Voice task processing method and device, electronic equipment and storage medium

Abstract

The present disclosure relates to a speech task processing method and device, electronic equipment and storage medium. The method comprises: obtaining to-be-processed speech information corresponding to a preset speech task; inputting the to-be-processed speech information into a speech feature extraction network comprising at least one time-frequency attention layer for feature extraction processing to obtain target speech feature information corresponding to the to-be-processed speech information, wherein any time-frequency attention layer is configured to extract speech feature information from the to-be-processed speech information based on learning the importance of time-frequency information in the to-be-processed speech information; and executing the preset speech task based on the target speech feature information to obtain a task processing result. The embodiments of the present disclosure can greatly improve the representation accuracy of speech feature information and the processing performance of speech tasks.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of artificial intelligence, and particularly relates to a speech task processing method and device, electronic equipment and storage medium. BACKGROUND

[0002] With the research and progress of artificial intelligence technology, artificial intelligence technology is researched and applied in multiple fields; a speech feature extraction network based on artificial intelligence technology is usually deployed at the front end of a speech processing system, such as an automatic speech recognition system, a multi-language translation system, etc. The speech feature extraction network can extract speech feature information from speech information, which can represent an input speech information and be used for subsequent downstream tasks, such as speaker recognition, dialect classification, and other high-order speech processing tasks.

[0003] In the related art, with the speech feature extraction network based on artificial intelligence technology, the network input speech information frame number is constantly increasing from a traditional shallow speech network to a deep learning network such as a time-delay neural network, although the amount of speech feature information that can be learned is improved, but with the increase of the frame number, the speech information of each frame needs to be averaged, which is obviously difficult to capture the truly useful features in a speech, so that the speech feature information extracted by the speech feature extraction network in the related art is difficult to accurately represent the input speech information, and further leads to the performance degradation of the downstream speech task. SUMMARY

[0004] The present disclosure provides a speech task processing method, device, electronic equipment and storage medium to at least solve the technical problems that the speech feature information cannot be accurately extracted in the related art, and the speech task processing performance is degraded. The technical solutions of the present disclosure are as follows:

[0005] According to a first aspect of an embodiment of the present disclosure, a speech task processing method is provided, comprising:

[0006] obtaining to-be-processed speech information corresponding to a preset speech task;

[0007] inputting the to-be-processed speech information into a speech feature extraction network comprising at least one time-frequency attention layer for feature extraction processing to obtain target speech feature information corresponding to the to-be-processed speech information, and any time-frequency attention layer is used to extract speech feature information from the to-be-processed speech information on the basis of learning the importance of time-frequency information in the to-be-processed speech information;

[0008] performing the preset speech task based on the target speech feature information to obtain a task processing result.

[0009] In an optional embodiment, the at least one time-frequency attention layer is one time-frequency attention layer, and the time-frequency attention layer includes a time domain attention layer, a frequency domain attention layer, a first weight fusion layer, and a weighted feature extraction layer; the feature extraction processing of the speech feature extraction network including at least one time-frequency attention layer by inputting the to-be-processed speech information includes:

[0010] The to-be-processed speech information is input into the time domain attention layer for feature weight learning processing in the time domain dimension to obtain time domain feature weight information, and the time domain feature weight information represents the importance of each frame of speech information in the to-be-processed speech information;

[0011] The to-be-processed speech information is input into the frequency domain attention layer for feature weight learning processing in the frequency domain dimension to obtain frequency domain feature weight information, and the frequency domain feature weight information represents the importance of each speech feature corresponding to each frame of speech information in the to-be-processed speech information;

[0012] The time domain feature weight information and the frequency domain feature weight information are input into the first weight fusion layer for time-frequency weight fusion processing to obtain time-frequency weight information; and the time-frequency weight information represents the importance of time-frequency information in the to-be-processed speech information;

[0013] The time-frequency weight information and the to-be-processed speech information are input into the weighted feature extraction layer for feature extraction processing to obtain the target speech feature information.

[0014] In an optional embodiment, the time domain attention layer includes a first average pooling layer, a first maximum pooling layer, a first feature concatenation layer, and a time domain weight learning layer, and the time domain weight learning layer is constructed based on a dilated convolution network.

[0015] The feature weight learning processing in the time domain dimension by inputting the to-be-processed speech information into the time domain attention layer to obtain time domain feature weight information includes:

[0016] The to-be-processed speech information is input into the first average pooling layer to compress along the frequency domain dimension to obtain global time domain feature information;

[0017] The to-be-processed speech information is input into the first maximum pooling layer to compress along the frequency domain dimension to obtain local time domain feature information;

[0018] The global time domain feature information and the local time domain feature information are input into the first feature concatenation layer for feature concatenation processing to obtain time domain concatenation feature information;

[0019] input the time domain splicing feature information into the time domain weight learning layer to perform time domain weight learning processing, and obtain the time domain feature weight information.

[0020] In an optional embodiment, the frequency domain attention layer comprises: a second average pooling layer, a second maximum pooling layer, a frequency domain weight learning layer, and a second weight fusion layer.

[0021] The inputting of the to-be-processed voice information into the frequency domain attention layer to perform frequency domain dimension feature weight learning processing to obtain frequency domain feature weight information comprises:

[0022] The to-be-processed voice information is input into the second average pooling layer to compress along the time domain dimension to obtain global frequency domain feature information.

[0023] The to-be-processed voice information is input into the second maximum pooling layer to compress along the time domain dimension to obtain local frequency domain feature information.

[0024] The global frequency domain feature information is input into the frequency domain weight learning layer to perform frequency domain weight learning processing to obtain first frequency domain weight information.

[0025] The local frequency domain feature information is input into the frequency domain weight learning layer to perform frequency domain weight learning processing to obtain second frequency domain weight information.

[0026] The first frequency domain weight information and the second frequency domain weight information are input into the second weight fusion layer to perform weight fusion processing to obtain the frequency domain feature weight information.

[0027] In an optional embodiment, the at least one time-frequency attention layer is a plurality of time-frequency attention layers connected in sequence; and the target voice feature information comprises voice feature information output by the plurality of time-frequency attention layers.

[0028] The inputting of the to-be-processed voice information into the voice feature extraction network comprising at least one time-frequency attention layer to perform feature extraction processing to obtain target voice feature information corresponding to the to-be-processed voice information comprises:

[0029] In a case where the current time-frequency attention layer is a first time-frequency attention layer in the plurality of time-frequency attention layers, the to-be-processed voice information is input into the current time-frequency attention layer to perform feature extraction processing on the to-be-processed voice information on the basis of learning the importance of time-frequency information in the to-be-processed voice information, and obtain voice feature information output by the first time-frequency attention layer; the current time-frequency attention layer is the time-frequency attention layer currently performing feature extraction processing, which is determined based on a connection order corresponding to the plurality of time-frequency attention layers.

[0030] In a case that the current time-frequency attention layer is any other time-frequency attention layer, input the previous speech feature information into any other time-frequency attention layer to perform feature extraction processing on the previous speech feature information on the basis of learning the importance of time-frequency information in the previous speech feature information, and obtain speech feature information output by any other time-frequency attention layer.

[0031] Any other time-frequency attention layer is any time-frequency attention layer other than the first time-frequency attention layer in the plurality of time-frequency attention layers; and the previous speech feature information is speech feature information output by a time-frequency attention layer preceding any other time-frequency attention layer.

[0032] In an optional embodiment, the speech feature extraction network further comprises at least one speech feature extraction layer arranged in sequence, and at least one time-frequency attention layer is arranged in sequence; at least one speech feature extraction layer and at least one time-frequency attention layer are connected in sequence alternately starting from a first speech feature extraction layer in at least one speech feature extraction layer; and the target speech feature information comprises speech feature information output by at least one time-frequency attention layer.

[0033] The feature extraction processing performed on the to-be-processed speech information by inputting the to-be-processed speech information into the speech feature extraction network comprising at least one time-frequency attention layer comprises:

[0034] In a case that the current speech feature extraction layer is the first speech feature extraction layer, input the to-be-processed speech information into the first speech feature extraction layer to perform feature extraction processing, and obtain first speech feature information, which is speech feature information output by the first speech feature extraction layer; the current speech feature extraction layer is determined based on a connection sequence corresponding to at least one speech feature extraction layer and at least one time-frequency attention layer, and is the speech feature extraction layer currently performing feature extraction processing.

[0035] Input the first speech feature information into a first time-frequency attention layer in at least one time-frequency attention layer to perform feature extraction processing on the first speech feature information on the basis of learning the importance of time-frequency information in the first speech feature information, and obtain speech feature information output by the first time-frequency attention layer.

[0036] In a case where the current speech feature extraction layer is any other speech feature extraction layer, speech feature information output by a previous time-frequency attention layer of any other speech feature extraction layer is input into any other speech feature extraction layer for feature extraction processing to obtain second speech feature information, which is speech feature information output by any other speech feature extraction layer;

[0037] The second speech feature information is input into a next time-frequency attention layer of any other speech feature extraction layer, so as to perform feature extraction processing on the second speech feature information on the basis of learning importance of time-frequency information in the second speech feature information, to obtain speech feature information output by the next time-frequency attention layer; any other speech feature extraction layer is any speech feature extraction layer in the at least one speech feature extraction layer except the first speech feature extraction layer.

