Dialect voice instruction recognition method and device, and electronic device

Abstract

The application discloses a dialect voice instruction recognition method and device and electronic equipment. It is related to the field of biological recognition. The method comprises the following steps: obtaining dialect voice information sent by a target object account; in the case that the target voiceprint feature of the dialect voice information matches the account identifier of the target object account, performing semantic recognition on the dialect voice information by using a dialect voice recognition model to obtain target text information corresponding to the dialect voice information, wherein the dialect voice recognition model is a neural network model pre-trained by using sample dialect voice information; obtaining a target instruction corresponding to the target text information, and sending response voice information for responding to the target instruction. Through the application, the problem that intelligent terminals or staff cannot accurately recognize dialect voice instructions in related technologies is solved.

Description

Technical Field

[0001] This application relates to the field of biometrics, and more specifically, to a method, apparatus, and electronic device for recognizing dialect voice commands. Background Technology

[0002] In banking scenarios, existing speech recognition technologies typically identify speech information based on standard language types. However, different regions have different dialect types, and the method of identifying speech information based on standard language types cannot accurately identify speech information from different dialect types. As a result, intelligent customer service or bank staff cannot respond accurately to users' dialect speech, making it difficult for intelligent customer service or bank staff to communicate with users and affecting the banking experience of users who speak dialects.

[0003] There is currently no effective solution to the problem that smart terminals or staff cannot accurately recognize dialect voice commands in related technologies. Summary of the Invention

[0004] The main objective of this application is to provide a method, device, and electronic device for recognizing dialect voice commands, so as to solve the problem that smart terminals or staff cannot accurately recognize dialect voice commands in related technologies.

[0005] To achieve the above objectives, according to one aspect of this application, a method for recognizing dialect speech commands is provided. The method includes: acquiring dialect speech information sent by a target account; when the target voiceprint features of the dialect speech information match the account identifier of the target account, performing semantic recognition on the dialect speech information using a dialect speech recognition model to obtain text information corresponding to the dialect speech information, wherein the dialect speech recognition model is a neural network model pre-trained using sample dialect speech information; acquiring a target command corresponding to the target text information, and sending response speech information to respond to the target command.

[0006] To achieve the above objectives, according to another aspect of this application, a dialect speech command recognition device is provided. The device includes: a first acquisition unit for acquiring dialect speech information sent by a target account; a recognition unit for performing semantic recognition on the dialect speech information using a dialect speech recognition model, when the target voiceprint features of the dialect speech information match the account identifier of the target account, to obtain text information corresponding to the dialect speech information, wherein the dialect speech recognition model is a neural network model pre-trained using sample dialect speech information; and a second acquisition unit for acquiring a target command corresponding to the target text information and sending response speech information in response to the target command.

[0007] Optionally, the above recognition unit includes: a noise reduction unit for performing noise reduction processing on dialect speech information; a first extraction unit for extracting speech features based on the noise-reduced dialect speech information to obtain target speech features; and a first input unit for inputting the target speech features into the dialect speech recognition model to obtain target text information corresponding to the dialect speech information.

[0008] Optionally, the first input unit includes: a dimensionality reduction unit for performing dimensionality reduction processing on the target speech features to obtain reference speech features; a third acquisition unit for acquiring N key sub-features from the M sub-features included in the reference speech features, where M and N are integers greater than or equal to 1, N is less than or equal to M, and the feature variance value of the key sub-features is greater than or equal to a target threshold; a first input sub-unit for inputting the N key sub-features into a dialect speech recognition model, wherein the dialect speech recognition model includes an N-layer feature recognition network, the first N-1 layers of the N-layer feature recognition network are connected to a ReLU activation network, and the Nth layer of the N-layer feature recognition network is connected to a maxout activation network; and a determination unit for determining the target text information corresponding to the dialect speech information based on the output of the maxout activation network.

[0009] Optionally, the first extraction unit includes a first extraction subunit, used to obtain the linear prediction dynamic difference coefficients of dialect speech information as target speech features;

[0010] The aforementioned recognition device further includes: a fourth acquisition unit for acquiring fbank features of dialect speech information; a second input unit for inputting the first voiceprint feature into the first convolution module of the voiceprint extraction model to obtain the first voiceprint feature; a third input unit for inputting the fbank feature into the second convolution module of the voiceprint extraction model to obtain multiple second voiceprint features, wherein the second convolution module includes multiple dilated convolution layers, and the receptive field of the second voiceprint feature is larger than that of the first voiceprint feature; a fusion unit for inputting the multiple second voiceprint features into the feature fusion module of the voiceprint extraction model for feature fusion processing; and a fifth acquisition unit for acquiring target voiceprint features that match the dialect speech information based on the third voiceprint feature.

[0011] The aforementioned recognition device further includes a sixth acquisition unit, used to acquire the target dialect type of the dialect speech information based on the account identifier of the target object account, and acquire the target dialect speech recognition model corresponding to the target dialect type, wherein the target dialect speech recognition model is a neural network model pre-trained based on sample dialect speech information of the target dialect type; and a seventh acquisition unit, used to acquire the target dialect speech recognition model, wherein the target dialect speech recognition model is a neural network model pre-trained based on sample dialect speech information of multiple dialect types.

[0012] Optionally, the second acquisition unit further includes a sending unit for sending response text information corresponding to the response voice information, wherein the response text information is used to display on the terminal corresponding to the target account.

[0013] To achieve the above objectives, according to another aspect of this application, a computer-readable storage medium is provided, wherein a computer program is stored in the computer program, which is configured to execute the aforementioned dialect speech instruction recognition method when running.

[0014] To achieve the above objectives, according to another aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the aforementioned dialect speech instruction recognition method.

[0015] To achieve the above objectives, according to another aspect of this application, a processor is provided for running a program, wherein the program executes the dialect speech instruction recognition method.

[0016] To achieve the above objectives, according to another aspect of this application, an electronic device is provided, including one or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are run by the one or more processors, the one or more processors execute the dialect speech instruction recognition method.

