A method, device, equipment and storage medium for determining speech recognition results

Abstract

The present application provides a method, device, equipment and storage medium for determining speech recognition results, wherein the method includes: obtaining candidate sets corresponding to target speech on multiple speech recognition systems, wherein the candidate set corresponding to the target speech on a speech recognition system includes several candidate recognition results obtained based on the speech recognition system; combining the obtained multiple candidate sets in pairs to obtain several groups of candidate sets; determining similarity information corresponding to the several groups of candidate sets, and determining the recognition result of the target speech from the multiple candidate sets based on the similarity information corresponding to the several groups of candidate sets, wherein the similarity information is sentence-level similarity information between the candidate recognition result contained in the first candidate set and the candidate recognition result contained in the second candidate set of the corresponding two candidate sets. The speech recognition result determination method provided by the present application can quickly determine a speech recognition result with higher accuracy.

Description

Technical Field

[0001] The present application relates to the field of speech recognition technology, and in particular to a speech recognition result determination method, device, equipment and storage medium. Background Art

[0002] Speech recognition is a technology that uses machines to automatically convert human speech into text. In recent years, with the widespread application of deep neural networks, speech recognition technology has made great progress.

[0003] The current method for determining speech recognition results is to recognize the target speech based on a speech recognition system to obtain a number of candidate recognition results, and then determine the candidate recognition result with the highest recognition score among the candidate recognition results as the recognition result of the target speech.

[0004] However, affected by certain factors (such as complex accents, dialects, distant speakers, noise, reverberation environments, semantic understanding, etc.), the recognition performance of speech recognition systems is unstable, which results in the above-mentioned solutions being insufficiently stable and sometimes failing to obtain highly accurate recognition results. Summary of the invention

[0005] In view of this, the present application provides a method, device, equipment and storage medium for determining speech recognition results, which are used to solve the problem that the existing speech recognition result determination method is not stable enough and sometimes cannot obtain a recognition result with high accuracy. The technical solution is as follows:

[0006] A method for determining a speech recognition result, comprising:

[0007] Acquire candidate sets corresponding to the target speech on multiple speech recognition systems respectively, wherein the candidate set corresponding to the target speech on one speech recognition system includes a number of candidate recognition results obtained based on the speech recognition system;

[0008] Combine the obtained multiple candidate sets in pairs to obtain several groups of candidate sets;

[0009] Determine similarity information corresponding to several groups of candidate sets, and determine the recognition result of the target speech from the multiple candidate sets based on the similarity information corresponding to the several groups of candidate sets, wherein the similarity information is sentence-level similarity information between the candidate recognition results contained in the first candidate set and the candidate recognition results contained in the second candidate set of the corresponding two candidate sets.

[0010] Optionally, obtain the candidate set corresponding to the target speech on a speech recognition system, including:

[0011] Obtaining multiple candidate recognition results output by the speech recognition system after recognizing the target speech;

[0012] Determining whether there is a candidate recognition result with a decoding abnormality among the multiple candidate recognition results;

[0013] If yes, the candidate recognition results with abnormal decoding are eliminated, and the remaining candidate recognition results form a candidate set corresponding to the target speech on the speech recognition system;

[0014] If not, the multiple candidate recognition results constitute a candidate set corresponding to the target speech on the speech recognition system.

[0015] Optionally, determining whether there is a candidate recognition result with a decoding abnormality among the multiple candidate recognition results includes:

[0016] Determine whether there is a candidate recognition result among the multiple candidate recognition results whose number of decoded words per second is greater than a preset word number threshold;

[0017] If so, determining that there is a candidate recognition result with a decoding abnormality among the multiple candidate recognition results;

[0018] If not, determining that there is no candidate recognition result with a decoding abnormality among the multiple candidate recognition results;

[0019] The step of eliminating candidate recognition results with abnormal decoding includes:

[0020] Eliminate candidate recognition results whose number of decoded words per second is greater than the preset word number threshold.

[0021] Optionally, determining the recognition result of the target speech from the multiple candidate sets based on the similarity information respectively corresponding to the multiple candidate sets includes:

[0022] Based on the similarity information respectively corresponding to the several groups of candidate sets, determine the candidate recognition result pairs respectively corresponding to the several groups of candidate sets to obtain several candidate recognition result pairs, wherein the candidate recognition result pairs include the two most similar candidate recognition results in the corresponding two candidate sets;

[0023] Based on the similarity information respectively corresponding to the plurality of candidate recognition result pairs, the recognition result of the target speech is determined, wherein the similarity information corresponding to a candidate recognition result pair is the similarity information of two candidate recognition results included in the candidate recognition result pair.

[0024] Optionally, the similarity information is a similarity measurement indicator;

[0025] The determining the recognition result of the target speech based on the similarity information respectively corresponding to the plurality of candidate recognition result pairs includes:

[0026] Based on the similarity measurement indicators respectively corresponding to the plurality of candidate recognition result pairs, determine the most similar candidate recognition result pair from the plurality of candidate recognition result pairs as the target candidate recognition result pair;

[0027] The recognition result of the target speech is determined based at least on the similarity measurement index corresponding to the target candidate recognition result pair.

[0028] Optionally, determining the recognition result of the target speech at least based on a similarity measurement indicator corresponding to the target candidate recognition result pair includes:

[0029] If the similarity measurement index corresponding to the target candidate recognition result pair is a target value that indicates that the two candidate recognition results are the same, any candidate recognition result in the target candidate recognition result pair is determined as the recognition result of the target speech;

[0030] If the similarity measurement index corresponding to the target candidate recognition result pair is not the target value, the recognition result of the target speech is determined based on the similarity measurement index corresponding to the target candidate recognition result pair and the number of words and the comprehensive score of each candidate recognition result in the target candidate recognition result pair, wherein the comprehensive score of a candidate recognition result is determined based on the acoustic score and the language score of the candidate recognition result.

[0031] Optionally, determining a comprehensive score of a candidate recognition result includes:

[0032] Determining a reasonableness representation value of the candidate recognition result according to the number of words in the candidate recognition result and the language score of the candidate recognition result;

[0033] Normalizing the reasonableness representation value of the candidate recognition result to a value within a target range to obtain a normalized reasonableness representation value, wherein the acoustic score of the candidate recognition result is within the target range;

[0034] The normalized rationality representation value is fused with the acoustic score of the candidate recognition result, and the fused score is used as the comprehensive score of the candidate recognition result.

[0035] Optionally, determining the recognition result of the target speech based on the similarity measurement index corresponding to the target candidate recognition result pair and the number of words and the comprehensive score of each candidate recognition result in the target candidate recognition result pair includes:

[0036] When the similarity measurement index of the target candidate recognition result pair and the preset similarity threshold satisfy a preset relationship:

[0037] If the number of words in the first candidate recognition result and the number of words in the second candidate recognition result in the target candidate recognition result pair are the same, determining the candidate recognition result with a larger comprehensive score in the target candidate recognition result pair as the recognition result of the target speech;

[0038] If the number of words in the first candidate recognition result is greater than the number of words in the second candidate recognition result, and the comprehensive score of the second candidate recognition result is greater than the comprehensive score of the first candidate recognition result, then determining the second candidate recognition result as the recognition result of the target speech;

[0039] If the number of words in the second candidate recognition result is greater than the number of words in the first candidate recognition result, and the comprehensive score of the first candidate recognition result is greater than the comprehensive score of the second candidate recognition result, the first candidate recognition result is determined as the recognition result of the target speech.

[0040] Optionally, the speech recognition result determination method further includes:

[0041] If it is determined that the recognition result of the target speech cannot be determined from the target candidate recognition result pair based on the similarity measurement index of the target candidate recognition result pair and the number of words and the comprehensive score of each candidate recognition result in the target candidate recognition result pair, then the candidate recognition result with the highest recognition score is obtained from each candidate set;

[0042] The candidate recognition result with the highest comprehensive score among the candidate recognition results with the highest recognition scores obtained from each candidate set is determined as the recognition result of the target speech.

