WO2023273579A1 - 模型的训练方法、语音识别方法、装置、介质及设备 - Google Patents

模型的训练方法、语音识别方法、装置、介质及设备 Download PDF

Info

Publication number
WO2023273579A1
WO2023273579A1 PCT/CN2022/089607 CN2022089607W WO2023273579A1 WO 2023273579 A1 WO2023273579 A1 WO 2023273579A1 CN 2022089607 W CN2022089607 W CN 2022089607W WO 2023273579 A1 WO2023273579 A1 WO 2023273579A1
Authority
WO
WIPO (PCT)
Prior art keywords
training
learning rate
model
preset
range
Prior art date
Application number
PCT/CN2022/089607
Other languages
English (en)
French (fr)
Inventor
董林昊
梁镇麟
蔡猛
马泽君
Original Assignee
北京有竹居网络技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京有竹居网络技术有限公司 filed Critical 北京有竹居网络技术有限公司
Publication of WO2023273579A1 publication Critical patent/WO2023273579A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00—Speech recognition
    • G10L15/06—Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice
    • G10L15/063—Training
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00—Speech recognition
    • G10L15/08—Speech classification or search
    • G10L15/16—Speech classification or search using artificial neural networks
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00—Speech recognition
    • G10L15/06—Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice
    • G10L15/063—Training
    • G10L2015/0635—Training updating or merging of old and new templates; Mean values; Weighting