[0017] This application employs the following steps: obtaining dialect voice information sent by a target account; when the target voiceprint features of the dialect voice information match the account identifier of the target account, using a dialect voice recognition model to perform semantic recognition on the dialect voice information to obtain text information corresponding to the dialect voice information, wherein the dialect voice recognition model is a neural network model pre-trained using sample dialect voice information; obtaining the target instruction corresponding to the target text information, and sending response voice information to respond to the target instruction, thus solving the problem in related technologies where smart terminals or staff cannot accurately recognize dialect voice instructions. This achieves the effect of accurately recognizing different dialect voice instructions and providing accurate responses to them. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0019] Figure 1 This is a schematic diagram of the hardware environment for a dialect voice command recognition method provided in an embodiment of this application;

[0020] Figure 2 This is a flowchart of a dialect speech command recognition method provided according to an embodiment of this application;

[0021] Figure 3 This is a flowchart of a method for feature extraction of dialect speech information according to an embodiment of this application;

[0022] Figure 4 This is a flowchart of a dialect speech recognition method based on an embodiment of this application;

[0023] Figure 5 This is an overall network structure diagram of a voiceprint extraction model provided according to an embodiment of this application;

[0024] Figure 6 This is a schematic diagram of a dialect speech command recognition method provided according to an embodiment of this application;

[0025] Figure 7 A schematic diagram of a dialect voice command recognition device provided in an embodiment of this application;

[0026] Figure 8 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0027] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0028] The dialect voice command recognition method, device, and electronic device in this application can be used in the field of biometrics, or in any field other than biometrics. The application field of the dialect voice command recognition method, device, and electronic device in this application is not limited.

[0029] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0030] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0031] It should be noted that all information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data used for analysis, etc.) involved in this disclosure are information and data authorized by the user or fully authorized by all parties. For example, this system has an interface with relevant users or organizations. Before obtaining relevant information, it is necessary to send an acquisition request to the aforementioned user or organization through the interface, and obtain the relevant information after receiving consent information from the aforementioned user or organization.

[0032] According to one aspect of the present invention, a method for recognizing dialect speech commands is provided. As an optional approach, the above-described method for recognizing dialect speech commands can be applied, but is not limited to, to applications such as... Figure 1 The hardware environment shown includes a dialect semantic command recognition system, wherein the dialect speech command recognition system may include, but is not limited to, terminal device 102, network 110, and server 104; the terminal device 102 includes a human-computer interaction screen, a display, a processor, and a memory. The human-computer interaction screen is used to display the application interface and the results of interactive operations, and also to provide a human-computer interaction interface to receive various interactive operations from users, including drag-and-drop operations and input operations. The server 104 includes a database and a processing engine.

[0033] Assumption Figure 1 Terminal device 102 is the terminal device used for login of the target object's account. The specific process in this embodiment is as follows:

[0034] S102, terminal device 102 sends dialect voice information to server 104 through network 110;

[0035] S104, Server 104 obtains dialect voice information sent by the target account;

[0036] It should be noted that the target account is used to indicate the account used to log in to the terminal device.

[0037] S106, when the target voiceprint features of dialect speech information match the account identifier of the target account, the dialect speech recognition model is used to perform semantic recognition on the dialect speech information to obtain text information corresponding to the dialect speech information. The dialect speech recognition model is a neural network model pre-trained using sample dialect speech information.

[0038] It should be noted that before matching the target voiceprint features with the account identifier of the target object's account, the dialect speech information can be preprocessed to extract clear speech information of the speaker to be acquired. The preprocessing operations for dialect speech information include noise reduction, smoothing, standardization, and silence removal to obtain clear speech information.

[0039] A dialect speech recognition model can be a model pre-trained on an original dialect speech recognition model using multiple speech data from different dialect types. This model can pre-store all existing dialect types to prevent inaccurate recognition of dialect speech data due to the absence of pre-stored dialect types. Furthermore, during the training of the original dialect speech recognition model, if new dialect types are not pre-stored or have not yet appeared when the model is trained, these new dialect types can be stored in the model for later recognition. The model can then be trained to recognize these new dialect types.

[0040] Semantic recognition can convert preprocessed dialect speech information into text information. The target text information can have different language types, and the language type corresponds to the region, such as Mandarin, English, French, etc.

[0041] S108, obtain the target instruction corresponding to the target text information; and send response voice information to respond to the target instruction.

[0042] It should be noted that the target instruction corresponding to the target text information can be a target instruction generated by adding keywords to the key information in the target text information. For example, if the target text information is "I want to apply for a bank card", the corresponding target instruction can be "The user wants to apply for a bank card". The intelligent customer service or terminal can inform the user and staff that "The user wants to apply for a bank card" through voice playback. The target instruction can also be an instruction that only includes the important information in the target text information. For example, the target instruction for the target text information "I want to apply for a bank card" is "Apply for a bank card". The response voice information can be voice information that instructs the user to perform the next operation based on the target instruction. For example, if the target text information is "I want to apply for a bank card", the corresponding voice information can be "Please click on the homepage". The intelligent customer service or terminal will then play the above response voice information through voice playback. After the user / user account performs the operation of clicking on the homepage, the response voice information "Please enter your name" will be played to instruct the user / user account to continue to enter relevant information.

[0043] As another optional implementation, when the terminal device 102 has powerful computing capabilities, the above-described S104 to S108 can also be performed by the terminal device 102. This is just an example, and no limitation is made in this embodiment.

[0044] Optionally, in this embodiment, the terminal device can be a terminal device configured with a target user client, which may include, but is not limited to, at least one of the following: mobile phone (such as Android phone, iOS phone, etc.), laptop computer, tablet computer, PDA, MID (Mobile Internet Devices), PAD, desktop computer, smart TV, vehicle terminal, VR / AR device, etc. The target user client can be a video user client, instant messaging user client, browser user client, educational user client, or other user client that supports providing shooting game tasks. The network can include, but is not limited to, wired networks and wireless networks. The wired network includes local area networks (LANs), metropolitan area networks (MANs), and wide area networks (WANs). The wireless network includes Bluetooth, Wi-Fi, and other networks that enable wireless communication. The server can be a single server, a server cluster consisting of multiple servers, or a cloud server. The above is merely an example, and no limitation is made in this embodiment.

[0045] The dialect voice command recognition method provided in this application obtains dialect voice information sent by a target account; when the target voiceprint features of the dialect voice information match the account identifier of the target account, a dialect voice recognition model is used to perform semantic recognition on the dialect voice information to obtain text information corresponding to the dialect voice information. The dialect voice recognition model is a neural network model pre-trained using sample dialect voice information. The method then obtains the target command corresponding to the target text information and sends response voice information to respond to the target command. This solves the problem in related technologies where smart terminals or staff cannot accurately recognize dialect voice commands. Consequently, it enables smart terminals or staff to make accurate responses and correct instructions based on the dialect voice information issued by the user.

[0046] As an optional implementation method, such as Figure 2 As shown, the method for recognizing dialect speech commands includes the following steps:

[0047] S202, Obtain dialect voice information sent by the target account;

[0048] S204, when the target voiceprint features of dialect speech information match the account identifier of the target account, the dialect speech recognition model is used to perform semantic recognition on the dialect speech information to obtain the text information corresponding to the dialect speech information. The dialect speech recognition model is a neural network model pre-trained using sample dialect speech information.