[0043] A speech recognition result determination device comprises: a candidate set acquisition module, a candidate set combination module, a similarity information determination module and a recognition result determination module;

[0044] The candidate set acquisition module is used to acquire candidate sets corresponding to the target speech on multiple speech recognition systems, wherein the candidate set corresponding to the target speech on a speech recognition system includes a number of candidate recognition results obtained based on the speech recognition system;

[0045] The candidate set combination module is used to combine the obtained multiple candidate sets in pairs to obtain several groups of candidate sets;

[0046] The similarity information determination module is used to determine the similarity information corresponding to the plurality of candidate sets, wherein the similarity information is sentence-level similarity information between the candidate recognition results included in the first candidate set and the candidate recognition results included in the second candidate set of the corresponding two candidate sets;

[0047] The recognition result determination module is used to determine the recognition result of the target speech from the multiple candidate sets based on the similarity information respectively corresponding to the multiple candidate sets.

[0048] A speech recognition result determination device, comprising: a memory and a processor;

[0049] The memory is used to store programs;

[0050] The processor is used to execute the program to implement each step of the method for determining the speech recognition result described in any one of items 1 to 9 above.

[0051] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements each step of the method for determining a speech recognition result described in any one of items 1 to 9 above.

[0052] The speech recognition result determination method, apparatus, device and storage medium provided in the present application first obtain candidate sets corresponding to the target speech on multiple speech recognition systems respectively, then combine the obtained multiple candidate sets in pairs to obtain several groups of candidate sets, then determine the similarity information corresponding to the several groups of candidate sets respectively, and finally determine the recognition result of the target speech from the multiple candidate sets based on the similarity information corresponding to the several groups of candidate sets respectively. Considering the differences and complementarity of multiple speech recognition systems, the present application determines the recognition result of the target speech from the candidate sets corresponding to the target speech on the multiple speech recognition systems. Compared with determining the recognition result of the target speech from the candidate recognition results obtained based on a single speech recognition system, the speech recognition result determination method provided by the present application is more stable and can improve the accuracy of the speech recognition result. When determining the recognition result of the target speech from the candidate sets corresponding to the target speech on the multiple speech recognition systems, in order to achieve the complementarity of the candidate recognition results of different speech recognition systems, and at the same time to determine the speech recognition result faster, the present application combines the candidate sets corresponding to the target speech on the multiple speech recognition systems in pairs, and based on the sentence-level similarity information corresponding to each group of candidate sets, determines the recognition result of the target speech from the candidate sets corresponding to the multiple speech recognition systems. The speech recognition result determination method provided by the present application can quickly determine a speech recognition result with a higher accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0054] Figure 1 A flowchart of a method for determining a speech recognition result provided in an embodiment of the present application;

[0055] Figure 2 A schematic diagram of a process for determining a recognition result of a target speech from multiple candidate sets based on similarity information corresponding to a plurality of candidate sets provided in an embodiment of the present application;

[0056] Figure 3 A schematic diagram of a process for determining a recognition result of a target speech based on similarity information corresponding to a plurality of candidate recognition result pairs provided in an embodiment of the present application;

[0057] Figure 4 A schematic diagram of the structure of a device for determining a speech recognition result provided in an embodiment of the present application;

[0058] Figure 5 A schematic diagram of the structure of a speech recognition result determination device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] In order to solve the problem that the current speech recognition result determination scheme is unstable and sometimes cannot obtain a high-accuracy recognition result, the applicant thought that for the same speech, the recognition results output by different speech recognition systems are quite different. When the recognition results of multiple speech recognition systems are complementary (for example, system A is unstable and does not output excellent recognition results, while system B is relatively stable and outputs excellent recognition results), the recognition results of these systems are integrated to improve the accuracy of the speech recognition results. Following the above ideas, the applicant conducted research. At first, the applicant thought of the following speech recognition result determination method:

[0061] A linear word transfer network WTN is established for the output of different speech recognition systems. The WTNs are sorted from small to large according to the word error rate WER, and the WTN corresponding to the minimum WER is selected as the benchmark WTN. The remaining WTNs are continuously aligned and merged using the benchmark WTN until all WTNs are merged into the benchmark WTN. After completing the alignment and merging of all WTNs, a voting scheme, such as word frequency, confidence, etc., is used to evaluate each branch point of the WTN, and then the word with the highest score (with the highest number of votes) is selected as the final recognition result.

[0062] The applicant studied the above method and found that there are some problems with the above method, which are mainly reflected in the following aspects: First, the above method is performed at the word level, and requires complex and tedious operations such as alignment, merging, voting, re-decoding, and word order connection. These operations take a long time. Therefore, the efficiency of determining the speech recognition results is low. If the number of fusion systems increases, the efficiency of determining the speech recognition results will be further reduced, which is very unfavorable for application scenarios with high response speed requirements; second, the above method can only be applied to the 1-Best candidate results output by different speech recognition systems, that is, the final speech recognition results are determined from the 1-Best candidate results output by each speech recognition system. However, in some cases, better recognition results may appear in other results in N-Best except 1-Best. In this case, the above method cannot obtain recognition results with higher accuracy.

[0063] In response to the problems existing in the above-mentioned method, the applicant continued to conduct research, and through continuous research, finally proposed a method for determining speech recognition results with better effects. The method is relatively stable and can determine accurate speech recognition results relatively quickly. The method can be applied to electronic devices with data processing capabilities. The electronic device can be a terminal used on the user side, such as a smart phone, PC, notebook, PAD, smart home appliance, car terminal, etc. The electronic device can also be a server on the network side. The server can be a single server, a server cluster composed of multiple servers, or a cloud computing server center. The electronic device can quickly determine accurate speech recognition results according to the speech recognition result determination method provided in this application.

[0064] Those skilled in the art should understand that the above-mentioned electronic devices are only examples, and other existing or future electronic devices or servers that are applicable to the present application should also be included in the protection scope of the present application and are included here by reference.

[0065] Next, the method for determining the speech recognition result provided by the present application is introduced through the following embodiments.

[0066] First embodiment

[0067] See also Figure 1 , which shows a flow chart of a method for determining a speech recognition result provided by an embodiment of the present application, which may include:

[0068] Step S101: Obtain candidate sets corresponding to the target speech in multiple speech recognition systems.

[0069] The target speech is the speech for which the speech recognition result is to be determined, and the candidate set corresponding to the target speech on a speech recognition system includes several candidate recognition results obtained based on the speech recognition system.

[0070] Multiple candidate sets may be obtained through step S101. For example, if there are three speech recognition systems, three candidate sets may be obtained through step S101.

[0071] Since the implementation method of obtaining each candidate set is the same, in this embodiment, a candidate set R is obtained. i (the candidate set corresponding to the target speech in the i-th speech recognition system) is taken as an example, to obtain the candidate set corresponding to the target speech in the i-th speech recognition system S i The corresponding candidate set R i The implementation method is introduced in this chapter.

[0072] Get the target speech in the i-th speech recognition system S i The corresponding candidate set R i There are many implementation methods, and this embodiment provides the following two optional implementation methods:

[0073] The first implementation method: obtaining the speech recognition system S i After recognizing the target speech, the speech recognition system S i The candidate set composed of multiple candidate recognition results outputted is used as the target speech in the speech recognition system S i The corresponding candidate set R i .

[0074] Among them, the speech recognition system S i The multiple candidate recognition results output after recognizing the target speech can be all the candidate recognition results or the N-best of all the candidate recognition results. Generally speaking, the speech recognition system S i After recognizing the target speech, N-best, i.e., the top N (N is an integer greater than or equal to 1) candidate recognition results will be output. It should be noted that N is set according to the specific application situation. The N of different speech recognition systems can be the same or different. In addition, the multiple candidate recognition results output by the speech recognition system are sorted. For example, the speech recognition system ranks the one with the highest recognition score in the first place, the one with the second highest recognition score in the second place, and so on.

