calculating the probability values ​​corresponding to a plurality of candidate words according to the weights corresponding to the decoding results and the candidate words corresponding to the decoding results;

[0014] A judgment module, used for judging whether the degree of confusion of the decoding path of the lattice decoding graph is greater than a threshold value according to the probability values ​​corresponding to the candidate words;

[0015] The determination module is used to determine that the speech recognition result of the current speech data is a noise misrecognition if the confusion degree of the decoding path of the lattice decoding diagram is greater than a threshold.

[0016] The present application also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0017] The present application also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0018] A detection method, device and computer equipment for noise misrecognition provided in the present application, a speech recognition system collects speech data, and recognizes and decodes the speech data to generate a lattice decoding diagram, wherein the lattice decoding diagram contains several decoding results. The speech recognition system assigns weights according to the arrangement order of each decoding result in the lattice decoding diagram, and sequentially according to the preset decreasing rule, to obtain the weights corresponding to each decoding result. The speech recognition system calculates the probability values ​​corresponding to several candidate words according to the weights corresponding to each decoding result and the candidate words corresponding to each decoding result. Finally, the speech recognition system determines whether the degree of confusion of the decoding path of the lattice decoding diagram is greater than a threshold value according to the probability values ​​corresponding to each candidate word. If the degree of confusion of the decoding path of the lattice decoding diagram is greater than the threshold, it is determined that the speech recognition result of the current speech data is a noise misrecognition. In the present application, the speech recognition system analyzes the degree of confusion of the decoding path by sorting the decoding results and the probability values ​​of the corresponding candidate words. If the current speech recognition result is the result of noise misrecognition, its decoding result will be relatively scattered, that is, the degree of confusion of the decoding path will be greater than the threshold. This method for detecting misidentification of noise is simple in calculation and does not require the introduction of additional discrimination models. It can greatly reduce the probability of misidentification of noise and significantly improve user experience and speech recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the steps of a method for detecting misidentification of noise in one embodiment of the present application;

[0020] Figure 2 It is an overall structural block diagram of a detection device for misidentification of noise in one embodiment of the present application;

[0021] Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present application.

[0022] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0024] Reference Figure 1 In one embodiment of the present application, a method for detecting misidentification of noise is provided, comprising:

[0025] S1: collecting voice data, and performing recognition and decoding on the voice data to generate a lattice decoding graph, wherein the lattice decoding graph includes a plurality of decoding results;

[0026] S2: according to the arrangement order of each decoding result in the lattice decoding diagram, weights are assigned in turn according to a preset decreasing rule to obtain weights corresponding to each decoding result;

[0027] S3: according to the weights corresponding to the decoding results and the candidate words corresponding to the decoding results, the probability values ​​corresponding to the candidate words are calculated;

[0028] S4: judging whether the degree of confusion of the decoding path of the lattice decoding graph is greater than a threshold value according to the probability values ​​corresponding to the candidate words;

[0029] S5: If the degree of confusion of the decoding path of the lattice decoding diagram is greater than a threshold, it is determined that the speech recognition result of the current speech data is a misrecognition of noise.