[0049] S206, obtain the target instruction corresponding to the target text information, and send response voice information to respond to the target instruction.

[0050] It should be noted that the aforementioned dialect voice information may include different dialects from multiple regions; the dialect voice information may also include multiple dialect types. Specifically, in scenarios where the target subject conducts banking business by telephone, there may be instances where the target subject does not switch dialect types in a timely manner. In such cases, the dialect voice information sent by the target subject may include multiple dialect types. The instances of untimely dialect type switching may include, but are not limited to, situations where part of the target subject's Mandarin is Sichuan dialect, or where the target subject's Sichuan dialect has a Henan accent, etc. This application does not limit the scope of such instances.

[0051] The aforementioned voiceprint features can be used to indicate the unique sound wave pattern produced by a user during vocalization. Each user's vocal cords, throat, and oral cavity structure are unique; therefore, the sound wave pattern produced by a user during vocalization is also unique. Voiceprint features can be identified and recognized using voiceprint recognition technology. Voiceprint recognition technology is a biometric technology based on sound features. It identifies and confirms an individual's identity by analyzing features such as the frequency, amplitude, duration, and harmonics of sound. Voiceprint features are unique, stable, and difficult to forge. Therefore, in banking scenarios, they can accurately verify user identities. Based on the target voiceprint features, information such as the user's age, gender, and dialect type can also be identified from dialect speech information. The account identifier for the target user's account can be pre-stored or created for the target user when the target user's account first sends dialect speech information. Semantic recognition can convert dialect speech information into a standard language type. The standard language type can be a language type corresponding to a region. Specifically, the standard language type includes, but is not limited to, multiple languages ​​such as Mandarin, English, and French. In specific application scenarios, the standard language type output by the dialect speech recognition model can be selected and configured according to different regions. For example, if the dialect speech recognition model is a model trained according to different standard language types, the corresponding model can be selected according to the specific region.

[0052] Before matching the target voiceprint features with the account identifier of the target account in S204 above, the dialect speech information can be preprocessed to extract clear speech information of the speaker to be acquired. The clear speech information obtained after the first preprocessing of the dialect speech information includes noise reduction processing, smoothing processing, standardization processing, and silence removal processing.

[0053] It should be noted that the aforementioned response voice information can be voice information that instructs the user / user account to perform the next operation based on the target instruction. For example, in a banking scenario, the target text information is "I want to apply for a bank card," and the corresponding response voice information could be "Please click on the homepage." Then, the above response voice information is played through intelligent customer service or terminal voice. After the user / user account performs the "click on the homepage" operation, the response voice information "Please enter your name" is played to instruct the user to continue entering relevant information, and so on.

[0054] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0055] In this embodiment, dialect voice information sent by the target account is obtained. When the target voiceprint features of the dialect voice information match the account identifier of the target account, a dialect voice recognition model is used to perform semantic recognition on the dialect voice information to obtain text information corresponding to the dialect voice information. The dialect voice recognition model is a neural network model pre-trained using sample dialect voice information. A target instruction corresponding to the target text information is obtained, and response voice information is sent to respond to the target instruction. This solves the problem in related technologies where smart terminals or staff cannot accurately recognize dialect voice instructions. This achieves the effect of smart terminals or staff being able to accurately respond and correctly instruct based on the dialect voice information sent by the user. This allows older users or users who do not speak standard language to conveniently and quickly handle the required business in banking scenarios.

[0056] As an optional implementation, when the voiceprint features of dialect speech information match the account identifier of the target object, the dialect speech recognition model is used to perform semantic processing on the dialect speech information to obtain target text information corresponding to the dialect speech information, including:

[0057] S1 performs noise reduction processing on dialect speech information;

[0058] S2, feature extraction is performed based on the denoised dialect speech information to obtain the target speech features;

[0059] S3 is used to input the target speech features into the dialect speech recognition model to obtain the target text information corresponding to the dialect speech information.

[0060] It should be noted that the noise reduction methods for dialect speech information in S1 above include at least one of the following: frequency domain-based noise reduction, time domain-based noise reduction, deep learning-based noise reduction, and feature enhancement-based noise reduction. Frequency domain-based noise reduction involves converting the speech information from the time domain to the frequency domain, removing the noise components in the spectrum, and then converting the speech information back to the time domain to obtain the denoised speech information. Common algorithms include Fourier transform and short-time Fourier transform. Time domain-based noise reduction directly processes the speech information in the time domain to remove noise. Common algorithms include Wiener filtering and waveform amplitude estimation. Deep learning-based noise reduction utilizes a deep learning model, treating the noise reduction task as an end-to-end classification or regression problem, directly learning noise removal patterns from the input speech information. Deep learning models can be autoencoder models and convolutional neural network models. Feature enhancement-based noise reduction extracts features from the speech information, such as MFCC and PLP, to remove noise. Common algorithms include MMSE and PSD.

[0061] It should be noted that feature extraction can be performed using Fbank or MFCC, or it can be performed by combining Fbank and MFCC with dialect speech information. Figure 3 The diagram shows a flowchart of feature extraction using Fbank and MFCC, including:

[0062] S301 receives preprocessed voice information;

[0063] S302, pre-emphasizes the received voice information to increase the energy of the high-frequency part of the voice information;

[0064] S303 performs frame-by-frame windowing on the pre-emphasis processed speech information to make the speech information more stable.

[0065] S304, DFT (Discrete Fourier Transform), transforms the speech information obtained by frame-by-frame windowing operation from the time domain to the frequency domain to obtain the spectrum, that is, the frequency distribution curve of the speech information;

[0066] S305, Mel filter bank: divides the spectrum into a fixed number of bandwidths, each bandwidth representing a Mel frequency, to simulate the auditory characteristics of the human ear;

[0067] S306, LOG: The purpose of calculating the LOG of the spectrum is to find the cepstrum. The spectrum of speech information can be understood as the product between the envelope and the formants. Calculating the LOG can separate these two terms. The envelope is a smooth curve passing through the main peak points of the spectrum of speech information, and the formants are some regions in the spectrum of speech information where the energy is relatively concentrated.

[0068] S307, Fbank: A front-end processing algorithm that processes speech information in a manner similar to the human ear, thereby improving the performance of speech recognition;

[0069] S308, IDFT (Inverse Discrete Fourier Transform), is used to transform speech information from the frequency domain to the time domain;

[0070] S309 removes noise and emotional influences from speech information and extracts feature values.

[0071] It is understood that when extracting features from speech information, the F-bank method can be used. If the F-bank method is used for feature extraction, then S301 to S307 above are executed; the MFCC method can also be used. If the MFCC method is used for feature extraction, then S301 to S306 and S308 to S309 above are executed; or a combination of F-bank and MFCC can be used for feature extraction. If a combination of the two is used for feature extraction, then S301 to S309 above are executed. The specific method of feature extraction is not limited here.