[0075] Assume that the speech recognition system S i After recognizing the target speech, output N i candidate recognition results, the first implementation method directly converts N i The candidate set composed of candidate recognition results is used as the target speech in the speech recognition system S i The corresponding candidate set R i It can be seen that if the target speech is obtained in the speech recognition system S through the first implementation method mentioned above i The corresponding candidate set R i , then the candidate set R i The speech recognition system S i All candidate recognition results output after recognizing the target speech.

[0076] Considering the speech recognition system S i Among the multiple candidate recognition results output after the target speech is recognized, there may be candidate recognition results with decoding abnormalities. In order to avoid the influence of the candidate recognition results with decoding abnormalities on the determination of the final recognition result and to improve the determination efficiency of the recognition result, this embodiment provides the following second implementation method:

[0077] First, obtain the speech recognition system S i After recognizing the target speech, multiple candidate recognition results (such as N-best candidate recognition results) are output; then the speech recognition system S is determined i Is there any recognition result with decoding anomaly among the multiple candidate recognition results output? i If there are recognition results with decoding anomalies among the multiple candidate recognition results output, the candidate recognition results with decoding anomalies will be eliminated, and the remaining candidate recognition results will form the target speech in the speech recognition system S i The corresponding candidate set R i ; If the speech recognition system S i If there is no candidate recognition result with decoding abnormality among the multiple candidate recognition results output, the speech recognition system S i The output of multiple candidate recognition results constitutes the target speech in the speech recognition system S i The corresponding candidate set R i .

[0078] For example, the speech recognition system S i After recognizing the target speech, output N i candidate recognition results, if N i If there are three candidate recognition results with abnormal decoding among the candidate recognition results, the three candidate recognition results with abnormal decoding will be eliminated, and the remaining N i-3 candidate recognition results constitute the target speech in the speech recognition system S i The corresponding candidate set R i , if N i There is no candidate recognition result with decoding anomaly among the candidate recognition results, then N i The candidate recognition results constitute the target speech in the speech recognition system S i The corresponding candidate set R i .

[0079] It can be seen that if the target speech is obtained in the speech recognition system S through the second implementation method mentioned above i The corresponding candidate set R i , then the candidate set R i The speech recognition system S i All (there are no candidate recognition results with decoding anomalies among the multiple candidate recognition results output) or some (there are candidate recognition results with decoding anomalies among the multiple candidate recognition results output) candidate recognition results output after recognizing the target speech.

[0080] In the second implementation method, after obtaining the speech recognition system S i After the target speech is recognized and multiple candidate recognition results are output, it is necessary to determine the speech recognition system S i Whether there are any recognition results with decoding anomalies among the multiple candidate recognition results outputted, this process is introduced below.

[0081] Determine the speech recognition system S i Whether there is a recognition result with decoding anomaly among the output multiple candidate recognition results may include: determining whether the speech recognition system S i Is there a number of candidate recognition results in which the number of decoded words per unit time is greater than the preset word number threshold e? thred Candidate recognition result; if so, determine the speech recognition system S i There are candidate recognition results with decoding anomalies among the multiple candidate recognition results output; if not, it is determined that the speech recognition system S i There is no candidate recognition result with abnormal decoding in the output multiple candidate recognition results. Among them, the preset word count threshold e thred Can be set according to specific application.

[0082] Among them, a candidate recognition result r ij The number of decoded words per unit time is based on the candidate recognition result r ij The decoding time t ij and candidate recognition result r ij The number of words q ij Determine, specific, candidate recognition results r ij The number of words decoded per unit time eij It can be expressed as q ij / t ij .

[0083] It should be noted that if the number of decoded words per unit time of a candidate recognition result is greater than the preset word number threshold e thred , it can be determined that the candidate recognition result is decoded abnormally. In the second implementation, the candidate recognition results with decoding abnormalities are eliminated, that is, the number of words decoded per unit time is greater than the preset word number threshold e thred The candidate recognition results are eliminated.

[0084] Step S102: Combine the obtained multiple candidate sets in pairs to obtain several groups of candidate sets.

[0085] Assuming that M candidate sets are obtained in step S101, by combining the M candidate sets in pairs, M(M-1) / 2 groups of candidate sets will be obtained.

[0086] Exemplarily, three candidate sets are obtained through step S101, namely R1, R2 and R3. Combining R1, R2 and R3 in pairs will obtain three candidate sets, namely (R1, R2), (R1, R3) and (R2, R3).

[0087] Step S103: Determine the similarity information corresponding to several groups of candidate sets.

[0088] The similarity information is sentence-level similarity information between the candidate recognition results included in the first candidate set and the candidate recognition results included in the second candidate set in the corresponding two candidate sets.

[0089] For example, a candidate set is (R i , R j ), where R i = {r i1 , r i2 , r i3}, R j = {r j1 , r j2 , r j3}, then (R i , R j ) The corresponding similarity information may include r i1 Respectively with r j1 、r j2 、r j3 The sentence-level similarity information, r i2 Respectively with r j1 、r j2 、r j3 The sentence-level similarity information, r i3 Respectively with r j1 、rj2 、r j3 The sentence-level similarity information.

[0090] Step S104: based on the similarity information corresponding to the plurality of candidate sets, determining the recognition result of the target speech from the plurality of candidate sets.

[0091] Considering the differences and complementarity of multiple speech recognition systems, this embodiment determines the recognition result of the target speech from the candidate sets corresponding to the target speech on the multiple speech recognition systems. Compared with determining the recognition result of the target speech from the candidate recognition results obtained based on a single speech recognition system, the speech recognition result determination method provided by this application is more stable and can improve the accuracy of the speech recognition result. When determining the recognition result of the target speech from the candidate sets corresponding to the target speech on the multiple speech recognition systems, in order to achieve the complementarity of the candidate recognition results of different speech recognition systems and to determine the speech recognition result faster, this application combines the candidate sets corresponding to the target speech on the multiple speech recognition systems in pairs, and based on the sentence-level similarity information corresponding to each group of candidate sets, determines the recognition result of the target speech from the candidate sets corresponding to the multiple speech recognition systems. The speech recognition result determination method provided by this embodiment can quickly determine a speech recognition result with a higher accuracy.

[0092] Second embodiment

[0093] This embodiment introduces the specific implementation process of "step S103: determining similarity information corresponding to several groups of candidate sets" in the above embodiment.

[0094] Optionally, the similarity information may be a similarity metric, which is an indicator that can measure the degree of similarity. Assume that a set of candidate sets (R i , R j ), then (R i , R j ) The corresponding similarity metrics include R i Each candidate recognition result is related to R j The similarity metric for each candidate recognition result in .

[0095] For example, R i = {r i1 , r i2}, R2={r j1 , r j2}, then (R i , R j ) The corresponding similarity metrics include R i The candidate recognition result r in i1 With R jThe candidate recognition result r in j1 Similarity measurement index, R i The candidate recognition result r in i1 With R j The candidate recognition result r in j2 Similarity measurement index, R i The candidate recognition result r in i2 With R j The candidate recognition result r in j1 Similarity measurement index, R i The candidate recognition result r in i2 With R j The candidate recognition result r in j2 Similarity measurement index.

[0096] It should be noted that the similarity measurement index can adopt the first measurement index or the second measurement index, wherein the smaller the first measurement index of the two candidate recognition results is, the more similar the two candidate recognition results are, and the larger the second measurement index of the two candidate recognition results is, the more similar the two candidate recognition results are. Optionally, the first measurement index can be but not limited to edit distance, Euclidean distance, etc., and the second measurement index can be but not limited to cosine similarity, etc.

