A method, device, equipment and storage medium for recognizing voice data

Abstract

An embodiment of the present invention discloses a method, apparatus, device, and storage medium for recognizing speech data, wherein the method includes: obtaining speech data input by a target user and location information of the target user; determining a target region language model corresponding to a target region to which the location information belongs from a plurality of region language models, wherein any region language model of the plurality of region language models is trained based on information points included in the any region; calling the target region language model and a universal language model to perform a first decoding process on the speech data to obtain N candidate recognition results; calling the target region language model and a neural network language model to perform a second decoding process on each of the N candidate recognition results to select a target recognition result from the N candidate recognition results, thereby improving the recognition accuracy of the speech data.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method, apparatus, device, and storage medium for recognizing voice data. Background Art

[0002] With the rapid development of key computer artificial intelligence technologies (automatic speech recognition, speech synthesis, and voiceprint recognition), numerous AI products and platforms have emerged in our lives. Enabling computers to hear, see, speak, and feel is the future direction of human-computer interaction, with speech recognition technology emerging as one of the most promising methods of human-computer interaction. As a common AI application, speech recognition technology has entered every aspect of our lives, such as map navigation, smart home appliances, voice search, and text-to-speech conversion.

[0003] Although speech recognition technology has made great progress in recent years, when speech data includes information points (points on a map that have no geographical significance but are meaningful, such as stores, bars, gas stations, etc.), the identification of information points in speech data still faces great challenges. Because the speech recognition models in existing speech data recognition methods are usually only combined with a network of information points in a specific region, the speech recognition models are not accurate in identifying information points when recognizing speech data that contains information points in other regions, which in turn leads to low recognition accuracy of speech data. Therefore, in the field of speech recognition technology, how to improve the recognition accuracy of speech data that includes information points has become a hot research issue today. Summary of the Invention

[0004] Embodiments of the present invention provide a method, apparatus, device, and storage medium for recognizing voice data, which can improve the accuracy of identifying information points in voice data, thereby improving the recognition accuracy of voice data.

[0005] In one aspect, an embodiment of the present invention provides a method for recognizing speech data, comprising:

[0006] Acquiring voice data input by a target user and location information of the target user;

[0007] Determining a target region language model corresponding to the target region to which the location information belongs from a plurality of region language models, wherein any region language model among the plurality of region language models is trained based on information points included in any region;

[0008] Calling the target region language model and the universal language model to perform a first decoding process on the speech data to obtain N candidate recognition results;

[0009] The target region language model and the neural network language model are called to perform a second decoding process on each of the N candidate recognition results to select a target recognition result from the N candidate recognition results.

[0010] In one aspect, an embodiment of the present invention provides a device for recognizing speech data, comprising:

[0011] an acquiring unit, configured to acquire voice data input by a target user and location information of the target user;

[0012] a determining unit, configured to determine a target region language model corresponding to the target region to which the location information belongs from a plurality of region language models, wherein any region language model among the plurality of region language models is trained based on information points included in any region;

[0013] a processing unit, configured to call the target region language model and the universal language model to perform a first decoding process on the speech data to obtain N candidate recognition results;

[0014] The processing unit is further configured to call the target region language model and the neural network language model to perform a second decoding process on each of the N candidate recognition results, so as to select a target recognition result from the N candidate recognition results.

[0015] In one embodiment, when the processing unit calls the target region language model and the neural network language model to perform a second decoding process on each of the N candidate recognition results to select a target recognition result from the N candidate recognition results, it is specifically configured to perform:

[0016] Obtaining a first fusion probability of each candidate recognition result among the N candidate recognition results, where the first fusion probability of any candidate recognition result is obtained during a first encoding process of the speech data;

[0017] calling the target region language model and the neural network language model to process each candidate recognition result respectively to obtain a second fusion probability of each candidate recognition result;

[0018] A target recognition result is selected from the N candidate recognition results based on the first fusion probability of each candidate recognition result and the second fusion probability of each candidate recognition result.

[0019] In one embodiment, when the processing unit selects the target recognition result from the N candidate recognition results based on the first fusion probability of each candidate recognition result and the second fusion probability of each candidate recognition result, it is specifically configured to perform:

[0020] Fusing the first fusion probability of each candidate recognition result with the second fusion probability of the corresponding candidate recognition result to obtain a target fusion probability of each candidate recognition result;

[0021] The candidate recognition result with the highest target fusion probability among the N candidate recognition results is used as the target recognition result.

[0022] In yet another embodiment, the processing unit, when fusing the first fusion probability of each candidate recognition result with the second fusion probability of the corresponding candidate recognition result to obtain the target fusion probability of each candidate recognition result, is further specifically configured to perform:

[0023] Obtaining an acoustic probability corresponding to any one of the N candidate recognition results, where the acoustic probability corresponding to any one of the candidate recognition results is obtained during a first decoding process of the speech data;

[0024] Obtaining an acoustic scaling parameter and a fusion ratio parameter, performing a logarithmic operation on the acoustic probability, multiplying the logarithmic operation result by the acoustic scaling parameter, and fusing the first fusion probability and the second fusion probability according to the fusion ratio parameter;

[0025] The multiplication result and the fusion processing result are added together to obtain the target fusion probability of any candidate recognition result.

[0026] In another embodiment, the N candidate recognition results include a target candidate recognition result, and the processing unit, when obtaining the first fusion probability of each candidate recognition result in the N candidate recognition results, is specifically configured to perform:

[0027] Calling the target regional language model to perform a first decoding process on the speech data to obtain a first language probability of the target candidate recognition result;

[0028] Calling the universal language model to perform a first decoding process on the speech data to obtain a second language probability of the target candidate recognition result;

[0029] fusing the first language probability of the target candidate recognition result and the second language probability of the target candidate recognition result using a first fusion parameter to obtain a fused language probability of the target candidate recognition result;

[0030] A first fusion probability of the target candidate recognition result is obtained based on the fusion language probability of the target candidate recognition result and the second language probability of the target candidate recognition result.

[0031] In another embodiment, when the processing unit calls the target region language model and the neural network language model to process each candidate recognition result respectively to obtain the second fusion probability of each candidate recognition result, it is specifically configured to perform:

[0032] Obtaining the first language probability of the target candidate recognition result;

[0033] Calling the neural network language model to perform a second decoding process on the target candidate recognition result to obtain a third language probability of the target candidate recognition result;

[0034] The first language probability and the third language probability are fused using a second fusion parameter to obtain a second fusion probability of the target candidate recognition result.

[0035] In another embodiment, before the determining unit determines the target regional language model corresponding to the speech data based on the location information, the determining unit may further be configured to:

[0036] Acquire information point data of a target area, the information point data including at least one information point and information point features of each of the at least one information point, wherein the target area includes a location indicated by the location information of the target user, and the information point features include one or both of the following: an information point name and an information point alias;

[0037] The regional language model corresponding to the target region is trained using the information point features to obtain a target regional language model corresponding to the target region.