[0072] By extracting features from dialect speech information in the above manner, the impact of factors such as aliasing, high-order harmonic distortion, and high frequency caused by the human vocal organs themselves and the equipment used to collect speech signals on the quality of the speech signal can be reduced. This ensures that the signal obtained from subsequent speech processing is more uniform and smooth, providing high-quality parameters for signal parameter extraction and improving the quality of speech processing.

[0073] The following is combined Figure 4 The dialect speech recognition model shown above provides a detailed explanation of S1 to S3, as follows: Figure 4 As shown, the specific steps include the following:

[0074] S402, Obtain dialect speech information. The obtained dialect speech information can be one or more pieces, and each piece of dialect speech information can include one or more dialect types.

[0075] S404 preprocesses dialect speech information, including noise reduction processing of dialect speech information.

[0076] S406, Feature extraction is performed on the preprocessed dialect speech information. The feature extraction operation here can refer to the above-mentioned steps. Figure 3 Detailed explanation.

[0077] S408, obtains multiple original dialect speech recognition models from the pattern library;

[0078] S410, by training the original dialect speech recognition model with dialect speech information of different dialect types, multiple dialect speech recognition models that match the dialect types can be obtained respectively.

[0079] S412, Select a dialect speech recognition model from the pattern library that matches the dialect type of the dialect speech information as the target dialect speech recognition model;

[0080] S414, By inputting the target speech features into the target dialect speech recognition model, the target text information corresponding to the dialect speech information can be obtained;

[0081] It should be noted that, in the S410 described above, when training the original dialect speech recognition model, dialect speech information of different dialect types can be used to train an original dialect speech recognition model, thereby obtaining a dialect speech recognition model that can recognize multiple dialect types. At this point, this dialect speech recognition model can be directly used as the target dialect recognition model, eliminating the need to select a corresponding dialect speech recognition model from a pattern library based on the dialect type of the dialect speech information. The target dialect recognition model can also have translation functions for different dialect types, allowing staff to choose according to their needs. For example, when the target dialect recognition model identifies the dialect type of the speech information as English, staff can choose to convert the English speech information into Chinese text, or convert the Sichuan dialect speech information into English text, and so on.

[0082] It is understood that the above S408 to S410 can be executed before S402 to S406, simultaneously with S402 to S406, or after S402 to S406. The specific execution order is not limited here.

[0083] In this embodiment, the dialect speech information is denoised; feature extraction is performed on the denoised dialect speech information to obtain target speech features; the target speech features are input into a dialect speech recognition model to obtain target text information corresponding to the dialect speech information. This not only obtains clear speech information but also extracts accurate target speech features through feature extraction, making the dialect speech recognition model more accurate in recognizing target speech features, thus obtaining more accurate target text information. This improves the accuracy of converting dialect speech information into target text information.

[0084] As an optional implementation, the above-mentioned inputting the target speech features into the dialect speech recognition model to obtain target text information corresponding to the dialect speech information includes:

[0085] S1, Dimensionality reduction of the target speech features is performed to obtain reference speech features;

[0086] S2, obtain N key sub-features from the M sub-features included in the reference speech features, where M and N are integers greater than or equal to 1, N is less than or equal to M, and the feature variance value of the key sub-features is greater than or equal to the target threshold.

[0087] S3, input N key sub-features into the dialect speech recognition model, wherein the dialect speech recognition model includes an N-layer feature recognition network, the first N-1 layers of the N-layer feature recognition network are connected to the ReLU activation network, and the Nth layer of the N-layer feature recognition network is connected to the maxout activation network.

[0088] S4. Determine the target text information corresponding to the dialect speech information based on the output of the maxout activation network.

[0089] It should be noted that in step S1 above, the dimensionality reduction process reduces the dimensionality of the data to increase sample density and reduce noise. There are two main dimensionality reduction methods: feature selection and feature extraction. Feature selection directly selects a portion of the original dimensions to participate in subsequent calculations and modeling processes, replacing all dimensions with the selected dimensions. The entire process does not generate new dimensions; that is, it selects smaller features from the existing features to achieve dimensionality reduction. Feature extraction, on the other hand, uses certain mathematical transformation methods to map data points from high-dimensional space to low-dimensional space, and then uses the mapped variable features to represent the original overall features. In other words, it achieves dimensionality reduction by combining existing features.

[0090] Before performing dimensionality reduction on the target speech features in step S1 above, a corpus can be established first. The corpus includes various dialect types, such as Cantonese, Sichuanese, Northeastern Mandarin, Henanese, and Shanghainese. Different speech information can also be used as training samples. After the training samples are disordered, they are normalized. The normalization process includes calculating the average frame and variance frame of all frames in the speech information, and then performing the above dimensionality reduction on the normalized speech information.

[0091] It should be noted that the operation in step S2 above is to select N sub-features with larger variance values ​​from the M sub-features after dimensionality reduction. Before selection, the M sub-features can be sorted from largest to smallest according to the size of the variance values, and the top N sub-features are selected as key sub-features.

[0092] It is understandable that the steps in S3 above can be performed after S2, which yields N key sub-features, and then a dialect speech recognition model with an N-layer feature recognition network is built based on the N key sub-features.

[0093] It is understandable that in step S4 above, both the ReLU activation network and the maxout activation network are used to map the input features to the output.

[0094] In this embodiment, reference speech features are obtained by dimensionality reduction of the target speech features. N key sub-features are then obtained from the M sub-features included in the reference speech features, where M and N are integers greater than or equal to 1, N is less than or equal to M, and the feature variance of the key sub-features is greater than or equal to a target threshold. These N key sub-features are input into a dialect speech recognition model, which includes an N-layer feature recognition network. The first N-1 layers of the N-layer feature recognition network are connected to a ReLU activation network, and the Nth layer is connected to a maxout activation network. The target text information corresponding to the dialect speech information is determined based on the output of the maxout activation network. The output of the dialect speech recognition model can be clearly and accurately observed. Furthermore, the above steps can be used to train the dialect speech recognition model to improve its accuracy.

[0095] As an optional implementation, the above-mentioned speech feature extraction based on the denoised dialect speech information to obtain the target speech features includes:

[0096] S1, obtain the linear prediction dynamic difference coefficients of dialect speech information as target speech features.

[0097] It should be noted that the aforementioned linear predictive dynamic difference coefficients achieve signal compression and reconstruction through the analysis and prediction of speech information.

[0098] Using the linear prediction dynamic difference coefficients obtained from the dialect speech information described above as target speech features can improve the accuracy of feature extraction, achieving higher precision in extracting speech features.