[0097] Among them, the edit distance between two candidate recognition results (i.e., a set of candidate sets) can be determined as follows:

[0098] Define d[k1][k2], where k1 corresponds to the first candidate recognition result of the two candidate recognition results, k1∈[1,n1], n1 represents the number of words in the first candidate recognition result, and k2 corresponds to the second candidate recognition result of the two candidate recognition results, k2∈[1,n2], n2 represents the number of words in the second candidate recognition result. For k1 from 1 to n1 and k2 from 1 to n2, calculate d[k1][k2] in sequence as follows:

[0099] When the first candidate recognition result at k1-1 is equal to the second candidate recognition result at k2-1, d[k1][k2]=d[k1-1][k2-1], otherwise:

[0100] d[k1][k2]=min{d[k1-1][k2],d[k1][k2-1],d[k1-1][k2-1]}+1

[0101] After calculation, we can finally get d[n1][n2], which is the edit distance between the two candidate recognition results.

[0102] For example, the first candidate recognition result of the two candidate recognition results is "my name is jack" and the second candidate recognition result is "my nick name tom". Then, the following can be obtained by calculation in the above manner:

[0103]

[0104] Since d[4][4]=3, the edit distance between the two candidate recognition results is 3.

[0105] Among them, the cosine similarity of two candidate recognition results can be determined as follows:

[0106] Determine the representation vector of each candidate recognition result, calculate the cosine similarity of the representation vectors of the two candidate recognition results, and use the calculated cosine similarity as the cosine similarity of the two candidate recognition results. The representation vector of a candidate recognition result can be obtained by converting each word included in the candidate recognition result into a corresponding word vector.

[0107] The calculation method for determining the Euclidean distance between two candidate recognition results is similar to the method for determining the cosine similarity. Both methods first determine the representation vector of each candidate recognition result, and then determine the Euclidean distance between the representation vectors of the two candidate recognition results. The calculated Euclidean distance is used as the Euclidean distance between the two candidate recognition results.

[0108] It should be noted that the similarity measurement indicators given above (ie, edit distance, cosine similarity, and Euclidean distance) are only examples, and other indicators that can measure the similarity between two candidate recognition results are also applicable to the present application.

[0109] Third embodiment

[0110] This embodiment introduces the specific implementation process of "step S104: determining the recognition result of the target speech from multiple candidate sets based on the similarity information corresponding to the multiple candidate sets" in the above embodiment.

[0111] See also Figure 2 , showing a flow chart of determining the recognition result of a target speech from a plurality of candidate sets based on the similarity information corresponding to a plurality of candidate sets, which may include:

[0112] Step S201: based on similarity information corresponding to several groups of candidate sets, respectively, determining candidate recognition result pairs corresponding to several groups of candidate sets, to obtain several candidate recognition result pairs.

[0113] The candidate recognition result pair includes the two most similar candidate recognition results in the corresponding two candidate sets.

[0114] For example, a candidate set is (R1, R2), if the candidate recognition result r in R1 13 and the candidate recognition result r in R2 21 The most similar, then (r 13 ,r 21 ) is determined as the target candidate recognition result pair corresponding to (R1, R2).

[0115] Specifically, based on the similarity information corresponding to several groups of candidate sets, the process of determining the candidate recognition result pairs corresponding to several groups of candidate sets can include: for each group of candidate sets, if the similarity measurement index adopts the first measurement index, then the candidate recognition result pair consisting of the two candidate recognition results with the smallest similarity measurement index (for example, the smallest edit distance) is determined as the candidate recognition result pair corresponding to the group of candidate sets; if the similarity measurement index adopts the second measurement index, then the candidate recognition result pair consisting of the two candidate recognition results with the largest similarity measurement index (for example, the largest cosine similarity) is determined as the candidate recognition result pair corresponding to the group of candidate sets.

[0116] For example, the similarity metric is edit distance. There are three candidate sets, namely (R1, R2), (R1, R3) and (R2, R3). Assume that R1 = {r 11 , r 12}, R2={r 21 , r 22}, R2={r 31 , r 32 , r 33}, for (R1, R2), the similarity information corresponding to (R1, R2) includes r 11 With r 21 The edit distance, r 11 With r 22 The edit distance, r 12 With r 21 The edit distance, r 12 With r 22 The edit distance, assuming r 12 With r 21 The edit distance of the candidate recognition result pair (r 12 , r 21 ) is determined as the candidate set for the candidate recognition result pair corresponding to (R1, R2). The same method can be used to determine the candidate recognition result pair corresponding to (R1, R3) and the candidate recognition result pair corresponding to (R2, R3) and (R2, R3). Finally, 3 candidate recognition result pairs are obtained.

[0117] It should be noted that for a set of candidate sets, if there are multiple pairs of the most similar candidate recognition results in the two candidate sets it contains (for example, there are multiple pairs of candidate recognition results with the smallest edit distance), then the pair of candidate recognition results with the largest sum of comprehensive scores is used as the pair of candidate recognition results corresponding to this set of candidate sets. Exemplarily, a set of candidate sets is (R1, R2), R1 = {r 11 , r 12}, R2 = {r 21 , r 22}, where the edit distance between r 11 and r 21 is d1, the edit distance between r 11 and r 22 is d2, the edit distance between r 12 and r 21 is d3, the edit distance between r 12 and r 22 is d4. Assuming d1 = d2 < d3 < d4, then the comprehensive scores of r 11 and r 21 are summed, and the sum of the scores is denoted as s1. The comprehensive scores of r 11 and r 22 are summed, and the sum of the scores is denoted as s2. If s1 > s2, then (r 11 , r 21 ) is determined as the pair of candidate recognition results corresponding to (R1, R2). If s1 < s2, then (r 11 , r 22 ) is determined as the pair of candidate recognition results corresponding to (R1, R2). If s1 = s2, then either (r 11 , r 21 ) or (r 11 , r 22 ) is determined as the pair of candidate recognition results corresponding to (R1, R2).

[0118] Step S202: Based on the similarity information corresponding to each of the several pairs of candidate recognition results, determine the recognition result of the target speech.

[0119] Among them, the similarity information corresponding to a pair of candidate recognition results is the similarity information of the two candidate recognition results included in this pair of candidate recognition results.

[0120] Please refer to Figure 3 , which shows a schematic flowchart of determining the recognition result of the target speech based on the similarity information corresponding to each of the several pairs of candidate recognition results, and may include:

[0121] Step S301: Based on the similarity metric indicators corresponding to each of the several pairs of candidate recognition results, determine the pair of the most similar candidate recognition results from the several pairs of candidate recognition results as the target pair of candidate recognition results.

[0122] Among them, the similarity metric index corresponding to a pair of candidate recognition results is the similarity metric index of the two candidate recognition results included in the pair of candidate recognition results. Exemplarily, if the edit distance is used as the similarity metric, the edit distance corresponding to a pair of candidate recognition results is the edit distance of the two candidate recognition results included in the pair of candidate recognition results.

[0123] Exemplarily, there are 3 candidate sets, and accordingly 3 pairs of candidate recognition results are obtained. Among them, the edit distance corresponding to the first pair of candidate recognition results is d1, the edit distance corresponding to the second pair of candidate recognition results is d2, and the edit distance corresponding to the third pair of candidate recognition results is d3. If d3 < d2 < d1, it indicates that the third pair of candidate recognition results is the most similar pair of recognition results, and the third pair of candidate recognition results is determined as the target pair of candidate recognition results.