[0030] In this embodiment, the speech recognition system monitors in real time whether the speech data sent by the user or the outside world is collected. If the speech data is collected, the speech data is decoded and recognized through the recognition and decoding tool (such as Kaldi) to generate a lattice decoding graph. The lattice decoding graph contains several decoding results. For example, when using the Kaldi tool for recognition, the number of decoding beams can be customized, but the actual generated decoding path (the decoding path corresponds to the decoding result) is less than or equal to this value; the decoding results in the lattice decoding graph are arranged in descending order from large to small according to their corresponding probability scores (the probability score of the decoding result is the product of each probability value on the decoding path corresponding to the decoding result). The speech recognition system first assigns weights according to the preset decreasing rules according to the arrangement order of each decoding result in the lattice decoding graph (that is, weights are assigned to each decoding result in order from large to small, and the weights corresponding to each decoding result are decreased in turn), so as to obtain the weights corresponding to each decoding result. Then, the speech recognition system adds up the weights corresponding to the several decoding results corresponding to the same candidate word to obtain the score of each candidate word; and adds up the weights corresponding to each decoding result to obtain the sum of weights. Then, the ratio of the score of each candidate word to the sum of weights is calculated to obtain the probability value corresponding to each candidate word. The speech recognition system calculates the entropy value that can characterize the degree of confusion (or discreteness) of the decoding path of the lattice decoding diagram based on the probability value corresponding to each candidate word and the preset algorithm, and then compares the entropy value with the preset entropy value threshold to determine whether the degree of confusion of the decoding path of the lattice decoding diagram is greater than the threshold. That is, if the entropy value is greater than the entropy value threshold, the speech recognition system determines that the degree of confusion of the decoding path of the lattice decoding diagram is greater than the threshold. If the recognition result of the speech recognition system for the current speech data belongs to normal speech recognition (that is, the speech data contains the correct wake-up word or command word, keyword), then the decoding result of the previous beam should be basically consistent (or it can be understood that the decoding result of the previous beam is relatively concentrated), that is, the decoding path of the lattice decoding graph is less chaotic (specifically, the calculated entropy value is less than the entropy value threshold). If the recognition result of the speech recognition system for the current speech data belongs to misrecognition of noise (that is, the speech data is ambient sound or meaningless word speech), the decoding result will be more scattered (that is, the decoding path of the lattice decoding graph is more chaotic, specifically, the calculated entropy value is greater than the entropy value threshold); therefore, the speech recognition system determines that the current speech recognition result of the speech data is misrecognition of noise, and no matter what the specific speech recognition result is, the speech recognition system will not give any feedback.

[0031] In this embodiment, the speech recognition system analyzes the degree of confusion of the decoding path by sorting the decoding results and the probability values ​​of the corresponding candidate words. If the current speech recognition result is the result of misrecognition of noise, the decoding result will be relatively scattered, that is, the degree of confusion of the decoding path will be greater than the threshold. This method of detecting misrecognition of noise is simple in calculation and does not require the introduction of additional discriminant models (such as secondary model judgment), which can effectively reduce the computing power consumption of the terminal device; moreover, the detection method can accurately detect the misrecognition of noise, thereby greatly reducing the probability of misrecognition of noise, and effectively improving the user experience and the accuracy of speech recognition.

[0032] Furthermore, the step of assigning weights according to a preset decreasing rule in accordance with the arrangement order of each decoding result in the lattice decoding diagram to obtain the weights corresponding to each decoding result includes:

[0033] S201: Substitute the arrangement order of each decoding result in the lattice decoding diagram into the first calculation formula in order, and calculate the weights corresponding to each decoding result respectively, wherein the first calculation formula is: , where Q is the weight corresponding to the decoding result, n is the order of arrangement corresponding to the decoding result, and x is a decreasing value;

[0034] S202: When the calculated weight corresponding to the first decoding result is less than the weight threshold, the weights of the first decoding result and each second decoding result are set to a fixed value, wherein the second decoding result is all decoding results sorted after the first decoding result in the lattice decoding diagram.

[0035] In this embodiment, the speech recognition system substitutes the arrangement sequence of each decoding in the lattice decoding diagram into the first calculation formula in the order of their corresponding arrangement. In order to calculate the weights corresponding to each decoding result, Q is the weight corresponding to the decoding result, n is the order of arrangement corresponding to the decoding result, and x is a decreasing value (the decreasing value can be customized according to actual needs). In the process of calculating the weights of each decoding result in turn, if the weight corresponding to the first decoding result (the first decoding result is one of all the decoding results) is less than the weight threshold, the weights corresponding to the first decoding result and all the decoding results after the first decoding result is sorted (that is, each second decoding result) are set to fixed values. The speech recognition system assigns weights to each decoding result according to the rules described in this embodiment to obtain the weights corresponding to each decoding result.

[0036] An example is given below to better illustrate the weight assignment rule in this embodiment. Assume that there are 7 decoding results of the lattice decoding graph, which are sorted from large to small as: r1, r2, r3, r4, r5, r6, r7, the decrement value x is set to 0.2, the weight threshold is set to 0.2, and the fixed value is set to 0.1; the weight of r1 is calculated to be 1, the weight of r2 is 0.8, the weight of r3 is 0.6, the weight of r4 is 0.4, and the weight of r5 is 0.2 by the first calculation formula; when the weight of r6 is calculated, if the calculation is performed according to the first calculation formula, the weight of r6 is 0, which is less than the weight threshold of 0.2, so the weights corresponding to r6 and r7 after sorting are set to a fixed value of 0.1.