Definitions

  • the present disclosure relates to the field of computer technology, and in particular, to a model training method, speech recognition method, device, medium and equipment.
  • the speech recognition can be performed by performing sequence alignment mapping through an alignment algorithm.
  • the model in order to improve the accuracy of the model for speech recognition, the model is usually trained in a multi-task learning manner, and the training method usually relies on super-large-scale speech-text training speech data for training.
  • the video memory on most training devices is limited, and it is difficult to support the training process of the speech recognition model under the ultra-large-scale training speech data.
  • the present disclosure provides a speech recognition model training method, the speech recognition model performs joint training on a preset model through a master node and a plurality of working nodes, and the method includes:
  • For the preset model in each of the working nodes obtain the training speech data corresponding to the preset model in the current iteration step, wherein the model parameters of the preset models in each of the working nodes are the same;
  • the learning rate increases to a first learning rate in a positive correlation with the number of iterations during the iterative process, and decreases from the first learning rate to a preset number of iterations per interval , until the second learning rate;
  • the model parameters of the preset model in the master node are updated according to the learning rate and the target gradient.
  • a speech recognition method comprising:
  • a speech recognition model training device performs joint training on a preset model through a master node and a plurality of working nodes, and the device includes:
  • An acquisition module configured to acquire the training voice data corresponding to the preset model in the current iteration step for each preset model in the working node, wherein the model of the preset model in each working node The parameters are the same;
  • a first determination module configured to determine the target gradient corresponding to the preset model in the master node in the iterative step according to the training voice data corresponding to each preset model
  • the second determining module is used to determine the learning rate corresponding to the iterative step, wherein the learning rate increases to the first learning rate in a positive correlation with the number of iterations during the iterative process, and increases from the first learning rate every Decreases the number of iterations at interval presets, up to the second learning rate;
  • An update module configured to update the model parameters of the preset model in the master node according to the learning rate and the target gradient.
  • a speech recognition device comprising:
  • a receiving module configured to receive voice data to be recognized
  • An input module configured to input the speech data into a speech recognition model to obtain the target text corresponding to the speech data, wherein the speech recognition model is trained by the speech recognition model training method described in the first aspect of the present disclosure income.
  • a computer-readable medium on which a computer program is stored, and when the program is executed by a processing device, the steps of the methods described in the first aspect and the second aspect of the present disclosure are implemented.
  • an electronic device including:
  • a processing device configured to execute the computer program in the storage device to implement the steps of the methods described in the first aspect and the second aspect of the present disclosure.
  • the speech recognition model is obtained through joint training of the preset model by the master node and multiple working nodes.
  • the preset model is obtained for each of the working nodes.
  • the speech recognition model can be obtained through the joint training of multiple nodes to the same preset model, and the pressure of large-scale training speech data can be distributed to multiple working nodes, so that the training method of the speech recognition model
  • the method can be applied to devices with limited video memory, and improves the application range of the training method of the speech recognition model.
  • the learning rate corresponding to the iterative step can be dynamically determined to improve the matching between the learning rate and the training process, improve the generalization and stability of the trained speech recognition model, and further improve the speech recognition model training efficiency and improve user experience.
  • Fig. 1 is the flowchart of the training method of the speech recognition model provided according to an embodiment of the present disclosure
  • FIG. 2 is a relationship diagram between a learning rate and the number of iterations provided according to an embodiment of the present disclosure
  • Fig. 3 is a correspondence diagram of working nodes and training voice data provided according to an embodiment of the present disclosure
  • FIG. 4 is a block diagram of a training device for a speech recognition model provided according to an embodiment of the present disclosure
  • Fig. 5 is a block diagram of a speech recognition device provided according to an embodiment of the present disclosure.
  • FIG. 6 shows a schematic structural diagram of an electronic device suitable for implementing an embodiment of the present disclosure.
  • the term “comprise” and its variations are open-ended, ie “including but not limited to”.
  • the term “based on” is “based at least in part on”.
  • the term “one embodiment” means “at least one embodiment”; the term “another embodiment” means “at least one further embodiment”; the term “some embodiments” means “at least some embodiments.” Relevant definitions of other terms will be given in the description below.
  • FIG. 1 is a flow chart of a speech recognition model training method according to an embodiment of the present disclosure.
  • the speech recognition model performs joint training on a preset model through a master node and multiple working nodes.
  • the training of the speech recognition model may be implemented in a parallel updating manner, and the training of the same model may be completed through multiple working nodes.
  • the master node is used to maintain the global model parameters of the model. As shown in Figure 1, the method includes:
  • step 11 for the preset model in each working node, the training voice data corresponding to the preset model in the current iteration step is obtained, wherein the model parameters of the preset models in each working node are the same.
  • the model parameters of the preset model in each working node can be synchronized from the global model parameters maintained in the master node, so as to ensure the consistency of the model parameters of the multiple working nodes when training in the current iteration step , that is, multiple working nodes are trained based on the same model parameters to simulate the training process of one node for the preset model.
  • the multiple working nodes can synchronize model parameters from the master node, and the master node can publish the global model parameters it maintains to each working node, and each working node can use the The model corresponding to the global model parameters is used as the preset model corresponding to the iterative step for training.
  • the training speech data can be divided in the embodiment of the present disclosure.
  • the multiple sets of training voice data corresponding to each iteration step can be randomly selected from the whole training voice data corresponding to the preset model, so that the randomness of the training voice data in each iteration step can be guaranteed, and the number of multiple sets of training voice data can be selected randomly. diversity.
  • step 12 according to the training voice data corresponding to each preset model, the target gradient corresponding to the preset model in the master node in the iterative step is determined.
  • the preset model of each working node based on the training voice data corresponding to the preset model in the working node, the preset model of the working node can be individually trained, and then based on each A gradient corresponding to the preset model in the working node in the iterative step, and a target gradient corresponding to the preset model in the master node in the iterative step is determined.
  • step 13 the learning rate corresponding to the iterative step is determined, wherein during the iterative process, the learning rate is positively correlated with the number of iterations and increases to the first learning rate, and the first learning rate is preset for each interval iteration The number of times decreases until the second learning rate.
  • the learning rate (Learning rate) is usually used as an important hyperparameter in supervised learning and deep learning, which determines whether the objective function can converge to the local minimum and when it can converge to the minimum.
  • learning rate is set too small, the convergence process will become very slow.
  • the learning rate is set too large, the gradient may oscillate back and forth near the minimum value, and may even fail to converge.
  • the learning rate corresponding to the current iteration step can be dynamically determined according to the iteration number of the training process, so as to improve the matching between the learning rate and the training process.
  • the learning rate in the iterative process, the learning rate is positively correlated with the number of iterations and increases to the first learning rate, that is, in the early stage of the iterative process, the learning rate of each iteration step can be used to quickly determine the model The initial position of convergence, and to ensure the generalization of the model, and then reduce the number of preset iterations per interval from the first learning rate to the second learning rate, that is, at the end of the iterative process, several iteration steps can be used to learn The optimal convergence position is determined around the possible convergence positions by reducing the rate, so that the model can converge quickly.
  • step 14 the model parameters of the preset model in the master node are updated according to the learning rate and the target gradient.
  • the method of updating the preset model based on the learning rate and gradient can adopt the model update method commonly used in this field to update the model parameters of the preset model in the master node,
  • the model parameters of the preset model can be updated based on the gradient descent method, which will not be repeated here.
  • the speech recognition model is obtained through joint training of the preset model by the master node and multiple working nodes.
  • the preset model is obtained for each of the working nodes.
  • the speech recognition model can be obtained through the joint training of multiple nodes to the same preset model, and the pressure of large-scale training speech data can be distributed to multiple working nodes, so that the training method of the speech recognition model
  • the method can be applied to devices with limited video memory, and improves the application range of the training method of the speech recognition model.
  • the learning rate corresponding to the iterative step can be dynamically determined to improve the matching between the learning rate and the training process, improve the generalization and stability of the trained speech recognition model, and further improve the speech recognition model training efficiency and improve user experience.
  • the number of iterations of the preset model is divided into a first range, a second range, and a third range in ascending order, and the three ranges are mutually do not coincide.
  • the first range is the initial stage of the preset model training
  • the second range is the middle stage of the preset model training
  • the third range is the end stage of the preset model training.
  • the learning rate is increased to the first learning rate in proportion to the number of iterations.
  • the learning rate increases in a positive correlation with the number of iterations, that is, the learning rate can be linearly increased as the number of iterations increases. increase, so that the learning rate of each iteration step can be increased evenly in the first range, avoiding the problem that the model is directly biased to a certain application scenario in the training speech data when a larger learning rate is directly adopted, so that the training In the initial preset model, some knowledge in more scenarios can be learned to ensure the reliability and effectiveness of the convergence location search and improve the generalization of the trained model.
  • the learning rate is the first learning rate.
  • the third range is divided into multiple sub-ranges at intervals of the preset number of iterations, the learning rate corresponding to each sub-range is the same, and the learning rate corresponding to the current sub-range is according to The preset attenuation rate is reduced to obtain the learning rate corresponding to the next sub-range until the learning rate corresponding to the next sub-range is the second learning rate.
  • the learning rate corresponding to the next sub-range can be determined by the following formula:
  • L' is used to represent the learning rate corresponding to the next sub-range
  • L is used to represent the learning rate corresponding to the current sub-range, and the number of iterations in the current sub-range is less than the number of iterations in the next sub-range;
  • d is used to represent the preset decay rate
  • n is used to represent the preset number of iterations
  • N is used to represent the total number of iterations corresponding to the training process
  • R is used to represent the ratio of the number of iterations corresponding to the third range to the total number of iterations.
  • the learning rate corresponding to the second range decreases exponentially, so that the speed of learning rate reduction can be gradually slowed down, which facilitates the precise exploration of the optimal solution position.
  • the preset model has determined the optimization space, then the learning rate can be reduced at intervals, and more precise exploration has been carried out in the optimization space , in order to determine the optimal solution and improve the convergence efficiency and accuracy of the model.
  • the learning rate corresponding to the current iteration step can be dynamically determined according to the number of iterations of the training process, and the matching between the learning rate and the training process can be improved. Moreover, in the iterative step corresponding to the first range, the learning rate increases in a positive correlation with the number of iterations, which can effectively improve the comprehensiveness of the knowledge that can be learned in the early stage of training, and ensure the stability and generality of the speech recognition model obtained through training. Chemical.
  • the optimization space can be quickly and accurately determined, and the convergence position can be accurately determined, which is convenient for improving the convergence efficiency of the preset model training process and improving the training efficiency of the speech recognition model and training accuracy to improve user experience.
  • the training voice data is divided into multiple batches of data, and the number of batches in each of the working nodes is the same.
  • each set of training voice data may be divided into M batches of data.
  • step 12 according to the training speech data corresponding to each preset model, an exemplary implementation manner of determining the target gradient corresponding to the preset model in the master node in the iterative step is as follows, and this step may include:
  • the training voice data in each working node is all divided into M batches, then in this iterative step, for each working node, the data of batch 1 can be input into the batch 1 first
  • the gradient corresponding to the batch of training voice data can be determined based on the output of the preset model and the annotation of the training voice data, that is, the batch gradient.
  • the data of batch 1 in the working node A1 can be input into the preset model of the working node A1, so that in the working node A1, the data in the batch 1 can be determined based on the preset model The gradient corresponding to the data.
  • the data of batch 1 in the working node A2 can be input into the preset model of the working node A2, so that in the working node A2, the data corresponding to the batch 1 can be determined based on the preset model Gradient, other working nodes are processed in the same way, so I won’t go into details here.
  • the process of determining the gradient based on the model can be a gradient calculation method commonly used in this field, which will not be repeated here.
  • the target gradient may be determined according to the batch gradient corresponding to each batch of training voice data in the plurality of working nodes.
  • the batch gradient corresponding to each batch of the working node can be accumulated to determine the overall gradient. Then, the target gradient can be determined based on the overall gradient of each working node, for example, the target gradient can be determined in a manner of averaging the overall gradient of each working node.
  • the corresponding batch gradient can be determined for each batch of training voice data in each working node, ensuring that each The synchronization and real-time performance of the gradient calculation of the training voice data in a working node improves the calculation efficiency of the target gradient, thereby improving the update efficiency of the preset model.
  • this step may include:
  • An average value of batch gradients corresponding to the same batch in the plurality of working nodes is determined as the gradient corresponding to the batch of the preset model in the master node.
  • the batch gradient corresponding to the batch can be directly averaged, that is, the batch gradient corresponding to the batch is determined at the working nodes A1-Ak.
  • the batch gradient corresponding to multiple working nodes A1-Ak can be directly averaged to obtain the overall gradient corresponding to the batch, which can improve the data accuracy of the batch to a certain extent. Consideration to ensure the reference of knowledge corresponding to the batch of data.
  • the sum of gradients corresponding to the preset model and each batch is determined as the target gradient.
  • the overall gradient of the model under the batch of training voice data can be determined for each batch of training voice data, further improving the efficiency of using multiple training voice data to train the same model.
  • Accuracy ensures the synchronization and real-time performance of preset model training in multiple working nodes, and can improve the training efficiency and accuracy of the speech recognition model to a certain extent, thereby improving the accuracy of the trained speech recognition model.
  • the present disclosure also provides a speech recognition method, the method comprising:
  • the speech recognition model in the training process of the speech recognition model, can be obtained through the joint training of multiple nodes on the same preset model, and at the same time, the learning rate corresponding to the iterative step can be dynamically determined during the iterative process, Improve the matching between the learning rate and the training process, improve the generalization and stability of the trained speech recognition model, and at the same time improve the accuracy of the trained speech recognition model to a certain extent, thus effectively improving the recognition
  • the accuracy of the target text improves the user experience.
  • FIG. 4 is a block diagram of a training device for a speech recognition model provided according to an embodiment of the present disclosure.
  • the speech recognition model carries out joint training to the preset model through the master node and a plurality of working nodes, and the device 40 includes:
  • the acquiring module 41 is configured to acquire the training speech data corresponding to the preset model in the current iteration step for each preset model in the working node, wherein the preset model in each working node The model parameters are the same;
  • the first determination module 42 is used to determine the target gradient corresponding to the preset model in the master node in the iterative step according to the training voice data corresponding to each of the preset models;
  • the second determining module 43 is used to determine the learning rate corresponding to the iterative step, wherein, during the iterative process, the learning rate is positively correlated with the number of iterations and increases to the first learning rate, and from the first learning rate Decrease the number of iterations per interval preset, up to the second learning rate;
  • An update module 44 configured to update the model parameters of the preset model in the master node according to the learning rate and the target gradient.
  • the number of iterations of the preset model is divided into a first range, a second range and a third range in ascending order, and the three ranges do not overlap with each other;
  • the learning rate increases to the first learning rate in direct proportion to the number of iterations; if the number of iterations belongs to the second range, the learning rate is the first learning rate; if the number of iterations belongs to the third range, the third range is divided into multiple sub-ranges at intervals of the preset number of iterations, and the learning rate corresponding to each sub-range is the same, currently The learning rate corresponding to the sub-range is reduced according to the preset attenuation rate to obtain the learning rate corresponding to the next sub-range until the learning rate corresponding to the next sub-range is the second learning rate.
  • the learning rate corresponding to the next sub-range is determined by the following formula:
  • L' is used to represent the learning rate corresponding to the next sub-range
  • L is used to represent the learning rate corresponding to the current sub-range, and the number of iterations in the current sub-range is less than the number of iterations in the next sub-range;
  • d is used to represent the preset decay rate
  • n is used to represent the preset number of iterations
  • N is used to represent the total number of iterations corresponding to the training process
  • R is used to represent the ratio of the number of iterations corresponding to the third range to the total number of iterations.
  • the training speech data is divided into multiple batches of data, and the number of batches in each of the working nodes is the same;
  • the first determination module 42 includes:
  • the first determining submodule is used to input the training speech data corresponding to the same batch in the plurality of working nodes into the preset models in the working nodes, and determine the relationship between each preset model and the batch The batch gradient corresponding to the second training speech data;
  • the second determining submodule is configured to determine the target gradient according to the batch gradient corresponding to each batch of training voice data in the plurality of working nodes.
  • the second determining submodule includes:
  • a third determining submodule configured to determine the average value of batch gradients corresponding to the same batch in the plurality of working nodes as the gradient corresponding to the batch in the preset model in the master node;
  • the fourth determining submodule is used to determine the sum of gradients corresponding to the preset model and each batch as the target gradient.
  • FIG. 5 it is a block diagram of a speech recognition device provided according to an embodiment of the present disclosure, and the device includes:
  • Receiving module 51 for receiving the voice data to be recognized
  • the input module 52 is configured to input the speech data into a speech recognition model to obtain the target text corresponding to the speech data, wherein the speech recognition model is performed by the speech recognition model training method provided in any embodiment of the present disclosure. training income.
  • FIG. 6 it shows a schematic structural diagram of an electronic device 600 suitable for implementing an embodiment of the present disclosure.
  • the terminal equipment in the embodiment of the present disclosure may include but not limited to such as mobile phone, notebook computer, digital broadcast receiver, PDA (personal digital assistant), PAD (tablet computer), PMP (portable multimedia player), vehicle terminal (such as mobile terminals such as car navigation terminals) and fixed terminals such as digital TVs, desktop computers and the like.
  • the electronic device shown in FIG. 6 is only an example, and should not limit the functions and application scope of the embodiments of the present disclosure.
  • an electronic device 600 may include a processing device (such as a central processing unit, a graphics processing unit, etc.) 601, which may be randomly accessed according to a program stored in a read-only memory (ROM) 602 or loaded from a storage device 608.
  • a processing device such as a central processing unit, a graphics processing unit, etc.
  • RAM memory
  • various programs and data necessary for the operation of the electronic device 600 are also stored.
  • the processing device 601, ROM 602, and RAM 603 are connected to each other through a bus 604.
  • An input/output (I/O) interface 605 is also connected to the bus 604 .
  • the following devices can be connected to the I/O interface 605: input devices 606 including, for example, a touch screen, touchpad, keyboard, mouse, camera, microphone, accelerometer, gyroscope, etc.; including, for example, a liquid crystal display (LCD), speaker, vibration an output device 607 such as a computer; a storage device 608 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 609.
  • the communication means 609 may allow the electronic device 600 to communicate with other devices wirelessly or by wire to exchange data. While FIG. 6 shows electronic device 600 having various means, it should be understood that implementing or having all of the means shown is not a requirement. More or fewer means may alternatively be implemented or provided.
  • embodiments of the present disclosure include a computer program product, which includes a computer program carried on a non-transitory computer readable medium, where the computer program includes program code for executing the method shown in the flowchart.
  • the computer program may be downloaded and installed from a network via communication means 609, or from storage means 608, or from ROM 602.
  • the processing device 601 When the computer program is executed by the processing device 601, the above-mentioned functions defined in the methods of the embodiments of the present disclosure are performed.
  • the above-mentioned computer-readable medium in the present disclosure may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two.
  • a computer readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to, electrical connections with one or more wires, portable computer diskettes, hard disks, random access memory (RAM), read-only memory (ROM), erasable Programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.
  • a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.
  • a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave carrying computer-readable program code therein. Such propagated data signals may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing.
  • a computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which can transmit, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device .
  • Program code embodied on a computer readable medium may be transmitted by any appropriate medium, including but not limited to wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.
  • the client and the server can communicate using any currently known or future network protocols such as HTTP (HyperText Transfer Protocol, Hypertext Transfer Protocol), and can communicate with digital data in any form or medium
  • HTTP HyperText Transfer Protocol
  • the communication eg, communication network
  • Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), internetworks (e.g., the Internet), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future developed network of.
  • the above-mentioned computer-readable medium may be included in the above-mentioned electronic device, or may exist independently without being incorporated into the electronic device.
  • the above-mentioned computer-readable medium carries one or more programs, and when the above-mentioned one or more programs are executed by the electronic device, the electronic device executes a speech recognition model training method, and the speech recognition model passes the master node and multiple Each working node performs joint training on the preset model, and the method includes: for each preset model in the working node, obtaining the training voice data corresponding to the preset model in the current iteration step, wherein each The model parameters of the preset models in the working nodes are the same; according to the training voice data corresponding to each of the preset models, determine the target gradient corresponding to the preset models in the master node in the iteration step; determine the The learning rate corresponding to the iterative step, wherein, in the iterative process, the learning rate increases to the first learning rate in a positive correlation with the number of iterations, and decreases from the first learning rate to the preset number of iterations per interval until A second learning rate: updating the model parameters of the preset model in
  • the above-mentioned computer-readable medium carries one or more programs, and when the above-mentioned one or more programs are executed by the electronic device, the electronic device is made to execute a speech recognition method, the method comprising: receiving the speech to be recognized data; input the speech data into a speech recognition model to obtain the target text corresponding to the speech data, wherein the speech recognition model is obtained by training the speech recognition model training method described in any embodiment of the present disclosure.
  • Computer program code for carrying out operations of the present disclosure may be written in one or more programming languages, or combinations thereof, including but not limited to object-oriented programming languages—such as Java, Smalltalk, C++, and Includes conventional procedural programming languages - such as "C" or similar programming languages.
  • the program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server.
  • the remote computer may be connected to the user computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (for example, using an Internet service provider to connected via the Internet).
  • LAN local area network
  • WAN wide area network
  • Internet service provider for example, using an Internet service provider to connected via the Internet.
  • each block in a flowchart or block diagram may represent a module, program segment, or portion of code that contains one or more logical functions for implementing specified executable instructions.
  • the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or they may sometimes be executed in the reverse order, depending upon the functionality involved.
  • each block of the block diagrams and/or flowchart illustrations, and combinations of blocks in the block diagrams and/or flowchart illustrations can be implemented by a dedicated hardware-based system that performs the specified functions or operations , or may be implemented by a combination of dedicated hardware and computer instructions.
  • the modules involved in the embodiments described in the present disclosure may be implemented by software or by hardware. Wherein, the name of the module does not constitute a limitation on the module itself under certain circumstances.
  • the obtaining module can also be described as "for each preset model in the working node, obtain the preset model in the The module of the training speech data corresponding to the current iteration step".
  • FPGAs Field Programmable Gate Arrays
  • ASICs Application Specific Integrated Circuits
  • ASSPs Application Specific Standard Products
  • SOCs System on Chips
  • CPLD Complex Programmable Logical device
  • a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device.
  • a machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium.
  • a machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing.
  • machine-readable storage media would include one or more wire-based electrical connections, portable computer discs, hard drives, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (EPROM or flash memory), optical fiber, compact disk read only memory (CD-ROM), optical storage, magnetic storage, or any suitable combination of the foregoing.
  • RAM random access memory
  • ROM read only memory
  • EPROM or flash memory erasable programmable read only memory
  • CD-ROM compact disk read only memory
  • magnetic storage or any suitable combination of the foregoing.
  • Example 1 provides a method for training a speech recognition model, where the speech recognition model performs joint training on a preset model through a master node and multiple working nodes, the method comprising:
  • For the preset model in each of the working nodes obtain the training speech data corresponding to the preset model in the current iteration step, wherein the model parameters of the preset models in each of the working nodes are the same;
  • the learning rate increases to a first learning rate in a positive correlation with the number of iterations during the iterative process, and decreases from the first learning rate to a preset number of iterations per interval , until the second learning rate;
  • the model parameters of the preset model in the master node are updated according to the learning rate and the target gradient.
  • Example 2 provides the method of Example 1, wherein the number of iterations of the preset model is divided into a first range, a second range, and a third range in ascending order, And the three ranges do not overlap with each other;
  • the learning rate increases to the first learning rate in direct proportion to the number of iterations; if the number of iterations belongs to the second range, the learning rate is the first learning rate; if the number of iterations belongs to the third range, the third range is divided into multiple sub-ranges at intervals of the preset number of iterations, and the learning rate corresponding to each sub-range is the same, currently The learning rate corresponding to the sub-range is reduced according to the preset attenuation rate to obtain the learning rate corresponding to the next sub-range until the learning rate corresponding to the next sub-range is the second learning rate.
  • Example 3 provides the method of Example 2, and the learning rate corresponding to the next sub-range is determined by the following formula:
  • L' is used to represent the learning rate corresponding to the next sub-range
  • L is used to represent the learning rate corresponding to the current sub-range, and the number of iterations in the current sub-range is less than the number of iterations in the next sub-range;
  • d is used to represent the preset decay rate
  • n is used to represent the preset number of iterations
  • N is used to represent the total number of iterations corresponding to the training process
  • R is used to represent the ratio of the number of iterations corresponding to the third range to the total number of iterations.
  • Example 4 provides the method of Example 1, the training speech data is divided into multiple batches of data, and the number of batches in each of the working nodes is the same;
  • determining the target gradient corresponding to the preset model in the master node in the iterative step includes:
  • the target gradient is determined according to the batch gradient corresponding to each batch of training speech data in the plurality of working nodes.
  • Example 5 provides the method of Example 4, wherein the target is determined according to the batch gradient corresponding to each batch of training voice data in the plurality of working nodes Gradients, including:
  • the sum of gradients corresponding to the preset model and each batch is determined as the target gradient.
  • Example 6 provides a speech recognition method, the method comprising:
  • Example 7 provides a speech recognition model training device, the speech recognition model performs joint training on the preset model through the master node and multiple working nodes, the device includes:
  • An acquisition module configured to acquire the training voice data corresponding to the preset model in the current iteration step for each preset model in the working node, wherein the model of the preset model in each working node The parameters are the same;
  • a first determination module configured to determine the target gradient corresponding to the preset model in the master node in the iterative step according to the training voice data corresponding to each preset model
  • the second determining module is used to determine the learning rate corresponding to the iterative step, wherein the learning rate increases to the first learning rate in a positive correlation with the number of iterations during the iterative process, and increases from the first learning rate every Decreases the number of iterations at interval presets, up to the second learning rate;
  • An update module configured to update the model parameters of the preset model in the master node according to the learning rate and the target gradient.
  • Example 8 provides a speech recognition device, the device comprising:
  • a receiving module configured to receive voice data to be recognized
  • An input module configured to input the speech data into a speech recognition model to obtain the target text corresponding to the speech data, wherein the speech recognition model is trained by the speech recognition model training method described in any embodiment of the present disclosure income.
  • Example 9 provides a computer-readable medium on which a computer program is stored, and when the program is executed by a processing device, the steps of the method described in any embodiment of the present disclosure are implemented.
  • Example 10 provides an electronic device, comprising:
  • a processing device configured to execute the computer program in the storage device to implement the steps of the method in any embodiment of the present disclosure.