[0099] As an optional implementation, the above-mentioned speech feature extraction based on the denoised dialect speech information to obtain the target speech features includes:

[0100] S1, obtain the fbank features of dialect speech information;

[0101] Understandably, the fbank method is used here for feature extraction, such as... Figure 3 As shown, it can be executed Figure 3 S301 to S307 in the series.

[0102] S2, input the fbank features into the first convolutional module in the voiceprint extraction model to obtain the first voiceprint features;

[0103] It should be noted that the input fbank features are 80-dimensional fbank features with a length of T. In reality, the audio is segmented into segments, with each segment lasting 3 seconds.

[0104] S3, input the first voiceprint feature into the second convolution module in the voiceprint extraction model to obtain multiple second voiceprint features. The second convolution module includes multiple dilated convolution layers, and the receptive field of the second voiceprint feature is larger than that of the first voiceprint feature.

[0105] S4, input multiple second voiceprint features into the feature fusion module in the voiceprint extraction model for feature fusion processing;

[0106] S5, based on the third voiceprint feature, obtains the target voiceprint feature that matches the dialect speech information.

[0107] The following is combined Figure 5 The network structure of the voiceprint extraction model shown above provides a detailed explanation of S1 to S5. An 80-dimensional fbank feature of length T is input into the first convolutional module 501 of the voiceprint extraction model to obtain the first voiceprint feature. The first convolutional module 501 is processed by a Conv1D+ReLU+BN module. ReLU is the activation function, and BN normalizes the input fbank feature, which can accelerate the training and convergence speed of the network. The obtained first voiceprint feature is input into the second convolutional module 502, which consists of multiple SE-Res2Block modules, to obtain the second voiceprint feature. The number of layers in the SE-Res2Block module in the second convolutional module 502 is the number of channels minus two. Figure 5 The network structure of the voiceprint extraction model shown lists three SE-Res2Block modules, which are actually implemented with five layers: [1024, 1024, 1024, 1024, 3072]. The SE-Res2Block module is also a dilated convolution with dilations of [1, 2, 3, 4, 1]. This allows for the generation of multiple second voiceprint features with different receptive fields. Inputting these multiple second voiceprint features into the feature fusion module 503 performs feature fusion on the multiple layers of second voiceprint features with different receptive fields. The input to the feature fusion module 503 is a concatenation of the outputs of each preceding SE-Res2Block module. In other words, each preceding SE-Res2Block module has certain preceding and following connections, i.e., context.

[0108] After feature fusion processing in feature fusion module 503, the method further includes: inputting the feature fusion result into probability pooling module 504 to perform probability pooling processing on the feature fusion result; and inputting the probability pooling processing result into linear combination module 505 to achieve linear combination of features, so that the input of each layer maintains the same distribution, and the output dimension is preset. The preset output dimension value in this embodiment is as follows: Figure 5The result of the linear combination processing is shown as 192; the linear combination processing result is input into the classification module 506, which can classify the linear combination result to obtain the target voiceprint features corresponding to multiple speakers and the number of speakers.

[0109] By processing the fbank features into the voiceprint extraction model as described above, the target voiceprint features can be obtained accurately not only when the dialect speech information includes a single dialect speech object (the speaker in the dialect speech information), but also when the dialect speech information includes multiple dialect speech objects, the number of dialect speech objects and the target voiceprint features matched with each dialect speech object can be obtained.

[0110] As an optional implementation, before performing semantic recognition on dialect speech information using a dialect speech recognition model to obtain target text information corresponding to the dialect speech information, the method further includes at least one of the following:

[0111] S1. Obtain the target dialect type of the dialect speech information based on the account identifier of the target object account, and obtain the target dialect speech recognition model corresponding to the target dialect type. The target dialect speech recognition model is a neural network model pre-trained based on the sample dialect speech information of the target dialect type.

[0112] S2, Obtain the target dialect speech recognition model, wherein the target dialect speech recognition model is a neural network model pre-trained based on sample dialect speech information of multiple dialect types.

[0113] It should be noted that in step S1 above, the target dialect speech recognition model can be a dialect speech recognition model corresponding to the target dialect type, or it can be a dialect speech recognition model that can recognize multiple dialect types. Even if the target dialect speech recognition model can recognize multiple dialect types, it still needs to be trained first. Before training, a corpus is established, which includes various dialect types, such as Cantonese, Sichuanese, Northeastern Mandarin, Henan dialect, and Shanghainese. Then, speech information from different dialect types is used as training samples, and the training samples are shuffled and normalized. The normalization process may include: calculating the average frame and variance frame for all frames in a speech information, then performing feature dimensionality reduction on the normalized data, and selecting N principal features based on the variance ratio of each principal component after dimensionality reduction, prioritizing features with larger variance ratios. An N-layer neural network is built based on the N principal features, with the first N-1 layers connected to a ReLU activation network and the Nth layer connected to a maxout activation network.

[0114] Understandably, in step S2 above, when training the original dialect speech recognition model, dialect speech information from different dialect types can be used to train an original dialect speech recognition model, thereby obtaining a dialect speech recognition model that can recognize multiple dialect types. This dialect speech recognition model can then be directly used as the target dialect recognition model, thus avoiding the need to select a dialect speech recognition model from a pattern library that corresponds to the dialect type of the dialect speech information.

[0115] By obtaining the target dialect type from the account identifier of the target object's speech information, and then obtaining the target dialect speech recognition model corresponding to the target dialect type, the target dialect speech recognition model is a neural network model pre-trained based on sample dialect speech information of the target dialect type. Alternatively, a target dialect speech recognition model can be obtained, which is a neural network model pre-trained based on sample dialect speech information of multiple dialect types. This allows for targeted recognition of dialect speech information of different types, achieving the technical effect of accurately obtaining text information that matches the dialect speech information of different types.

[0116] As an optional implementation, when acquiring the target instruction corresponding to the target text information and sending response voice information for responding to the target instruction, the method further includes:

[0117] S1, send response text information corresponding to the response voice information, wherein the response text information is used to display on the terminal corresponding to the target account.