[0038] In an optional embodiment, the performing of the preset speech task based on the target speech feature information to obtain a task processing result includes:

[0039] The target speech feature information is input into a task processing network corresponding to the preset speech task for task processing to obtain the task processing result.

[0040] In an optional embodiment, the speech feature extraction network is obtained by training in the following manner:

[0041] Sample speech information corresponding to the preset speech task and a task processing result label corresponding to the sample speech information are obtained;

[0042] The sample speech information is input into a to-be-trained speech feature extraction network including at least one to-be-trained time-frequency attention layer for feature extraction processing to obtain sample speech feature information corresponding to the sample speech information, and any to-be-trained time-frequency attention layer is used to extract speech feature information from the sample speech information on the basis of learning importance of time-frequency information in the sample speech information;

[0043] The sample speech feature information is input into a task processing network corresponding to the preset speech task for task processing to obtain a predicted task processing result;

[0044] The to-be-trained speech feature extraction network is trained based on the task processing result label and the predicted task processing result to obtain the speech feature extraction network.

[0045] According to a second aspect of the embodiments of the present disclosure, a speech task processing apparatus is provided, including:

[0046] The voice information to be processed acquisition module is configured to perform acquisition of voice information to be processed corresponding to a preset voice task;

[0047] The first feature extraction processing module is configured to perform feature extraction processing on the voice information to be processed input into a voice feature extraction network comprising at least one time-frequency attention layer, to obtain target voice feature information corresponding to the voice information to be processed, any time-frequency attention layer being used to extract voice feature information from the voice information to be processed on the basis of learning the importance of time-frequency information in the voice information to be processed;

[0048] The voice task processing module is configured to perform the preset voice task based on the target voice feature information, to obtain a task processing result.

[0049] In one specific embodiment, the at least one time-frequency attention layer is one time-frequency attention layer, the time-frequency attention layer comprising a time domain attention layer, a frequency domain attention layer, a first weight fusion layer and a weighted feature extraction layer, and the first feature extraction processing module comprising:

[0050] The first feature weight learning unit is configured to perform time domain dimension feature weight learning on the voice information to be processed input into the time domain attention layer, to obtain time domain feature weight information representing the importance of each frame of voice information in the voice information to be processed;

[0051] The second feature weight learning unit is configured to perform frequency domain dimension feature weight learning on the voice information to be processed input into the frequency domain attention layer, to obtain frequency domain feature weight information representing the importance of each voice feature corresponding to each frame of voice information in the voice information to be processed;

[0052] The time-frequency weight fusion processing unit is configured to perform time-frequency weight fusion processing on the time domain feature weight information and the frequency domain feature weight information input into the first weight fusion layer, to obtain time-frequency weight information representing the importance of time-frequency information in the voice information to be processed;

[0053] The first feature extraction processing unit is configured to perform feature extraction processing on the time-frequency weight information and the voice information to be processed input into the weighted feature extraction layer, to obtain the target voice feature information.

[0054] In one optional embodiment, the time domain attention layer comprises a first average pooling layer, a first maximum pooling layer, a first feature concatenation layer and a time domain weight learning layer, the time domain weight learning layer being constructed based on a dilated convolution network.

[0055] The first feature weight learning unit comprises:

[0056] The first compression unit is configured to perform compression along the frequency domain dimension to obtain global time domain feature information by inputting the to-be-processed voice information into the first average pooling layer;

[0057] The second compression unit is configured to perform compression along the frequency domain dimension to obtain local time domain feature information by inputting the to-be-processed voice information into the first maximum pooling layer;

[0058] The feature splicing processing unit is configured to perform feature splicing processing by inputting the global time domain feature information and the local time domain feature information into the first feature splicing layer to obtain time domain splicing feature information;

[0059] The time domain weight learning unit is configured to perform time domain weight learning by inputting the time domain splicing feature information into the time domain weight learning layer to obtain the time domain feature weight information.

[0060] In an optional embodiment, the frequency domain attention layer comprises: a second average pooling layer, a second maximum pooling layer, a frequency domain weight learning layer, and a second weight fusion layer;

[0061] The second feature weight learning unit comprises:

[0062] The third compression unit is configured to perform compression along the time domain dimension to obtain global frequency domain feature information by inputting the to-be-processed voice information into the second average pooling layer;

[0063] The fourth compression unit is configured to perform compression along the time domain dimension to obtain local frequency domain feature information by inputting the to-be-processed voice information into the second maximum pooling layer;

[0064] The first frequency domain weight learning unit is configured to perform frequency domain weight learning by inputting the global frequency domain feature information into the frequency domain weight learning layer to obtain first frequency domain weight information;

[0065] The second frequency domain weight learning unit is configured to perform frequency domain weight learning by inputting the local frequency domain feature information into the frequency domain weight learning layer to obtain second frequency domain weight information;

[0066] The weight fusion processing unit is configured to perform weight fusion processing by inputting the first frequency domain weight information and the second frequency domain weight information into the second weight fusion layer to obtain the frequency domain feature weight information.

[0067] In an optional embodiment, the at least one time-frequency attention layer is a plurality of time-frequency attention layers connected in sequence; and the target speech feature information comprises speech feature information output by the plurality of time-frequency attention layers.

[0068] The first feature extraction processing module comprises:

[0069] The second feature extraction processing unit is configured to perform, in a case where the current time-frequency attention layer is a first time-frequency attention layer in the plurality of time-frequency attention layers, input of the to-be-processed speech information into the current time-frequency attention layer to perform feature extraction processing on the to-be-processed speech information on a basis of learning an importance degree of time-frequency information in the to-be-processed speech information, to obtain speech feature information output by the first time-frequency attention layer; the current time-frequency attention layer is the time-frequency attention layer currently performing feature extraction processing, which is determined based on a connection sequence corresponding to the plurality of time-frequency attention layers.

[0070] The third feature extraction processing unit is configured to perform, in a case where the current time-frequency attention layer is any other time-frequency attention layer, input of previous speech feature information into any of the other time-frequency attention layers to perform feature extraction processing on the previous speech feature information on a basis of learning an importance degree of time-frequency information in the previous speech feature information, to obtain speech feature information output by any of the other time-frequency attention layers.

[0071] Any of the other time-frequency attention layers is any of the time-frequency attention layers in the plurality of time-frequency attention layers except the first time-frequency attention layer; and the previous speech feature information is speech feature information output by a previous time-frequency attention layer of any of the other time-frequency attention layers.

[0072] In an optional embodiment, the speech feature extraction network further comprises at least one speech feature extraction layer arranged in sequence, and the at least one time-frequency attention layer is arranged in sequence; the at least one speech feature extraction layer and the at least one time-frequency attention layer are connected in sequence alternately starting from a first speech feature extraction layer in the at least one speech feature extraction layer; and the target speech feature information comprises speech feature information output by the at least one time-frequency attention layer.

[0073] The first feature extraction processing module comprises:

[0074] The fourth feature extraction processing unit is configured to perform the following: inputting the previous speech feature information into the first speech feature extraction layer for feature extraction processing, to obtain first speech feature information, when the current speech feature extraction layer is the first speech feature extraction layer; and inputting the speech information to be processed into the current speech feature extraction layer for feature extraction processing, to obtain the first speech feature information, when the current speech feature extraction layer is any other speech feature extraction layer; the first speech feature information is speech feature information output by the first speech feature extraction layer; the current speech feature extraction layer is determined based on a connection order corresponding to at least one speech feature extraction layer and at least one time-frequency attention layer, and the speech feature extraction layer currently performing feature extraction processing;

[0075] The fifth feature extraction processing unit is configured to perform the following: inputting the first speech feature information into a first time-frequency attention layer in the at least one time-frequency attention layer, to perform feature extraction processing on the first speech feature information based on learning an importance of time-frequency information in the first speech feature information, to obtain speech feature information output by the first time-frequency attention layer;

[0076] The sixth feature extraction processing unit is configured to perform the following: inputting speech feature information output by a previous time-frequency attention layer of any other speech feature extraction layer into any other speech feature extraction layer for feature extraction processing, to obtain second speech feature information, when the current speech feature extraction layer is any other speech feature extraction layer; the second speech feature information is speech feature information output by any other speech feature extraction layer;

[0077] The seventh feature extraction processing unit is configured to perform the following: inputting the second speech feature information into a next time-frequency attention layer of any other speech feature extraction layer, to perform feature extraction processing on the second speech feature information based on learning an importance of time-frequency information in the second speech feature information, to obtain speech feature information output by the next time-frequency attention layer; any other speech feature extraction layer is any speech feature extraction layer in the at least one speech feature extraction layer except the first speech feature extraction layer.

[0078] In an optional embodiment, the speech task processing module is specifically configured to perform the following: inputting the target speech feature information into a task processing network corresponding to the preset speech task for task processing, to obtain the task processing result.

[0079] In an optional embodiment, the speech feature extraction network is obtained by training the following modules:

[0080] The sample obtaining module is configured to perform obtaining sample voice information corresponding to the preset voice task and a task processing result label corresponding to the sample voice information.

