Method, device, equipment and readable storage medium for acquiring target object voice
By optimizing the TasNet network and using convolutional neural networks and permutation problem algorithms, we solved the problems of high computational complexity and poor sensitivity of mixed speech signals in the speech separation process, and achieved efficient target object speech separation in noisy environments.
Patent Information
- Application Number
- CN202210601636.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-30
- Publication Date
- 2025-09-05
- Estimated Expiration
- 2042-05-30
AI Technical Summary
The existing TasNet network has problems in the speech separation process, such as the convolution kernel size affecting the output length, high computational complexity, and poor sensitivity to mixed speech signals. In particular, it cannot effectively separate the target object's speech in noisy environments.
By optimizing the time-domain audio separation network, adopting the convolutional neural network model framework and the permutation problem algorithm, optimizing the encoder and separator, using the void convolution and deep recurrent separable convolution algorithms, and combining the permutation problem algorithm for speech separation, the time-varying and computational complexity problems of mixed speech signals are solved.
It improves the robustness and time efficiency of speech separation, can effectively separate the target object's speech in noisy environments, is suitable for situations with an uncertain number of people, and improves the time complexity and separation accuracy.
Smart Images
Figure CN115206336B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of speech separation technology, and in particular to a method, apparatus, device and readable storage medium for acquiring target object speech. Background Art
[0002] Currently, there are several issues with using the original TasNet network for speech separation. First, kernel size affects the output length. If the convolution kernel is too small, the neural network framework cannot control the length of the processing required to separate mixed speech. Second, if a long short-term memory neural network is used, the framework has a large number of parameters, which increases computational complexity and overall training time complexity. Third, the sensitivity to mixed speech signals varies. Because the original framework lacks time-invariant properties, when processing a mixed speech signal, the starting point of the mixed speech signal varies, resulting in different separation results each time. Summary of the Invention
[0003] The object of the present invention is to provide a method, apparatus, device and readable storage medium for acquiring target object speech, so as to improve the above-mentioned problem.
[0004] In order to achieve the above objectives, the embodiments of the present application provide the following technical solutions:
[0005] In one aspect, an embodiment of the present application provides a method for acquiring a target object's voice, the method comprising:
[0006] Acquire mixed voice data, where the mixed voice data is a mixture of voices of multiple objects, including the target object;
[0007] Based on the convolutional neural network model framework, the time domain audio separation network is optimized to obtain the optimized time domain audio separation network;
[0008] Based on the permutation problem algorithm, the mixed speech data is processed using the optimized time-domain audio separation network to obtain speech data containing only the speech of the target object.
[0009] In a second aspect, an embodiment of the present application provides a device for acquiring target object speech, the device comprising an acquisition module, an optimization module, and a separation module.
[0010] An acquisition module, configured to acquire mixed voice data, wherein the mixed voice data is formed by mixing voices of multiple objects, including the target object;
[0011] An optimization module is used to optimize the time-domain audio separation network based on the convolutional neural network model framework to obtain an optimized time-domain audio separation network;
[0012] The separation module is used to process the mixed speech data using the optimized time-domain audio separation network based on the permutation problem algorithm to obtain speech data containing only the speech of the target object.
[0013] In a third aspect, embodiments of the present application provide a device for acquiring a target object's speech, the device comprising a memory and a processor. The memory is configured to store a computer program; the processor is configured to implement the steps of the target object's speech acquisition method when executing the computer program.
[0014] In a fourth aspect, an embodiment of the present application provides a readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the above-mentioned method for acquiring the target object voice are implemented.
[0015] The beneficial effects of the present invention are:
[0016] 1. The original time-domain audio separation network (TasNet) had some issues. By optimizing the original TasNet neural network framework, these issues have been resolved. As a result, the optimized TasNet neural network framework has a certain degree of robustness in speech separation and is suitable for general use.
[0017] 2. In this paper, a speech separation algorithm based on the permutation problem is proposed based on the permutation problem algorithm and the optimized time-domain audio separation network. This method solves the problem of separating the target human voice in a noisy environment and achieves a significant improvement in time efficiency compared to the original TasNet framework.