[0037] Furthermore, the step of calculating the probability values ​​corresponding to a plurality of candidate words respectively according to the weights respectively corresponding to the decoding results and the candidate words respectively corresponding to the decoding results includes:

[0038] S301: summing up the weights corresponding to the plurality of decoding results corresponding to the same candidate word to obtain the scores corresponding to the candidate words; and summing up the weights corresponding to the decoding results to obtain the sum of the weights;

[0039] S302: Calculate the ratio of the score of each candidate word to the total weight to obtain the probability value corresponding to each candidate word.

[0040] In this embodiment, the speech recognition system adds up the weights corresponding to the several decoding results corresponding to the same candidate word to obtain the scores corresponding to the candidate words, and adds up the weights corresponding to the decoding results to obtain the weight sum. Then, the speech recognition system calculates the ratio of the score of each candidate word to the weight sum to obtain the probability value corresponding to each candidate word.

[0041] Based on the example listed in the embodiment of steps S201-S202, this embodiment is illustrated. Assuming that the decoding results corresponding to candidate word A are r1 and r3, the score of candidate word A is: r1+r3=1+0.6=1.6; the sum of weights is: 1+0.8+0.6+0.4+0.2+0.1+0.1=3.2; the probability value of candidate word A is: 1.6÷3.2=0.5.

[0042] Furthermore, the step of judging whether the degree of confusion of the decoding path of the lattice decoding graph is greater than a threshold value according to the probability values ​​corresponding to the candidate words includes:

[0043] S401: Substitute the probability values ​​corresponding to the candidate words into the second calculation formula to calculate the entropy value, wherein the entropy value represents the degree of confusion of the decoding path of the lattice decoding graph, and the second calculation formula is: , S is the entropy value, P i Represents the probability value of the i-th candidate word;

[0044] S402: Determine whether the entropy value is greater than an entropy value threshold;

[0045] S403: If the entropy value is greater than the entropy value threshold, it is determined that the degree of confusion of the decoding path of the lattice decoding graph is greater than the threshold;

[0046] S404: If the entropy value is less than the entropy value threshold, it is determined that the degree of confusion of the decoding path of the lattice decoding graph is less than the threshold.

[0047] In this embodiment, after calculating the probability value corresponding to each candidate word, the speech recognition system substitutes the probability value of each candidate word into the second calculation formula to calculate the entropy value, so as to analyze and obtain the degree of confusion of the decoding path of the lattice decoding graph; wherein the second calculation formula is: , S represents the entropy value, P i Characterize the probability value of the i-th candidate word. For example, there are 3 candidate words in total, the probability value of candidate word A is 0.5, the probability value of candidate word B is 0.3, and the probability value of candidate word C is 0.2, then the entropy value S=-(0.5×log(0.5)+0.3×log(0.3)+0.2×log(0.2)). The speech recognition system calls the preset entropy value threshold and compares the entropy value calculated for the current time with the entropy value threshold. If the entropy value is greater than the entropy value threshold, it means that the decoding path of the current lattice decoding graph is relatively chaotic, and the speech recognition system determines that the degree of chaos of the decoding path of the lattice decoding graph is greater than the threshold. If the entropy value is less than the entropy value threshold, it means that the decoding path of the current lattice decoding graph is basically the same, and the speech recognition system determines that the degree of chaos of the decoding path of the lattice decoding graph is less than the threshold.

[0048] Furthermore, the step of identifying and decoding the voice data to generate a lattice decoding graph includes:

[0049] S101: using a speech recognition tool to recognize and decode the speech data to obtain the lattice decoding graph;

[0050] S102: Arrange the decoding results contained in the lattice decoding graph in descending order according to their respective probability scores.

[0051] In this embodiment, after the speech recognition system monitors and collects the speech data input from the outside (the speech data may be human voice, environmental sound or other types of sound), it uses a speech recognition tool (such as Kaldi) to recognize and decode the speech data, thereby obtaining a lattice decoding graph (taking Kaldi as an example, the general process of Kaldi is: acoustic model + language model, specifically DNN output PDF-ID, combined with hclg's WFS, decoding to generate a lattice decoding graph, and searching the optimal N best in the lattice decoding graph as the recognition result). The speech recognition system arranges the decoding results contained in the lattice decoding graph in descending order according to their respective probability scores (the probability score of the decoding result is the product of each probability value on the decoding path corresponding to the decoding result), so that the subsequent speech recognition system can assign weights to each decoding result.