[0038] In one aspect, an embodiment of the present invention further provides a voice data recognition device, the recognition device including an input interface and an output interface, and further including:

[0039] a processor adapted to implement one or more instructions;

[0040] A computer storage medium storing one or more instructions adapted to be loaded and executed by the processor:

[0041] Acquire voice data input by a target user and location information of the target user; determine a target regional language model corresponding to the voice data from a plurality of regional language models based on the location information, wherein any one of the plurality of regional language models is trained based on information points included in the region corresponding to any one of the regional language models; call the target regional language model and a universal language model to perform a first decoding process on the voice data to obtain N candidate recognition results; call the target regional language model and a neural network language model to perform a second decoding process on each of the N candidate recognition results to select a target recognition result from the N candidate recognition results.

[0042] In one aspect, an embodiment of the present invention provides a computer storage medium, characterized in that the computer storage medium stores one or more instructions, wherein the one or more instructions are suitable for being loaded and executed by a processor:

[0043] Acquire voice data input by a target user and location information of the target user; determine a target regional language model corresponding to the voice data from a plurality of regional language models based on the location information, wherein any one of the plurality of regional language models is trained based on information points included in the region corresponding to any one of the regional language models; call the target regional language model and a universal language model to perform a first decoding process on the voice data to obtain N candidate recognition results; call the target regional language model and a neural network language model to perform a second decoding process on each of the N candidate recognition results to select a target recognition result from the N candidate recognition results.

[0044] In one aspect, an embodiment of the present invention provides a computer program product or computer program. The computer program product includes a computer program stored in a computer storage medium. A processor of a terminal reads the computer program from the computer storage medium and executes the computer program, causing a speech data recognition device to perform:

[0045] Acquire voice data input by a target user and location information of the target user; determine a target regional language model corresponding to the voice data from a plurality of regional language models based on the location information, wherein any one of the plurality of regional language models is trained based on information points included in the region corresponding to any one of the regional language models; call the target regional language model and a universal language model to perform a first decoding process on the voice data to obtain N candidate recognition results; call the target regional language model and a neural network language model to perform a second decoding process on each of the N candidate recognition results to select a target recognition result from the N candidate recognition results.

[0046] In an embodiment of the present invention, a voice data recognition device obtains the voice data input by a target user and the location information of the target user, and further determines a target regional language model corresponding to the target region to which the location information belongs from a plurality of regional language models, and any of the plurality of regional language models is trained based on the information points included in any region. It should be understood that a regional language model is established separately for each region based on the information points included in each region, which can solve the problem of inaccurate information point recognition in traditional voice data recognition methods because the general language model is usually only combined with the information point network of a specific region. Furthermore, the voice data recognition device combines the target regional language model with the general language model and the neural network language model respectively to perform a first decoding process and a second decoding process on the voice data, and selects the recognition result of the voice data through multiple decoding processes, which can further improve the recognition accuracy of the voice data. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0048] Figure 1a is a schematic diagram of a voice data recognition system provided by an embodiment of the present invention;

[0049] Figure 1b is a schematic flow chart of a speech data recognition process provided by an embodiment of the present invention;

[0050] Figure 1c This is a schematic diagram of an application scenario of a voice data recognition method provided by an embodiment of the present invention;

[0051] Figure 2 1 is a schematic diagram of a method for recognizing voice data provided by an embodiment of the present invention;

[0052] Figure 3 Schematic diagram of another method for recognizing voice data provided by an embodiment of the present invention;

[0053] Figure 4 1 is a schematic diagram of a voice data recognition device provided by an embodiment of the present invention;

[0054] Figure 5 The figure is a schematic diagram of the architecture of a voice data recognition device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0055] An embodiment of the present invention provides a method for recognizing speech data. The method can be applied to various speech recognition scenarios, such as query scenarios for information points in map navigation applications and in-vehicle navigation systems. The embodiment of the present invention dynamically combines a target region language model with a general language model to perform a first decoding process on the speech data to obtain N candidate recognition results. The target region language model refers to a region language model corresponding to the region where the target user generating the speech data is located. The target region language model is trained based on information points included in the target region. Therefore, it can be seen that dynamically combining the target region language model with the general language model to perform a decoding process can improve the accuracy of identifying information points in the speech data. The target region language model is then dynamically combined with a neural network language model and applied to a second decoding process of the speech data to determine a target recognition result from the N candidate recognition results. This solves the speech recognition problem of information points with the same or similar pronunciation in different regions and can significantly improve the recognition accuracy of map information points in the user's region. Information points can be any non-geographically meaningful but meaningful points on a map, such as stores, bars, gas stations, hospitals, and bus stations.

[0056] Based on the above-mentioned voice data recognition method, an embodiment of the present invention provides a voice data recognition system. The structural diagram of the voice data recognition system can be found in Figure 1a ,like Figure 1a As shown, the system includes a terminal 101 and a server 102. The terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, an intelligent robot, a smart watch, etc., but is not limited thereto. The server 102 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal 101 and the server 102 can be directly or indirectly connected through wired or wireless communication methods, and this application does not limit this.

[0057] In one embodiment, the voice data recognition method provided in the embodiment of the present application can be applied in a server or in a terminal. The embodiment of the present application is described by taking the voice data recognition method applied in a server as an example. Figure 1bAs shown, the server obtains the voice data input by the target user at the terminal and the location information of the terminal, uses location services to convert the location information into regional information, and uses the regional information to determine the target regional language model used in the voice data recognition process. Different regions correspond to different target regional language models. For example, Beijing uses regional language model 1, Chongqing uses regional language model 1, and Shenzhen uses regional language model 3. The server combines the target regional language model with the universal language model and performs a first decoding process on the voice data to obtain N candidate recognition results corresponding to the voice data. The server then dynamically combines the target regional language model with the neural network language model and performs a second decoding process on the N candidate recognition results to determine the target recognition result from the N candidate recognition results. The server then outputs the target recognition result to the terminal.

[0058] In an exemplary scenario, when the voice data is recognized by the server, taking a map navigation application running in a terminal as an example, the map navigation application can query the location through the voice data input by the user, and the map navigation application can call the location service component in the terminal to obtain the location information of the terminal. Figure 1c As shown, Figure 1c 1 is a schematic diagram of an application scenario for voice data recognition provided by an embodiment of the present invention. The target user 111 clicks the voice input button 1121 in the map navigation APP (Application) running in the terminal 112 to open the voice signal input box 1122 to obtain the voice data input by the user. Assume that the voice data input by the target user 111 is "I want to go to Square A", and the "Beijing Road" where the target user 111 is located is in Beijing. Then, the map navigation APP calls the location service component to obtain the location information of the target user as "Beijing City", and after the voice data is the voice signal "I want to go to Square A", the location information is sent to the server. The server determines through the location information that the target regional language model is the regional language model a suitable for recognizing the Beijing dialect, and then combines the regional language model a with the general language model to perform a first decoding process on the voice data to obtain N candidate recognition results. For example, assuming that N is 2, the N candidate recognition probabilities are "I want to go to Square A" and "I want to go to the A2 Square parking garage"; the server can then combine the regional language model a with the neural network language model to perform a second decoding process on the above two candidate recognition results, determine that the recognition result of the voice data is "I want to go to Square A", and then input the recognition result into the terminal 112 for display.