[0018] 3. Based on the environmental characteristics of ticket purchase in a noisy environment, this paper proposes a speech separation algorithm under uncertain number of people, which solves the shortcoming of the TasNet framework that speech separation can only be performed after the number of people is determined.
[0019] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or understood by practicing the embodiments of the present invention. The purposes and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.
[0021] Figure 1 1 is a flow chart of a method for acquiring a target object's voice according to an embodiment of the present invention;
[0022] Figure 2 Schematic diagram of the structure of the target object voice acquisition device according to an embodiment of the present invention;
[0023] Figure 3 It is a schematic diagram of the structure of the target object voice acquisition device described in an embodiment of the present invention. DETAILED DESCRIPTION
[0024] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0025] It should be noted that similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are used only to distinguish the description and should not be understood as indicating or implying relative importance.
[0026] Example 1
[0027] like Figure 1 As shown, this embodiment provides a method for acquiring target object speech, which includes step S1, step S2, step S3, step S4 and step S5.
[0028] Step S1: Acquire mixed voice data, where the mixed voice data is a mixture of voices of multiple objects, including the target object;
[0029] Step S2: Optimizing the time-domain audio separation network based on the convolutional neural network model framework to obtain an optimized time-domain audio separation network;
[0030] Step S3: Based on the permutation problem algorithm, the optimized time-domain audio separation network is used to process the mixed speech data to obtain speech data containing only the target object's speech.
[0031] This embodiment first optimizes the time-domain audio separation network (TasNet network). In the encoder, an optimization from LSTM network to CNN network is proposed, and the model is optimized using optimizations such as repeated multi-layer convolution. The influence of different parameters on the network framework is tested to determine the optimal parameters of the framework, ensuring that the best effect of speech separation is achieved. Then, based on the environmental characteristics of ticket purchase in a noisy environment, a speech separation algorithm under an uncertain number of people is proposed, which solves the shortcoming that the TasNet framework can only perform speech separation after the number of people is determined. Finally, in order to extract the target voice, a speech separation optimization algorithm based on PermutationIssue (permutation problem) is proposed, which optimizes the time complexity compared to the previous method.
[0032] In a specific embodiment of the present disclosure, step S2 may further include step S21 and step S22.
[0033] Step S21: Based on the convolutional neural network model framework, the separator in the time-domain audio separation network is optimized, wherein the separator in the time-domain audio separation network is replaced with a convolutional automatic separator; the convolutional automatic separator in the time-domain audio separation network is subjected to a preset number of dilated convolutions; and the convolution algorithm in the time-domain audio separation network is replaced with a deep cyclic separable convolution algorithm.
[0034] Step S22: Optimize the separator in the time-domain audio separation network and determine the optimal parameters to obtain the optimized time-domain audio separation network.
[0035] The original time-domain audio separation network consists of three parts: an encoder, a separator, and a decoder. In this embodiment, in order to solve the shortcomings of the long short-term memory neural network in the TasNet separator, a method for optimizing the classifier in the TasNet network based on the convolutional neural network model framework is proposed. This method can solve the problem that the long short-term memory neural network framework cannot control the length of mixed speech separation. At the same time, it also has good time invariance, which can solve the problem that the long short-term memory neural network cannot adapt to the starting point of mixed speech.
[0036] The advantage of dilated convolution is not only that it reduces the time complexity of model processing, but also reduces the number of model parameters. Repeated dilated convolution has a good effect on expanding the receptive field. Here, TasNet can reduce the number of framework model parameters and reduce the time complexity of operation by optimizing the convolution operation. In addition, in this embodiment, the optimal parameters are determined by manual parameter adjustment.
[0037] In a specific embodiment of the present disclosure, the step S3 may further include step S31, step S32, step S33, step S34 and step S35.
[0038] Step S31: Processing the mixed speech data using the optimized time-domain audio separation network, wherein the mixed speech data is converted into a vector having a preset feature dimension;
[0039] Step S32: subjecting the vector with the preset feature dimension to N consecutive layers of dilated convolution, and repeating this for M times, to obtain a first processing result, where N and M are both positive integers;
[0040] Step S33: applying the first processing result to the feature map processed by the encoder in the optimized time-domain audio separation network to obtain a second processing result, wherein the second processing result includes a first separation feature map representing the first speech and a second separation feature map representing the second speech;
[0041] Step S34: sorting each voice in the mixed voice data based on the distance between each voice in the mixed voice data and the microphone to obtain sorted voice data;
[0042] Step S35: Obtain the speech data of the target object based on the sorted speech data, the first separation feature map, and the second separation feature map.