[0124] It should be noted that if there are multiple pairs of candidate recognition results that are the most similar among several pairs of candidate recognition results, the pair of candidate recognition results with the highest comprehensive score is determined as the target pair of candidate recognition results. Exemplarily, there are 3 candidate sets, and accordingly 3 pairs of candidate recognition results are obtained. Among them, the edit distance corresponding to the first pair of candidate recognition results is d1, the edit distance corresponding to the second pair of candidate recognition results is d2, and the edit distance corresponding to the third pair of candidate recognition results is d3. If d3 = d2 < d1, the comprehensive scores of the two candidate recognition results included in the second pair of candidate recognition results are summed, and the comprehensive scores of the two candidate recognition results included in the third pair of candidate recognition results are summed. If the sum of the comprehensive scores of the two candidate recognition results included in the second pair of candidate recognition results is greater than the sum of the comprehensive scores of the two candidate recognition results included in the third pair of candidate recognition results, the second pair of candidate recognition results is determined as the target pair of candidate recognition results. If the sum of the comprehensive scores of the two candidate recognition results included in the third pair of candidate recognition results is greater than the sum of the comprehensive scores of the two candidate recognition results included in the second pair of candidate recognition results, the third pair of candidate recognition results is determined as the target pair of candidate recognition results. If the sum of the comprehensive scores of the two candidate recognition results included in the third pair of candidate recognition results is the same as the sum of the comprehensive scores of the two candidate recognition results included in the second pair of candidate recognition results, any one of the second pair of candidate recognition results and the third pair of candidate recognition results is determined as the target pair of candidate recognition results.

[0125] Step S302: Determine the recognition result of the target speech based at least on the similarity metric index corresponding to the target pair of candidate recognition results.

[0126] Specifically, determining the recognition result of the target speech based at least on the similarity metric index corresponding to the target pair of candidate recognition results includes:

[0127] Step S302a: If the similarity measurement index corresponding to the target candidate recognition result pair is a target value that indicates that the two candidate recognition results are the same, any candidate recognition result in the target candidate recognition result pair is determined as the recognition result of the target speech.

[0128] Exemplarily, the similarity measurement indicator uses the edit distance, and the target value representing that two candidate recognition results are the same is 0. If the edit distance corresponding to the target candidate recognition result pair is 0, it means that the two candidate recognition results are the same, and any candidate recognition result in the target candidate recognition result pair can be determined as the recognition result of the target speech.

[0129] Exemplarily, the similarity measurement indicator uses cosine similarity, and the target value representing that two candidate recognition results are the same is 1. If the cosine similarity corresponding to the target candidate recognition result pair is 1, it means that the two candidate recognition results are the same, and any candidate recognition result in the target candidate recognition result pair can be determined as the recognition result of the target speech.

[0130] Step S302b: If the similarity measurement index corresponding to the target candidate recognition result pair is not the target value, the recognition result of the target speech is determined based on the similarity measurement index corresponding to the target candidate recognition result pair and the number of words and the comprehensive score of each candidate recognition result in the target candidate recognition result pair.

[0131] Among them, the comprehensive score of a candidate recognition result is determined based on the acoustic score and language score of the candidate recognition result. It should be noted that each candidate recognition result has two corresponding scores, namely the acoustic score (the score of the acoustic model) and the language score (the score of the language model). The acoustic score and the language score are output together by the corresponding speech recognition system when outputting the candidate recognition result. For example, a speech recognition system outputs 10 candidate recognition results. When outputting these 10 candidate recognition results, it also outputs the acoustic score and language score of each candidate recognition result.

[0132] There are many implementation methods for determining the comprehensive score of a candidate recognition result based on the acoustic score and language score of the candidate recognition result. This embodiment provides the following two optional implementation methods:

[0133] The first implementation method:

[0134] Step a1: Processing the language score of the candidate recognition result into a value within a target range to obtain a processed language score.

[0135] Among them, the acoustic score of the candidate recognition result is within the target range.

[0136] Step a2: Fusing the processed language score with the acoustic score of the candidate recognition result, and using the fused score as the comprehensive score of the candidate recognition result.

[0137] There are many ways to merge the processed language score with the acoustic score of the candidate recognition result. For example, the processed language score and the acoustic score of the candidate recognition result can be directly summed, or the processed language score and the acoustic score of the candidate recognition result can be weightedly summed.

[0138] The second implementation method:

[0139] Step b1: Determine a reasonableness representation value of the candidate recognition result according to the number of words in the candidate recognition result and the language score of the candidate recognition result.

[0140] Assume that the candidate recognition result is the jth candidate recognition result r in the candidate set corresponding to the target speech on the i-th speech recognition system. ij , then the candidate recognition result r can be determined according to the following formula ij The reasonableness of the value z ij :

[0141] z ij =q ij / ln(ppl ij +∈) (1)

[0142] Among them, q ij Represents the candidate recognition result r ij The number of words, that is, the candidate recognition result r ij The number of words contained, ppl ij Represents the candidate recognition result r ij The language score of ∈ is 10 -6 .

[0143] Step b2: normalize the reasonableness representation value of the candidate recognition result to a value within a target range to obtain a normalized reasonableness representation value.

[0144] Among them, the acoustic score of the candidate recognition result is within the target range.

[0145] Specifically, the candidate recognition result r is normalized as shown in the following formula ij The reasonableness of the value z ij Normalize to values ​​within the target range:

[0146]

[0147] Where c represents the smoothing factor, which is an adjustable parameter, and l ij That is, it represents the normalized reasonable degree representation value.

[0148] Step b3: Fusing the normalized rationality representation value with the acoustic score of the candidate recognition result, and using the fused score as the comprehensive score of the candidate recognition result.

[0149] In one possible implementation, the normalized reasonableness representation value and the acoustic score of the candidate recognition result may be directly summed, and the summed score is used as the comprehensive score of the candidate recognition result; in another possible implementation, the normalized reasonableness representation value and the acoustic score of the candidate recognition result may be weighted summed, and the weighted summed score is used as the comprehensive score of the candidate recognition result, that is:

[0150] s ij =αl ij +βa ij (3)

[0151] Among them, s ij That is, the candidate recognition result r ij The comprehensive score of ij Represents the candidate recognition result r ij The acoustic score of , α represents the language model score factor, β represents the acoustic model score factor, α+β=1, α and β are adjustable parameters.

[0152] Compared with the first implementation, the second implementation also considers the number of words in the candidate recognition result. Since the quality of the candidate recognition result is related to the number of words, the comprehensive score determined by combining the number of words in the candidate recognition result can better reflect the quality of the candidate recognition result. In view of this, the second implementation is preferred to determine the comprehensive score of the candidate recognition result.

[0153] Next, the implementation process of "determining the recognition result of the target speech based on the similarity measurement indicators corresponding to the target candidate recognition result pair and the number of words and the comprehensive score of each candidate recognition result in the target candidate recognition result pair" in step S302b is introduced.

[0154] Based on the similarity measurement index corresponding to the target candidate recognition result pair and the word count and comprehensive score of each candidate recognition result in the target candidate recognition result pair, the process of determining the recognition result of the target speech may include:

[0155] When the similarity measurement index of the target candidate recognition result pair and the preset similarity threshold satisfy the preset relationship:

[0156] (1) If the number of words in the first candidate recognition result and the number of words in the second candidate recognition result in the target candidate recognition result pair are the same, the candidate recognition result with the larger comprehensive score in the target candidate recognition result pair is determined as the recognition result of the target speech.

[0157] It should be noted that the number of words in the first candidate recognition result and the second candidate recognition result in the target candidate recognition result pair is the same, which means that the number of replacements, deletions, and insertions that may occur in the first candidate recognition result and the second candidate recognition result is the same. In this case, the candidate recognition result with a larger comprehensive score between the first candidate recognition result and the second candidate recognition result is determined as the recognition result of the target speech.

[0158] (2) If the number of words in the first candidate recognition result is greater than the number of words in the second candidate recognition result, and the comprehensive score of the second candidate recognition result is greater than the comprehensive score of the first candidate recognition result, the second candidate recognition result is determined as the recognition result of the target speech.