[0059] See Figure 2 , Figure 2 is a schematic diagram of a method for recognizing speech data provided by an embodiment of the present invention, such as Figure 2 As shown, the method includes:

[0060] S201: Acquire voice data input by a target user and location information of the target user.

[0061] In one embodiment, the server may obtain the location information of the target user by instructing the map navigation APP to call the location service component in the terminal, wherein the map navigation APP runs in the terminal, and the location information of the target user is the location information of the terminal.

[0062] For example, assume that the target user is located in Beijing and is about to use the map navigation app installed on their smartphone to direct themselves to Square A. After the map navigation app detects that the target user has opened the voice input box, it displays a voice data box on the smartphone and obtains the voice data "I want to go to Square A" input by the target user through the voice input box. The map navigation app then calls the location service component on the smartphone and obtains the smartphone's location information as "Street A, District A, Beijing". The map navigation app then sends the voice data "I want to go to Square A" and the location information "Street A, District A, Beijing" to the server, causing the server to use "Street A, District A, Beijing" as the target user's location information.

[0063] S202: Determine a target region language model corresponding to the target region to which the location information belongs from a plurality of region language models.

[0064] In one embodiment, any of the multiple regional language models is trained based on the information points included in any region. That is, different regions correspond to different regional language models, and each regional language model is used to recognize (decode) speech data generated in that region. This avoids the problem of inaccurate information point recognition caused by the inclusion of the same information points in different regions. In specific embodiments, the regional language model can be an N-gram language model, a recursive neural network type neural network language model, a convolutional neural network type neural network language model, or the like. Among them, different regions can be understood as different provinces and cities within a country. For example, China can be divided into 34 provincial regions. Different regions can also be understood as a dialect area. For example, Chongqing, Sichuan and Guizhou with the same dialect are regarded as one region. Then, based on the name of the information point in a certain region, the regional language model corresponding to the region is trained to obtain the regional language model corresponding to each region. Any regional language model is used to process the voice data generated in any region. In this way, multiple regional language models are established, which solves the problem in traditional voice data recognition methods that users cannot perform cross-regional information point queries because the general language model is usually only combined with the information point network of a specific region. While realizing cross-regional information point queries, the recognition accuracy of voice data (especially voice data containing information points) can also be improved.

[0065] Furthermore, because the target region language model is trained based on the information points included in the target region, when using the target region language model to recognize speech data containing information points, as long as the speech data contains the information points, there are no requirements for the sentence structure of the input speech data. Therefore, embodiments of the present invention can support recognition of speech data input using any sentence structure, enriching application scenarios.

[0066] In one embodiment, the method of determining the target regional language model corresponding to the target region to which the location information belongs from a plurality of regional language models includes: mapping the location information to a regional index, and determining the target region to which the location information belongs based on the regional index; and selecting a regional language model corresponding to the target region from a plurality of regional language models as the target regional language model. Specifically, after the server obtains the location information of the target user, the location information can be mapped to a regional index, so that the server determines the target region where the target user is located based on the regional index, and then determines the target regional language model based on the target region where the target user is located. The target regional language model will be used in the first decoding process and the second decoding process of the voice data. The first decoding process is used to obtain N candidate recognition results corresponding to the voice data. For example, the N candidate recognition results can be specifically 20 candidate recognition results. The second decoding process is used to determine the target recognition result from the N candidate recognition results, for example, the target recognition result is determined from the 20 candidate recognition results.

[0067] Assume that the server obtains the target user's location information as "Street A, District A, Beijing" and the speech data is "I want to go to Square A." The server first maps "Street A, District A, Beijing" to the regional index "Beijing." Based on "Beijing," the server determines that the target regional language model is regional language model a for the Beijing dialect. The server then applies regional language model a to the first and second decoding processes of the speech data "I want to go to Square A."

[0068] S203: Call the target region language model and the universal language model to perform a first decoding process on the speech data to obtain N candidate recognition results.

[0069] In one embodiment, the universal language model can be an N-gram model (N-gram model) using Kneser-Ney smoothing, such as a 5-gram model. This universal language model can be used to decode speech data in any region. The N candidate recognition results can then be all recognition results obtained by calling the target region language model and the universal language model to perform a first decoding process on the speech data. For example, if a total of 20 recognition results are obtained after calling the target region language model and the universal language model to perform a first decoding process on the speech data, the N candidate recognition results can refer to these 20 recognition results.

[0070] In other embodiments, the N candidate recognition results may be N recognition results selected from all recognition results, for example, N recognition results with the highest first confidence rankings are selected from all recognition results as candidate recognition results. Specifically, after the server obtains the first confidence of each candidate recognition result among the M (M>N) candidate recognition results, the server may sort the first confidence of each candidate recognition result from high to low and select the candidate recognition result corresponding to each first confidence among the top N first confidences as the N candidate recognition results.

[0071] In an embodiment of the present invention, the first confidence level of each recognition result can be reflected by the first fusion probability, that is, among the N candidate recognition results, each candidate recognition result corresponds to a first fusion probability, and the first fusion probability corresponding to any candidate recognition result is obtained when the target regional language model and the general language model are called to perform the first decoding process on the speech data. Assume that the N candidate recognition results include the target candidate recognition result. The following takes the target candidate recognition result as an example to introduce how to obtain the first fusion probability of the target candidate recognition result when calling the target regional language model and the general language model to perform the first decoding process on the speech data.

[0072] The server performs a first decoding process on the speech data by calling the target regional language model to obtain the first language probability p corresponding to the target candidate recognition result w l (w), and then the first decoding process is performed on the speech data by calling the general language model to obtain the second language probability p corresponding to the target candidate recognition result w g (w). Then, the server uses the first fusion parameter to calculate the first language probability p corresponding to the target candidate recognition result w. l (w) and the second language probability p g (w) is fused to obtain the fused language probability log p′1(w) of the target candidate recognition result w, and then the first fused probability log p1(w) of the target recognition candidate result is determined according to the fused language probability and the second language probability of the target candidate recognition result w.

[0073] The first fusion parameter may be, for example, α (α is a constant greater than zero, usually close to 1, such as α=1.1), which is used to control the second language probability p obtained by the general language model. g (w) and the first language probability p obtained by the target regional language model lThe fusion ratio of (w), the larger the value of α, the higher the probability of fusion of tail information points (information points searched by a small number of users, usually a large number of information points, so these information points are usually not recognized during the recognition process), and thus the higher the chance of being selected when selecting homophones in the recognition process (selecting a word from all the word results corresponding to the same pronunciation unit, such as selecting one of the words "wo", "wo", "wo" and so on corresponding to the pronunciation unit "wò").