[0081] The second feature extraction processing module is configured to perform feature extraction processing on the sample voice information by inputting the sample voice information into a to-be-trained voice feature extraction network including at least one to-be-trained time-frequency attention layer, to obtain sample voice feature information corresponding to the sample voice information, and any to-be-trained time-frequency attention layer is used to extract voice feature information from the sample voice information on the basis of learning importance of time-frequency information in the sample voice information.

[0082] The task processing module is configured to perform task processing on the sample voice feature information by inputting the sample voice feature information into a task processing network corresponding to the preset voice task, to obtain a predicted task processing result.

[0083] The network training module is configured to perform training on the to-be-trained voice feature extraction network based on the task processing result label and the predicted task processing result, to obtain the voice feature extraction network.

[0084] According to a third aspect of the embodiments of the present disclosure, an electronic device is provided, including a processor, a memory for storing instructions executable by the processor, and wherein the processor is configured to execute the instructions to implement the method according to any one of the first aspect.

[0085] According to a fourth aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, when instructions in the storage medium are executed by a processor of an electronic device, the electronic device can perform the method according to any one of the voice task processing methods of the embodiments of the present disclosure.

[0086] According to a fifth aspect of the embodiments of the present disclosure, a computer program product containing instructions is provided, when the computer program product is executed on a computer, the computer can perform the method according to any one of the first aspect.

[0087] The technical solutions provided by the embodiments of the present disclosure at least bring the following beneficial effects:

[0088] In the voice task processing process, the preset voice task corresponding to the to-be-processed voice information is input into the voice feature extraction network including at least one time-frequency attention layer for feature extraction processing. Based on the learning of the importance of the time-frequency information in the to-be-processed voice information by at least one time-frequency attention layer, the target voice feature information can be extracted from the to-be-processed voice information, the important information in the time domain and the frequency domain in the to-be-processed voice information can be simultaneously focused on, and then the important feature information (important frequency domain feature information) can be effectively extracted from each frame of voice information of the to-be-processed voice information, and the important voice information (important time domain feature information) in each frame of the to-be-processed voice information can also be effectively extracted, which greatly improves the representation accuracy of the voice feature information, and the preset voice task is executed based on the target voice feature information, which can effectively guarantee the accuracy of the voice task processing result and greatly improve the processing performance of the voice task.

[0089] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0090] The accompanying drawings, which are incorporated into the specification and constitute part of it, illustrate embodiments consistent with the present disclosure and, together with the specification, serve to explain the principles of the present disclosure and do not limit the present disclosure.

[0091] Figure 1 is a schematic diagram of an application environment according to an exemplary embodiment;

[0092] Figure 2 is a flowchart of a voice task processing method according to an exemplary embodiment;

[0093] Figure 3 is a flowchart of a voice task processing method according to an exemplary embodiment;

[0094] Figure 4 is a flowchart of a voice task processing method according to an exemplary embodiment;

[0095] Figure 5 is a flowchart of a voice task processing method according to an exemplary embodiment;

[0096] Figure 6is a schematic diagram of a speech feature extraction process incorporating a time-frequency attention layer according to an example embodiment;

[0097] Figure 7 is a flowchart of a speech feature extraction network training process according to an example embodiment;

[0098] Figure 8 is a result schematic diagram of a speech feature extraction network according to an example embodiment;

[0099] Figure 9 is a block diagram of a speech task processing apparatus according to an example embodiment;

[0100] Figure 10 is a block diagram of an electronic device for speech task processing according to an example embodiment;

[0101] Figure 11 is a block diagram of an electronic device for speech task processing according to an example embodiment. DETAILED DESCRIPTION

[0102] In order for those skilled in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be described clearly and completely below with reference to the drawings.

[0103] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily have to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The implementation described in the following example embodiments does not represent all implementations consistent with the present disclosure. Rather, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0104] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for display, analyzed data, etc.) involved in the present disclosure are all information and data authorized by the user or authorized by all parties.

[0105] Please refer to Figure 1 , Figure 1 is a schematic diagram of an application environment according to an example embodiment, which can include a terminal 100 and a server 200.

[0106] In an optional embodiment, the terminal 100 can be used to provide a voice task processing service for any user. Specifically, the terminal 100 can include, but is not limited to, an electronic device such as a smart phone, a desktop computer, a tablet computer, a notebook computer, a smart speaker, a digital assistant, an augmented reality (AR) / virtual reality (VR) device, a smart wearable device, and the like, or a software such as an application program running on the electronic device.

[0107] In an optional embodiment, the server 200 can provide a background service for the terminal 100. The server can pre-train a voice feature extraction network, and then can execute a downstream voice task in combination with voice feature information extracted by the voice feature extraction network. Specifically, the server 200 can be a standalone physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0108] In addition, it should be noted that Figure 1 The application environment shown is only one application environment provided by the present disclosure, and other application environments can also be included in actual applications.

[0109] In the embodiments of the present disclosure, the terminal 100 and the server 200 described above can be connected directly or indirectly through wired or wireless communication, and the present disclosure is not limited in this regard.

[0110] Figure 2 is a flowchart of a voice task processing method according to an exemplary embodiment, which can be applied to an electronic device such as a terminal or a server, as shown in Figure 2 The voice task processing method can include the following steps:

[0111] In step S201, the voice information to be processed corresponding to a preset voice task is obtained.

[0112] In a specific embodiment, the preset speech task can be set according to actual application requirements, such as a dialect recognition task, a translation task, a speaker recognition task, etc. Specifically, the speech information to be processed in the preset speech task processing process can be converted into preset size speech information to obtain the to-be-processed speech information corresponding to the preset speech task. Specifically, the preset size can be the size of the input information corresponding to the speech feature extraction network; for example, N*C*T (preset size), wherein N is the batchsize (i.e., the number of speech information processed by the network at a time, in the case of a to-be-processed speech information being a piece of speech information, N = 1), T is the number of frames contained in a piece of speech information, and C is the dimension of each frame.

[0113] In step S203, the to-be-processed speech information is input into the speech feature extraction network including at least one time-frequency attention layer for feature extraction processing to obtain target speech feature information corresponding to the to-be-processed speech information.

[0114] In a specific embodiment, any time-frequency attention layer can be used to extract speech feature information from the to-be-processed speech information on the basis of learning the importance of the time-frequency information in the to-be-processed speech information; specifically, the time-frequency information in the to-be-processed speech information can include time domain information (information in the time domain dimension) and frequency domain information (information in the frequency domain dimension) in the to-be-processed speech information.

[0115] In an optional embodiment, the at least one time-frequency attention layer can be one time-frequency attention layer, which can optionally include a time domain attention layer, a frequency domain attention layer, a first weight fusion layer, and a weighted feature extraction layer. Correspondingly, as shown in Figure 3 The input of the to-be-processed speech information into the speech feature extraction network including at least one time-frequency attention layer for feature extraction processing to obtain target speech feature information corresponding to the to-be-processed speech information can include the following steps:

[0116] In step S301, the to-be-processed speech information is input into the time domain attention layer for time domain dimension feature weight learning processing to obtain time domain feature weight information;

[0117] In step S303, the to-be-processed speech information is input into the frequency domain attention layer for frequency domain dimension feature weight learning processing to obtain frequency domain feature weight information;

[0118] In step S305, the time domain feature weight information and the frequency domain feature weight information are input into the first weight fusion layer for time-frequency weight fusion processing to obtain time-frequency weight information;

[0119] In step S307, the time-frequency weight information and the to-be-processed speech information are input into the weighted feature extraction layer for feature extraction processing to obtain target speech feature information.

[0120] In one specific embodiment, each frame of speech information in the speech information to be processed can be information in the time domain dimension (T dimension) of the speech information to be processed (time domain information); the time domain feature weight information can represent the importance of each frame of speech information in the speech information to be processed. Specifically, the greater the time domain feature weight information corresponding to any frame of speech information, the higher the importance of the frame of speech information in the speech information to be processed.

[0121] In one optional embodiment, the time domain attention layer can include a first average pooling layer, a first maximum pooling layer, a first feature concatenation layer, and a time domain weight learning layer, which is constructed based on a dilated convolutional network; optionally, as shown in Figure 4 The input of the speech information to be processed into the time domain attention layer for time domain dimension feature weight learning processing to obtain the time domain feature weight information can include the following steps:

[0122] In step S3011, the speech information to be processed is input into the first average pooling layer, compressed along the frequency domain dimension, and global time domain feature information is obtained;

[0123] In step S3013, the speech information to be processed is input into the first maximum pooling layer, compressed along the frequency domain dimension, and local time domain feature information is obtained;

[0124] In step S3015, the global time domain feature information and the local time domain feature information are input into the first feature concatenation layer for feature concatenation processing, and time domain concatenation feature information is obtained;

[0125] In step S3017, the time domain concatenation feature information is input into the time domain weight learning layer for time domain weight learning processing, and time domain feature weight information is obtained.

[0126] In one specific embodiment, the semantic feature (C-dimensional speech feature) corresponding to each frame of speech information in the speech information to be processed can be information in the frequency domain dimension (frequency domain information) of the speech information to be processed. Specifically, compression along the frequency domain dimension (C dimension) can effectively exclude the interference of information in the frequency domain dimension (C dimension) on T-dimensional information (time domain information).