[0052] Reference Figure 2 In one embodiment of the present application, a device for detecting misidentification of noise is also provided, comprising:

[0053] Recognition module 1, used for collecting voice data, and performing recognition and decoding on the voice data to generate a lattice decoding graph, wherein the lattice decoding graph includes a plurality of decoding results;

[0054] Assignment module 2, used for assigning weights according to the arrangement order of each decoding result in the lattice decoding diagram and according to a preset decreasing rule, to obtain the weights corresponding to each decoding result;

[0055] A calculation module 3 is used to calculate the probability values ​​corresponding to a plurality of candidate words according to the weights corresponding to the decoding results and the candidate words corresponding to the decoding results;

[0056] A judgment module 4 is used to judge whether the degree of confusion of the decoding path of the lattice decoding graph is greater than a threshold value according to the probability values ​​corresponding to the candidate words;

[0057] The determination module 5 is used to determine that the speech recognition result of the current speech data is a noise misrecognition if the confusion degree of the decoding path of the lattice decoding diagram is greater than a threshold.

[0058] Furthermore, the assignment module 2 includes:

[0059] The first calculation unit is used to substitute the arrangement order of each decoding result in the lattice decoding diagram into the first calculation formula in sequence according to the arrangement order, and calculate the weights corresponding to each decoding result respectively, wherein the first calculation formula is: , where Q is the weight corresponding to the decoding result, n is the order of arrangement corresponding to the decoding result, and x is a decreasing value;

[0060] A setting unit is used to set the weights of the first decoding result and each second decoding result to a fixed value when the weight corresponding to the calculated first decoding result is less than the weight threshold, wherein the second decoding result is all decoding results sorted after the first decoding result in the lattice decoding diagram.

[0061] Furthermore, the calculation module 3 includes:

[0062] A second calculation unit is used to sum the weights corresponding to the plurality of decoding results corresponding to the same candidate word to obtain the scores corresponding to the candidate words; and to sum the weights corresponding to the decoding results to obtain the weight sum;

[0063] The third calculation unit is used to calculate the ratio of the score of each candidate word to the total weight, so as to obtain the probability value corresponding to each candidate word.

[0064] Furthermore, the judging module 4 comprises:

[0065] The fourth calculation unit is used to substitute the probability values ​​corresponding to each of the candidate words into the second calculation formula to calculate the entropy value, wherein the entropy value represents the degree of confusion of the decoding path of the lattice decoding graph, and the second calculation formula is: , S is the entropy value, P i Represents the probability value of the i-th candidate word;

[0066] A judging unit, used to judge whether the entropy value is greater than an entropy value threshold;

[0067] A first determination unit, configured to determine that the degree of confusion of the decoding path of the lattice decoding graph is greater than a threshold value if the entropy value is greater than an entropy value threshold value;

[0068] The second determination unit is configured to determine that the degree of confusion of the decoding path of the lattice decoding graph is less than a threshold value if the entropy value is less than an entropy value threshold value.

[0069] Furthermore, the identification module 1 includes:

[0070] A decoding unit, used to use a speech recognition tool to recognize and decode the speech data to obtain the lattice decoding graph;

[0071] A sorting unit is used to arrange the decoding results contained in the lattice decoding graph in descending order according to their respective probability scores.

[0072] In this embodiment, each module and unit in the device for detecting misidentification of noise is used to correspondingly execute each step in the above-mentioned method for detecting misidentification of noise, and its specific implementation process will not be described in detail here.

[0073] The present embodiment provides a detection device for misidentification of noise, wherein the speech recognition system collects speech data, recognizes and decodes the speech data, and generates a lattice decoding diagram, wherein the lattice decoding diagram contains several decoding results. The speech recognition system assigns weights according to the arrangement order of each decoding result in the lattice decoding diagram, and sequentially according to the preset decreasing rule, to obtain the weights corresponding to each decoding result. The speech recognition system calculates the probability values ​​corresponding to several candidate words according to the weights corresponding to each decoding result and the candidate words corresponding to each decoding result. Finally, the speech recognition system determines whether the degree of confusion of the decoding path of the lattice decoding diagram is greater than a threshold value according to the probability values ​​corresponding to each candidate word. If the degree of confusion of the decoding path of the lattice decoding diagram is greater than the threshold, it is determined that the speech recognition result of the current speech data is misidentified as noise. In the present application, the speech recognition system analyzes the degree of confusion of the decoding path by sorting the decoding results and the probability values ​​of the corresponding candidate words. If the current speech recognition result is the result of misidentification of noise, its decoding result will be relatively scattered, that is, the degree of confusion of the decoding path will be greater than the threshold. This method for detecting misidentification of noise is simple in calculation and does not require the introduction of additional discrimination models. It can greatly reduce the probability of misidentification of noise and significantly improve user experience and speech recognition accuracy.