[0074] In one embodiment, a first fusion parameter is used to fuse the first language probability of the target candidate recognition result and the second language probability of the target candidate recognition result to obtain the fused language probability of the target candidate recognition result, which may include: the server calculates the first language probability p corresponding to the target candidate recognition result w based on the first fusion parameter α. l (w) and the second language probability p g (w) is subjected to log-linear interpolation to complete the fusion process and obtain the fused language probability log p′1(w). The log-linear interpolation method can be specifically shown in formula (1):

[0075] log p′1(w)=log p l (w)-αlog p g (w)......Formula (1)

[0076] As can be seen from the above description, the server uses the formula (1) in the embodiment of the present invention to determine the fusion language probability of each candidate recognition result, which can improve the recognition probability of the tail information point.

[0077] In the first decoding process, the first fusion probability log p1(w) of the target candidate recognition result w can be determined by taking the log p corresponding to the target candidate recognition result. g The maximum value of (w) and log p′1(w) is achieved, and the implementation method can be specifically shown in formula (2):

[0078] log p1(w)=max(log p g (w),log p′1(w)).....Formula (2)

[0079] In this embodiment, after the server obtains the fusion language probability of the target candidate recognition result, it selects the first fusion probability of the target candidate recognition result based on the fusion language probability and the first language probability of the target candidate recognition result, thereby ensuring the recognition performance of the entire speech data recognition model for the head information points (information points searched by a large number of users, usually a small number of information points, and therefore often preferred when selecting homophones) and the robustness of the entire speech data recognition model.

[0080] For example, assume that the server uses the target region language model and the universal language model to perform a first decoding process on the speech data "I want to go to Plaza A" and obtains the target candidate recognition result "I want to go to Plaza A parking garage". If the first language probability of the target candidate recognition result in the target region language model is a, and the second language probability of the target candidate recognition result in the universal language model is b, then the server processes the first language probability and the second language probability of the target candidate recognition result as shown in the above formula (1) to obtain the fused language probability c of the target candidate recognition result. Then, the fused language probability c and the second language probability b are processed as shown in the above formula (2) to obtain the first fused probability d of the target candidate recognition result. That is, the larger value of b and c is selected as the first fused probability d of the target candidate recognition result.

[0081] In another embodiment, the server may also call the acoustic model, the target region language model, and the universal language model to perform a first decoding process on the speech data to obtain N candidate recognition results and the first language probability p of each candidate recognition result w in the N candidate recognition results. l (w), second language probability p g (w) and acoustic probability p a (w).

[0082] S204 , calling the target region language model and the neural network language model to perform a second decoding process on each of the N candidate recognition results, so as to select a target recognition result from the N candidate recognition results.

[0083] In one embodiment, a neural network language model is used to predict the probability that a certain word in any candidate recognition result appears after the previous word, so as to obtain the language probability (the third language probability) corresponding to any candidate recognition result. For example, for the candidate recognition result of "I want to go to Square A", the neural network language model can be used to predict the probability that the word "want" appears after the word "I", the probability that the word "go" appears after the word "want",..., the probability that the word "square" appears after the word "broad", and finally obtain the third language probability corresponding to the entire sentence "I want to go to Square A". Then, the target recognition result can be determined by the server based on the target fusion probability of all candidate recognition results in the second decoding process. For example, the candidate recognition result with the maximum target fusion probability is used as the target recognition result. Among them, the target fusion probability can be obtained by the server based on the acoustic probability, the first confidence level, and the second confidence level of any candidate recognition result, and the second confidence level can be obtained by the server based on the third language probability and the first language probability of any candidate recognition result.

[0084] In a specific embodiment, the second confidence level of any candidate recognition result can be represented by a second fusion probability. Among them, each candidate recognition result corresponds to a second confidence level (second fusion probability), and the second confidence level (second fusion probability) of any candidate recognition result is obtained in the second decoding process. Below, taking the target candidate recognition result w as an example, introduce how to obtain the second fusion probability and the target fusion probability of the target candidate recognition result w when calling the target regional language model and the neural network language model to perform the second decoding process on the target candidate recognition result w.

[0085] First, the server uses the target candidate recognition result w as the input data of the neural network language model, and performs the second decoding process on the target candidate recognition result w by calling the neural network language model to obtain the third language probability p n (w) of the target candidate recognition result w. Then, the server uses the second fusion parameter β to perform a fusion process on the third language probability p n (w) of the target candidate recognition result w and the first language probability p l (w) of the target candidate recognition result w to obtain the second fusion probability log p2(w) of the target candidate recognition result w. Among them, the second fusion parameter β is a constant used to balance the target regional language model and the neural network language model. Exemplarily, the server can use the second fusion parameter β (such as β = 0.2) to perform a linear logarithmic interpolation process on the third language probability p n (w) of the target candidate recognition result w and the first language probability p l (w) of the target candidate recognition result w as shown in formula (3) to achieve the above fusion process.

[0086] log p2(w)=βlog p l (w)+(1-β)log p n (w)...Formula (3)

[0087] When the server obtains the second fusion probability log p2(w) of the target candidate recognition result w, it uses the fusion ratio parameter λ and the acoustic scaling parameter η to calculate the first fusion probability log p1(w) of the target candidate recognition result and the acoustic probability p a (w) obtains the target fusion probability S(w) of the target candidate recognition result. For example, the server performs a logarithmic operation on the acoustic probability and logs the logarithmic operation result log p a (w) is multiplied by the acoustic scaling parameter η to obtain an operation result, and the first fusion probability p1(w) and the second fusion probability p2(w) of the target candidate recognition result w are fused through the fusion ratio parameter λ to obtain a fusion result; then the multiplication operation result and the fusion result of the target candidate recognition result w are added to obtain the target fusion probability S(w) of the target candidate recognition result w.

[0088] The first fusion probability and the second fusion probability are fused according to the fusion ratio parameter λ, including: performing a logarithmic operation on the fusion ratio parameter λ of the first fusion probability to obtain log p1(w), and performing a logarithmic operation on the second fusion probability p2(w) to obtain log p2(w). Finally, log p1(w) and log p2(w) are fused according to the fusion ratio parameter. According to the above description, the target fusion probability of the target candidate recognition result can be expressed as shown in formula (4).

[0089] S(w)=ηlog p a (w)+λlog p1(w)+(1-λ)log p2(w)...Formula (4)

[0090] Among them, the acoustic scaling parameter η is a constant used to balance the acoustic model and the language model, and the second fusion parameter λ is used to control the fusion ratio of the first fusion probability and the second fusion probability. For example, η = 0.84 and λ = 0.5. Then, the server can obtain the target fusion probability of each candidate recognition result in the N candidate recognition results through the above method. Then, when the target fusion probability of the target candidate recognition result is the largest among the N candidate recognition results, the server can determine that the target candidate recognition result is the target recognition result.

[0091] In this embodiment, when the server selects a target recognition result, for any candidate recognition result among the N candidate recognition results, not only does it use the second fusion parameter to fuse the first language probability and the third language probability of any recognition result to obtain the second fusion probability, but it also uses the fusion ratio parameter and the acoustic scaling parameter to fuse the first fusion probability, the second fusion probability and the acoustic probability of any candidate recognition result to obtain the target fusion probability. This fusion method further improves the recognition accuracy of the speech data.