Landscapes

  • Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Evolutionary Computation (AREA)
  • Machine Translation (AREA)

Abstract

提供了一种模型的训练方法、语音识别方法、装置、计算机可读介质及电子设备，该语音识别模型通过主节点和多个工作节点对预设模型进行联合训练获得。该语音识别模型训练方法包括：针对每一工作节点中的预设模型，获取预设模型在当前的迭代步骤对应的训练语音数据（11）；根据每一预设模型对应的训练语音数据，确定主节点中的预设模型在该迭代步骤对应的目标梯度（12）；确定该迭代步骤对应的学习率（13）；根据该学习率和目标梯度对主节点中的预设模型的模型参数进行更新（14）。由此，在迭代过程中可以动态确定与该迭代步骤对应的学习率，提高学习率与训练过程中的匹配性，提高训练所得的语音识别模型泛化性和稳定性的同时，提升训练效率。

Description

模型的训练方法、语音识别方法、装置、介质及设备
相关申请的交叉引用
本申请要求于2021年06月30日提交的，申请号为202110736548.2、发明名称为“模型的训练方法、语音识别方法、装置、介质及设备”的中国专利申请的优先权，该申请的全部内容通过引用结合在本申请中。
技术领域
本公开涉及计算机技术领域，具体地，涉及一种模型的训练方法、语音识别方法、装置、介质及设备。
背景技术
随着深度学习的兴起，各种完全依赖于神经网络进行端到端建模的方法逐渐兴起。在进行语音识别时，由于输入的语音数据和输出的文本数据的长度不同，可以通过对齐算法进行序列对齐映射的方式进行语音识别。
相关技术中，为了提高模型对语音识别的准确度，通常会采用多任务学习的方式对模型进行训练，在该训练方式中通常会依赖超大规模的语音-文本的训练语音数据进行训练。然而大部分训练的设备上的显存有限，难以支持超大规模的训练语音数据下的语音识别模型的训练过程。
发明内容
提供该发明内容部分以便以简要的形式介绍构思，这些构思将在后面的具体实施方式部分被详细描述。该发明内容部分并不旨在标识要求保护的技术方案的关键特征或必要特征，也不旨在用于限制所要求的保护的技术方案的范围。
第一方面，本公开提供一种语音识别模型训练方法，所述语音识别模型通过主节点和多个工作节点对预设模型进行联合训练，所述方法包括：
针对每一所述工作节点中的预设模型，获取所述预设模型在当前的迭代步骤对应的训练语音数据，其中，每一所述工作节点中的预设模型的模型参数相同；
根据每一所述预设模型对应的训练语音数据，确定所述主节点中的预设模型在所述迭代步骤对应的目标梯度；
确定所述迭代步骤对应的学习率，其中，在迭代过程中所述学习率与迭代次数成正相关关系增大至第一学习率，并从所述第一学习率每间隔预设迭代次数减小，直至第二学习 率；
根据所述学习率和所述目标梯度对所述主节点中的预设模型的模型参数进行更新。
第二方面，提供一种语音识别方法，所述方法包括：
接收待识别的语音数据；
将所述语音数据输入语音识别模型，获得所述语音数据对应的目标文本，其中，所述语音识别模型是通过本公开第一方面所述的语音识别模型的训练方法进行训练所得。
第三方面，提供一种语音识别模型训练装置，所述语音识别模型通过主节点和多个工作节点对预设模型进行联合训练，所述装置包括：
获取模块，用于针对每一所述工作节点中的预设模型，获取所述预设模型在当前的迭代步骤对应的训练语音数据，其中，每一所述工作节点中的预设模型的模型参数相同；
第一确定模块，用于根据每一所述预设模型对应的训练语音数据，确定所述主节点中的预设模型在所述迭代步骤对应的目标梯度；
第二确定模块，用于确定所述迭代步骤对应的学习率，其中，在迭代过程中所述学习率与迭代次数成正相关关系增大至第一学习率，并从所述第一学习率每间隔预设迭代次数减小，直至第二学习率；
更新模块，用于根据所述学习率和所述目标梯度对所述主节点中的预设模型的模型参数进行更新。
第四方面，提供一种语音识别装置，所述装置包括：
接收模块，用于接收待识别的语音数据；
输入模块，用于将所述语音数据输入语音识别模型，获得所述语音数据对应的目标文本，其中，所述语音识别模型是通过本公开第一方面所述的语音识别模型的训练方法进行训练所得。
第五方面，提供一种计算机可读介质，其上存储有计算机程序，该程序被处理装置执行时实现本公开第一方面和第二方面所述方法的步骤。
第六方面，提供一种电子设备，包括：
存储装置，其上存储有计算机程序；
处理装置，用于执行所述存储装置中的所述计算机程序，以实现本公开第一方面和第二方面所述方法的步骤。
在上述技术方案中，通过主节点和多个工作节点对预设模型进行联合训练获得语音识别模型，在该训练过程中，针对每一所述工作节点中的预设模型，获取所述预设模型在当前的迭代步骤对应的训练语音数据；根据每一所述预设模型对应的训练语音数据，确定所述主节点中的预设模型在所述迭代步骤对应的目标梯度；确定所述迭代步骤对应的学习率， 从而可以根据所述学习率和所述目标梯度对所述主节点中的预设模型的模型参数进行更新。由此，通过上述技术方案，可以通过多个节点对同一预设模型的联合训练获得语音识别模型，则可以将大规模训练语音数据的压力分配至多个工作节点，使得该语音识别模型的训练方法可以应用于显存有限的设备中，提高该语音识别模型的训练方法的使用范围。同时，在迭代过程中可以动态确定与该迭代步骤对应的学习率，提高学习率与训练过程中的匹配性，提高训练所得的语音识别模型泛化性和稳定性的同时，进一步提升语音识别模型的训练效率，提升用户使用体验。
本公开的其他特征和优点将在随后的具体实施方式部分予以详细说明。
附图说明
结合附图并参考以下具体实施方式，本公开各实施例的上述和其他特征、优点及方面将变得更加明显。贯穿附图中，相同或相似的附图标记表示相同或相似的元素。应当理解附图是示意性的，原件和元素不一定按照比例绘制。在附图中：
图1是根据本公开的一种实施方式提供的语音识别模型的训练方法的流程图；
图2是根据本公开的一种实施方式提供的学习率与迭代次数的关系图；
图3是根据本公开的一种实施方式提供的工作节点和训练语音数据的对应关系图；
图4是根据本公开的一种实施方式提供的语音识别模型的训练装置的框图；
图5是根据本公开的一种实施方式提供的语音识别装置的框图；
图6示出了适于用来实现本公开实施例的电子设备的结构示意图。
具体实施方式
下面将参照附图更详细地描述本公开的实施例。虽然附图中显示了本公开的某些实施例，然而应当理解的是，本公开可以通过各种形式来实现，而且不应该被解释为限于这里阐述的实施例，相反提供这些实施例是为了更加透彻和完整地理解本公开。应当理解的是，本公开的附图及实施例仅用于示例性作用，并非用于限制本公开的保护范围。
应当理解，本公开的方法实施方式中记载的各个步骤可以按照不同的顺序执行，和/或并行执行。此外，方法实施方式可以包括附加的步骤和/或省略执行示出的步骤。本公开的范围在此方面不受限制。
本文使用的术语“包括”及其变形是开放性包括，即“包括但不限于”。术语“基于”是“至少部分地基于”。术语“一个实施例”表示“至少一个实施例”；术语“另一实施例”表示“至少一个另外的实施例”；术语“一些实施例”表示“至少一些实施例”。其他术语的相关定义将在下文描述中给出。
需要注意，本公开中提及的“第一”、“第二”等概念仅用于对不同的装置、模块或单元进行区分，并非用于限定这些装置、模块或单元所执行的功能的顺序或者相互依存关系。
需要注意，本公开中提及的“一个”、“多个”的修饰是示意性而非限制性的，本领域技术人员应当理解，除非在上下文另有明确指出，否则应该理解为“一个或多个”。
本公开实施方式中的多个装置之间所交互的消息或者信息的名称仅用于说明性的目的，而并不是用于对这些消息或信息的范围进行限制。
图1所示，为根据本公开的一种实施方式提供的语音识别模型的训练方法的流程图，所述语音识别模型通过主节点和多个工作节点对预设模型进行联合训练。在本公开实施例中，可以采用并行更新的方式实现语音识别模型的训练，通过多个工作节点完成对同一模型的训练。主节点用于维护该模型的全局模型参数。如图1所示，所述方法包括：
在步骤11中，针对每一工作节点中的预设模型，获取预设模型在当前的迭代步骤对应的训练语音数据，其中，每一所述工作节点中的预设模型的模型参数相同。
其中，每一工作节点中的预设模型的模型参数可以从主节点中维护的全局模型参数中同步而来，从而保证该多个工作节点在当前的迭代步骤进行训练时的模型参数的一致性，即多个工作节点基于同一模型参数进行训练，以模拟一个节点对预设模型的训练过程。
在该实施例中，在每一迭代步骤中该多个工作节点可以从该主节点同步模型参数，主节点可以将其维护的全局模型参数发布给各个工作节点，则每一工作节点可以以该全局模型参数对应的模型作为该迭代步骤对应的预设模型进行训练。
在一种可能的实施例中，在对模型进行训练的过程中需要依赖于超大规模的训练语音数据，为了使得训练语音数据与显存相匹配，在本公开实施例中可以将训练语音数据切分为多组训练语音数据，每组训练语音数据对应于一个工作节点。每一迭代步骤对应的多组训练语音数据可以从该预设模型对应的训练语音数据整体中进行随机选择，从而可以保证每一迭代步骤中训练语音数据的随机性，以及多组训练语音数据的多样性。
在步骤12中，根据每一预设模型对应的训练语音数据，确定主节点中的预设模型在所述迭代步骤对应的目标梯度。
其中，在本公开实施例中，在每一工作节点的预设模型中，可以基于该工作节点中预设模型对应的训练语音数据，对该工作节点的预设模型进行单独训练，之后基于每一工作节点中的预设模型在该迭代步骤对应的梯度，确定主节点中的预设模型在所述迭代步骤对应的目标梯度。
在步骤13中，确定迭代步骤对应的学习率，其中，在迭代过程中所述学习率与迭代次数成正相关关系增大至第一学习率，并从所述第一学习率每间隔预设迭代次数减小，直至第二学习率。
学习率(Learning rate)通常作为监督学习以及深度学习中重要的超参数，其决定着目标函数能否收敛到局部最小值以及何时收敛到最小值。当学习率设置过小时，收敛过程将变得十分缓慢。而当学习率设置过大时，梯度可能会在最小值附近来回震荡，甚至可能无法收敛。
因此，在本公开的实施例中，可以根据训练过程的迭代次数动态确定当前迭代步骤对应的学习率，提高学习率与训练过程中的匹配性。在该实施例中，在迭代过程中所述学习率与迭代次数成正相关关系增大至第一学习率，即在迭代过程初期中可以采用每一迭代步骤学习率增大的方式快速确定出模型收敛的初步位置，并保证模型的泛化性，之后从所述第一学习率每间隔预设迭代次数减小，直至第二学习率，即在迭代过程末期中可以采用间隔几次迭代步骤学习率减小的方式以在可能的收敛位置周围确定出最优收敛位置，以便模型可以快速收敛。
在步骤14中，根据学习率和目标梯度对主节点中的预设模型的模型参数进行更新。
其中，在确定出学习率和目标梯度后，基于学习率和梯度对预设模型进行更新的方式可以采用本领域常用的模型模型更新方式对对主节点中的预设模型的模型参数进行更新，如可以基于梯度下降法对预设模型的模型参数进行更新，在此不再赘述。
在上述技术方案中，通过主节点和多个工作节点对预设模型进行联合训练获得语音识别模型，在该训练过程中，针对每一所述工作节点中的预设模型，获取所述预设模型在当前的迭代步骤对应的训练语音数据；根据每一所述预设模型对应的训练语音数据，确定所述主节点中的预设模型在所述迭代步骤对应的目标梯度；确定所述迭代步骤对应的学习率，从而可以根据所述学习率和所述目标梯度对所述主节点中的预设模型的模型参数进行更新。由此，通过上述技术方案，可以通过多个节点对同一预设模型的联合训练获得语音识别模型，则可以将大规模训练语音数据的压力分配至多个工作节点，使得该语音识别模型的训练方法可以应用于显存有限的设备中，提高该语音识别模型的训练方法的使用范围。同时，在迭代过程中可以动态确定与该迭代步骤对应的学习率，提高学习率与训练过程中的匹配性，提高训练所得的语音识别模型泛化性和稳定性的同时，进一步提升语音识别模型的训练效率，提升用户使用体验。
在一种可能的实施例中，如图2所示，所述预设模型的迭代次数按照由小至大的顺序被划分为第一范围、第二范围和第三范围，且三个范围互不重合。