[0043] In a specific embodiment of the present disclosure, step S35 may further include step S351 and step S352.
[0044] Step S51: performing weighted processing on the separators in the optimized time-domain audio separation network according to the first separation feature map and the second separation feature map, and then performing parameter optimization processing on the separators to obtain a second optimized time-domain audio separation network;
[0045] Step S52: Use the second optimized time-domain audio separation network to extract the sorted voice data to obtain the voice data of the target object closest to the microphone.
[0046] In this embodiment, after the mixed speech data is input into the optimized time-domain audio separation network, the encoder converts it into a vector with a feature dimension of 512 dimensions; then the optimized separator obtains the feature vector, passes it through 8 consecutive layers of dilated convolution to expand the receptive field, and repeats the process of continuous dilated convolution 3 times to ensure that a sufficiently long speech signal can be recognized. After processing, two mask values are output, and these two mask values are applied to the feature map processed by the encoder to obtain a separated feature map representing the two segments of speech.
[0047] In this embodiment, the voice data ranked first is extracted to obtain the voice data of the target object.
[0048] In this embodiment, a speech separation algorithm based on the permutation problem is proposed based on the permutation problem algorithm and the optimized time-domain audio separation network. The above method solves the problem of separating the target human voice in a noisy environment, and the time efficiency is greatly improved compared with the original TasNet framework.
[0049] In a specific embodiment of the present disclosure, after step S2, step S4, step S5 and step S6 may be further included.
[0050] Step S4: Processing the mixed speech data using the optimized time-domain audio separation network, wherein the mixed speech data is converted into a vector having a preset feature dimension;
[0051] Step S5: obtaining a first speech and a second speech based on the vector having the preset feature dimension, wherein the first speech contains only the speech of one subject, and the second speech contains the speech of multiple subjects;
[0052] Step S6: Determine whether one of the first speech and the second speech contains sounds of multiple objects. If so, return to the step of converting the mixed speech data into a vector with a preset feature dimension; if not, the process ends.
[0053] In a specific embodiment of the present disclosure, step S5 may further include step S51, step S52 and step S53.
[0054] Step S51: subjecting the vector with the preset feature dimension to N consecutive layers of dilated convolution, and repeating this for M times, to obtain a first processing result, where N and M are both positive integers;
[0055] Step S52: applying the first processing result to the feature map processed by the encoder in the optimized time-domain audio separation network to obtain a second processing result, wherein the second processing result includes a first separation feature map representing the first speech and a second separation feature map representing the second speech;
[0056] Step S53: Analyze the first separation feature map and the second separation feature map to obtain the first speech and the second speech.
[0057] In this embodiment, after the mixed speech data is input into the optimized time-domain audio separation network, the encoder converts it into a vector with a feature dimension of 512;
[0058] The optimized separator then obtains the feature vector, passes it through 8 consecutive layers of dilated convolution to expand the receptive field, and repeats the process of continuous dilated convolution 3 times to ensure that a sufficiently long speech signal can be recognized. After processing, two mask values are output, and these two mask values are applied to the feature map processed by the encoder to obtain a separated feature map representing the two segments of speech; the decoder obtains the two feature maps and parses out a speech containing only one person's voice and a mixed speech containing multiple people's voices; after separating the two voices, it is determined whether one of the two voices contains multiple different people's voices. If so, return to the step of converting the mixed speech data into a vector with a feature dimension of 512 dimensions; if not, the process ends.
[0059] This embodiment proposes a speech separation algorithm for an uncertain number of people based on the environmental characteristics of ticket purchase in a noisy environment. This is the calculation method in this embodiment. This method overcomes the shortcoming of the TasNet framework that speech separation can only be performed after the number of people is determined.
[0060] Example 2
[0061] like Figure 2 As shown, this embodiment provides a device for acquiring target object speech, which includes an acquisition module 701, an optimization module 702 and a separation module 703.