[0118] It should be noted that the target instruction corresponding to the target text information can be a target instruction generated by adding keywords to the key information in the target text information. For example, in a banking scenario, if the target text information is "I want to apply for a bank card," the target instruction could be "The user wants to apply for a bank card." If the user is conducting business offline, the intelligent customer service or terminal can also inform bank staff via voice playback that "The user wants to apply for a bank card." The target instruction can also be an instruction that only includes the important information in the target text information. For example, in a banking scenario, if the target text information is "I want to apply for a bank card," the corresponding target instruction is "Apply for a bank card." The response voice information can be voice information that instructs the user to perform the next operation based on the target instruction. For example, in a banking scenario, if the target text information is "I want to apply for a bank card," the corresponding response voice information could be "Please click on the homepage," and then the intelligent customer service or terminal can play the aforementioned response via voice playback. In response to voice information, after the user / user account performs the click on the homepage operation, the user can provide a second dialect voice message (e.g., the second dialect voice message is "I have clicked on the homepage"). If the user account is required to send a second dialect voice message, the first dialect voice message sent by the user account is the first dialect voice message. At this time, voiceprint recognition can be performed again to confirm the user's identity, or dialect voice recognition can be performed directly to recognize the second dialect language to obtain the second text information. Then, the second response voice message "Please enter your name" corresponding to the second command will be played, instructing the user to continue performing the operation corresponding to the second command. After the user / user account performs the click on the homepage operation, the user / user account may also not need to provide a second dialect voice message again. The second command can be triggered directly after the user performs the click on the homepage operation, instructing the user / user account to continue to the next operation.

[0119] By sending and responding to the voice information, the response text information is displayed on the terminal corresponding to the target account. This not only allows the smart terminal or staff to clearly understand the user's needs, but also enables accurate guidance to the user based on those needs. This greatly improves the user experience and ensures smooth human-computer interaction (between the user and the smart terminal) and smooth communication between people (between the user and the staff).

[0120] The application scenarios of this application can be human-computer interaction scenarios in banking business scenarios, human-to-human communication scenarios where this application solution is used as a translation function, or human-computer interaction scenarios in other scenarios. The following will illustrate these scenarios in conjunction with... Figure 6To illustrate this application, user 602 provides dialect voice information to terminal device 604. The device receiving the dialect voice information provided by the user can be terminal device 604 or intelligent customer service 610. After receiving the dialect voice information, the terminal device inputs the dialect voice information into voiceprint extraction model 606. Voiceprint extraction model 606 is used to verify the user's identity. After the user's identity verification is successful, dialect voice recognition model 608 processes the dialect voice information and converts the dialect voice information into accurate target text information corresponding to the dialect voice information. Then, the target text information is sent to intelligent customer service 610.

[0121] For example, user 602 provides terminal device 604 with dialect voice information, "I want to change the mobile phone number linked to my bank card." Voiceprint extraction model 606 verifies user 602's identity based on the dialect voice information. Identity verification can be triggered when user 602 provides a portion of the dialect voice information or when voiceprint extraction model 604 receives a portion of the dialect voice information; alternatively, it can be triggered when user 602 provides the complete dialect voice information or when voiceprint extraction model 604 receives the complete dialect voice information. The specific triggering method for identity verification is not limited in this application. After successful identity verification, dialect voice recognition model 608 converts the dialect voice information into target text information, then obtains the target instruction corresponding to the target text information, and sends the response voice message used to respond to the target instruction and the corresponding response text information to intelligent customer service 610. Intelligent customer service 610 then displays the response text information and plays the response voice information.

[0122] Through the above embodiments of this application, not only can the identity of the user be quickly confirmed, but business can also be processed quickly for users in different regions. This avoids the problem of not being able to accurately and timely process business for users due to the inability to accurately obtain the user's needs based on the user's dialect voice information, and greatly improves the user's experience.

[0123] This application also provides a dialect voice command recognition device. It should be noted that the dialect voice command recognition device of this application can be used to execute the dialect voice command recognition method provided in this application. The following describes the dialect voice command recognition device provided in this application.

[0124] Figure 7 This is a schematic diagram of a dialect voice command recognition device according to an embodiment of this application. Figure 7 As shown, the device includes:

[0125] The first acquisition unit 702 is used to acquire dialect voice information sent by the target account;

[0126] The recognition unit 704 is used to perform semantic recognition of dialect speech information by using a dialect speech recognition model when the target voiceprint features of dialect speech information match the account identifier of the target object account, and obtain text information corresponding to the dialect speech information. The dialect speech recognition model is a neural network model pre-trained using sample dialect speech information.

[0127] The second acquisition unit 706 is used to acquire the target instruction corresponding to the target text information and send response voice information in response to the target instruction.

[0128] Optionally, in the dialect speech command recognition device provided in the embodiments of this application, the recognition unit 704 further includes a noise reduction unit for performing noise reduction processing on the dialect speech information; a first extraction unit for extracting speech features based on the noise-reduced dialect speech information to obtain target speech features; and a first input unit for inputting the target speech features into the dialect speech recognition model to obtain target text information corresponding to the dialect speech information.

[0129] Optionally, the first input unit includes: a dimensionality reduction unit for performing dimensionality reduction processing on the target speech features to obtain reference speech features; a third acquisition unit for acquiring N key sub-features from the M sub-features included in the reference speech features, where M and N are integers greater than or equal to 1, N is less than or equal to M, and the feature variance value of the key sub-features is greater than or equal to a target threshold; a first input sub-unit for inputting the N key sub-features into a dialect speech recognition model, wherein the dialect speech recognition model includes an N-layer feature recognition network, the first N-1 layers of the N-layer feature recognition network are connected to a ReLU activation network, and the Nth layer of the N-layer feature recognition network is connected to a maxout activation network; and a determination unit for determining the target text information corresponding to the dialect speech information based on the output of the maxout activation network.

[0130] Optionally, the first extraction unit includes a first extraction subunit, used to obtain the linear prediction dynamic difference coefficients of dialect speech information as target speech features.

[0131] Optionally, the aforementioned dialect speech command recognition device further includes: a fourth acquisition unit for acquiring fbank features of dialect speech information; a second input unit for inputting the first voiceprint feature into the first convolution module of the voiceprint extraction model to obtain the first voiceprint feature; a third input unit for inputting the fbank feature into the second convolution module of the voiceprint extraction model to obtain multiple second voiceprint features, wherein the second convolution module includes multiple dilated convolution layers, and the receptive field of the second voiceprint feature is larger than that of the first voiceprint feature; a fusion unit for inputting the multiple second voiceprint features into the feature fusion module of the voiceprint extraction model for feature fusion processing; and a fifth acquisition unit for acquiring target voiceprint features that match the dialect speech information based on the third voiceprint feature.

[0132] Optionally, the aforementioned dialect speech command recognition device further includes: a sixth acquisition unit, used to acquire the target dialect type of the dialect speech information based on the account identifier of the target object account, and acquire the target dialect speech recognition model corresponding to the target dialect type, wherein the target dialect speech recognition model is a neural network model pre-trained based on sample dialect speech information of the target dialect type; and a seventh acquisition unit, used to acquire the target dialect speech recognition model, wherein the target dialect speech recognition model is a neural network model pre-trained based on sample dialect speech information of multiple dialect types.