[0074] Reference Figure 3 In an embodiment of the present application, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data such as thresholds. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a detection method for misidentification of noise is implemented.

[0075] The processor executes the following steps of the method for detecting misidentification of noise:

[0076] S1: collecting voice data, and performing recognition and decoding on the voice data to generate a lattice decoding graph, wherein the lattice decoding graph includes a plurality of decoding results;

[0077] S2: according to the arrangement order of each decoding result in the lattice decoding diagram, weights are assigned in turn according to a preset decreasing rule to obtain weights corresponding to each decoding result;

[0078] S3: according to the weights corresponding to the decoding results and the candidate words corresponding to the decoding results, the probability values ​​corresponding to the candidate words are calculated;

[0079] S4: judging whether the degree of confusion of the decoding path of the lattice decoding graph is greater than a threshold value according to the probability values ​​corresponding to the candidate words;

[0080] S5: If the degree of confusion of the decoding path of the lattice decoding diagram is greater than a threshold, it is determined that the speech recognition result of the current speech data is a misrecognition of noise.

[0081] Furthermore, the step of assigning weights according to a preset decreasing rule in accordance with the arrangement order of each decoding result in the lattice decoding diagram to obtain the weights corresponding to each decoding result includes:

[0082] S201: Substitute the arrangement order of each decoding result in the lattice decoding diagram into the first calculation formula in order, and calculate the weights corresponding to each decoding result respectively, wherein the first calculation formula is: , where Q is the weight corresponding to the decoding result, n is the order of arrangement corresponding to the decoding result, and x is a decreasing value;

[0083] S202: When the calculated weight corresponding to the first decoding result is less than the weight threshold, the weights of the first decoding result and each second decoding result are set to a fixed value, wherein the second decoding result is all decoding results sorted after the first decoding result in the lattice decoding diagram.

[0084] Furthermore, the step of calculating the probability values ​​corresponding to a plurality of candidate words respectively according to the weights respectively corresponding to the decoding results and the candidate words respectively corresponding to the decoding results includes:

[0085] S301: summing up the weights corresponding to the plurality of decoding results corresponding to the same candidate word to obtain the scores corresponding to the candidate words; and summing up the weights corresponding to the decoding results to obtain the sum of the weights;

[0086] S302: Calculate the ratio of the score of each candidate word to the total weight to obtain the probability value corresponding to each candidate word.

[0087] Furthermore, the step of judging whether the degree of confusion of the decoding path of the lattice decoding graph is greater than a threshold value according to the probability values ​​corresponding to the candidate words includes:

[0088] S401: Substitute the probability values ​​corresponding to the candidate words into the second calculation formula to calculate the entropy value, wherein the entropy value represents the degree of confusion of the decoding path of the lattice decoding graph, and the second calculation formula is: , S is the entropy value, P i Represents the probability value of the i-th candidate word;

[0089] S402: Determine whether the entropy value is greater than an entropy value threshold;

[0090] S403: If the entropy value is greater than the entropy value threshold, it is determined that the degree of confusion of the decoding path of the lattice decoding graph is greater than the threshold;

[0091] S404: If the entropy value is less than the entropy value threshold, it is determined that the degree of confusion of the decoding path of the lattice decoding graph is less than the threshold.

[0092] Furthermore, the step of identifying and decoding the voice data to generate a lattice decoding graph includes:

[0093] S101: using a speech recognition tool to recognize and decode the speech data to obtain the lattice decoding graph;

[0094] S102: Arrange the decoding results contained in the lattice decoding graph in descending order according to their respective probability scores.