[0062] An acquisition module 701 is configured to acquire mixed speech data, where the mixed speech data is a mixture of speech from multiple objects, including the target object.
[0063] An optimization module 702 is configured to optimize the time-domain audio separation network based on a convolutional neural network model framework to obtain an optimized time-domain audio separation network;
[0064] The separation module 703 is configured to process the mixed speech data using the optimized time-domain audio separation network based on a permutation problem algorithm to obtain speech data containing only the speech of the target object.
[0065] This embodiment first optimizes the time-domain audio separation network (TasNet network). In the encoder, an optimization from LSTM network to CNN network is proposed, and the model is optimized using optimizations such as repeated multi-layer convolution. The influence of different parameters on the network framework is tested to determine the optimal parameters of the framework, ensuring that the best effect of speech separation is achieved. Then, based on the environmental characteristics of ticket purchase in a noisy environment, a speech separation algorithm under an uncertain number of people is proposed, which solves the shortcoming that the TasNet framework can only perform speech separation after the number of people is determined. Finally, in order to extract the target voice, a speech separation optimization algorithm based on PermutationIssue (permutation problem) is proposed, which optimizes the time complexity compared to the previous method.
[0066] In a specific embodiment of the present disclosure, the optimization module 702 further includes a separation unit 7021 and a parameter adjustment unit 7022 .
[0067] The separation unit 7021 is used to optimize the separator in the time-domain audio separation network based on the convolutional neural network model framework, wherein the separator in the time-domain audio separation network is replaced with a convolutional automatic separator; the convolutional automatic separator in the time-domain audio separation network is subjected to a preset number of dilated convolution processes; and the convolution algorithm in the time-domain audio separation network is replaced with a deep cyclic separable convolution algorithm.
[0068] The parameter adjustment unit 7022 is used to optimize the separator in the time domain audio separation network and determine the optimal parameters to obtain the optimized time domain audio separation network.
[0069] In a specific embodiment of the present disclosure, the separation module 703 further includes a conversion unit 7031 , a first calculation unit 7032 , a second calculation unit 7033 , a sorting unit 7034 and a third calculation unit 7035 .
[0070] a conversion unit 7031, configured to process the mixed speech data using the optimized time-domain audio separation network, wherein the mixed speech data is converted into a vector having a preset feature dimension;
[0071] A first computing unit 7032 is configured to perform N consecutive layers of dilated convolution on the vector having the preset feature dimension, and repeat the process M times to obtain a first processing result, where N and M are both positive integers;
[0072] A second computing unit 7033 is configured to apply the first processing result to the feature map processed by the encoder in the optimized time-domain audio separation network to obtain a second processing result, where the second processing result includes a first separation feature map representing the first speech and a second separation feature map representing the second speech;
[0073] a sorting unit 7034, configured to sort each voice in the mixed voice data based on a distance between each voice in the mixed voice data and the microphone, to obtain sorted voice data;
[0074] The third calculation unit 7035 is used to obtain the speech data of the target object based on the sorted speech data, the first separation feature map and the second separation feature map.
[0075] In a specific embodiment of the present disclosure, the third calculation unit 7035 further includes a processing subunit 70351 and an extraction subunit 70352.
[0076] The processing subunit 70351 is configured to perform weighted processing on the separators in the optimized time-domain audio separation network according to the first separation feature map and the second separation feature map, and then perform parameter optimization processing on the separators to obtain a second optimized time-domain audio separation network.
[0077] The extraction subunit 70352 is configured to extract the sorted speech data using the second optimized time-domain audio separation network to obtain the speech data of the target object closest to the microphone.
[0078] In a specific embodiment of the present disclosure, the optimization module 702 further includes a processing module 704 , a calculation module 705 and a judgment module 706 .
[0079] a processing module 704 for processing the mixed speech data using the optimized time-domain audio separation network, wherein the mixed speech data is converted into a vector having a preset feature dimension;
[0080] A calculation module 705 is configured to obtain a first speech and a second speech based on the vector having a preset feature dimension, wherein the first speech includes only the speech of one subject, and the second speech includes the speech of multiple subjects;
[0081] a determination module 706 for determining whether one of the first speech and the second speech contains sounds of multiple objects, and if so, returning to the step of converting the mixed speech data into a vector having a preset feature dimension;
[0082] If not, the process ends.