在本公开实施例中，第一范围为该预设模型进行训练的初期，第二范围为该预设模型训练的中间阶段，第三范围为该预设模型进行训练的末期。每一范围对应的迭代次数可以根据该预设模型在训练过程对应的迭代总次数确定，示例地，迭代总次数可以是N次，其中可以分别设置第一范围、第二范围、第三范围对应的比例分别为α1、α2和α3，且 α1+α2+α3＝1，从而可以对第一范围、第二范围和第三范围进行划分。
若所述迭代次数属于所述第一范围，所述学习率与所述迭代次数成正比例关系增大至所述第一学习率。
其中，在超大规模的训练语音数据下，该训练语音数据的多样性是比较大的，在第一范围中，学习率与迭代次数成正相关关系增大，即随着迭代次数增加学习率可以线性增加，从而可以在第一范围中使得每一迭代步骤的学习率均匀增加，避免直接采用较大的学习率时导致模型直接偏置到该训练语音数据中的某一应用场景的问题，使得训练初期预设模型中可以学习到更多场景下的一些知识，保证收敛位置查找的可靠性和有效性，提高训练所得的模型的泛化性。
若所述迭代次数属于所述第二范围，所述学习率为所述第一学习率。
在通过第一范围对应的迭代次数的迭代训练获得一个知识较全面的预设模型后，在第二范围对应的迭代次数对应的迭代步骤中，则可以直接以第一范围中最大的学习率进行更新，在保证模型泛化性的基础上以较大的步长对优化空间进行探索，以便快速确定出优化空间，获得局部最优解。
若所述迭代次数属于所述第三范围，所述第三范围以所述预设迭代次数间隔划分为多个子范围，每一子范围对应的学习率相同，当前的子范围对应的学习率按照预设衰减率进行减小获得下一子范围对应的学习率，直至下一子范围对应的学习率为所述第二学习率。
在一种可能的实施例中，针对第三范围，可以通过以下公式确定下一子范围对应的学习率：
Figure PCTCN2022089607-appb-000001
其中，L’用于表示下一子范围对应的学习率；
L用于表示当前的子范围对应的学习率，该当前的子范围中的迭代次数小于下一子范围的迭代次数；
d用于表示所述预设衰减率；
n用于表示所述预设迭代次数；
N用于表示训练过程对应的迭代总次数；
R用于表示所述第三范围对应的迭代次数占所述迭代总次数的比例。
其中，针对第三范围中的第一个子范围，其对应的学习率为第二范围对应的学习率。通过上述方案，针对每一子范围而言，每一子范围对应的学习率以指数关系降低，从而可以使得学习率降低的速度逐渐减慢，便于对最优解位置的精确探索。
之后在第三范围对应的迭代次数对应的迭代步骤中，若继续以较大的学习率进行更新，则会由于更新过程中跳动的步长过大导致收敛波动。因此，在其对应的迭代步骤中，基于 第二范围对应的迭代步骤，预设模型已经确定该优化空间，则此时可以间隔减小学习率，已在该优化空间中进一步进行更加精确的探索，以便确定最优解，提高模型的收敛效率和收敛准确度。
由此，通过上述技术方案，可以根据训练过程的迭代次数动态确定当前迭代步骤对应的学习率，提高学习率与训练过程中的匹配性。并且，在第一范围对应的迭代步骤中，学习率与迭代次数成正相关关系增大，可以有效提高训练前期所能学习到的知识的全面性，保证训练获得的语音识别模型的稳定性和泛化性。并且结合第二范围和第三范围的迭代步骤，既可以快速且准确的确定出优化空间，又可以精确确定出收敛位置，便于提高预设模型训练过程的收敛效率，提高语音识别模型的训练效率和训练准确度，提升用户使用体验。
在一种可能的实施例中，所述训练语音数据被划分为多个批次的数据，每一所述工作节点中批次的数量相同。示例地，针对上文所述的每组训练语音数据，可以将每组训练语音数据划分为M个批次的数据。
相应地，在步骤12中，根据每一预设模型对应的训练语音数据，确定主节点中的预设模型在迭代步骤对应的目标梯度的示例性实现方式如下，该步骤可以包括：
将所述多个工作节点中对应于同一批次的训练语音数据，分别输入该工作节点中的预设模型，并确定每一所述预设模型与所述批次的训练语音数据所对应的批次梯度。
示例地，如图3所示，每一工作节点中的训练语音数据均被划分为M个批次，则在该迭代步骤中，针对每一工作节点，可以先将批次1的数据输入该工作节点的预设模型中，从而可以基于该预设模型的输出以及训练语音数据的标注确定该批次的训练语音数据对应的梯度，即为批次梯度。
具体地，针对工作节点A1，可以将该工作节点A1中的批次1的数据输入工作节点A1的预设模型，从而在工作节点A1中，可以基于该预设模型确定该批次1中的数据对应的梯度。针对工作节点A2，可以将该工作节点A2中的批次1的数据输入工作节点A2的预设模型，从而在工作节点A2中，可以基于该预设模型确定该批次1中的数据对应的梯度，其他工作节点的处理方式相同，在此不再赘述。
其中，基于模型确定梯度的过程可以本领域中常用的梯度计算方式，在此不再赘述。
之后，可以根据所述多个工作节点中的每一批次的训练语音数据所对应的批次梯度，确定所述目标梯度。
作为示例，针对每一工作节点，可以针对每一批次的训练语音数据计算出批次梯度之后，对该工作节点的每一批次对应的批次梯度进行累加，确定该工作节点对应的整体梯度。之后可以基于每一工作节点的整体梯度，确定目标梯度，如可以对每一工作节点的整体梯度进行平均的方式确定该目标梯度。
由此，通过上述技术方案，在确定主节点中的预设模型对应的目标梯度时，可以针对每一工作节点中的每一批次的训练语音数据确定其对应的批次梯度，保证对每一工作节点中的训练语音数据的梯度计算的同步性和实时性，提高目标梯度的计算效率，进而提高预设模型更新的效率。
作为另一示例，所述根据所述多个工作节点中的每一批次的训练语音数据所对应的批次梯度，确定所述目标梯度的另一示例性实现方式如下，该步骤可以包括：
将所述多个工作节点中与同一批次对应的批次梯度的平均值，确定为所述主节点中所述预设模型与所述批次对应的梯度。
在该实施例中，在每一工作节点中确定出同一批次对应的批次梯度后，则可以直接对该批次对应的批次梯度进行平均，即在工作节点A1-Ak均确定出批次1对应的批次梯度后，可以直接对多个工作节点A1-Ak对应的批次梯度进行平均处理，以获得该批次对应的整体梯度，可以在一定程度上提高该批次的数据的考量，保证对该批次的数据对应的知识的参考。
之后，将所述预设模型与每一批次对应的梯度之和确定为所述目标梯度。
由此，通过上述技术方案，可以针对每一批次的训练语音数据，均确定出该模型在该批次的训练语音数据下的整体梯度，进一步提高采用多训练语音数据对同一模型进行训练的准确性，保证多个工作节点中预设模型训练的同步性和实时性，并且可以在一定程度上提高语音识别模型的训练效率和训练准确性，进而提高训练所得的语音识别模型的准确率。
本公开还提供一种语音识别方法，所述方法包括：
接收待识别的语音数据；
将所述语音数据输入语音识别模型，获得所述语音数据对应的目标文本，其中，所述语音识别模型是通过上文任一所述的语音识别模型的训练方法进行训练所得。
通过上述技术方案，在语音识别模型的训练过程中，可以通过多个节点对同一预设模型的联合训练获得语音识别模型，同时，在迭代过程中可以动态确定与该迭代步骤对应的学习率，提高学习率与训练过程中的匹配性，提高训练所得的语音识别模型泛化性和稳定性的同时，并且可以在一定程度上提高训练所得的语音识别模型的准确性，从而可以有效提高识别出的目标文本的准确性，提高用户使用体验。
图4所示，为根据本公开的一种实施方式提供的语音识别模型的训练装置的框图。所述语音识别模型通过主节点和多个工作节点对预设模型进行联合训练，该装置40包括：
获取模块41，用于针对每一所述工作节点中的预设模型，获取所述预设模型在当前的迭代步骤对应的训练语音数据，其中，每一所述工作节点中的预设模型的模型参数相同；
第一确定模块42，用于根据每一所述预设模型对应的训练语音数据，确定所述主节点 中的预设模型在所述迭代步骤对应的目标梯度；
第二确定模块43，用于确定所述迭代步骤对应的学习率，其中，在迭代过程中所述学习率与迭代次数成正相关关系增大至第一学习率，并从所述第一学习率每间隔预设迭代次数减小，直至第二学习率；
更新模块44，用于根据所述学习率和所述目标梯度对所述主节点中的预设模型的模型参数进行更新。
可选地，所述预设模型的迭代次数按照由小至大的顺序被划分为第一范围、第二范围和第三范围，且三个范围互不重合；
若所述迭代次数属于所述第一范围，所述学习率与所述迭代次数成正比例关系增大至所述第一学习率；若所述迭代次数属于所述第二范围，所述学习率为所述第一学习率；若所述迭代次数属于所述第三范围，所述第三范围以所述预设迭代次数间隔划分为多个子范围，每一子范围对应的学习率相同，当前的子范围对应的学习率按照预设衰减率进行减小获得下一子范围对应的学习率，直至下一子范围对应的学习率为所述第二学习率。
可选地，通过以下公式确定下一子范围对应的学习率：
Figure PCTCN2022089607-appb-000002
其中，L’用于表示下一子范围对应的学习率；
L用于表示当前的子范围对应的学习率，该当前的子范围中的迭代次数小于下一子范围的迭代次数；
d用于表示所述预设衰减率；
n用于表示所述预设迭代次数；
N用于表示训练过程对应的迭代总次数；
R用于表示所述第三范围对应的迭代次数占所述迭代总次数的比例。
可选地，所述训练语音数据被划分为多个批次的数据，每一所述工作节点中批次的数量相同；
所述第一确定模块42，包括：
第一确定子模块，用于将所述多个工作节点中对应于同一批次的训练语音数据，分别输入该工作节点中的预设模型，并确定每一所述预设模型与所述批次的训练语音数据所对应的批次梯度；
第二确定子模块，用于根据所述多个工作节点中的每一批次的训练语音数据所对应的批次梯度，确定所述目标梯度。
可选地，所述第二确定子模块，包括：
第三确定子模块，用于将所述多个工作节点中与同一批次对应的批次梯度的平均值， 确定为所述主节点中所述预设模型与所述批次对应的梯度；
第四确定子模块，用于将所述预设模型与每一批次对应的梯度之和确定为所述目标梯度。
图5所示，为根据本公开的一种实施方式提供的语音识别装置的框图，所述装置包括：
接收模块51，用于接收待识别的语音数据；
输入模块52，用于将所述语音数据输入语音识别模型，获得所述语音数据对应的目标文本，其中，所述语音识别模型是通过本公开任意实施例所提供的语音识别模型的训练方法进行训练所得。
关于上述实施例中的装置，其中各个模块执行操作的具体方式已经在有关该方法的实施例中进行了详细描述，此处将不做详细阐述说明。
下面参考图6，其示出了适于用来实现本公开实施例的电子设备600的结构示意图。本公开实施例中的终端设备可以包括但不限于诸如移动电话、笔记本电脑、数字广播接收器、PDA(个人数字助理)、PAD(平板电脑)、PMP(便携式多媒体播放器)、车载终端(例如车载导航终端)等等的移动终端以及诸如数字TV、台式计算机等等的固定终端。图6示出的电子设备仅仅是一个示例，不应对本公开实施例的功能和使用范围带来任何限制。
如图6所示，电子设备600可以包括处理装置(例如中央处理器、图形处理器等)601，其可以根据存储在只读存储器(ROM)602中的程序或者从存储装置608加载到随机访问存储器(RAM)603中的程序而执行各种适当的动作和处理。在RAM 603中，还存储有电子设备600操作所需的各种程序和数据。处理装置601、ROM 602以及RAM 603通过总线604彼此相连。输入/输出(I/O)接口605也连接至总线604。
通常，以下装置可以连接至I/O接口605：包括例如触摸屏、触摸板、键盘、鼠标、摄像头、麦克风、加速度计、陀螺仪等的输入装置606；包括例如液晶显示器(LCD)、扬声器、振动器等的输出装置607；包括例如磁带、硬盘等的存储装置608；以及通信装置609。通信装置609可以允许电子设备600与其他设备进行无线或有线通信以交换数据。虽然图6示出了具有各种装置的电子设备600，但是应理解的是，并不要求实施或具备所有示出的装置。可以替代地实施或具备更多或更少的装置。