[0127] In one specific embodiment, taking the size of the speech information to be processed as N*C*T as an example, the speech information to be processed is input into the first average pooling layer and the first maximum pooling layer, respectively, compressed along the frequency domain dimension, and the global time domain feature information and the local time domain feature information The size of each of the global time domain feature information and the local time domain feature information is N*1*T; further, after the global time domain feature information and the local time domain feature information are processed by the first feature splicing layer for feature splicing, the size of the time domain splicing feature information obtained can be N*2*T. Optionally, in order to save the calculation amount, the time domain weight learning layer can be a one-dimensional hole convolution network (the size of the convolution kernel is 3) with 2 input channels and 1 output channel, and correspondingly, the network parameters of the time domain weight learning layer (the one-dimensional hole convolution network) can be 2*1*3, and correspondingly, the size of the time domain feature weight information output by the time domain weight learning layer can be N*1*T.

[0128] In the above embodiment, in the process of learning the importance of the time domain information in the to-be-processed speech information, the to-be-processed speech information is input into the first average pooling layer and the first maximum pooling layer respectively, and is compressed along the frequency domain dimension, so that the multi-scale time domain feature information (the global time domain feature information and the local time domain feature information) can be obtained while effectively eliminating the interference of the frequency domain information on the time domain information, and then, combined with the time domain weight learning layer constructed based on the hole convolution network, the time domain weight learning processing can be performed on the basis of expanding the feature receptive field, so that the accuracy of the learned time domain feature weight information is greatly improved.

[0129] In one specific embodiment, the frequency domain feature weight information can represent the importance of each speech feature corresponding to each frame of speech information in the to-be-processed speech information; specifically, the greater the frequency domain feature weight information corresponding to each speech feature in each frame of speech information, the higher the importance of the speech feature in the corresponding frame of speech information.

[0130] In one optional embodiment, the frequency domain attention layer can include a second average pooling layer, a second maximum pooling layer, a frequency domain weight learning layer, and a second weight fusion layer; optionally, as shown in Figure 5 The frequency domain feature weight information obtained by inputting the to-be-processed speech information into the frequency domain attention layer for frequency domain dimension feature weight learning processing can include the following steps:

[0131] In step S3031, the to-be-processed speech information is input into the second average pooling layer and compressed along the time domain dimension to obtain global frequency domain feature information;

[0132] In step S3033, the to-be-processed speech information is input into the second maximum pooling layer and compressed along the time domain dimension to obtain local frequency domain feature information;

[0133] In step S3035, the global frequency domain feature information is input into the frequency domain weight learning layer for frequency domain weight learning processing to obtain first frequency domain weight information;

[0134] In step S3037, the local frequency domain feature information is input into the frequency domain weight learning layer for frequency domain weight learning processing, to obtain second frequency domain weight information;

[0135] In step S3039, the first frequency domain weight information and the second frequency domain weight information are input into the second weight fusion layer for weight fusion processing, to obtain frequency domain feature weight information.

[0136] In one specific embodiment, compression along the time domain dimension (T dimension) can effectively eliminate the interference of information in the time domain dimension (T dimension) on the C dimension information (time domain information).

[0137] In one specific embodiment, taking the size of the to-be-processed speech information as N*C*T as an example, the to-be-processed speech information is input into the second average pooling layer and the second maximum pooling layer respectively, and compression along the time domain dimension is performed, to obtain global frequency domain feature information and local frequency domain feature information , both of which have a size of N*1*T. Optionally, in order to save the amount of calculation, the same network (frequency domain weight learning layer) can be used to process the global frequency domain feature information and the local frequency domain feature information of different scales. Optionally, in actual application, two networks can also be used to process the global frequency domain feature information and the local frequency domain feature information.

[0138] In one specific embodiment, the above-mentioned frequency domain weight learning layer can include a first full connection layer, a nonlinear layer and a second full connection layer connected in sequence. Optionally, the size of the network parameters corresponding to the first full connection layer can be C*(C / r), where r is the compression ratio; the nonlinear layer can be an activation function ReLU, and the size of the network parameters corresponding to the second full connection layer can be (C / r)*C. Correspondingly, optionally, the size corresponding to the first frequency domain weight information and the second frequency domain weight information can be N*C. Further, in the second weight fusion layer, the first frequency domain weight information and the second frequency domain weight information can be added to obtain the frequency domain feature weight information.

[0139] In the above-mentioned embodiment, in the process of learning the importance of the frequency domain information in the to-be-processed speech information, the to-be-processed speech information is input into the second average pooling layer and the second maximum pooling layer respectively, and compression along the time domain dimension is performed, which can effectively eliminate the interference of the time domain information on the frequency domain information while obtaining multi-scale frequency domain feature information (global frequency domain feature information and local frequency domain feature information), ensuring the comprehensiveness and effectiveness of the frequency domain features in the subsequent frequency domain weight learning process. In addition, combining a frequency domain weight learning layer to perform frequency domain weight learning on the global frequency domain feature information and the local frequency domain feature information can effectively reduce the amount of calculation and greatly improve the processing efficiency.

[0140] In one specific embodiment, the time-frequency weight information can represent the importance of the time-frequency information in the to-be-processed speech information; specifically, since the sizes of the time-domain feature weight information and the frequency-domain feature weight information are inconsistent, in the time-frequency second weight fusion layer, the time-domain feature weight information and the frequency-domain feature weight information can be added through a broadcast mechanism, and combined with the activation function sigmoid, each element in the added weight information is normalized to 0-1, to obtain the time-frequency weight information. Then, in the weighted feature extraction layer, the time-frequency weight information can be multiplied by the corresponding position elements in the to-be-processed speech information to obtain the target speech feature information.

[0141] In one specific embodiment, as shown in Figure 6 Figure 6 is a schematic diagram of a speech feature extraction process combined with a time-frequency attention layer according to an exemplary embodiment. Specifically, the time-frequency attention layer can include a time-domain attention layer, a frequency-domain attention layer, a first weight fusion layer, and a weighted feature extraction layer; optionally, the to-be-processed speech information H can be input into the first max-pooling layer and the first average-pooling layer in the time-domain attention layer, and compressed along the frequency domain dimension, to obtain local time-domain feature information and global time-domain feature information Then, and are input into the first feature concatenation layer in the time-domain attention layer for feature concatenation processing, to obtain time-domain concatenated feature information Then, is input into the time-domain weight learning layer in the time-domain attention layer for time-domain weight learning processing, to obtain time-domain feature weight information Ht. In addition, the to-be-processed speech information H can be input into the second average-pooling layer and the second max-pooling layer in the frequency-domain attention layer, and compressed along the time domain dimension, to obtain global frequency-domain feature information and local frequency-domain feature information Then, and are input into the frequency-domain weight learning layer in the frequency-domain attention layer for frequency-domain weight learning processing, to obtain first frequency-domain weight information Hfatt1 and second frequency-domain weight information Hfatt2; then, Hfatt1 and Hfatt are input into the second weight fusion layer in the frequency-domain attention layer for weight fusion processing, to obtain frequency-domain feature weight information Hf. Further, Ht and Hf can be input into the first weight fusion layer for time-frequency weight fusion processing, to obtain time-frequency weight information Hatt; then, Hatt and the to-be-processed speech information H are input into the weighted feature extraction layer for feature extraction processing, to obtain the target speech feature information Hout.

[0142] ​In the above embodiment, in the process of extracting the speech feature information, the time domain attention layer is combined to learn the time domain feature weight information representing the importance of each frame of speech information in the to-be-processed speech information, the frequency domain attention layer is combined to learn the frequency domain feature weight information representing the importance of each speech feature corresponding to each frame of speech information in the to-be-processed speech information, the time domain feature weight information and the frequency domain feature weight information are input into the first weight fusion layer to perform time-frequency weight fusion processing, and the time-frequency weight information representing the importance of the time-frequency information in the to-be-processed speech information is obtained. Then, the time-frequency weight information and the to-be-processed speech information are input into the weighted feature extraction layer to extract the target speech feature information from the to-be-processed speech information based on the importance of each frame of speech information and the importance of each speech feature in each frame of speech information. The representation accuracy of the speech feature information is greatly improved.

[0143] In an optional embodiment, the at least one time-frequency attention layer can be a plurality of time-frequency attention layers connected in sequence; and the target speech feature information can include speech feature information output by the plurality of time-frequency attention layers.

[0144] In the case where the current time-frequency attention layer is the first time-frequency attention layer of the plurality of time-frequency attention layers, the to-be-processed speech information is input into the current time-frequency attention layer to extract features of the to-be-processed speech information based on the importance of the time-frequency information, and obtain speech feature information output by the first time-frequency attention layer.

[0145] In the case where the current time-frequency attention layer is any other time-frequency attention layer, the previous speech feature information is input into any other time-frequency attention layer to extract features of the previous speech feature information based on the importance of the time-frequency information, and obtain speech feature information output by any other time-frequency attention layer.

[0146] In a specific embodiment, the current time-frequency attention layer can be a time-frequency attention layer determined based on the connection order of the plurality of time-frequency attention layers, any other time-frequency attention layer can be any time-frequency attention layer except the first time-frequency attention layer of the plurality of time-frequency attention layers, and the previous speech feature information can be speech feature information output by a previous time-frequency attention layer of any other time-frequency attention layer.