[0159] It should be noted that the number of words in the first candidate recognition result of the target candidate recognition result pair is greater than the number of words in the second candidate recognition result, which means that the first candidate recognition result may have more insertion errors than the second candidate recognition result. In this case, the second candidate recognition result with fewer insertion errors and a higher comprehensive score is taken as the recognition result of the target speech.

[0160] (3) If the number of words in the second candidate recognition result is greater than the number of words in the first candidate recognition result, and the comprehensive score of the first candidate recognition result is greater than the comprehensive score of the second candidate recognition result, the first candidate recognition result is determined as the recognition result of the target speech.

[0161] It should be noted that the number of words in the second candidate recognition result of the target candidate recognition result pair is greater than the number of words in the first candidate recognition result, indicating that the second candidate recognition result may have more insertion errors than the first candidate recognition result. In this case, the first candidate recognition result with fewer insertion errors and a higher comprehensive score is taken as the recognition result of the target speech.

[0162] Exemplarily, the target candidate recognition result pair is (r ip 、r jq ), (r ip 、r jq )The corresponding similarity measurement index is (r ip 、r jq ) corresponds to the edit distance, i.e. r ip With r jq The edit distance is based on (r ip 、r jq ) and the word count and comprehensive score of each candidate recognition result in the target candidate recognition result pair, the process of determining the recognition result of the target speech includes: ip 、r jq ) is less than the preset edit distance threshold, if r ip The number of words and r jq If the number of words is the same, rip 、r jq Any one of the above is determined as the recognition result of the target speech; if r ip The number of words is greater than r jq The number of words, and r jq The comprehensive score is greater than r ip The comprehensive score of the input is r, which has fewer errors and a higher comprehensive score. jq Determine the recognition result of the target speech; if r jq The number of words is greater than r ip The number of words, r ip The comprehensive score is greater than r jq The comprehensive score of the input is r, which has fewer errors and a higher comprehensive score. ip The recognition result of the target speech is determined.

[0163] It should be noted that: if the similarity measurement index of the target candidate recognition result pair is not the target value, and the similarity measurement index of the target candidate recognition result pair does not meet the preset conditions with the preset similarity threshold (for example, the edit distance of the target candidate recognition result pair is not less than the preset edit distance threshold), it is determined that the recognition result of the target speech cannot be determined from the target candidate recognition result pair; when the similarity measurement index of the target candidate recognition result pair is not the target value, and the similarity measurement index of the target candidate recognition result pair meets the preset conditions with the preset similarity threshold, if the number of words in the first candidate recognition result is greater than the number of words in the second candidate recognition result If the number of words in the second candidate recognition result is greater than the number of words in the first candidate recognition result, and the comprehensive score of the first candidate recognition result is greater than or equal to the comprehensive score of the second candidate recognition result, it is determined that the recognition result of the target speech cannot be determined from the target candidate recognition result pair; when the similarity measurement index of the target candidate recognition result pair is not the target value, and the similarity measurement index of the target candidate recognition result pair and the preset similarity threshold meet the preset conditions, if the number of words in the second candidate recognition result is greater than the number of words in the first candidate recognition result, and the comprehensive score of the second candidate recognition result is greater than or equal to the comprehensive score of the first candidate recognition result, it is determined that the recognition result of the target speech cannot be determined from the target candidate recognition result pair.

[0164] When it is determined that the recognition result of the target speech cannot be determined from the target candidate recognition result pair, the present application adopts the following strategy to determine the recognition result of the target speech: obtain the candidate recognition result with the highest recognition score from each candidate set (such as 1-best output by each speech recognition system); and determine the candidate recognition result with the highest comprehensive score among the candidate recognition results with the highest recognition scores obtained from each candidate set as the recognition result of the target speech.

[0165] It can be known from the above-mentioned first and second embodiments that the method for determining the speech recognition result provided in the embodiment of the present application can obtain the N-Best candidate recognition results output by multiple different speech recognition systems respectively, and the candidate recognition results with decoding abnormalities in the N-Best candidate recognition results output by each speech recognition system can be eliminated to obtain the candidate sets corresponding to the target speech on different speech recognition systems respectively, and the candidate sets corresponding to the target speech on different speech recognition systems can also be combined in pairs, and for each group of candidate sets, the two most similar candidate recognition results are found from the two candidate sets, so as to obtain the candidate recognition result pairs corresponding to each group of candidate sets respectively, and then the most similar candidate recognition result pairs are found from the candidate recognition result pairs corresponding to each group of candidate sets respectively, and finally the recognition result of the target speech is determined based on the relevant information of the most similar candidate recognition result pairs (such as the similarity measurement index corresponding to the most similar candidate recognition result pairs, the number of times and the comprehensive score of each candidate recognition result contained in the most similar candidate recognition result pairs). The present application determines the recognition result of the target speech from the candidate sets corresponding to the target speech on different speech recognition systems. Since the candidate recognition results output by different speech recognition systems are complementary, compared with determining the recognition result of the target speech from the candidate recognition results output by a single speech recognition system, the speech recognition result determination method provided by the present application is more stable and can improve the accuracy of the speech recognition result. When determining the recognition result of the target speech from the candidate sets corresponding to the target speech on multiple speech recognition systems, in order to achieve the complementarity of the candidate recognition results of different speech recognition systems and to determine the speech recognition result faster, the present application combines the candidate sets corresponding to the target speech on multiple speech recognition systems in pairs, and based on the sentence-level similarity information corresponding to each group of candidate sets, determines the recognition result of the target speech from the candidate sets corresponding to the multiple target speech on multiple speech recognition systems. The speech recognition result determination method provided by the present application can quickly determine a speech recognition result with a high accuracy.

[0166] Fourth embodiment

[0167] An embodiment of the present application also provides a speech recognition result determination device. The speech recognition result determination device provided in the embodiment of the present application is described below. The speech recognition result determination device described below and the speech recognition result determination method described above can refer to each other.

[0168] See also Figure 4 , shows a schematic diagram of the structure of the speech recognition result determination device provided in an embodiment of the present application, which may include a candidate set acquisition module 401, a candidate set combination module 402, a similarity information determination module 403 and a recognition result determination module 404. Among them:

[0169] The candidate set acquisition module 401 is used to acquire candidate sets corresponding to the target speech on multiple speech recognition systems. The candidate set corresponding to the target speech on a speech recognition system includes several candidate recognition results obtained based on the speech recognition system.

[0170] The candidate set combining module 402 is used to combine the obtained multiple candidate sets in pairs to obtain several groups of candidate sets.

[0171] The similarity information determination module 403 is used to determine the similarity information corresponding to the plurality of candidate sets, wherein the similarity information is sentence-level similarity information between the candidate recognition results included in the first candidate set and the candidate recognition results included in the second candidate set of the two corresponding candidate sets.

[0172] The recognition result determination module 404 is used to determine the recognition result of the target speech from the multiple candidate sets based on the similarity information respectively corresponding to the multiple candidate sets.

[0173] Optionally, when acquiring candidate sets corresponding to the target speech on multiple speech recognition systems, the candidate set acquisition module 401 is specifically used to:

[0174] Obtaining multiple candidate recognition results output by the speech recognition system after recognizing the target speech;

[0175] Determining whether there is a candidate recognition result with a decoding abnormality among the multiple candidate recognition results;

[0176] If yes, the candidate recognition results with abnormal decoding are eliminated, and the remaining candidate recognition results form a candidate set corresponding to the target speech on the speech recognition system;

[0177] If not, the multiple candidate recognition results constitute a candidate set corresponding to the target speech on the speech recognition system.