[0133] Optionally, the second acquisition unit 706 further includes a sending unit for sending response text information corresponding to the response voice information, wherein the response text information is used to display on the terminal corresponding to the target account.

[0134] The dialect voice command recognition device provided in this application embodiment solves the problem in related technologies where smart terminals or staff cannot accurately recognize dialect voice commands. It achieves the effect of accurately recognizing different dialect voice commands and responding accurately to them, thereby improving the efficiency of business processing and quickly confirming the identity of the user, thus improving the user experience.

[0135] The dialect speech command recognition device includes a processor and a memory. The aforementioned units are all stored in the memory as program units, and the processor executes the aforementioned program units stored in the memory to realize the corresponding functions.

[0136] The processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and adjusting kernel parameters can address the problem of smart terminals or staff being unable to accurately recognize dialectal voice commands in related technologies.

[0137] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0138] This invention provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements a method for recognizing dialect speech commands.

[0139] This invention provides a processor for running a program, wherein the program executes a method for recognizing dialect speech commands during runtime.

[0140] like Figure 8 As shown, this embodiment of the invention provides an electronic device, which includes a processor 804, a memory 802 and a program stored in the memory and executable on the processor, as well as a transmission device 806, a display 808 and a connection bus 810. The device described herein can be a server, PC, PAD, mobile phone, etc. When the processor executes the program, it performs the following steps:

[0141] Obtain dialect voice information sent by the target account; if the target voiceprint features of the dialect voice information match the account identifier of the target account, use a dialect voice recognition model to perform semantic recognition on the dialect voice information to obtain text information corresponding to the dialect voice information. The dialect voice recognition model is a neural network model pre-trained using sample dialect voice information; obtain the target command corresponding to the target text information and send response voice information to respond to the target command.

[0142] When the voiceprint features of dialect speech information match the account identifier of the target account, the dialect speech recognition model is used to perform semantic recognition on the dialect speech information to obtain the target text information corresponding to the dialect speech information. This includes: denoising the dialect speech information; extracting speech features based on the denoised dialect speech information to obtain the target speech features; and inputting the target speech features into the dialect speech recognition model to obtain the target text information corresponding to the dialect speech information.

[0143] The process of inputting target speech features into a dialect speech recognition model to obtain target text information corresponding to the dialect speech information includes: dimensionality reduction of the target speech features to obtain reference speech features; extraction of N key sub-features from M sub-features included in the reference speech features, where M and N are integers greater than or equal to 1, N is less than or equal to M, and the feature variance of the key sub-features is greater than or equal to a target threshold; inputting the N key sub-features into the dialect speech recognition model, wherein the dialect speech recognition model includes an N-layer feature recognition network, the first N-1 layers of the N-layer feature recognition network are connected to a ReLU activation network, and the Nth layer of the N-layer feature recognition network is connected to a maxout activation network; and determining the target text information corresponding to the dialect speech information based on the output of the maxout activation network.

[0144] Based on the denoised dialect speech information, speech features are extracted to obtain target speech features, including: obtaining the linear prediction dynamic difference coefficients of dialect speech information as target speech features;

[0145] Before using a dialect speech recognition model to perform semantic recognition on dialect speech information and obtain the target text information corresponding to the dialect speech information, the method further includes: obtaining the fbank features of the dialect speech information; inputting the fbank features into the first convolutional module of the speaker extraction model to obtain the first speaker features; inputting the first speaker features into the second convolutional module of the speaker extraction model to obtain multiple second speaker features, wherein the second convolutional module includes multiple dilated convolutional layers, and the receptive field of the second speaker features is larger than that of the first speaker features; inputting the multiple second speaker features into the feature fusion module of the speaker extraction model for feature fusion processing to obtain the third speaker features; and obtaining the target speaker features that match the dialect speech information based on the third speaker features.

[0146] Before using a dialect speech recognition model to perform semantic recognition on dialect speech information and obtain the target text information corresponding to the dialect speech information, the method includes one of the following: obtaining the target dialect type of the dialect speech information based on the account identifier of the target object account; obtaining the target dialect speech recognition model corresponding to the target dialect type, wherein the target dialect speech recognition model is a neural network model pre-trained based on sample dialect speech information of the target dialect type; or obtaining the target dialect speech recognition model, wherein the target dialect speech recognition model is a neural network model pre-trained based on sample dialect speech information of multiple dialect types.

[0147] When acquiring the target instruction corresponding to the target text information and sending response voice information to respond to the target instruction, the method further includes: sending response text information corresponding to the response voice information, wherein the response text information is used to display on the terminal corresponding to the target object account.

[0148] This application also provides a computer program product, which, when executed on a data processing device, is suitable for executing a program that initializes the following method steps:

[0149] Obtain dialect voice information sent by the target account; if the target voiceprint features of the dialect voice information match the account identifier of the target account, use a dialect voice recognition model to perform semantic recognition on the dialect voice information to obtain text information corresponding to the dialect voice information. The dialect voice recognition model is a neural network model pre-trained using sample dialect voice information; obtain the target command corresponding to the target text information and send response voice information to respond to the target command.

[0150] When the voiceprint features of dialect speech information match the account identifier of the target account, the dialect speech recognition model is used to perform semantic recognition on the dialect speech information to obtain the target text information corresponding to the dialect speech information. This includes: denoising the dialect speech information; extracting speech features based on the denoised dialect speech information to obtain the target speech features; and inputting the target speech features into the dialect speech recognition model to obtain the target text information corresponding to the dialect speech information.

[0151] The process of inputting target speech features into a dialect speech recognition model to obtain target text information corresponding to the dialect speech information includes: dimensionality reduction of the target speech features to obtain reference speech features; extraction of N key sub-features from M sub-features included in the reference speech features, where M and N are integers greater than or equal to 1, N is less than or equal to M, and the feature variance of the key sub-features is greater than or equal to a target threshold; inputting the N key sub-features into the dialect speech recognition model, wherein the dialect speech recognition model includes an N-layer feature recognition network, the first N-1 layers of the N-layer feature recognition network are connected to a ReLU activation network, and the Nth layer of the N-layer feature recognition network is connected to a maxout activation network; and determining the target text information corresponding to the dialect speech information based on the output of the maxout activation network.

[0152] Based on the denoised dialect speech information, speech features are extracted to obtain target speech features, including: obtaining the linear prediction dynamic difference coefficients of dialect speech information as target speech features;

[0153] Before using a dialect speech recognition model to perform semantic recognition on dialect speech information and obtain the target text information corresponding to the dialect speech information, the method further includes: obtaining the fbank features of the dialect speech information; inputting the fbank features into the first convolutional module of the speaker extraction model to obtain the first speaker features; inputting the first speaker features into the second convolutional module of the speaker extraction model to obtain multiple second speaker features, wherein the second convolutional module includes multiple dilated convolutional layers, and the receptive field of the second speaker features is larger than that of the first speaker features; inputting the multiple second speaker features into the feature fusion module of the speaker extraction model for feature fusion processing to obtain the third speaker features; and obtaining the target speaker features that match the dialect speech information based on the third speaker features.