[0147] In a specific embodiment, the above-mentioned inputting the to-be-processed speech information into the current time-frequency attention layer to perform feature extraction processing on the to-be-processed speech information on the basis of learning the importance of the time-frequency information in the to-be-processed speech information, and obtaining the specific refinement of the speech feature information output by the first time-frequency attention layer can refer to the specific refinement of the target speech feature information corresponding to the to-be-processed speech information obtained by inputting the to-be-processed speech information into the speech feature extraction network including at least one time-frequency attention layer for feature extraction processing in the case that the at least one time-frequency attention layer is one time-frequency attention layer, and details are not repeated here.

[0148] In a specific embodiment, the above-mentioned inputting the previous speech feature information into any other time-frequency attention layer to perform feature extraction processing on the previous speech feature information on the basis of learning the importance of the time-frequency information in the previous speech feature information, and obtaining the specific refinement of the speech feature information output by any other time-frequency attention layer can refer to the specific refinement of the target speech feature information corresponding to the to-be-processed speech information obtained by inputting the to-be-processed speech information into the speech feature extraction network including at least one time-frequency attention layer for feature extraction processing in the case that the at least one time-frequency attention layer is one time-frequency attention layer, that is, replacing the initial input of the to-be-processed speech information with the previous speech feature information, and details are not repeated here.

[0149] In the above-mentioned embodiments, the speech feature information is extracted in combination with the multiple time-frequency attention layers connected in sequence, which can effectively extract important feature information from each frame of speech information of the to-be-processed speech information under different receptive fields, and can also effectively extract important speech information of each frame of the to-be-processed speech information under different receptive fields, greatly improving the representation accuracy and effectiveness of the speech feature information.

[0150] In an optional embodiment, the above-mentioned speech feature extraction network can further include at least one speech feature extraction layer arranged in sequence, and the at least one time-frequency attention layer is also arranged in sequence; the at least one speech feature extraction layer and the at least one time-frequency attention layer are alternately connected in sequence with the first speech feature extraction layer in the at least one speech feature extraction layer as a starting layer; the target speech feature information includes the speech feature information output by the at least one time-frequency attention layer.

[0151] Correspondingly, the above-mentioned inputting the to-be-processed speech information into the speech feature extraction network including the at least one time-frequency attention layer for feature extraction processing to obtain the target speech feature information corresponding to the to-be-processed speech information can include:

[0152] In a case where the current speech feature extraction layer is the first speech feature extraction layer, the speech information to be processed is input into the first speech feature extraction layer for feature extraction processing to obtain first speech feature information, which is speech feature information output by the first speech feature extraction layer; the current speech feature extraction layer is determined based on a connection order corresponding to at least one speech feature extraction layer and at least one time-frequency attention layer, and the speech feature extraction layer currently performing feature extraction processing;

[0153] The first speech feature information is input into a first time-frequency attention layer in the at least one time-frequency attention layer, so as to perform feature extraction processing on the first speech feature information on the basis of learning the importance of time-frequency information in the first speech feature information, and obtain speech feature information output by the first time-frequency attention layer;

[0154] In a case where the current speech feature extraction layer is any other speech feature extraction layer, the speech feature information output by the previous time-frequency attention layer of any other speech feature extraction layer is input into any other speech feature extraction layer for feature extraction processing to obtain second speech feature information; the second speech feature information is speech feature information output by any other speech feature extraction layer;

[0155] The second speech feature information is input into the next time-frequency attention layer of any other speech feature extraction layer, so as to perform feature extraction processing on the second speech feature information on the basis of learning the importance of time-frequency information in the second speech feature information, and obtain speech feature information output by the next time-frequency attention layer;

[0156] In a specific embodiment, any other speech feature extraction layer is any speech feature extraction layer in the at least one speech feature extraction layer except the first speech feature extraction layer. Specifically, any speech feature extraction layer in the at least one speech feature extraction layer can be a bottleneck layer. The next time-frequency attention layer of any other speech feature extraction layer refers to a time-frequency attention layer connected after the other speech feature extraction layer.

[0157] In addition, it needs to be noted that, as the number of network layers increases, the receptive field of features in the current network layer also gradually increases, and accordingly, the hole rate of the time-domain weight learning layer based on the dilated convolution in the plurality of time-frequency attention layers can also gradually increase, which can improve the scale of capturing speech features, and further improve the effectiveness and accuracy of the learned speech features.

[0158] In the above embodiments, a speech feature extraction layer is arranged before each time-frequency attention layer, so that deeper speech feature information can be extracted from the speech feature extraction layer, and the video attention layer can effectively extract important feature information from each frame of speech information under different receptive fields, and can effectively extract important speech information of each frame of speech information to be processed under different receptive fields, thereby greatly improving the representation accuracy and effectiveness of the speech feature information.

[0159] In an optional embodiment, the speech feature extraction network can further include an initial speech feature extraction layer connected before the first speech feature extraction layer. Optionally, a time-delay neural network capable of modeling long-time dependence of speech signals can be used as the initial speech feature extraction layer.

[0160] In an optional embodiment, the speech feature extraction network can further include a second feature splicing layer connected with each time-frequency attention layer of the at least one time-frequency attention layer. Specifically, the second feature splicing layer can be used for splicing processing of speech feature information (target speech feature information) output by the at least one time-frequency attention layer.

[0161] In an optional embodiment, the speech feature extraction network can further include an intermediate speech feature extraction layer connected after the second feature splicing layer. Optionally, the intermediate speech feature extraction layer can be a time-delay neural network.

[0162] In an optional embodiment, the speech feature extraction network can further include an attentive statistic pooling layer connected after the intermediate speech feature extraction layer. Specifically, in the case of multiple time-frequency attention layers, the target speech feature information includes speech feature information output by the multiple time-frequency attention layers, and accordingly, the attentive statistic pooling layer can be used for statistic pooling of the speech feature information in the time domain, thereby better ensuring the effectiveness of the extracted speech feature information.

[0163] In an optional embodiment, as shown in FIG. 8, the speech feature extraction network can be trained by the following steps: Figure 7

[0164] In step S701, sample speech information corresponding to a preset speech task and a task processing result label corresponding to the sample speech information are obtained.

[0165] ​In step S703, the sample speech information is input into a to-be-trained speech feature extraction network including at least one to-be-trained time-frequency attention layer for feature extraction processing, to obtain sample speech feature information corresponding to the sample speech information, and any to-be-trained time-frequency attention layer is configured to extract speech feature information from the sample speech information on the basis of learning the importance of time-frequency information in the sample speech information.

[0166] In step S705, the sample speech feature information is input into a task processing network corresponding to a preset speech task for task processing, to obtain a predicted task processing result.

[0167] In step S707, the to-be-trained speech feature extraction network is trained based on the task processing result label and the predicted task processing result, to obtain the speech feature extraction network.

[0168] In one specific embodiment, the sample speech information can be speech information used for training the speech feature extraction network; the task processing result label corresponding to the sample speech information can correspond to the preset speech task. Optionally, in the case where the preset speech task is a dialect recognition task, the task processing result label can be a label representing a dialect type corresponding to the sample speech information. Optionally, there are multiple preset dialect types, and the task processing result label can include labels corresponding to the multiple preset dialect types, each label being a probability that the dialect type of the sample speech information belongs to the corresponding preset dialect type. Optionally, if the dialect type of the sample speech information is dialect type A, the label corresponding to the dialect type A in the task processing result label is 1, and the labels corresponding to other dialect types are 0. In the case where the preset speech task is a speaker recognition task, the task processing result label can be a label representing a speaker recognition result corresponding to the sample speech information. Optionally, there are multiple preset speakers, and the task processing result label can include labels corresponding to the multiple preset speakers, each label being a probability that the speaker corresponding to the sample speech information belongs to the corresponding preset speaker. Optionally, if the speaker of the sample speech information is speaker B, the label corresponding to speaker B in the task processing result label is 1, and the labels corresponding to other speakers are 0.

[0169] In one specific embodiment, the specific refinement of the above-mentioned inputting the sample speech information into the to-be-trained speech feature extraction network including at least one to-be-trained time-frequency attention layer for feature extraction processing, to obtain sample speech feature information corresponding to the sample speech information, can refer to the specific refinement of the above-mentioned inputting the to-be-processed speech information into the speech feature extraction network including at least one time-frequency attention layer for feature extraction processing, to obtain target speech feature information corresponding to the to-be-processed speech information, which will not be repeated here.

[0170] In a specific embodiment, the task processing network corresponding to the preset voice task can be a network for converting the sample voice feature information into a task processing result corresponding to the preset voice task (predicted task processing result). Correspondingly, the predicted task processing result can be a task processing result obtained by predicting the task processing result based on the sample voice feature information extracted by the to-be-trained voice feature extraction network. Optionally, the task processing network can be Bn-linear (batch normalization layer + linear normalization layer). Specifically, the batch normalization layer can improve the convergence speed in the training process of the to-be-trained voice feature extraction network, and the linear normalization layer can be a fully connected layer.