[0095] An embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, a method for detecting misidentification of noise is implemented. The method for detecting misidentification of noise is specifically:

[0096] S1: collecting voice data, and performing recognition and decoding on the voice data to generate a lattice decoding graph, wherein the lattice decoding graph includes a plurality of decoding results;

[0097] S2: according to the arrangement order of each decoding result in the lattice decoding diagram, weights are assigned in turn according to a preset decreasing rule to obtain weights corresponding to each decoding result;

[0098] S3: according to the weights corresponding to the decoding results and the candidate words corresponding to the decoding results, the probability values ​​corresponding to the candidate words are calculated;

[0099] S4: judging whether the degree of confusion of the decoding path of the lattice decoding graph is greater than a threshold value according to the probability values ​​corresponding to the candidate words;

[0100] S5: If the degree of confusion of the decoding path of the lattice decoding diagram is greater than a threshold, it is determined that the speech recognition result of the current speech data is a misrecognition of noise.

[0101] Furthermore, the step of assigning weights according to a preset decreasing rule in accordance with the arrangement order of each decoding result in the lattice decoding diagram to obtain the weights corresponding to each decoding result includes:

[0102] S201: Substitute the arrangement order of each decoding result in the lattice decoding diagram into the first calculation formula in order, and calculate the weights corresponding to each decoding result respectively, wherein the first calculation formula is: , where Q is the weight corresponding to the decoding result, n is the order of arrangement corresponding to the decoding result, and x is a decreasing value;

[0103] S202: When the calculated weight corresponding to the first decoding result is less than the weight threshold, the weights of the first decoding result and each second decoding result are set to a fixed value, wherein the second decoding result is all decoding results sorted after the first decoding result in the lattice decoding diagram.

[0104] Furthermore, the step of calculating the probability values ​​corresponding to a plurality of candidate words respectively according to the weights respectively corresponding to the decoding results and the candidate words respectively corresponding to the decoding results includes:

[0105] S301: summing up the weights corresponding to the plurality of decoding results corresponding to the same candidate word to obtain the scores corresponding to the candidate words; and summing up the weights corresponding to the decoding results to obtain the sum of the weights;

[0106] S302: Calculate the ratio of the score of each candidate word to the total weight to obtain the probability value corresponding to each candidate word.

[0107] Furthermore, the step of judging whether the degree of confusion of the decoding path of the lattice decoding graph is greater than a threshold value according to the probability values ​​corresponding to the candidate words includes:

[0108] S401: Substitute the probability values ​​corresponding to the candidate words into the second calculation formula to calculate the entropy value, wherein the entropy value represents the degree of confusion of the decoding path of the lattice decoding graph, and the second calculation formula is: , S is the entropy value, P i Represents the probability value of the i-th candidate word;

[0109] S402: Determine whether the entropy value is greater than an entropy value threshold;

[0110] S403: If the entropy value is greater than the entropy value threshold, it is determined that the degree of confusion of the decoding path of the lattice decoding graph is greater than the threshold;

[0111] S404: If the entropy value is less than the entropy value threshold, it is determined that the degree of confusion of the decoding path of the lattice decoding graph is less than the threshold.

[0112] Furthermore, the step of identifying and decoding the voice data to generate a lattice decoding graph includes:

[0113] S101: using a speech recognition tool to recognize and decode the speech data to obtain the lattice decoding graph;

[0114] S102: Arrange the decoding results contained in the lattice decoding graph in descending order according to their respective probability scores.

[0115] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0116] It should be noted that, in this article, the terms "include", "comprises" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, device, first object or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, first object or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, first object or method including the element.

[0117] The above description is only a preferred embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly used in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for detecting misidentification of noise, characterized in that: include: Collecting voice data, and performing recognition and decoding on the voice data to generate a lattice decoding graph, wherein the lattice decoding graph includes a plurality of decoding results; According to the arrangement order of each of the decoding results in the lattice decoding diagram, weights are assigned in turn according to a preset decreasing rule to obtain weights corresponding to each of the decoding results; According to the weights corresponding to the decoding results and the candidate words corresponding to the decoding results, the probability values ​​corresponding to the candidate words are calculated; According to the probability values ​​corresponding to the candidate words, determining whether the degree of confusion of the decoding path of the lattice decoding graph is greater than a threshold; If the degree of confusion of the decoding path of the lattice decoding graph is greater than a threshold, it is determined that the speech recognition result of the current speech data is a misrecognition of noise.