[0092] For example, assuming that the two (N=2) candidate recognition results corresponding to the voice data "I want to go to A Square" are "I want to go to A Square Parking Garage" and "I want to go to A Square", and the target fusion probability of the two candidate recognition results is as follows: the target fusion probability of "I want to go to A Square Parking Garage" is "-100" (logarithmic probability value), and the target fusion probability of "I want to go to A Square" is "-90" (logarithmic probability value), then the server will take the candidate recognition result "I want to go to A Square" as the target recognition result.

[0093] In an embodiment of the present application, in an embodiment of the present invention, a voice data recognition device obtains the voice data input by the target user and the location information of the target user, and further determines a target regional language model corresponding to the target region to which the location information belongs from a plurality of regional language models, and any of the plurality of regional language models is trained based on the information points included in any region. It should be understood that a regional language model is established separately for each region based on the information points included in each region, which can solve the problem of inaccurate information point recognition in traditional voice data recognition methods because the general language model is usually only combined with the information point network of a specific region. Furthermore, the voice data recognition device combines the target regional language model with the general language model and the neural network language model respectively to perform a first decoding process and a second decoding process on the voice data, and selects the recognition result of the voice data through multiple decoding processes, which can further improve the recognition accuracy of the voice data.

[0094] See Figure 3 , Figure 3 FIG. 1 is a schematic diagram of another method for recognizing speech data provided by an embodiment of the present invention. Figure 3 As shown, the method includes:

[0095] S301: Acquire voice data input by a target user and location information of the target user.

[0096] In one embodiment, some feasible implementations of step S301 are already described in Figure 2 The relevant steps in the embodiments are described in detail and will not be repeated here.

[0097] S302: Determine a target region language model corresponding to the target region to which the location information belongs from a plurality of region language models.

[0098] In one embodiment, since the method of decoding speech data using only a general language model in the prior art (general language model modeling method) is difficult to effectively describe all information point data, this embodiment adopts a solution of language modeling by region to improve the recognition rate of information points. Specifically, the server can train the regional language model before determining the target regional language model to obtain the target regional language model. Specifically, the server can obtain information point data of the target region, the information point data includes at least one information point, and information point features of each information point in the at least one information point, wherein the target region includes the location indicated by the location information of the target user, and the information point features include one or two of the following: information point name, information point alias, and the server then uses the information point features to train the regional language model corresponding to the target region to obtain the target region language model corresponding to the target region.

[0099] In a specific embodiment, the target region is the area indicated by the target user's location information. For example, if the target user's location information is "Block B, City A," the corresponding target region is City A. In this case, the information point data consists of the name of each information point and its alias within City A, for example, information point X and its alias Information Point X1, and information point Y and its alias Information Point Y1. The server then uses this information point data to train a regional language model corresponding to City A, obtaining a target regional language model corresponding to City A. This target regional language model is suitable for recognizing and processing the dialect speech data used in City A.

[0100] S303: Call the target region language model and the universal language model to perform a first decoding process on the speech data to obtain N candidate recognition results.

[0101] In one embodiment, the specific process of the first decoding process can be described as follows: the server can preset a database, in which there is a network relationship composed of all words, and the network relationship is used to instruct the language model to sort individual words to form candidate recognition results. The server can call the acoustic model to perform a first decoding process on the voice data to obtain P pronunciation sequences corresponding to the voice data, and the acoustic probability corresponding to each pronunciation sequence in the P pronunciation sequences, wherein the pronunciation sequence is composed of pronunciation units corresponding to a single word. The server then calls the target region language model and the general language model to perform a first decoding process on the word corresponding to any pronunciation sequence in the P pronunciation sequences, and obtains Q candidate recognition results, and the first language probability corresponding to each candidate recognition result in the Q candidate recognition results, wherein the Q candidate recognition results each have an acoustic probability, and the acoustic probability is the acoustic probability of any pronunciation sequence corresponding to the Q candidate recognition results.

[0102] Specifically, the target region language model and the universal language model can search for the corresponding word in the database, as well as the next word following the word, using any pronunciation unit and network relationship in any pronunciation sequence. The first language probability of any candidate recognition result can be calculated by calculating the probabilities of all the words that make up the candidate recognition result in the target region language model. The second language probability of any candidate recognition result can be calculated by calculating the probabilities of all the words that make up the candidate recognition result in the universal language model.

[0103] For example, assuming that the existing voice data is a pronunciation sequence {wo, yao, qu, A, guang, chang} corresponding to "I want to go to A Square", then the server can perform a word search based on any pronunciation unit in the pronunciation sequence after obtaining the pronunciation unit in the acoustic model. For example, according to the pronunciation unit "wo", the target area language model and the general language model are called to find the words "I", "nest", "hold", etc. in the database, and obtain the corresponding probabilities of "I", "nest", and "hold" respectively; then the server searches for the words that may be arranged after "I" (the second word) based on the network relationship in the database, and the corresponding probabilities of these words, for example, "want", "invite", "can", "is", etc. are found, and then the server determines that the second word may be "want", "invite" based on the pronunciation unit "yao", and obtains the corresponding probability A of "want" in the target area language model and the corresponding probability a in the general language model, as well as the corresponding probability B of "invite" in the target area language model and the corresponding probability b in the general language model. In this way, two candidate recognition results corresponding to the entire pronunciation sequence are finally obtained: "I want to go to A Square", "I want to go to A Square Parking Garage". The first language probability of "I want to go to A Square" can be calculated by calculating the probabilities of the words "I", "to", "go", "A", "guang", and "field" in the target region language model; the second language probability of "I want to go to A Square" can be calculated by calculating the probabilities of the words "I", "to", "go", "A", "guang", and "field" in the general language model.

[0104] S304: Obtain a first fusion probability of each candidate recognition result among the N candidate recognition results.

[0105] In one embodiment, the specific implementation method is described in detail in step S203 and will not be repeated here.

[0106] S305 , calling the target region language model and the neural network language model to process each candidate recognition result respectively, and obtaining a second fusion probability of each candidate recognition result.

[0107] S306 : Select a target recognition result from the N candidate recognition results based on the first fusion probability of each candidate recognition result and the second fusion probability of each candidate recognition result.

[0108] In one embodiment, the specific implementation of steps S305 to S306 can be found in the detailed description of step S204 and will not be repeated here.

[0109] After the experiment, it was found that after the server obtained the target user's voice data test set, the voice data test set was processed using the traditional voice data recognition method and the voice data recognition method of this application to obtain the character error rate corresponding to the different voice data recognition methods. The experimental results are shown in the following table:

[0110] Test data The error rate of traditional recognition methods The error rate of this application method Test Set 1 5.02％ 3.90％ Test Set 2 4.38％ 3.82％

[0111] Test Sets 1 and 2 were both collected from voice data corresponding to users in 31 regions across China through the Tencent Maps app. A lower character error rate indicates better performance and higher recognition accuracy for the corresponding voice data recognition method. As can be seen from the table above, the voice data recognition method provided in this application can significantly improve the recognition accuracy of information points.