[0171] In a specific embodiment, the above training of the to-be-trained voice feature extraction network based on the task processing result label and the predicted task processing result can include: calculating task loss information between the task processing result label and the predicted task processing result based on a preset loss function; adjusting network parameters in the to-be-trained voice feature extraction network according to the task loss information, repeatedly inputting sample voice information into the to-be-trained voice feature extraction network including at least one to-be-trained time-frequency attention layer for feature extraction processing based on the adjusted to-be-trained voice feature extraction network, obtaining sample voice feature information corresponding to the sample voice information, and adjusting the network parameters in the to-be-trained voice feature extraction network according to the task loss information until a preset convergence condition is met; and taking the to-be-trained voice feature extraction network corresponding to the preset convergence condition as the trained voice feature extraction network.

[0172] In a specific embodiment, the preset loss function can be set according to actual application requirements, such as an exponential loss function, a cross-entropy loss function, etc. The above task loss information can represent the accuracy of the feature extraction of the to-be-trained voice feature extraction network. The above preset convergence condition can be that the task loss information is less than or equal to a preset loss threshold, or the number of training iteration steps reaches a preset number, etc. Specifically, the preset loss threshold and the preset number can be set according to the network precision and training speed requirements in actual applications.

[0173] In the above embodiments, in the process of training the speech feature extraction network corresponding to the preset speech task, the sample speech information corresponding to the preset speech task is input into the to-be-trained speech feature extraction network including at least one to-be-trained time-frequency attention layer for feature extraction processing, the sample speech feature information can be extracted from the sample speech information on the basis of learning the importance of the time-frequency information in the sample speech information, the important information in the time domain and the frequency domain in the sample speech information is simultaneously focused on, and then the important feature information (important frequency domain feature information) can be effectively extracted from each frame of speech information of the sample speech information, and meanwhile the important speech information (important time domain feature information) in each frame of the sample speech information can also be effectively extracted, the representation accuracy of the speech feature information is greatly improved, the effectiveness of the prediction task processing result obtained by performing task processing based on the sample speech feature information is ensured, and then the speech feature extraction performance of the trained speech feature extraction network can be improved.

[0174] In addition, it should be noted that in actual application, the number of layers and the structure of each network layer in the speech feature extraction network can be set according to actual needs.

[0175] In step S205, the preset speech task is performed based on the target speech feature information, and a task processing result is obtained.

[0176] In one specific embodiment, different speech tasks can correspond to different task processing results. Optionally, in the case where the preset speech task is a dialect recognition task, the task processing result can be information representing a dialect type corresponding to the to-be-processed speech information; in the case where the preset speech task is a speaker recognition task, the task processing result can be information representing a speaker recognition result corresponding to the to-be-processed speech information.

[0177] In one optional embodiment, the above performing the preset speech task based on the target speech feature information to obtain a task processing result can include: inputting the target speech feature information into a task processing network corresponding to the preset speech task for task processing to obtain the task processing result.

[0178] In the above embodiments, the target speech feature information is extracted from the to-be-processed speech information on the basis of learning the importance of the time-frequency information in the to-be-processed speech information, and then input into the task processing network corresponding to the preset speech task for task processing, which can effectively ensure the accuracy of the speech task processing result and greatly improve the processing performance of the speech task.

[0179] In one specific embodiment, as shown in Figure 8 , the target speech feature information is extracted from the to-be-processed speech information on the basis of learning the importance of the time-frequency information in the to-be-processed speech information, and then input into the task processing network corresponding to the preset speech task for task processing, which can effectively ensure the accuracy of the speech task processing result and greatly improve the processing performance of the speech task. Figure 8is a result schematic diagram of a speech feature extraction network provided according to an exemplary embodiment. Wherein the speech feature extraction network can include an initial speech feature extraction layer, M speech feature extraction layers + time-frequency attention layers, a second feature splicing layer, an intermediate speech feature extraction layer and an attention summary layer. Correspondingly, in the process of speech task processing, the speech information to be processed can be input into the initial speech feature extraction layer for speech feature extraction processing, then the speech feature information output by the initial speech feature extraction layer is input into the first speech feature extraction layer for speech feature extraction processing, then the speech feature information output by the first speech feature extraction layer is input into the first time-frequency attention layer, so as to perform speech feature extraction processing on the basis of learning the importance degree of time-frequency information, and the same is true for the following M-1 speech feature extraction layers + time-frequency attention layers, then the speech feature information output by the M time-frequency attention layers can be input into the second feature splicing layer for speech feature splicing processing, then the speech feature information output by the second feature splicing layer is input into the intermediate speech feature extraction layer for speech feature extraction processing, then the speech feature information output by the intermediate speech feature extraction layer is input into the attention summary layer for speech feature summary, and correspondingly, the above performing the preset speech task based on the target speech feature information to obtain the task processing result can include performing the preset speech task based on the speech feature information output by the attention summary layer to obtain the task processing result.

[0180] From the above technical solutions provided by the embodiments of the present specification, in the process of speech task processing, the speech information to be processed corresponding to the preset speech task is input into the speech feature extraction network including at least one time-frequency attention layer for feature extraction processing, the target speech feature information can be extracted from the speech information to be processed based on at least one time-frequency attention layer learning the importance degree of time-frequency information in the speech information to be processed, the important information in the time domain and the frequency domain in the speech information to be processed is simultaneously focused on, and then the important feature information (important frequency domain feature information) can be effectively extracted from each frame of speech information of the speech information to be processed, and the important speech information (important time domain feature information) in each frame of the speech information to be processed can also be effectively extracted, which greatly improves the representation accuracy of the speech feature information, and performing the preset speech task based on the target speech feature information can effectively ensure the accuracy of the speech task processing result, and greatly improves the processing performance of the speech task.

[0181] Figure 9 is a block diagram of a speech task processing device according to an exemplary embodiment. Referring to Figure 9 , the device comprises:

[0182] The speech information to be processed acquisition module 910 is configured to perform acquisition of the speech information to be processed corresponding to the preset speech task.

[0183] The first feature extraction processing module 920 is configured to perform feature extraction processing on the to-be-processed voice information input into the voice feature extraction network including at least one time-frequency attention layer to obtain target voice feature information corresponding to the to-be-processed voice information, and any time-frequency attention layer is used to extract voice feature information from the to-be-processed voice information on the basis of learning the importance of time-frequency information in the to-be-processed voice information.

[0184] The voice task processing module 930 is configured to perform a preset voice task based on the target voice feature information to obtain a task processing result.

[0185] In one specific embodiment, the at least one time-frequency attention layer is one time-frequency attention layer, and the time-frequency attention layer includes a time domain attention layer, a frequency domain attention layer, a first weight fusion layer, and a weighted feature extraction layer. The first feature extraction processing module 920 includes:

[0186] The first feature weight learning unit is configured to perform feature weight learning in a time domain dimension by inputting the to-be-processed voice information into the time domain attention layer to obtain time domain feature weight information, which represents the importance of each frame of voice information in the to-be-processed voice information.

[0187] The second feature weight learning unit is configured to perform feature weight learning in a frequency domain dimension by inputting the to-be-processed voice information into the frequency domain attention layer to obtain frequency domain feature weight information, which represents the importance of each voice feature corresponding to each frame of voice information in the to-be-processed voice information.

[0188] The time-frequency weight fusion processing unit is configured to perform time-frequency weight fusion processing by inputting the time domain feature weight information and the frequency domain feature weight information into the first weight fusion layer to obtain time-frequency weight information, which represents the importance of time-frequency information in the to-be-processed voice information.

[0189] The first feature extraction processing unit is configured to perform feature extraction processing by inputting the time-frequency weight information and the to-be-processed voice information into the weighted feature extraction layer to obtain the target voice feature information.

[0190] In one optional embodiment, the time domain attention layer includes a first average pooling layer, a first maximum pooling layer, a first feature concatenation layer, and a time domain weight learning layer, and the time domain weight learning layer is constructed based on a dilated convolution network.

[0191] The first feature weight learning unit includes:

[0192] The first compression unit is configured to input the to-be-processed voice information into a first average pooling layer and compress along a frequency domain dimension to obtain global time domain feature information.

[0193] The second compression unit is configured to input the to-be-processed voice information into a first maximum pooling layer and compress along a frequency domain dimension to obtain local time domain feature information.

[0194] The feature splicing processing unit is configured to input the global time domain feature information and the local time domain feature information into a first feature splicing layer for feature splicing processing to obtain time domain splicing feature information.

[0195] The time domain weight learning unit is configured to input the time domain splicing feature information into a time domain weight learning layer for time domain weight learning to obtain time domain feature weight information.

[0196] In an optional embodiment, the frequency domain attention layer includes a second average pooling layer, a second maximum pooling layer, a frequency domain weight learning layer, and a second weight fusion layer.

[0197] The second feature weight learning unit includes:

[0198] The third compression unit is configured to input the to-be-processed voice information into a second average pooling layer and compress along a time domain dimension to obtain global frequency domain feature information.

[0199] The fourth compression unit is configured to input the to-be-processed voice information into a second maximum pooling layer and compress along a time domain dimension to obtain local frequency domain feature information.

[0200] The first frequency domain weight learning unit is configured to input the global frequency domain feature information into a frequency domain weight learning layer for frequency domain weight learning to obtain first frequency domain weight information.