特别地，根据本公开的实施例，上文参考流程图描述的过程可以被实现为计算机软件程序。例如，本公开的实施例包括一种计算机程序产品，其包括承载在非暂态计算机可读介质上的计算机程序，该计算机程序包含用于执行流程图所示的方法的程序代码。在这样的实施例中，该计算机程序可以通过通信装置609从网络上被下载和安装，或者从存储装置608被安装，或者从ROM 602被安装。在该计算机程序被处理装置601执行时，执行本公开实施例的方法中限定的上述功能。
需要说明的是，本公开上述的计算机可读介质可以是计算机可读信号介质或者计算机可读存储介质或者是上述两者的任意组合。计算机可读存储介质例如可以是——但不限于——电、磁、光、电磁、红外线、或半导体的系统、装置或器件，或者任意以上的组合。计算机可读存储介质的更具体的例子可以包括但不限于：具有一个或多个导线的电连接、便携式计算机磁盘、硬盘、随机访问存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、光纤、便携式紧凑磁盘只读存储器(CD-ROM)、光存储器件、磁存储器件、或者上述的任意合适的组合。在本公开中，计算机可读存储介质可以是任何包含或存储程序的有形介质，该程序可以被指令执行系统、装置或者器件使用或者与其结合使用。而在本公开中，计算机可读信号介质可以包括在基带中或者作为载波一部分传播的数据信号，其中承载了计算机可读的程序代码。这种传播的数据信号可以采用多种形式，包括但不限于电磁信号、光信号或上述的任意合适的组合。计算机可读信号介质还可以是计算机可读存储介质以外的任何计算机可读介质，该计算机可读信号介质可以发送、传播或者传输用于由指令执行系统、装置或者器件使用或者与其结合使用的程序。计算机可读介质上包含的程序代码可以用任何适当的介质传输，包括但不限于：电线、光缆、RF(射频)等等，或者上述的任意合适的组合。
在一些实施方式中，客户端、服务器可以利用诸如HTTP(HyperText Transfer Protocol，超文本传输协议)之类的任何当前已知或未来研发的网络协议进行通信，并且可以与任意形式或介质的数字数据通信(例如，通信网络)互连。通信网络的示例包括局域网(“LAN”)，广域网(“WAN”)，网际网(例如，互联网)以及端对端网络(例如，ad hoc端对端网络)，以及任何当前已知或未来研发的网络。
上述计算机可读介质可以是上述电子设备中所包含的；也可以是单独存在，而未装配入该电子设备中。
上述计算机可读介质承载有一个或者多个程序，当上述一个或者多个程序被该电子设备执行时，使得该电子设备执行一种语音识别模型训练方法，所述语音识别模型通过主节点和多个工作节点对预设模型进行联合训练，所述方法包括：针对每一所述工作节点中的预设模型，获取所述预设模型在当前的迭代步骤对应的训练语音数据，其中，每一所述工作节点中的预设模型的模型参数相同；根据每一所述预设模型对应的训练语音数据，确定所述主节点中的预设模型在所述迭代步骤对应的目标梯度；确定所述迭代步骤对应的学习率，其中，在迭代过程中所述学习率与迭代次数成正相关关系增大至第一学习率，并从所述第一学习率每间隔预设迭代次数减小，直至第二学习率；根据所述学习率和所述目标梯度对所述主节点中的预设模型的模型参数进行更新。
或者，上述计算机可读介质承载有一个或者多个程序，当上述一个或者多个程序被该 电子设备执行时，使得该电子设备执行一种语音识别方法，所述方法包括：接收待识别的语音数据；将所述语音数据输入语音识别模型，获得所述语音数据对应的目标文本，其中，所述语音识别模型是通过本公开任意实施例所述的语音识别模型的训练方法进行训练所得。
可以以一种或多种程序设计语言或其组合来编写用于执行本公开的操作的计算机程序代码，上述程序设计语言包括但不限于面向对象的程序设计语言—诸如Java、Smalltalk、C++，还包括常规的过程式程序设计语言——诸如“C”语言或类似的程序设计语言。程序代码可以完全地在用户计算机上执行、部分地在用户计算机上执行、作为一个独立的软件包执行、部分在用户计算机上部分在远程计算机上执行、或者完全在远程计算机或服务器上执行。在涉及远程计算机的情形中，远程计算机可以通过任意种类的网络——包括局域网(LAN)或广域网(WAN)——连接到用户计算机，或者，可以连接到外部计算机(例如利用因特网服务提供商来通过因特网连接)。
附图中的流程图和框图，图示了按照本公开各种实施例的系统、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段、或代码的一部分，该模块、程序段、或代码的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。也应当注意，在有些作为替换的实现中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个接连地表示的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合，可以用执行规定的功能或操作的专用的基于硬件的系统来实现，或者可以用专用硬件与计算机指令的组合来实现。
描述于本公开实施例中所涉及到的模块可以通过软件的方式实现，也可以通过硬件的方式来实现。其中，模块的名称在某种情况下并不构成对该模块本身的限定，例如，获取模块还可以被描述为“针对每一所述工作节点中的预设模型，获取所述预设模型在当前的迭代步骤对应的训练语音数据的模块”。
本文中以上描述的功能可以至少部分地由一个或多个硬件逻辑部件来执行。例如，非限制性地，可以使用的示范类型的硬件逻辑部件包括：现场可编程门阵列(FPGA)、专用集成电路(ASIC)、专用标准产品(ASSP)、片上系统(SOC)、复杂可编程逻辑设备(CPLD)等等。
在本公开的上下文中，机器可读介质可以是有形的介质，其可以包含或存储以供指令执行系统、装置或设备使用或与指令执行系统、装置或设备结合地使用的程序。机器可读介质可以是机器可读信号介质或机器可读储存介质。机器可读介质可以包括但不限于电子 的、磁性的、光学的、电磁的、红外的、或半导体系统、装置或设备，或者上述内容的任何合适组合。机器可读存储介质的更具体示例会包括基于一个或多个线的电气连接、便携式计算机盘、硬盘、随机存取存储器(RAM)、只读存储器(ROM)、可擦除可编程只读存储器(EPROM或快闪存储器)、光纤、便捷式紧凑盘只读存储器(CD-ROM)、光学储存设备、磁储存设备、或上述内容的任何合适组合。
根据本公开的一个或多个实施例，示例1提供了一种语音识别模型训练方法，所述语音识别模型通过主节点和多个工作节点对预设模型进行联合训练，所述方法包括：
针对每一所述工作节点中的预设模型，获取所述预设模型在当前的迭代步骤对应的训练语音数据，其中，每一所述工作节点中的预设模型的模型参数相同；
根据每一所述预设模型对应的训练语音数据，确定所述主节点中的预设模型在所述迭代步骤对应的目标梯度；
确定所述迭代步骤对应的学习率，其中，在迭代过程中所述学习率与迭代次数成正相关关系增大至第一学习率，并从所述第一学习率每间隔预设迭代次数减小，直至第二学习率；
根据所述学习率和所述目标梯度对所述主节点中的预设模型的模型参数进行更新。
根据本公开的一个或多个实施例，示例2提供了示例1的方法，所述预设模型的迭代次数按照由小至大的顺序被划分为第一范围、第二范围和第三范围，且三个范围互不重合；
若所述迭代次数属于所述第一范围，所述学习率与所述迭代次数成正比例关系增大至所述第一学习率；若所述迭代次数属于所述第二范围，所述学习率为所述第一学习率；若所述迭代次数属于所述第三范围，所述第三范围以所述预设迭代次数间隔划分为多个子范围，每一子范围对应的学习率相同，当前的子范围对应的学习率按照预设衰减率进行减小获得下一子范围对应的学习率，直至下一子范围对应的学习率为所述第二学习率。
根据本公开的一个或多个实施例，示例3提供了示例2的方法，通过以下公式确定下一子范围对应的学习率：
Figure PCTCN2022089607-appb-000003
其中，L’用于表示下一子范围对应的学习率；
L用于表示当前的子范围对应的学习率，该当前的子范围中的迭代次数小于下一子范围的迭代次数；
d用于表示所述预设衰减率；
n用于表示所述预设迭代次数；
N用于表示训练过程对应的迭代总次数；
R用于表示所述第三范围对应的迭代次数占所述迭代总次数的比例。
根据本公开的一个或多个实施例，示例4提供了示例1的方法，所述训练语音数据被划分为多个批次的数据，每一所述工作节点中批次的数量相同；
所述根据每一所述预设模型对应的训练语音数据，确定所述主节点中的预设模型在所述迭代步骤对应的目标梯度，包括：
将所述多个工作节点中对应于同一批次的训练语音数据，分别输入该工作节点中的预设模型，并确定每一所述预设模型与所述批次的训练语音数据所对应的批次梯度；
根据所述多个工作节点中的每一批次的训练语音数据所对应的批次梯度，确定所述目标梯度。
根据本公开的一个或多个实施例，示例5提供了示例4的方法，所述根据所述多个工作节点中的每一批次的训练语音数据所对应的批次梯度，确定所述目标梯度，包括：
将所述多个工作节点中与同一批次对应的批次梯度的平均值，确定为所述主节点中所述预设模型与所述批次对应的梯度；
将所述预设模型与每一批次对应的梯度之和确定为所述目标梯度。
根据本公开的一个或多个实施例，示例6提供了一种语音识别方法，所述方法包括：
接收待识别的语音数据；
将所述语音数据输入语音识别模型，获得所述语音数据对应的目标文本，其中，所述语音识别模型是通过本公开任意实施例所述的语音识别模型的训练方法进行训练所得。
根据本公开的一个或多个实施例，示例7提供了一种语音识别模型训练装置，所述语音识别模型通过主节点和多个工作节点对预设模型进行联合训练，所述装置包括：
获取模块，用于针对每一所述工作节点中的预设模型，获取所述预设模型在当前的迭代步骤对应的训练语音数据，其中，每一所述工作节点中的预设模型的模型参数相同；
第一确定模块，用于根据每一所述预设模型对应的训练语音数据，确定所述主节点中的预设模型在所述迭代步骤对应的目标梯度；
第二确定模块，用于确定所述迭代步骤对应的学习率，其中，在迭代过程中所述学习率与迭代次数成正相关关系增大至第一学习率，并从所述第一学习率每间隔预设迭代次数减小，直至第二学习率；
更新模块，用于根据所述学习率和所述目标梯度对所述主节点中的预设模型的模型参数进行更新。
根据本公开的一个或多个实施例，示例8提供了一种语音识别装置，所述装置包括：
接收模块，用于接收待识别的语音数据；
输入模块，用于将所述语音数据输入语音识别模型，获得所述语音数据对应的目标文本，其中，所述语音识别模型是通过本公开任意实施例所述的语音识别模型的训练方法进 行训练所得。
根据本公开的一个或多个实施例，示例9提供了一种计算机可读介质，其上存储有计算机程序，该程序被处理装置执行时实现本公开任意实施例所述方法的步骤。
根据本公开的一个或多个实施例，示例10提供了一种电子设备，包括：
存储装置，其上存储有计算机程序；
处理装置，用于执行所述存储装置中的所述计算机程序，以实现本公开任意实施例所述方法的步骤。
以上描述仅为本公开的较佳实施例以及对所运用技术原理的说明。本领域技术人员应当理解，本公开中所涉及的公开范围，并不限于上述技术特征的特定组合而成的技术方案，同时也应涵盖在不脱离上述公开构思的情况下，由上述技术特征或其等同特征进行任意组合而形成的其它技术方案。例如上述特征与本公开中公开的(但不限于)具有类似功能的技术特征进行互相替换而形成的技术方案。
此外，虽然采用特定次序描绘了各操作，但是这不应当理解为要求这些操作以所示出的特定次序或以顺序次序执行来执行。在一定环境下，多任务和并行处理可能是有利的。同样地，虽然在上面论述中包含了若干具体实现细节，但是这些不应当被解释为对本公开的范围的限制。在单独的实施例的上下文中描述的某些特征还可以组合地实现在单个实施例中。相反地，在单个实施例的上下文中描述的各种特征也可以单独地或以任何合适的子组合的方式实现在多个实施例中。
尽管已经采用特定于结构特征和/或方法逻辑动作的语言描述了本主题，但是应当理解所附权利要求书中所限定的主题未必局限于上面描述的特定特征或动作。相反，上面所描述的特定特征和动作仅仅是实现权利要求书的示例形式。关于上述实施例中的装置，其中各个模块执行操作的具体方式已经在有关该方法的实施例中进行了详细描述，此处将不做详细阐述说明。