[0112] In an embodiment of the present application, the server establishes a separate regional language model for each region based on the information points of each region, and determines the target regional language model to be used in the recognition process of the voice data through the user's location information. This solves the problem of inaccurate information point recognition caused by the general language model usually only being combined with the information point network of a specific region in traditional voice data recognition methods; and by combining the target regional language model with the general language model and the neural network language model respectively to perform the first decoding process and the second decoding process on the voice data, the recognition accuracy of the voice data input by the user is further improved.

[0113] Based on the description of the embodiment of the voice data recognition method, the embodiment of the present invention further discloses a voice data recognition device, which can be a computer program (including program code) running on the server mentioned above. The voice data recognition device can execute Figure 2 or Figure 3 See the method shown in Figure 4 The speech data recognition device may include: an acquisition unit 401, a determination unit 402 and a processing unit 403.

[0114] An acquisition unit 401 is configured to acquire voice data input by a target user and location information of the target user;

[0115] A determining unit 402 is configured to determine a target region language model corresponding to the target region to which the location information belongs from a plurality of region language models, wherein any region language model among the plurality of region language models is trained based on information points included in the region;

[0116] The processing unit 403 is configured to call the target region language model and the universal language model to perform a first decoding process on the speech data to obtain N candidate recognition results;

[0117] The processing unit 403 is further configured to call the target region language model and the neural network language model to perform a second decoding process on each of the N candidate recognition results, so as to select a target recognition result from the N candidate recognition results.

[0118] In one embodiment, when the processing unit 403 calls the target region language model and the neural network language model to perform a second decoding process on each of the N candidate recognition results to select a target recognition result from the N candidate recognition results, it is specifically configured to perform:

[0119] Obtaining a first fusion probability of each candidate recognition result among the N candidate recognition results, where the first fusion probability of any candidate recognition result is obtained during a first encoding process of the speech data;

[0120] calling the target region language model and the neural network language model to process each candidate recognition result respectively to obtain a second fusion probability of each candidate recognition result;

[0121] A target recognition result is selected from the N candidate recognition results based on the first fusion probability of each candidate recognition result and the second fusion probability of each candidate recognition result.

[0122] In another embodiment, when the processing unit 403 selects the target recognition result from the N candidate recognition results based on the first fusion probability of each candidate recognition result and the second fusion probability of each candidate recognition result, it is specifically configured to perform:

[0123] Fusing the first fusion probability of each candidate recognition result with the second fusion probability of the corresponding candidate recognition result to obtain a target fusion probability of each candidate recognition result;

[0124] The candidate recognition result with the highest target fusion probability among the N candidate recognition results is used as the target recognition result.

[0125] In another embodiment, when the processing unit 403 fuses the first fusion probability of each candidate recognition result with the second fusion probability of the corresponding candidate recognition result to obtain the target fusion probability of each candidate recognition result, it is further specifically configured to perform:

[0126] Obtaining an acoustic probability corresponding to any one of the N candidate recognition results, where the acoustic probability corresponding to any one of the candidate recognition results is obtained during a first decoding process of the speech data;

[0127] Obtaining an acoustic scaling parameter and a fusion ratio parameter, performing a logarithmic operation on the acoustic probability, multiplying the logarithmic operation result by the acoustic scaling parameter, and fusing the first fusion probability and the second fusion probability according to the fusion ratio parameter;

[0128] The multiplication result and the fusion processing result are added together to obtain the target fusion probability of any candidate recognition result.

[0129] In another embodiment, the N candidate recognition results include a target candidate recognition result, and the processing unit 403 is specifically configured to perform the following when obtaining the first fusion probability of each candidate recognition result in the N candidate recognition results:

[0130] Calling the target regional language model to perform a first decoding process on the speech data to obtain a first language probability of the target candidate recognition result;

[0131] Calling the universal language model to perform a first decoding process on the speech data to obtain a second language probability of the target candidate recognition result;

[0132] fusing the first language probability of the target candidate recognition result and the second language probability of the target candidate recognition result using a first fusion parameter to obtain a fused language probability of the target candidate recognition result;

[0133] A first fusion probability of the target candidate recognition result is obtained based on the fusion language probability of the target candidate recognition result and the second language probability of the target candidate recognition result.

[0134] In another embodiment, when the processing unit 403 calls the target region language model and the neural network language model to process each candidate recognition result respectively to obtain the second fusion probability of each candidate recognition result, it is specifically configured to perform:

[0135] Obtaining the first language probability of the target candidate recognition result;

[0136] Calling the neural network language model to perform a second decoding process on the target candidate recognition result to obtain a third language probability of the target candidate recognition result;

[0137] The first language probability of the target candidate recognition result and the third language probability of the target candidate recognition result are fused using a second fusion parameter to obtain a second fusion probability of the target candidate recognition result.

[0138] In another embodiment, before the determining unit 402 determines the target regional language model corresponding to the speech data from multiple regional language models based on the location information, the determining unit 402 may further be configured to:

[0139] Acquire information point data included in the target area, wherein the information point data includes at least one information point and information point features of each of the at least one information point, wherein the information point features of any information point include one or both of the following: an information point name and an information point alias;

[0140] The information point features of each information point are used to train a regional language model to obtain a target regional language model corresponding to the target region.

[0141] According to one embodiment of the present invention, Figure 2 and Figure 3 The steps involved in the voice data recognition method shown can be Figure 4 The recognition of the speech data is performed by various units in the device shown. For example, Figure 2 The step S201 can be performed by Figure 4 The speech data recognition device is implemented by acquiring unit 401, and step S202 can be performed by Figure 4 The determination unit 402 in the speech data recognition device shown in FIG. 1 is executed, and both step S203 and step S204 can be performed by Figure 4 The processing unit 403 in the speech data recognition device shown in FIG. Figure 3 The steps S301 and S304 shown can be performed by Figure 4 The acquisition unit 401 in the speech data recognition device shown in FIG. 1 is executed, and step S302 can be performed by Figure 4 The determination unit 402 in the speech data recognition device shown in FIG. 4 is executed, and steps S303, S305 and S306 can all be performed by Figure 4 The speech data recognition is performed by the processing unit 403 in the apparatus shown.

[0142] According to another embodiment of the present invention, Figure 4The various units in the voice data recognition device shown are divided based on logical functions. The above-mentioned various units can be individually or all merged into one or more other units to form a structure, or one (or some) of the units can be further divided into multiple functionally smaller units to form a structure, which can achieve the same operation without affecting the realization of the technical effects of the embodiments of the present invention. In other embodiments of the present invention, the above-mentioned voice data recognition device can also include other units. In actual applications, these functions can also be assisted by other units and can be achieved by the collaboration of multiple units.