[0083] In a specific implementation of the present disclosure, the calculation module 705 further includes a fourth calculation unit 7051 , a fifth calculation unit 7052 and an analysis unit 7053 .
[0084] a fourth computing unit 7051 configured to perform N consecutive layers of dilated convolution on the vector having the preset feature dimension, and repeat the process M times to obtain a first processing result, where N and M are both positive integers;
[0085] a fifth computing unit 7052, configured to apply the first processing result to the feature map processed by the encoder in the optimized time-domain audio separation network to obtain a second processing result, where the second processing result includes a first separated feature map representing the first speech and a second separated feature map representing the second speech;
[0086] The parsing unit 7053 is used to parse the first separation feature map and the second separation feature map to obtain the first speech and the second speech.
[0087] It should be noted that, regarding the apparatus in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated on here.
[0088] Example 3
[0089] Corresponding to the above method embodiment, the embodiment of the present disclosure further provides a device for acquiring target object voice. The device for acquiring target object voice described below and the method for acquiring target object voice described above can refer to each other.
[0090] Figure 3 FIG. 8 is a block diagram of a device 800 for acquiring target object speech according to an exemplary embodiment.
[0091] like Figure 3 As shown, the target object speech acquisition device 800 may include: a processor 801, a memory 802. The target object speech acquisition device 800 may also include one or more of a multimedia component 803, an input / output (I / O) interface 804, and a communication component 805.
[0092] The processor 801 is used to control the overall operation of the target object voice acquisition device 800 to complete all or part of the steps in the target object voice acquisition method described above. The memory 802 is used to store various types of data to support the operation of the target object voice acquisition device 800. This data may include, for example, instructions for any application or method operating on the target object voice acquisition device 800, as well as application-related data such as contact information, sent and received messages, pictures, audio, video, etc. The memory 802 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touch screen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signal may be further stored in the memory 802 or transmitted via the communication component 805. The audio component also includes at least one speaker for outputting audio signals. The I / O interface 804 provides an interface between the processor 801 and other interface modules, and the above-mentioned other interface modules can be a keyboard, a mouse, buttons, etc. These buttons can be virtual buttons or physical buttons. The communication component 805 is used for wired or wireless communication between the acquisition device 800 of the target object voice and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G or 4G, or a combination of one or more of them, so the corresponding communication component 805 may include: Wi-Fi module, Bluetooth module, NFC module.
[0093] In an exemplary embodiment, the target object speech acquisition device 800 can be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to execute the above-mentioned target object speech acquisition method.
[0094] In another exemplary embodiment, a computer-readable storage medium including program instructions is further provided. When the program instructions are executed by a processor, the steps of the above-mentioned method for acquiring the target object speech are implemented.
[0095] For example, the computer-readable storage medium may be the aforementioned memory 802 including program instructions, and the aforementioned program instructions may be executed by the processor 801 of the target object speech acquisition device 800 to complete the aforementioned target object speech acquisition method.
[0096] Example 4
[0097] Corresponding to the above method embodiment, the embodiment of the present disclosure further provides a readable storage medium. The readable storage medium described below and the target object speech acquisition method described above can refer to each other.
[0098] A readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for acquiring the target object speech of the above method embodiment.
[0099] The readable storage medium may specifically be any readable storage medium that can store program code, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.
[0100] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.