Claims (10)

  1. 一种语音识别模型训练方法，其特征在于，所述语音识别模型通过主节点和多个工作节点对预设模型进行联合训练，所述方法包括：
    针对每一所述工作节点中的预设模型，获取所述预设模型在当前的迭代步骤对应的训练语音数据，其中，每一所述工作节点中的预设模型的模型参数相同；
    根据每一所述预设模型对应的训练语音数据，确定所述主节点中的预设模型在所述迭代步骤对应的目标梯度；
    确定所述迭代步骤对应的学习率，其中，在迭代过程中所述学习率与迭代次数成正相关关系增大至第一学习率，并从所述第一学习率每间隔预设迭代次数减小，直至第二学习率；
    根据所述学习率和所述目标梯度对所述主节点中的预设模型的模型参数进行更新。
  2. 根据权利要求1所述的方法，其特征在于，所述预设模型的迭代次数按照由小至大的顺序被划分为第一范围、第二范围和第三范围，且三个范围互不重合；
    若所述迭代次数属于所述第一范围，所述学习率与所述迭代次数成正比例关系增大至所述第一学习率；若所述迭代次数属于所述第二范围，所述学习率为所述第一学习率；若所述迭代次数属于所述第三范围，所述第三范围以所述预设迭代次数间隔划分为多个子范围，每一子范围对应的学习率相同，当前的子范围对应的学习率按照预设衰减率进行减小获得下一子范围对应的学习率，直至下一子范围对应的学习率为所述第二学习率。
  3. 根据权利要求2所述的方法，其特征在于，通过以下公式确定下一子范围对应的学习率：
    Figure PCTCN2022089607-appb-100001
    其中，L’用于表示下一子范围对应的学习率；
    L用于表示当前的子范围对应的学习率，该当前的子范围中的迭代次数小于下一子范围的迭代次数；
    d用于表示所述预设衰减率；
    n用于表示所述预设迭代次数；
    N用于表示训练过程对应的迭代总次数；
    R用于表示所述第三范围对应的迭代次数占所述迭代总次数的比例。
  4. 根据权利要求1所述的方法，其特征在于，所述训练语音数据被划分为多个批次的数据，每一所述工作节点中批次的数量相同；
    所述根据每一所述预设模型对应的训练语音数据，确定所述主节点中的预设模型在所述迭代步骤对应的目标梯度，包括：
    将所述多个工作节点中对应于同一批次的训练语音数据，分别输入该工作节点中的预设模型，并确定每一所述预设模型与所述批次的训练语音数据所对应的批次梯度；
    根据所述多个工作节点中的每一批次的训练语音数据所对应的批次梯度，确定所述目标梯度。
  5. 根据权利要求4所述的方法，其特征在于，所述根据所述多个工作节点中的每一批次的训练语音数据所对应的批次梯度，确定所述目标梯度，包括：
    将所述多个工作节点中与同一批次对应的批次梯度的平均值，确定为所述主节点中所述预设模型与所述批次对应的梯度；
    将所述预设模型与每一批次对应的梯度之和确定为所述目标梯度。
  6. 一种语音识别方法，其特征在于，所述方法包括：
    接收待识别的语音数据；
    将所述语音数据输入语音识别模型，获得所述语音数据对应的目标文本，其中，所述语音识别模型是通过权利要求1-5中任一项所述的语音识别模型的训练方法进行训练所得。
  7. 一种语音识别模型训练装置，其特征在于，所述语音识别模型通过主节点和多个工作节点对预设模型进行联合训练，所述装置包括：
    获取模块，用于针对每一所述工作节点中的预设模型，获取所述预设模型在当前的迭代步骤对应的训练语音数据，其中，每一所述工作节点中的预设模型的模型参数相同；
    第一确定模块，用于根据每一所述预设模型对应的训练语音数据，确定所述主节点中的预设模型在所述迭代步骤对应的目标梯度；
    第二确定模块，用于确定所述迭代步骤对应的学习率，其中，在迭代过程中所述学习率与迭代次数成正相关关系增大至第一学习率，并从所述第一学习率每间隔预设迭代次数减小，直至第二学习率；
    更新模块，用于根据所述学习率和所述目标梯度对所述主节点中的预设模型的模型参数进行更新。
  8. 一种语音识别装置，其特征在于，所述装置包括：
    接收模块，用于接收待识别的语音数据；
    输入模块，用于将所述语音数据输入语音识别模型，获得所述语音数据对应的目标文本，其中，所述语音识别模型是通过权利要求1-5中任一项所述的语音识别模型的训练方法进行训练所得。
  9. 一种计算机可读介质，其上存储有计算机程序，其特征在于，该程序被处理装置执行时实现权利要求1-6中任一项所述方法的步骤。
  10. 一种电子设备，其特征在于，包括：
    存储装置，其上存储有计算机程序；
    处理装置，用于执行所述存储装置中的所述计算机程序，以实现权利要求1-6中任一项所述方法的步骤。
PCT/CN2022/089607 2021-06-30 2022-04-27 模型的训练方法、语音识别方法、装置、介质及设备 WO2023273579A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110736548.2 2021-06-30
CN202110736548.2A CN113327598B (zh) 2021-06-30 2021-06-30 模型的训练方法、语音识别方法、装置、介质及设备