[0143] According to another embodiment of the present invention, the program can be executed by running a program on a general-purpose computing device such as a computer including a central processing unit (CPU), a random access memory (RAM), a read-only memory (ROM), and other processing elements and storage elements. Figure 2 or Figure 3 A computer program (including program code) for each step of the corresponding method shown in FIG. Figure 4 The speech data recognition device shown in and the speech data recognition method of the embodiment of the present invention are implemented. The computer program can be recorded on a computer storage medium, for example, and loaded into the above-mentioned computing device through the computer storage medium and run therein.

[0144] In an embodiment of the present invention, a processing unit establishes a separate regional language model for each region based on the information points of each region, and determines the target regional language model to be used in the recognition process of the voice data based on the user's location information through a determination unit, thereby solving the problem of inaccurate information point recognition caused by the fact that a general language model is usually only combined with an information point network in a specific region in traditional voice data recognition methods; the processing unit also combines the target regional language model with the general language model and the neural network language model respectively to perform a first decoding process and a second decoding process on the voice data, thereby further improving the recognition accuracy of the voice data input by the user.

[0145] Based on the description of the above method embodiment and apparatus embodiment, the present invention also provides a device for recognizing speech data. Figure 5 The speech data recognition device includes at least a processor 501, an input interface 502, an output interface 503 and a computer storage medium 504, and the processor 501, input interface 502, output interface 503 and computer storage medium 504 in the computer device can be connected via a bus or other means.

[0146] The computer storage medium 504 is a memory device in a computer device, used to store programs and data. It is understood that the computer storage medium 504 here can include both the built-in storage medium in the voice data recognition device and, of course, the extended storage medium supported by the computer device. The computer storage medium 504 provides a storage space that stores the operating system of the computer device. In addition, the storage space also stores one or more instructions suitable for being loaded and executed by the processor 501. These instructions can be one or more computer programs (including program code). It should be noted that the computer storage medium here can be a high-speed RAM memory or a non-volatile memory (non-volatile memory), such as at least one disk storage; optionally, it can also be at least one computer storage medium located away from the aforementioned processor. The processor 501 (or CPU (Central Processing Unit)) is the computing core and control core of the computer device, which is suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to implement the corresponding method flow or corresponding function.

[0147] In one embodiment, the processor 501 may load and execute one or more instructions stored in the computer storage medium 504 to implement the above-mentioned Figure 2 and Figure 3 The corresponding method steps in the embodiment of the method for recognizing speech data shown in FIG. 1 are as follows; in a specific implementation, one or more instructions in the computer storage medium 504 are loaded by the processor 501 and execute the following steps:

[0148] Acquiring voice data input by a target user and location information of the target user;

[0149] Determining a target region language model corresponding to the target region to which the location information belongs from a plurality of region language models, wherein any one of the plurality of region language models is trained based on information points included in the region;

[0150] Calling the target region language model and the universal language model to perform a first decoding process on the speech data to obtain N candidate recognition results;

[0151] The target region language model and the neural network language model are called to perform a second decoding process on each of the N candidate recognition results to select a target recognition result from the N candidate recognition results.

[0152] In one embodiment, the instruction in the computer storage medium 504 for calling the target regional language model and the neural network language model to perform a second decoding process on each of the N candidate recognition results to select a target recognition result from the N candidate recognition results is specifically loaded and executed by the processor 501:

[0153] Obtaining a first fusion probability of each candidate recognition result among the N candidate recognition results, where the first fusion probability of any candidate recognition result is obtained during a first encoding process of the speech data;

[0154] calling the target region language model and the neural network language model to process each candidate recognition result respectively to obtain a second fusion probability of each candidate recognition result;

[0155] A target recognition result is selected from the N candidate recognition results based on the first fusion probability of each candidate recognition result and the second fusion probability of each candidate recognition result.

[0156] In another embodiment, the instruction for selecting a target recognition result from the N candidate recognition results based on the first fusion probability of each candidate recognition result and the second fusion probability of each candidate recognition result in the computer storage medium 504 is specifically loaded and executed by the processor 501:

[0157] Fusing the first fusion probability of each candidate recognition result with the second fusion probability of the corresponding candidate recognition result to obtain a target fusion probability of each candidate recognition result;

[0158] The candidate recognition result with the highest target fusion probability among the N candidate recognition results is used as the target recognition result.

[0159] In another embodiment, the instruction in the computer storage medium 504 for fusing the first fusion probability of each candidate recognition result with the second fusion probability of the corresponding candidate recognition result to obtain the target fusion probability of each candidate recognition result is specifically loaded and executed by the processor 501:

[0160] Obtaining an acoustic probability corresponding to any one of the N candidate recognition results, where the acoustic probability corresponding to any one of the candidate recognition results is obtained during a first decoding process of the speech data;

[0161] Obtaining an acoustic scaling parameter and a fusion ratio parameter, performing a logarithmic operation on the acoustic probability, multiplying the logarithmic operation result by the acoustic scaling parameter, and fusing the first fusion probability and the second fusion probability according to the fusion ratio parameter;

[0162] The multiplication result and the fusion processing result are added together to obtain the target fusion probability of any candidate recognition result.

[0163] In another embodiment, the N candidate recognition results include a target candidate recognition result, and the first fusion general instruction for obtaining each candidate recognition result in the N candidate recognition results in the computer storage medium 504 is specifically loaded and executed by the processor 501:

[0164] Calling the target regional language model to perform a first decoding process on the speech data to obtain a first language probability of the target candidate recognition result;

[0165] Calling the universal language model to perform a first decoding process on the speech data to obtain a second language probability of the target candidate recognition result;

[0166] fusing the first language probability of the target candidate recognition result and the second language probability of the target candidate recognition result using a first fusion parameter to obtain a fused language probability of the target candidate recognition result;

[0167] A first fusion probability of the target candidate recognition result is obtained based on the fusion language probability of the target candidate recognition result and the second language probability of the target candidate recognition result.

[0168] In another embodiment, the computer storage medium 504 calls the target region language model and the neural network language model to process each candidate recognition result respectively to obtain a second fusion probability instruction for each candidate recognition result, which is specifically loaded and executed by the processor 501:

[0169] Obtaining the first language probability of the target candidate recognition result;

[0170] Calling the neural network language model to perform a second decoding process on the target candidate recognition result to obtain a third language probability of the target candidate recognition result;

[0171] The first language probability of the target candidate recognition result and the third language probability of the target candidate recognition result are fused using a second fusion parameter to obtain a second fusion probability of the target candidate recognition result.

[0172] In yet another embodiment, before the computer storage medium loads the instruction for determining, based on the location information, from multiple regional language models a target regional language model corresponding to the speech data, the processor 501 is further configured to load and execute the following steps:

[0173] Acquire information point data included in the target area, wherein the information point data includes at least one information point and information point features of each of the at least one information point, wherein the information point features of any information point include one or both of the following: an information point name and an information point alias;

[0174] The information point features of each information point are used to train a regional language model to obtain a target regional language model corresponding to the target region.