[0154] Before using a dialect speech recognition model to perform semantic recognition on dialect speech information and obtain the target text information corresponding to the dialect speech information, the method includes one of the following: obtaining the target dialect type of the dialect speech information based on the account identifier of the target object account; obtaining the target dialect speech recognition model corresponding to the target dialect type, wherein the target dialect speech recognition model is a neural network model pre-trained based on sample dialect speech information of the target dialect type; or obtaining the target dialect speech recognition model, wherein the target dialect speech recognition model is a neural network model pre-trained based on sample dialect speech information of multiple dialect types.

[0155] When acquiring the target instruction corresponding to the target text information and sending response voice information to respond to the target instruction, the method further includes: sending response text information corresponding to the response voice information, wherein the response text information is used to display on the terminal corresponding to the target object account.

[0156] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0157] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0158] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0159] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0160] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0161] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0162] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0163] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0164] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0165] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for recognizing dialect speech commands, characterized in that, include: Obtain dialect voice messages sent by the target account; Obtain the fbank features of the dialect speech information; The fbank features are input into the first convolutional module of the voiceprint extraction model to obtain the first voiceprint features. The first convolutional module is processed by a Conv1D+ReLU+BN module, where ReLU is the activation function and BN is the normalization function for the fbank features. The first voiceprint feature is input into the second convolution module in the voiceprint extraction model to obtain multiple second voiceprint features. The second convolution module includes multiple dilated convolution layers. The number of dilated convolution layers is positively correlated with the number of channels in the second convolution module. The receptive field of the second voiceprint feature is larger than that of the first voiceprint feature. Multiple second voiceprint features are input into the feature fusion module of the voiceprint extraction model for feature fusion processing to obtain third voiceprint features; The third voiceprint feature is subjected to probability pooling to obtain the probability pooling result; The probability pooling results are linearly combined to obtain a linear processing result; The linear processing results are classified to obtain target voiceprint features that match the dialect speech information; When the target voiceprint features of the dialect speech information match the account identifier of the target object account, the dialect speech recognition model is used to perform semantic recognition on the dialect speech information to obtain target text information corresponding to the dialect speech information. The dialect speech recognition model is a neural network model pre-trained using sample dialect speech information. Obtain the target instruction corresponding to the target text information, and send response voice information in response to the target instruction.

2. The method according to claim 1, characterized in that, When the voiceprint features of the dialect speech information match the account identifier of the target account, the dialect speech recognition model is used to perform semantic recognition on the dialect speech information to obtain target text information corresponding to the dialect speech information, including: The dialect speech information is subjected to noise reduction processing; Based on the denoised dialect speech information, speech features are extracted to obtain the target speech features; The target speech features are input into the dialect speech recognition model to obtain the target text information corresponding to the dialect speech information.

3. The method according to claim 2, characterized in that, The step of inputting the target speech features into the dialect speech recognition model to obtain the target text information corresponding to the dialect speech information includes: The target speech features are subjected to dimensionality reduction processing to obtain reference speech features; N key sub-features are obtained from the M sub-features included in the reference speech features, wherein M and N are integers greater than or equal to 1, N is less than or equal to M, and the feature variance value of the key sub-features is greater than or equal to a target threshold. The N key sub-features are input into the dialect speech recognition model, wherein the dialect speech recognition model includes an N-layer feature recognition network, the first N-1 layers of the N-layer feature recognition network are connected to a ReLU activation network, and the Nth layer of the N-layer feature recognition network is connected to a maxout activation network. The target text information corresponding to the dialect speech information is determined based on the output of the maxout activation network.

4. The method according to claim 2, characterized in that, The step of extracting speech features from the dialect speech information after noise reduction to obtain target speech features includes: obtaining the linear prediction dynamic difference coefficients of the dialect speech information as the target speech features.

5. The method according to claim 1, characterized in that, Before performing semantic recognition on the dialect speech information using a dialect speech recognition model to obtain the target text information corresponding to the dialect speech information, the method further includes one of the following: The target dialect type of the dialect speech information is obtained based on the account identifier of the target object account, and the target dialect speech recognition model corresponding to the target dialect type is obtained. The target dialect speech recognition model is a neural network model pre-trained based on the sample dialect speech information of the target dialect type. Obtain a target dialect speech recognition model, wherein the target dialect speech recognition model is a neural network model pre-trained based on the sample dialect speech information of multiple dialect types.

6. The method according to claim 1, characterized in that, When acquiring the target instruction corresponding to the target text information and sending response voice information in response to the target instruction, the method further includes: Send response text information corresponding to the response voice information, wherein the response text information is used to display on the terminal corresponding to the target object account.

7. A dialect voice command recognition device, characterized in that, include: The first acquisition unit is used to acquire dialect voice information sent by the target account; The recognition unit is used to perform semantic recognition on the dialect speech information using a dialect speech recognition model when the target voiceprint features of the dialect speech information match the account identifier of the target object account, so as to obtain target text information corresponding to the dialect speech information. The dialect speech recognition model is a neural network model pre-trained using sample dialect speech information. The second acquisition unit is used to acquire the target instruction corresponding to the target text information and send response voice information in response to the target instruction. The device is further configured to acquire fbank features of the dialect speech information; input the fbank features into a first convolutional module in the voiceprint extraction model to obtain a first voiceprint feature, wherein the first convolutional module is processed by a Conv1D+ReLU+BN module, where ReLU is an activation function and BN is a normalization process for the fbank features; input the first voiceprint feature into a second convolutional module in the voiceprint extraction model to obtain multiple second voiceprint features, wherein the second convolutional module includes multiple dilated convolutional layers, the number of dilated convolutional layers is positively correlated with the number of channels in the second convolutional module, and the receptive field of the second voiceprint features is larger than that of the first voiceprint features; input the multiple second voiceprint features into a feature fusion module in the voiceprint extraction model for feature fusion processing to obtain a third voiceprint feature; perform probability pooling processing on the third voiceprint feature to obtain a probability pooling processing result; perform linear combination of the probability pooling result to obtain a linear processing result; and classify the linear processing result to obtain a target voiceprint feature matching the dialect speech information.

8. A processor, characterized in that, The processor is used to run a program, wherein the program executes the method according to any one of claims 1 to 6 when it runs.

9. An electronic device, characterized in that, It includes one or more processors and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method of any one of claims 1 to 6.

Images