2. The method for detecting misidentification of noise according to claim 1, characterized in that: The step of assigning weights according to a preset decreasing rule in the order in which the decoding results are arranged in the lattice decoding diagram to obtain weights corresponding to the decoding results includes: Substitute the arrangement order of each decoding result in the lattice decoding diagram into the first calculation formula in order, and calculate the weights corresponding to each decoding result respectively, wherein the first calculation formula is: , where Q is the weight corresponding to the decoding result, n is the order of arrangement corresponding to the decoding result, and x is a decreasing value; When the calculated weight corresponding to the first decoding result is less than the weight threshold, the weights of the first decoding result and each second decoding result are set to a fixed value, wherein the second decoding result is all decoding results sorted after the first decoding result in the lattice decoding diagram.

3. The method for detecting misidentification of noise according to claim 1, characterized in that: The step of calculating the probability values ​​corresponding to a plurality of candidate words respectively according to the weights respectively corresponding to the decoding results and the candidate words respectively corresponding to the decoding results comprises: The weights corresponding to the plurality of decoding results corresponding to the same candidate word are added together to obtain the scores corresponding to the candidate words; and the weights corresponding to the decoding results are added together to obtain the weight sum; The ratio of the score of each candidate word to the total weight is calculated respectively to obtain the probability value corresponding to each candidate word.

4. The method for detecting misidentification of noise according to claim 1, characterized in that: The step of judging whether the degree of confusion of the decoding path of the lattice decoding graph is greater than a threshold value according to the probability values ​​corresponding to the candidate words respectively comprises: Substitute the probability values ​​corresponding to the candidate words into the second calculation formula to calculate the entropy value, wherein the entropy value represents the degree of confusion of the decoding path of the lattice decoding graph. The second calculation formula is: , S is the entropy value, P i Represents the probability value of the i-th candidate word; Determine whether the entropy value is greater than an entropy value threshold; If the entropy value is greater than the entropy value threshold, it is determined that the degree of confusion of the decoding path of the lattice decoding graph is greater than the threshold; If the entropy value is less than the entropy value threshold, it is determined that the degree of confusion of the decoding path of the lattice decoding graph is less than the threshold.

5. The method for detecting misidentification of noise according to claim 1, characterized in that: The step of identifying and decoding the voice data to generate a lattice decoding graph comprises: Using a speech recognition tool to recognize and decode the speech data to obtain the lattice decoding graph; Arrange the decoding results contained in the lattice decoding graph in descending order according to their respective probability scores.

6. A device for detecting misidentification of noise, characterized in that: include: A recognition module, used to collect voice data, recognize and decode the voice data, and generate a lattice decoding graph, wherein the lattice decoding graph includes a plurality of decoding results; An assignment module, used to assign weights according to a preset decreasing rule in accordance with the arrangement order of each decoding result in the lattice decoding diagram, to obtain weights corresponding to each decoding result; A calculation module, used for calculating the probability values ​​corresponding to a plurality of candidate words according to the weights corresponding to the decoding results and the candidate words corresponding to the decoding results; A judgment module, used for judging whether the degree of confusion of the decoding path of the lattice decoding graph is greater than a threshold value according to the probability values ​​corresponding to the candidate words; The determination module is used to determine that the speech recognition result of the current speech data is a noise misrecognition if the degree of confusion of the decoding path of the lattice decoding diagram is greater than a threshold.

7. The device for detecting misidentification of noise according to claim 6, characterized in that: The assignment module comprises: The first calculation unit is used to substitute the arrangement order of each decoding result in the lattice decoding diagram into the first calculation formula in sequence according to the arrangement order, and calculate the weights corresponding to each decoding result respectively, wherein the first calculation formula is: , where Q is the weight corresponding to the decoding result, n is the order of arrangement corresponding to the decoding result, and x is a decreasing value; A setting unit is used to set the weights of the first decoding result and each second decoding result to a fixed value when the weight corresponding to the calculated first decoding result is less than the weight threshold, wherein the second decoding result is all decoding results sorted after the first decoding result in the lattice decoding diagram.

8. The device for detecting misidentification of noise according to claim 6, characterized in that: The computing module comprises: A second calculation unit is used to sum the weights corresponding to the plurality of decoding results corresponding to the same candidate word to obtain the scores corresponding to the candidate words; and to sum the weights corresponding to the decoding results to obtain the weight sum; The third calculation unit is used to calculate the ratio of the score of each candidate word to the total weight, so as to obtain the probability value corresponding to each candidate word.

9. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images