[0175] In an embodiment of the present invention, the processor establishes a separate regional language model for each region based on the information points of each region, and determines the target regional language model to be used in the recognition process of the voice data based on the user's location information. This solves the problem of inaccurate information point recognition caused by the general language model usually only being combined with the information point network of a specific region in traditional voice data recognition methods. The processor also combines the target regional language model with the general language model and the neural network language model respectively to perform a first decoding process and a second decoding process on the voice data, further improving the recognition accuracy of the voice data input by the user.

[0176] The present application also provides a computer storage medium that stores a computer program for the above-mentioned voice data recognition method. The computer program includes program instructions. When one or more processors load and execute the program instructions, the description of the voice data recognition method in the embodiment can be implemented, which will not be repeated here. The description of the beneficial effects of adopting the same method will not be repeated here. It is understandable that the program instructions can be deployed on one or more devices that can communicate with each other for execution.

[0177] It should be noted that, according to one aspect of the present application, a computer program product or computer program is also provided, the computer program product or computer program including computer instructions, the computer instructions being stored in a computer-readable storage medium. The processor in the speech data recognition device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the speech data recognition device performs the above-mentioned Figure 2 and Figure 3 The method for recognizing speech data shown in the embodiment is provided in various optional manners.

[0178] Furthermore, it should be understood that the above disclosure is only a preferred embodiment of the present invention and certainly cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for recognizing speech data, characterized in that: include: Acquire voice data input by a target user and call a location service component in a terminal to acquire location information of the target user; Determining a target region language model corresponding to the target region to which the location information belongs from a plurality of region language models, wherein any region language model among the plurality of region language models is trained based on information points included in the region, and different regions correspond to different region language models; Calling the target regional language model and the universal language model to perform a first decoding process on the speech data to obtain N candidate recognition results; including: calling the target regional language model to perform a first decoding process on the speech data to obtain a first language probability corresponding to the target candidate recognition result, calling the universal language model to perform a first decoding process on the speech data to obtain a second language probability corresponding to the target candidate recognition result, using a first fusion parameter to fuse the first language probability and the second language probability to obtain a fused language probability of the target candidate recognition result, determining a first fused probability of the target candidate recognition result based on the fused language probability and the second language probability, a first confidence of the candidate recognition result determined by the first fused probability, and selecting N recognition results with the highest first confidence ranking from all recognition results as the N candidate recognition results; wherein each candidate recognition result corresponds to a first fused probability, and the first fusion parameter is used to control the fusion ratio of the first language probability and the second language probability; Calling the target regional language model and the neural network language model to perform a second decoding process on each of the N candidate recognition results to select a target recognition result from the N candidate recognition results; including: calling the neural network language model to perform a second decoding process on the target candidate recognition result to obtain a third language probability of the target candidate recognition result, using a second fusion parameter to fuse the third language probability and the first language probability to obtain a second fusion probability of the target candidate recognition result, and selecting a target recognition result from the N candidate recognition results based on the first fusion probability of each candidate recognition result and the second fusion probability of each candidate recognition result; wherein the second fusion parameter is a constant and is used to balance the target regional language model and the neural network language model.

2. The method according to claim 1, characterized in that The selecting a target recognition result from the N candidate recognition results based on the first fusion probability of each candidate recognition result and the second fusion probability of each candidate recognition result includes: Fusing the first fusion probability of each candidate recognition result with the second fusion probability of the corresponding candidate recognition result to obtain a target fusion probability of each candidate recognition result; The candidate recognition result with the highest target fusion probability among the N candidate recognition results is used as the target recognition result.

3. The method according to claim 2, characterized in that The fusing the first fusion probability of each candidate recognition result and the second fusion probability of the corresponding candidate recognition result to obtain the target fusion probability of each candidate recognition result includes: Obtaining an acoustic probability corresponding to any one of the N candidate recognition results, where the acoustic probability corresponding to any one of the candidate recognition results is obtained during a first decoding process of the speech data; Obtaining an acoustic scaling parameter and a fusion ratio parameter, performing a logarithmic operation on the acoustic probability, multiplying the logarithmic operation result by the acoustic scaling parameter, and fusing the first fusion probability and the second fusion probability according to the fusion ratio parameter; The multiplication result and the fusion processing result are added together to obtain the target fusion probability of any candidate recognition result.

4. The method according to claim 1, wherein Before determining the target regional language model corresponding to the speech data from a plurality of regional language models based on the location information, the method further includes: Acquire information point data included in the target area, wherein the information point data includes at least one information point and information point features of each of the at least one information point, wherein the information point features of any information point include one or both of the following: an information point name and an information point alias; The information point features of each information point are used to train a regional language model to obtain a target regional language model corresponding to the target region.

5. A speech data recognition device, characterized in that: include: An acquiring unit, configured to acquire voice data input by a target user and to invoke a location service component in a terminal to acquire location information of the target user; a determining unit, configured to determine a target region language model corresponding to the target region to which the location information belongs from a plurality of region language models, wherein any region language model among the plurality of region language models is trained based on information points included in the region, and different regions correspond to different region language models; A processing unit, configured to call the target regional language model and the universal language model to perform a first decoding process on the speech data to obtain N candidate recognition results; the processing unit includes: calling the target regional language model to perform a first decoding process on the speech data to obtain a first language probability corresponding to the target candidate recognition result, calling the universal language model to perform a first decoding process on the speech data to obtain a second language probability corresponding to the target candidate recognition result, using a first fusion parameter to fuse the first language probability and the second language probability to obtain a fused language probability of the target candidate recognition result, determining a first fused probability of the target candidate recognition result based on the fused language probability and the second language probability, a first confidence of the candidate recognition result being determined by the first fused probability, and selecting N recognition results with the highest first confidence ranking from all recognition results as the N candidate recognition results; wherein each candidate recognition result corresponds to a first fused probability, and the first fusion parameter is used to control the fusion ratio of the first language probability and the second language probability; The processing unit is further configured to call the target regional language model and the neural network language model to perform a second decoding process on each of the N candidate recognition results to select a target recognition result from the N candidate recognition results. This process includes: calling the neural network language model to perform a second decoding process on the target candidate recognition result to obtain a third language probability of the target candidate recognition result, using a second fusion parameter to fuse the third language probability with the first language probability to obtain a second fused probability of the target candidate recognition result, and selecting a target recognition result from the N candidate recognition results based on the first fused probability of each candidate recognition result and the second fused probability of each candidate recognition result. The second fusion parameter is a constant and is configured to balance the target regional language model and the neural network language model.

6. A voice data recognition device, characterized in that: The device includes an input interface and an output interface, and also includes: a processor adapted to implement one or more instructions; A computer storage medium storing one or more instructions, wherein the one or more instructions are suitable for being loaded by the processor and executed by the method for recognizing speech data according to any one of claims 1 to 4.

7. A computer storage medium, characterized in that The computer storage medium stores one or more instructions, and the one or more instructions are suitable for being loaded by a processor and executing the voice data recognition method according to any one of claims 1 to 4.

Images