[0201] The second frequency domain weight learning unit is configured to input the local frequency domain feature information into the frequency domain weight learning layer for frequency domain weight learning to obtain second frequency domain weight information.

[0202] The weight fusion processing unit is configured to input the first frequency domain weight information and the second frequency domain weight information into a second weight fusion layer for weight fusion processing to obtain frequency domain feature weight information.

[0203] In an optional embodiment, the at least one time-frequency attention layer is a plurality of time-frequency attention layers connected in sequence; and the target voice feature information includes voice feature information output by the plurality of time-frequency attention layers.

[0204] The first feature extraction processing module 920 includes:

[0205] The second feature extraction processing unit is configured to perform, in a case where the current time-frequency attention layer is a first time-frequency attention layer in the plurality of time-frequency attention layers, inputting the to-be-processed speech information into the current time-frequency attention layer to perform feature extraction processing on the to-be-processed speech information on a basis of learning an importance degree of time-frequency information in the to-be-processed speech information, to obtain speech feature information output by the first time-frequency attention layer; the current time-frequency attention layer is a time-frequency attention layer currently performing feature extraction processing, which is determined based on a connection order corresponding to the plurality of time-frequency attention layers;

[0206] The third feature extraction processing unit is configured to perform, in a case where the current time-frequency attention layer is any other time-frequency attention layer, inputting the previous speech feature information into any other time-frequency attention layer to perform feature extraction processing on the previous speech feature information on a basis of learning an importance degree of time-frequency information in the previous speech feature information, to obtain speech feature information output by any other time-frequency attention layer;

[0207] The any other time-frequency attention layer is any time-frequency attention layer in the plurality of time-frequency attention layers except the first time-frequency attention layer; and the previous speech feature information is speech feature information output by a previous time-frequency attention layer of the any other time-frequency attention layer.

[0208] In an optional embodiment, the speech feature extraction network further includes at least one speech feature extraction layer arranged in sequence, and the at least one time-frequency attention layer is arranged in sequence; the at least one speech feature extraction layer and the at least one time-frequency attention layer are connected in sequence in an alternating manner with a first speech feature extraction layer in the at least one speech feature extraction layer as a starting layer; and the target speech feature information includes speech feature information output by the at least one time-frequency attention layer.

[0209] The first feature extraction processing module 920 includes:

[0210] The fourth feature extraction processing unit is configured to perform, in a case where the current time-frequency attention layer is any other time-frequency attention layer, inputting the previous speech feature information into any other time-frequency attention layer to perform feature extraction processing on the previous speech feature information on a basis of learning an importance degree of time-frequency information in the previous speech feature information, to obtain speech feature information output by any other time-frequency attention layer; and in a case where the current speech feature extraction layer is the first speech feature extraction layer, inputting the to-be-processed speech information into the first speech feature extraction layer to perform feature extraction processing, to obtain first speech feature information, which is speech feature information output by the first speech feature extraction layer; the current speech feature extraction layer is a speech feature extraction layer currently performing feature extraction processing, which is determined based on a connection order corresponding to the at least one speech feature extraction layer and the at least one time-frequency attention layer.

[0211] The fifth feature extraction processing unit is configured to perform inputting the first speech feature information into a first time-frequency attention layer of the at least one time-frequency attention layer, to perform feature extraction processing on the first speech feature information on the basis of learning an importance degree of time-frequency information in the first speech feature information, and to obtain speech feature information output by the first time-frequency attention layer;

[0212] The sixth feature extraction processing unit is configured to perform, in a case where the current speech feature extraction layer is any other speech feature extraction layer, inputting speech feature information output by a previous time-frequency attention layer of the any other speech feature extraction layer into the any other speech feature extraction layer for feature extraction processing, to obtain second speech feature information, the second speech feature information being speech feature information output by the any other speech feature extraction layer;

[0213] The seventh feature extraction processing unit is configured to perform inputting the second speech feature information into a next time-frequency attention layer of the any other speech feature extraction layer, to perform feature extraction processing on the second speech feature information on the basis of learning an importance degree of time-frequency information in the second speech feature information, and to obtain speech feature information output by the next time-frequency attention layer; the any other speech feature extraction layer being any speech feature extraction layer other than the first speech feature extraction layer in the at least one speech feature extraction layer.

[0214] In an optional embodiment, the speech task processing module 930 is specifically configured to perform inputting the target speech feature information into a task processing network corresponding to the preset speech task for task processing, to obtain a task processing result.

[0215] In an optional embodiment, the speech feature extraction network is obtained by training the following modules:

[0216] The sample acquisition module is configured to perform acquiring sample speech information corresponding to the preset speech task and a task processing result label corresponding to the sample speech information;

[0217] The second feature extraction processing module is configured to perform inputting the sample speech information into a to-be-trained speech feature extraction network including the at least one to-be-trained time-frequency attention layer for feature extraction processing, to obtain sample speech feature information corresponding to the sample speech information, and any to-be-trained time-frequency attention layer is used to extract speech feature information from the sample speech information on the basis of learning an importance degree of time-frequency information in the sample speech information;

[0218] The task processing module is configured to perform inputting the sample speech feature information into a task processing network corresponding to the preset speech task for task processing, to obtain a predicted task processing result;

[0219] The network training module is configured to train the speech feature extraction network to be trained based on the task processing result label and the predicted task processing result.

[0220] As to the apparatus in the above-mentioned embodiments, the specific manner in which each layer performs the operation has been described in detail in the embodiments of the method, and thus will not be described in detail here.

[0221] Figure 10 is a block diagram of an electronic device for speech task processing according to an example embodiment. The electronic device can be a terminal, and its internal structure can be as shown in Figure 10 The electronic device includes a processor, a memory, a network interface, a display screen, and an input device connected through a system bus. The processor of the electronic device is configured to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The network interface of the electronic device is configured to communicate with external terminals through a network connection. The computer program is executed by the processor to implement a speech task processing method. The display screen of the electronic device can be a liquid crystal display screen or an electronic ink display screen. The input device of the electronic device can be a touch layer overlaid on the display screen, or a key, trackball, or touchpad provided on the housing of the electronic device, or an external keyboard, touchpad, or mouse, etc.

[0222] Figure 11 is a block diagram of another electronic device for speech task processing according to an example embodiment. The electronic device can be a server, and its internal structure can be as shown in Figure 11 The electronic device includes a processor, a memory, and a network interface connected through a system bus. The processor of the electronic device is configured to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The network interface of the electronic device is configured to communicate with external terminals through a network connection. The computer program is executed by the processor to implement a speech task processing method.

[0223] Those skilled in the art can understand that Figure 10 or Figure 11The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present disclosure, and does not constitute a limitation on the electronic device to which the scheme of the present disclosure is applied. The specific electronic device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0224] In an example embodiment, an electronic device is also provided, comprising: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the voice task processing method in the embodiments of the present disclosure.

[0225] In an example embodiment, a computer readable storage medium is also provided, when the instructions in the storage medium are executed by the processor of an electronic device, the electronic device is enabled to perform the voice task processing method in the embodiments of the present disclosure.

[0226] In an example embodiment, a computer program product containing instructions is also provided, when it is run on a computer, the computer is enabled to perform the voice task processing method in the embodiments of the present disclosure.

[0227] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium and, when executed, can include the processes of the above-mentioned embodiments of the methods. Any reference to memory, storage, databases, or other media in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0228] Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the features disclosed herein. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the disclosure being indicated by the following claims.

[0229] It should be understood that the present disclosure is not limited to the precise structures as herein described and illustrated in the drawings, and that various modifications and changes can be made without departing from its scope. The scope of the present disclosure is limited only by the claims that follow.

Claims

1. A method of voice task processing, the method comprising: The method comprises: obtaining to-be-processed voice information corresponding to a preset voice task; inputting the to-be-processed voice information into a voice feature extraction network comprising at least one time-frequency attention layer for feature extraction processing to obtain target voice feature information corresponding to the to-be-processed voice information, wherein any time-frequency attention layer is configured to extract voice feature information from the to-be-processed voice information on the basis of learning the importance of time-frequency information in the to-be-processed voice information; the voice feature extraction network comprises at least one voice feature extraction layer arranged in sequence; at least one voice feature extraction layer and at least one time-frequency attention layer are alternately connected in sequence with the first voice feature extraction layer in the at least one voice feature extraction layer as a starting layer; the voice feature extraction network further comprises an initial voice feature extraction layer connected before the first voice feature extraction layer, a second feature splicing layer connected with each time-frequency attention layer, an intermediate voice feature extraction layer connected after the second feature splicing layer, and an attention summary layer connected after the intermediate voice feature extraction layer; the intermediate voice feature extraction layer is a time-delay neural network; the attention summary layer is configured to summarize voice feature information in the time domain; the inputting of the to-be-processed voice information into the voice feature extraction network comprising at least one time-frequency attention layer for feature extraction processing to obtain target voice feature information corresponding to the to-be-processed voice information comprises: in the case that the current voice feature extraction layer is the first voice feature extraction layer, inputting the to-be-processed voice information into the first voice feature extraction layer to obtain first voice feature information; inputting the first voice feature information into the first time-frequency attention layer in the at least one time-frequency attention layer to extract features of the first voice feature information on the basis of learning the importance of time-frequency information in the first voice feature information, and obtaining voice feature information output by the first time-frequency attention layer; in the case that the current voice feature extraction layer is any other voice feature extraction layer, inputting voice feature information output by the previous time-frequency attention layer of any other voice feature extraction layer into any other voice feature extraction layer to obtain second voice feature information; inputting the second voice feature information into the next time-frequency attention layer of any other voice feature extraction layer to extract features of the second voice feature information on the basis of learning the importance of time-frequency information in the second voice feature information, and obtaining voice feature information output by the next time-frequency attention layer; performing the preset voice task based on the target voice feature information to obtain a task processing result.