[0178] Optionally, when determining whether there is a candidate recognition result with a decoding abnormality among the multiple candidate recognition results, the candidate set acquisition module 401 is specifically configured to:

[0179] Determine whether there is a candidate recognition result among the multiple candidate recognition results whose number of decoded words per second is greater than a preset word number threshold;

[0180] If so, determining that there is a candidate recognition result with a decoding abnormality among the multiple candidate recognition results;

[0181] If not, determining that there is no candidate recognition result with a decoding abnormality among the multiple candidate recognition results;

[0182] When eliminating candidate recognition results with abnormal decoding, the candidate set acquisition module 401 is specifically used to:

[0183] Eliminate candidate recognition results whose number of decoded words per second is greater than the preset word number threshold.

[0184] Optionally, the recognition result determination module 404 may include: a candidate recognition result pair determination submodule and a recognition result determination submodule.

[0185] A candidate recognition result pair determination submodule, used to determine the candidate recognition result pairs corresponding to the plurality of candidate sets respectively based on the similarity information corresponding to the plurality of candidate sets respectively, so as to obtain a plurality of candidate recognition result pairs, wherein the candidate recognition result pairs include the two most similar candidate recognition results in the corresponding two candidate sets;

[0186] The recognition result determination submodule is used to determine the recognition result of the target speech based on the similarity information corresponding to the several candidate recognition result pairs, wherein the similarity information corresponding to a candidate recognition result pair is the similarity information of the two candidate recognition results contained in the candidate recognition result pair.

[0187] Optionally, the similarity information is a similarity measurement indicator, and the recognition result determination submodule is specifically used to:

[0188] Based on the similarity measurement indicators respectively corresponding to the plurality of candidate recognition result pairs, determine the most similar candidate recognition result pair from the plurality of candidate recognition result pairs as the target candidate recognition result pair;

[0189] The recognition result of the target speech is determined based at least on the similarity measurement index corresponding to the target candidate recognition result pair.

[0190] Optionally, when the recognition result determination submodule determines the recognition result of the target speech based at least on the similarity measurement index corresponding to the target candidate recognition result pair, it is specifically used to:

[0191] If the similarity measurement index corresponding to the target candidate recognition result pair is a target value that indicates that the two candidate recognition results are the same, any candidate recognition result in the target candidate recognition result pair is determined as the recognition result of the target speech;

[0192] If the similarity measurement index corresponding to the target candidate recognition result pair is not the target value, the recognition result of the target speech is determined based on the similarity measurement index corresponding to the target candidate recognition result pair and the number of words and the comprehensive score of each candidate recognition result in the target candidate recognition result pair, wherein the comprehensive score of a candidate recognition result is determined based on the acoustic score and the language score of the candidate recognition result.

[0193] Optionally, the speech recognition result determination device provided in the present application may further include a comprehensive score determination module. When determining the comprehensive score of a candidate recognition result, the comprehensive score determination module is specifically used to:

[0194] Determining a reasonableness representation value of the candidate recognition result according to the number of words in the candidate recognition result and the language score of the candidate recognition result;

[0195] Normalizing the reasonableness representation value of the candidate recognition result to a value within a target range to obtain a normalized reasonableness representation value, wherein the acoustic score of the candidate recognition result is within the target range;

[0196] The normalized rationality representation value is fused with the acoustic score of the candidate recognition result, and the fused score is used as the comprehensive score of the candidate recognition result.

[0197] Optionally, when determining the recognition result of the target speech based on the similarity measurement index corresponding to the target candidate recognition result pair and the number of words and the comprehensive score of each candidate recognition result in the target candidate recognition result pair, the recognition result determination submodule is specifically used to:

[0198] When the similarity measurement index of the target candidate recognition result pair and the preset similarity threshold satisfy a preset relationship:

[0199] If the number of words in the first candidate recognition result and the number of words in the second candidate recognition result in the target candidate recognition result pair are the same, determining the candidate recognition result with a larger comprehensive score in the target candidate recognition result pair as the recognition result of the target speech;

[0200] If the number of words in the first candidate recognition result is greater than the number of words in the second candidate recognition result, and the comprehensive score of the second candidate recognition result is greater than the comprehensive score of the first candidate recognition result, then determining the second candidate recognition result as the recognition result of the target speech;

[0201] If the number of words in the second candidate recognition result is greater than the number of words in the first candidate recognition result, and the comprehensive score of the first candidate recognition result is greater than the comprehensive score of the second candidate recognition result, the first candidate recognition result is determined as the recognition result of the target speech.

[0202] Optionally, the recognition result determination submodule is further used for:

[0203] If it is determined that the recognition result of the target speech cannot be determined from the target candidate recognition result pair based on the similarity measurement index of the target candidate recognition result pair and the number of words and the comprehensive score of each candidate recognition result in the target candidate recognition result pair, then the candidate recognition result with the highest recognition score is obtained from each candidate set;

[0204] The candidate recognition result with the highest comprehensive score among the candidate recognition results with the highest recognition scores obtained from each candidate set is determined as the recognition result of the target speech.

[0205] Considering the differences and complementarity of multiple speech recognition systems, the speech recognition determination device provided in this embodiment determines the recognition result of the target speech from the candidate sets corresponding to the target speech on multiple speech recognition systems. Compared with determining the recognition result of the target speech from the candidate recognition results obtained based on a single speech recognition system, the speech recognition determination device provided in this application determines the speech recognition result in a more stable manner and can improve the accuracy of the speech recognition result. When determining the recognition result of the target speech from the candidate sets corresponding to the target speech on multiple speech recognition systems, in order to achieve the complementarity of the candidate recognition results of different speech recognition systems and to determine the speech recognition result more quickly, the speech recognition determination device provided in this embodiment combines the candidate sets corresponding to the target speech on multiple speech recognition systems in pairs, and based on the sentence-level similarity information corresponding to each group of candidate sets, determines the recognition result of the target speech from the candidate sets corresponding to the multiple target speech on multiple speech recognition systems. The speech recognition result determination device provided in this embodiment can quickly determine a speech recognition result with a higher accuracy.

[0206] Fifth embodiment

[0207] The present application also provides a device for determining a speech recognition result. Figure 5 , showing a schematic diagram of the structure of the speech recognition result determination device, the public opinion analysis device may include: at least one processor 501, at least one communication interface 502, at least one memory 503 and at least one communication bus 504;

[0208] In the embodiment of the present application, the number of the processor 501, the communication interface 502, the memory 503, and the communication bus 504 is at least one, and the processor 501, the communication interface 502, and the memory 503 communicate with each other through the communication bus 504;

[0209] The processor 501 may be a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention;

[0210] The memory 503 may include a high-speed RAM memory, and may also include a non-volatile memory, such as at least one disk memory;

[0211] The memory stores a program, and the processor can call the program stored in the memory, wherein the program is used to:

[0212] Acquire candidate sets corresponding to the target speech on multiple speech recognition systems respectively, wherein the candidate set corresponding to the target speech on one speech recognition system includes a number of candidate recognition results obtained based on the speech recognition system;

[0213] Combine the obtained multiple candidate sets in pairs to obtain several groups of candidate sets;

[0214] Determine the similarity information respectively corresponding to the several groups of candidate sets, and determine the recognition result of the target speech from the multiple candidate sets based on the similarity information respectively corresponding to the several groups of candidate sets, wherein the similarity information is sentence-level similarity information between the candidate recognition results contained in the first candidate set and the candidate recognition results contained in the second candidate set of the corresponding two candidate sets.

[0215] Optionally, the detailed functions and extended functions of the program may refer to the above description.

[0216] Sixth embodiment

[0217] The embodiment of the present application further provides a computer-readable storage medium, which may store a program suitable for execution by a processor, wherein the program is used to:

[0218] Acquire candidate sets corresponding to the target speech on multiple speech recognition systems respectively, wherein the candidate set corresponding to the target speech on one speech recognition system includes a number of candidate recognition results obtained based on the speech recognition system;

[0219] Combine the obtained multiple candidate sets in pairs to obtain several groups of candidate sets;

[0220] Determine the similarity information respectively corresponding to the several groups of candidate sets, and determine the recognition result of the target speech from the multiple candidate sets based on the similarity information respectively corresponding to the several groups of candidate sets, wherein the similarity information is sentence-level similarity information between the candidate recognition results contained in the first candidate set and the candidate recognition results contained in the second candidate set of the corresponding two candidate sets.