Claims
1. A method for acquiring target object speech, characterized in that: include: Acquire mixed voice data, where the mixed voice data is a mixture of voices of multiple objects, including the target object; Based on the convolutional neural network model framework, the time domain audio separation network is optimized to obtain the optimized time domain audio separation network; Based on the permutation problem algorithm, the mixed speech data is processed using the optimized time-domain audio separation network to obtain speech data containing only the speech of the target object; Among them, based on the convolutional neural network model framework, the time domain audio separation network is optimized to obtain the optimized time domain audio separation network, including: Based on the convolutional neural network model framework, the separator in the time-domain audio separation network is optimized, wherein the separator in the time-domain audio separation network is replaced with a convolutional automatic separator; Performing a preset number of dilated convolutions on the convolution automatic separator in the time-domain audio separation network; Replacing the convolution algorithm in the time-domain audio separation network with a deep recurrent separable convolution algorithm; After optimizing the separator in the time-domain audio separation network, optimal parameters are determined to obtain the optimized time-domain audio separation network; The method of processing the mixed speech data using the optimized time-domain audio separation network based on the permutation problem algorithm to obtain speech data containing only the speech of the target object includes: Processing the mixed speech data using the optimized time-domain audio separation network, wherein the mixed speech data is converted into a vector having a preset feature dimension; The vector with the preset feature dimension is subjected to N consecutive layers of dilated convolution, and the convolution is repeated M times to obtain a first processing result, where N and M are both positive integers; Applying the first processing result to the feature map processed by the encoder in the optimized time-domain audio separation network to obtain a second processing result, wherein the second processing result includes a first separation feature map representing the first speech and a second separation feature map representing the second speech; sorting each voice in the mixed voice data based on the distance between each voice in the mixed voice data and the microphone to obtain sorted voice data; Obtaining speech data of the target object based on the sorted speech data, the first separation feature map, and the second separation feature map; The step of obtaining the speech data of the target object based on the sorted speech data, the first separation feature map, and the second separation feature map includes: performing weighted processing on the separators in the optimized time-domain audio separation network according to the first separation feature map and the second separation feature map, and then performing parameter optimization processing on the separators to obtain a second optimized time-domain audio separation network; The sorted voice data is extracted using the second optimized time-domain audio separation network to obtain the voice data of the target object closest to the microphone.
2. A device for acquiring target object speech, characterized in that: include: An acquisition module, configured to acquire mixed voice data, wherein the mixed voice data is formed by mixing voices of multiple objects, including the target object; An optimization module is used to optimize the time-domain audio separation network based on the convolutional neural network model framework to obtain an optimized time-domain audio separation network; A separation module, configured to process the mixed speech data using the optimized time-domain audio separation network based on a permutation problem algorithm to obtain speech data containing only the speech of the target subject; Among them, the optimization module includes: A separation unit, configured to optimize a separator in the time-domain audio separation network based on a convolutional neural network model framework, wherein the separator in the time-domain audio separation network is replaced with a convolutional automatic separator; Performing a preset number of dilated convolutions on the convolution automatic separator in the time-domain audio separation network; Replacing the convolution algorithm in the time-domain audio separation network with a deep recurrent separable convolution algorithm; A parameter adjustment unit, configured to optimize the separator in the time-domain audio separation network and determine the optimal parameters to obtain the optimized time-domain audio separation network; The separation module includes: a conversion unit, configured to process the mixed speech data using the optimized time-domain audio separation network, wherein the mixed speech data is converted into a vector having a preset feature dimension; A first computing unit is configured to perform N consecutive layers of dilated convolution on the vector having the preset feature dimension, and repeat the process M times to obtain a first processing result, where N and M are both positive integers; a second computing unit, configured to apply the first processing result to a feature map processed by an encoder in the optimized time-domain audio separation network to obtain a second processing result, wherein the second processing result includes a first separation feature map representing the first speech and a second separation feature map representing the second speech; a sorting unit, configured to sort each voice in the mixed voice data based on a distance between each voice in the mixed voice data and a microphone, to obtain sorted voice data; a third computing unit, configured to obtain the speech data of the target object based on the sorted speech data, the first separation feature map, and the second separation feature map; The third computing unit includes: a processing subunit, configured to perform weighted processing on the separator in the optimized time-domain audio separation network according to the first separation feature map and the second separation feature map, and then perform parameter adjustment and optimization processing on the separator to obtain a second optimized time-domain audio separation network; The extraction subunit is used to extract the sorted voice data using the second optimized time domain audio separation network to obtain the voice data of the target object closest to the microphone.
3. A device for acquiring target object voice, characterized in that: include: memory for storing computer programs; A processor is configured to implement the steps of the method for acquiring the target object speech as claimed in claim 1 when executing the computer program.
4. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method for acquiring the target object speech as claimed in claim 1 are implemented.