Publications (1)

Publication Number Publication Date
WO2023273579A1 true WO2023273579A1 (zh) 2023-01-05

Family

ID=77423535

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/089607 WO2023273579A1 (zh) 2021-06-30 2022-04-27 模型的训练方法、语音识别方法、装置、介质及设备

Country Status (2)

Country Link
CN (1) CN113327598B (zh)
WO (1) WO2023273579A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116663639A (zh) * 2023-07-31 2023-08-29 浪潮电子信息产业股份有限公司 一种梯度数据同步方法、系统、装置及介质
CN116755941A (zh) * 2023-08-21 2023-09-15 之江实验室 一种模型训练的方法、装置、存储介质及电子设备

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113327598B (zh) * 2021-06-30 2023-11-14 北京有竹居网络技术有限公司 模型的训练方法、语音识别方法、装置、介质及设备
CN114267328A (zh) * 2021-12-14 2022-04-01 北京达佳互联信息技术有限公司 一种语音合成模型的训练方法、装置、设备以及存储介质
CN114841341B (zh) * 2022-04-25 2023-04-28 北京百度网讯科技有限公司 图像处理模型训练及图像处理方法、装置、设备和介质
CN115188371A (zh) * 2022-07-13 2022-10-14 合肥讯飞数码科技有限公司 一种语音识别模型训练方法、语音识别方法及相关设备
CN116013256B (zh) * 2022-12-19 2024-01-30 镁佳(北京)科技有限公司 一种语音识别模型构建及语音识别方法、装置及存储介质

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150019214A1 (en) * 2013-07-10 2015-01-15 Tencent Technology (Shenzhen) Company Limited Method and device for parallel processing in model training
CN104463324A (zh) * 2014-11-21 2015-03-25 长沙马沙电子科技有限公司 一种基于大规模高性能集群的卷积神经网络并行处理方法
CN110379416A (zh) * 2019-08-15 2019-10-25 腾讯科技（深圳）有限公司 一种神经网络语言模型训练方法、装置、设备及存储介质
CN110705705A (zh) * 2019-09-25 2020-01-17 浪潮电子信息产业股份有限公司 卷积神经网络模型同步训练方法、集群及可读存储介质
CN111052155A (zh) * 2017-09-04 2020-04-21 华为技术有限公司 异步梯度平均的分布式随机梯度下降法
CN111382844A (zh) * 2020-03-11 2020-07-07 华南师范大学 一种深度学习模型的训练方法及装置
CN112529029A (zh) * 2019-09-18 2021-03-19 华为技术有限公司 信息处理方法、神经网络的训练方法、装置及存储介质
US20210089887A1 (en) * 2019-09-24 2021-03-25 Apple Inc. Variance-Based Learning Rate Control For Training Machine-Learning Models
CN113327598A (zh) * 2021-06-30 2021-08-31 北京有竹居网络技术有限公司 模型的训练方法、语音识别方法、装置、介质及设备

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9824692B1 (en) * 2016-09-12 2017-11-21 Pindrop Security, Inc. End-to-end speaker recognition using deep neural network
CN110135573B (zh) * 2018-02-02 2023-10-03 阿里巴巴集团控股有限公司 一种深度学习模型的训练方法、计算设备以及系统
CN111160531B (zh) * 2019-12-30 2023-09-22 北京迈格威科技有限公司 神经网络模型的分布式训练方法、装置及电子设备
CN111354345B (zh) * 2020-03-11 2021-08-31 北京字节跳动网络技术有限公司 生成语音模型和语音识别的方法、装置、设备以及介质
CN112560327B (zh) * 2020-11-13 2024-03-12 盐城工学院 一种基于深度梯度下降森林的轴承剩余寿命预测方法
CN112435656B (zh) * 2020-12-11 2024-03-01 平安科技（深圳）有限公司 模型训练方法、语音识别方法、装置、设备及存储介质
CN113032117A (zh) * 2021-03-10 2021-06-25 北京百度网讯科技有限公司 深度学习框架的训练方法、装置及存储介质

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150019214A1 (en) * 2013-07-10 2015-01-15 Tencent Technology (Shenzhen) Company Limited Method and device for parallel processing in model training
CN104463324A (zh) * 2014-11-21 2015-03-25 长沙马沙电子科技有限公司 一种基于大规模高性能集群的卷积神经网络并行处理方法
CN111052155A (zh) * 2017-09-04 2020-04-21 华为技术有限公司 异步梯度平均的分布式随机梯度下降法
CN110379416A (zh) * 2019-08-15 2019-10-25 腾讯科技（深圳）有限公司 一种神经网络语言模型训练方法、装置、设备及存储介质
CN112529029A (zh) * 2019-09-18 2021-03-19 华为技术有限公司 信息处理方法、神经网络的训练方法、装置及存储介质
US20210089887A1 (en) * 2019-09-24 2021-03-25 Apple Inc. Variance-Based Learning Rate Control For Training Machine-Learning Models
CN110705705A (zh) * 2019-09-25 2020-01-17 浪潮电子信息产业股份有限公司 卷积神经网络模型同步训练方法、集群及可读存储介质
CN111382844A (zh) * 2020-03-11 2020-07-07 华南师范大学 一种深度学习模型的训练方法及装置
CN113327598A (zh) * 2021-06-30 2021-08-31 北京有竹居网络技术有限公司 模型的训练方法、语音识别方法、装置、介质及设备

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116663639A (zh) * 2023-07-31 2023-08-29 浪潮电子信息产业股份有限公司 一种梯度数据同步方法、系统、装置及介质
CN116663639B (zh) * 2023-07-31 2023-11-03 浪潮电子信息产业股份有限公司 一种梯度数据同步方法、系统、装置及介质
CN116755941A (zh) * 2023-08-21 2023-09-15 之江实验室 一种模型训练的方法、装置、存储介质及电子设备
CN116755941B (zh) * 2023-08-21 2024-01-09 之江实验室 一种节点故障感知的分布式模型训练的方法及装置

Also Published As

Publication number Publication date
CN113327598B (zh) 2023-11-14
CN113327598A (zh) 2021-08-31

Similar Documents

Publication Publication Date Title
WO2023273579A1 (zh) 模型的训练方法、语音识别方法、装置、介质及设备
WO2022121801A1 (zh) 信息处理方法、装置和电子设备
WO2023273985A1 (zh) 一种语音识别模型的训练方法、装置及设备
WO2022127620A1 (zh) 语音唤醒方法、装置、电子设备及存储介质
WO2023143016A1 (zh) 特征提取模型的生成方法、图像特征提取方法和装置
CN111597825B (zh) 语音翻译方法、装置、可读介质及电子设备
CN111858381B (zh) 应用程序容错能力测试方法、电子设备及介质
WO2023029850A1 (zh) 一种数据处理方法、装置、电子设备和介质
WO2023103653A1 (zh) 键值匹配方法、装置、可读介质及电子设备
WO2022250609A1 (zh) 数据保护方法、网络结构训练方法、装置、介质及设备
WO2022228067A1 (zh) 语音处理方法、装置和电子设备
WO2022116819A1 (zh) 模型训练方法及装置、机器翻译方法及装置、设备、存储介质
WO2023014298A2 (zh) 神经网络模型的构建方法及装置
CN112464039B (zh) 树形结构的数据显示方法、装置、电子设备和介质
CN113191257B (zh) 笔顺检测方法、装置和电子设备
CN113392018A (zh) 流量分发方法、装置、存储介质及电子设备
CN113220281A (zh) 一种信息生成方法、装置、终端设备及存储介质
WO2023155713A1 (zh) 用于标记说话人的方法、装置和电子设备
WO2023045870A1 (zh) 网络模型的压缩方法、图像生成方法、装置、设备及介质
WO2023130925A1 (zh) 字体识别方法、装置、可读介质及电子设备
WO2022134968A1 (zh) 模型的训练方法、语音识别方法、装置、介质及设备
WO2023011397A1 (zh) 一种生成声学特征、语音模型训练、语音识别方法及装置
WO2023000782A1 (zh) 获取视频热点的方法、装置、可读介质和电子设备
WO2022121800A1 (zh) 声源定位方法、装置和电子设备
CN111680754B (zh) 图像分类方法、装置、电子设备及计算机可读存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22831407

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 22831407

Country of ref document: EP

Kind code of ref document: A1