[0221] Optionally, the detailed functions and extended functions of the program may refer to the above description.

[0222] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0223] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0224] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for determining a speech recognition result, characterized in that: include: Acquire candidate sets corresponding to the target speech on multiple speech recognition systems respectively, wherein the candidate set corresponding to the target speech on one speech recognition system includes a number of candidate recognition results obtained based on the speech recognition system; Combine the obtained multiple candidate sets in pairs to obtain several groups of candidate sets; Determine the similarity information corresponding to the plurality of candidate sets respectively, and determine the candidate recognition result pairs corresponding to the plurality of candidate sets respectively based on the similarity information corresponding to the plurality of candidate sets respectively, so as to obtain a plurality of candidate recognition result pairs, wherein the candidate recognition result pairs include the two most similar candidate recognition results in the corresponding two candidate sets; the similarity information is sentence-level similarity information between the candidate recognition results included in the first candidate set and the candidate recognition results included in the second candidate set in the corresponding two candidate sets; Based on the similarity measurement indicators respectively corresponding to the plurality of candidate recognition result pairs, determine the most similar candidate recognition result pair from the plurality of candidate recognition result pairs as the target candidate recognition result pair; If the similarity measurement index corresponding to the target candidate recognition result pair is a target value that indicates that the two candidate recognition results are the same, any candidate recognition result in the target candidate recognition result pair is determined as the recognition result of the target speech; If the similarity measurement index corresponding to the target candidate recognition result pair is not the target value, the recognition result of the target speech is determined based on the similarity measurement index corresponding to the target candidate recognition result pair and the number of words and the comprehensive score of each candidate recognition result in the target candidate recognition result pair, wherein the comprehensive score of a candidate recognition result is determined based on the acoustic score and the language score of the candidate recognition result.

2. The method for determining a speech recognition result according to claim 1, characterized in that: Obtain the candidate set corresponding to the target speech in a speech recognition system, including: Obtaining multiple candidate recognition results output by the speech recognition system after recognizing the target speech; Determining whether there is a candidate recognition result with a decoding abnormality among the multiple candidate recognition results; If yes, the candidate recognition results with abnormal decoding are eliminated, and the remaining candidate recognition results form a candidate set corresponding to the target speech on the speech recognition system; If not, the multiple candidate recognition results constitute a candidate set corresponding to the target speech on the speech recognition system.

3. The method for determining a speech recognition result according to claim 2, characterized in that: The determining whether there is a candidate recognition result with a decoding abnormality among the multiple candidate recognition results includes: Determine whether there is a candidate recognition result among the multiple candidate recognition results whose number of decoded words per second is greater than a preset word number threshold; If so, determining that there is a candidate recognition result with a decoding abnormality among the multiple candidate recognition results; If not, determining that there is no candidate recognition result with a decoding abnormality among the multiple candidate recognition results; The step of eliminating candidate recognition results with abnormal decoding includes: Eliminate candidate recognition results whose number of decoded words per second is greater than the preset word number threshold.

4. The method for determining a speech recognition result according to claim 1, characterized in that: Determine the comprehensive score of a candidate recognition result, including: Determining a reasonableness representation value of the candidate recognition result according to the number of words in the candidate recognition result and the language score of the candidate recognition result; Normalizing the reasonableness representation value of the candidate recognition result to a value within a target range to obtain a normalized reasonableness representation value, wherein the acoustic score of the candidate recognition result is within the target range; The normalized rationality representation value is fused with the acoustic score of the candidate recognition result, and the fused score is used as the comprehensive score of the candidate recognition result.

5. The method for determining a speech recognition result according to claim 1, characterized in that: The determining the recognition result of the target speech based on the similarity measurement index corresponding to the target candidate recognition result pair and the number of words and the comprehensive score of each candidate recognition result in the target candidate recognition result pair includes: When the similarity measurement index of the target candidate recognition result pair and the preset similarity threshold satisfy a preset relationship: If the number of words in the first candidate recognition result and the number of words in the second candidate recognition result in the target candidate recognition result pair are the same, determining the candidate recognition result with a larger comprehensive score in the target candidate recognition result pair as the recognition result of the target speech; If the number of words in the first candidate recognition result is greater than the number of words in the second candidate recognition result, and the comprehensive score of the second candidate recognition result is greater than the comprehensive score of the first candidate recognition result, then determining the second candidate recognition result as the recognition result of the target speech; If the number of words in the second candidate recognition result is greater than the number of words in the first candidate recognition result, and the comprehensive score of the first candidate recognition result is greater than the comprehensive score of the second candidate recognition result, the first candidate recognition result is determined as the recognition result of the target speech.

6. The method for determining a speech recognition result according to claim 5, characterized in that: Also includes: If it is determined that the recognition result of the target speech cannot be determined from the target candidate recognition result pair based on the similarity measurement index of the target candidate recognition result pair and the number of words and the comprehensive score of each candidate recognition result in the target candidate recognition result pair, then the candidate recognition result with the highest recognition score is obtained from each candidate set; The candidate recognition result with the highest comprehensive score among the candidate recognition results with the highest recognition scores obtained from each candidate set is determined as the recognition result of the target speech.

7. A device for determining a speech recognition result, characterized in that: include: A candidate set acquisition module, a candidate set combination module, a similarity information determination module and a recognition result determination module; The candidate set acquisition module is used to acquire candidate sets corresponding to the target speech on multiple speech recognition systems, wherein the candidate set corresponding to the target speech on a speech recognition system includes a number of candidate recognition results obtained based on the speech recognition system; The candidate set combination module is used to combine the obtained multiple candidate sets in pairs to obtain several groups of candidate sets; The similarity information determination module is used to determine the similarity information corresponding to the plurality of candidate sets, wherein the similarity information is sentence-level similarity information between the candidate recognition results included in the first candidate set and the candidate recognition results included in the second candidate set of the corresponding two candidate sets; The recognition result determination module is used to determine the recognition result of the target speech from the multiple candidate sets based on the similarity information respectively corresponding to the multiple candidate sets; Wherein, the recognition result determination module includes: a candidate recognition result pair determination submodule and a recognition result determination submodule; The candidate recognition result pair determination submodule is used to determine the candidate recognition result pairs corresponding to the plurality of candidate sets respectively based on the similarity information corresponding to the plurality of candidate sets respectively, so as to obtain a plurality of candidate recognition result pairs, wherein the candidate recognition result pairs include the two most similar candidate recognition results in the corresponding two candidate sets; A recognition result determination submodule is used to determine the most similar candidate recognition result pair from the several candidate recognition result pairs based on the similarity measurement indicators respectively corresponding to the several candidate recognition result pairs, as the target candidate recognition result pair; if the similarity measurement indicator corresponding to the target candidate recognition result pair is a target value that characterizes that two candidate recognition results are the same, then any candidate recognition result in the target candidate recognition result pair is determined as the recognition result of the target speech; if the similarity measurement indicator corresponding to the target candidate recognition result pair is not the target value, then based on the similarity measurement indicator corresponding to the target candidate recognition result pair and the number of words and the comprehensive score of each candidate recognition result in the target candidate recognition result pair, the recognition result of the target speech is determined, wherein the comprehensive score of a candidate recognition result is determined based on the acoustic score and the language score of the candidate recognition result.

8. A device for determining a speech recognition result, characterized in that: include: Memory and processor; The memory is used to store programs; The processor is used to execute the program to implement each step of the method for determining a speech recognition result according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, each step of the method for determining a speech recognition result according to any one of claims 1 to 6 is implemented.

Images