2. The voice task processing method of claim 1, wherein, The at least one time-frequency attention layer is one time-frequency attention layer, and the time-frequency attention layer comprises a time domain attention layer, a frequency domain attention layer, a first weight fusion layer and a weighted feature extraction layer; the feature extraction processing of the speech feature extraction network comprising at least one time-frequency attention layer on the to-be-processed speech information comprises: The to-be-processed speech information is input into the time domain attention layer for feature weight learning processing in the time domain dimension, so as to obtain time domain feature weight information; the time domain feature weight information represents the importance of each frame of speech information in the to-be-processed speech information; The to-be-processed speech information is input into the frequency domain attention layer for feature weight learning processing in the frequency domain dimension, so as to obtain frequency domain feature weight information; the frequency domain feature weight information represents the importance of each speech feature corresponding to each frame of speech information in the to-be-processed speech information; The time domain feature weight information and the frequency domain feature weight information are input into the first weight fusion layer for time-frequency weight fusion processing, so as to obtain time-frequency weight information; the time-frequency weight information represents the importance of time-frequency information in the to-be-processed speech information; The time-frequency weight information and the to-be-processed speech information are input into the weighted feature extraction layer for feature extraction processing, so as to obtain the target speech feature information.

3. The voice task processing method of claim 2, wherein, The time domain attention layer comprises a first average pooling layer, a first maximum pooling layer, a first feature splicing layer and a time domain weight learning layer, and the time domain weight learning layer is constructed based on a hollow convolutional network; The to-be-processed speech information is input into the time domain attention layer for feature weight learning processing in the time domain dimension, so as to obtain time domain feature weight information, which comprises: The to-be-processed speech information is input into the first average pooling layer and compressed along the frequency domain dimension, so as to obtain global time domain feature information; The to-be-processed speech information is input into the first maximum pooling layer and compressed along the frequency domain dimension, so as to obtain local time domain feature information; The global time domain feature information and the local time domain feature information are input into the first feature splicing layer for feature splicing processing, so as to obtain time domain splicing feature information; The time domain splicing feature information is input into the time domain weight learning layer for time domain weight learning processing, so as to obtain the time domain feature weight information.

4. The voice task processing method of claim 2, wherein, The frequency domain attention layer comprises a second average pooling layer, a second maximum pooling layer, a frequency domain weight learning layer and a second weight fusion layer; The to-be-processed speech information is input into the frequency domain attention layer for feature weight learning processing in the frequency domain dimension, so as to obtain frequency domain feature weight information, which comprises: The to-be-processed speech information is input into the second average pooling layer and compressed along the time domain dimension, so as to obtain global frequency domain feature information; The to-be-processed speech information is input into the second maximum pooling layer and compressed along the time domain dimension, so as to obtain local frequency domain feature information; The global frequency domain feature information is input into the frequency domain weight learning layer for frequency domain weight learning processing, so as to obtain first frequency domain weight information; inputting the first frequency domain weight information and the second frequency domain weight information into the second weight fusion layer for weight fusion processing to obtain the frequency domain feature weight information. The at least one time-frequency attention layer is a plurality of time-frequency attention layers connected in sequence; the target speech feature information comprises speech feature information output by the plurality of time-frequency attention layers; 5. The voice task processing method of claim 1, wherein, The method comprises the following steps: In a case where the current time-frequency attention layer is a first time-frequency attention layer of the plurality of time-frequency attention layers, inputting the to-be-processed speech information into the current time-frequency attention layer to perform feature extraction processing on the to-be-processed speech information on the basis of learning the importance of time-frequency information in the to-be-processed speech information, to obtain speech feature information output by the first time-frequency attention layer; the current time-frequency attention layer is the time-frequency attention layer currently performing feature extraction processing, which is determined based on the connection order of the plurality of time-frequency attention layers; In a case where the current time-frequency attention layer is any other time-frequency attention layer, inputting previous speech feature information into any other time-frequency attention layer to perform feature extraction processing on the previous speech feature information on the basis of learning the importance of time-frequency information in the previous speech feature information, to obtain speech feature information output by any other time-frequency attention layer; Any other time-frequency attention layer is any time-frequency attention layer of the plurality of time-frequency attention layers except the first time-frequency attention layer; the previous speech feature information is speech feature information output by a time-frequency attention layer preceding any other time-frequency attention layer. The at least one time-frequency attention layer is arranged in sequence; the target speech feature information comprises speech feature information output by the at least one time-frequency attention layer; 6. The voice task processing method according to any one of claims 1 to 5, characterized in that, The first speech feature information is speech feature information output by the first speech feature extraction layer; the current speech feature extraction layer is the speech feature extraction layer currently performing feature extraction processing, which is determined based on the connection order of the at least one speech feature extraction layer and the at least one time-frequency attention layer; The second speech feature information is speech feature information output by any other speech feature extraction layer; Any other speech feature extraction layer is any speech feature extraction layer of the at least one speech feature extraction layer except the first speech feature extraction layer. The method comprises the following steps:

7. The voice task processing method according to any one of claims 1 to 5, characterized by, inputting the target speech feature information into a task processing network corresponding to the preset speech task to perform task processing, to obtain the task processing result. The speech feature extraction network is trained in the following manner:

8. The voice task processing method according to any one of claims 1 to 5, characterized by, ​ obtain sample voice information corresponding to the preset voice task and a task processing result label corresponding to the sample voice information; input the sample voice information into a to-be-trained voice feature extraction network including at least one to-be-trained time-frequency attention layer for feature extraction processing to obtain sample voice feature information corresponding to the sample voice information, and any to-be-trained time-frequency attention layer is used to extract voice feature information from the sample voice information on the basis of learning the importance of time-frequency information in the sample voice information; input the sample voice feature information into the task processing network corresponding to the preset voice task for task processing to obtain a predicted task processing result; train the to-be-trained voice feature extraction network based on the task processing result label and the predicted task processing result to obtain the voice feature extraction network.

9. A voice task processing apparatus characterized by comprising: comprise: a to-be-processed voice information acquisition module configured to perform obtaining to-be-processed voice information corresponding to a preset voice task; a first feature extraction processing module configured to perform inputting the to-be-processed voice information into a voice feature extraction network including at least one time-frequency attention layer for feature extraction processing to obtain target voice feature information corresponding to the to-be-processed voice information, any time-frequency attention layer is used to extract voice feature information from the to-be-processed voice information on the basis of learning the importance of time-frequency information in the to-be-processed voice information; the voice feature extraction network includes at least one voice feature extraction layer arranged in sequence; at least one voice feature extraction layer and at least one time-frequency attention layer are alternately connected in sequence with the first voice feature extraction layer in at least one voice feature extraction layer as a starting layer; the voice feature extraction network further includes an initial voice feature extraction layer connected before the first voice feature extraction layer, a second feature splicing layer connected with each time-frequency attention layer, an intermediate voice feature extraction layer connected after the second feature splicing layer, and an attention summary layer connected after the intermediate voice feature extraction layer; the intermediate voice feature extraction layer is a time-delay neural network; The attention summary layer is configured to summarize the speech feature information in the time domain; the first feature extraction processing module comprises: a fourth feature extraction processing unit configured to execute, in a case where the current speech feature extraction layer is the first speech feature extraction layer, inputting the to-be-processed speech information into the first speech feature extraction layer to obtain first speech feature information; a fifth feature extraction processing unit configured to execute inputting the first speech feature information into a first time-frequency attention layer in the at least one time-frequency attention layer, to perform feature extraction processing on the first speech feature information on the basis of learning the importance of time-frequency information in the first speech feature information, and obtain speech feature information output by the first time-frequency attention layer; a sixth feature extraction processing unit configured to execute, in a case where the current speech feature extraction layer is any other speech feature extraction layer, inputting speech feature information output by a previous time-frequency attention layer of any other speech feature extraction layer into any other speech feature extraction layer to obtain second speech feature information; and a seventh feature extraction processing unit configured to execute inputting the second speech feature information into a next time-frequency attention layer of any other speech feature extraction layer, to perform feature extraction processing on the second speech feature information on the basis of learning the importance of time-frequency information in the second speech feature information, and obtain speech feature information output by the next time-frequency attention layer. The speech task processing module is configured to execute the preset speech task based on the target speech feature information to obtain a task processing result.

10. An electronic device, comprising: It comprises: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the speech task processing method according to any one of claims 1 to 8.

11. A computer readable storage medium, characterized in that, When the instructions in the storage medium are executed by the processor of the electronic device, the electronic device can execute the speech task processing method according to any one of claims 1 to 8.

Images