WO2023165538A1 - 语音识别方法、装置、计算机可读介质及电子设备 - Google Patents

语音识别方法、装置、计算机可读介质及电子设备 Download PDF

Info

Publication number
WO2023165538A1
WO2023165538A1 PCT/CN2023/079156 CN2023079156W WO2023165538A1 WO 2023165538 A1 WO2023165538 A1 WO 2023165538A1 CN 2023079156 W CN2023079156 W CN 2023079156W WO 2023165538 A1 WO2023165538 A1 WO 2023165538A1
Authority
WO
WIPO (PCT)
Prior art keywords
multilingual
training
training model
parameters
language
Prior art date
Application number
PCT/CN2023/079156
Other languages
English (en)
French (fr)
Inventor
卢怡宙
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 WO2023165538A1 publication Critical patent/WO2023165538A1/zh

Links

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

Definitions

  • the present disclosure relates to the field of computer technology, and in particular, relates to a voice recognition method, a voice recognition device, a computer-readable medium, electronic equipment, a computer program product, and a computer program.
  • the current cross-lingual representation learning method does not take into account the diversity of pronunciation between different languages, and still uses the model structure of monolingual representation learning.
  • the model does not have a module dedicated to modeling specific language characteristics, so it often faces language problems. problems of mutual interference. This problem will become more serious when the number of languages increases and unsupervised data increases.
  • this multilingual pre-training model is used for downstream multilingual speech recognition tasks, the recognition performance of large languages (such as Chinese and English) will drop significantly. , compared with the single-language pre-training model, there is a big gap.
  • the present disclosure provides a speech recognition method, including: acquiring a target speech signal to be recognized including multiple languages; recognizing the semantics of the target speech signal through a speech recognition model that integrates sparse sub-networks of various languages; The sparse sub-network is obtained by performing parameter pruning on a multilingual pre-training model, and the multilingual pre-training model is obtained by training according to speech signals containing the multiple languages.
  • the present disclosure provides a speech recognition device, including: an acquisition module, used to acquire a target speech signal to be recognized including multiple languages; a recognition module, used for speech recognition through fusion of sparse sub-networks of various languages mold type to recognize the semantics of the target speech signal; the sparse sub-network is obtained by performing parameter pruning on the multilingual pre-training model, and the multilingual pre-training model is obtained according to the speech signal training containing the multiple languages of.
  • the present disclosure provides a computer-readable medium on which a computer program is stored, and when the computer program is executed by a processor, the steps of the aforementioned speech recognition method are realized.
  • the present disclosure provides an electronic device, including: a memory, on which a computer program is stored; and a processor, configured to execute the computer program in the memory, so as to implement the steps of the aforementioned speech recognition method.
  • the embodiments of the present disclosure further provide a computer program product, the computer program product includes a computer program carried on a computer-readable medium, and the program code included in the computer program can be used to implement the steps in the first aspect above .
  • the embodiments of the present disclosure further provide a computer program, which implements the steps in the first aspect when executed by a processor.
  • the target speech signal to be recognized including multiple languages is obtained, and the semantics of the target speech signal is recognized by a speech recognition model that integrates sparse sub-networks of various languages.
  • the sparse sub-network is a multilingual pre-training model.
  • the multilingual pre-training model is obtained by parameter pruning processing based on the training of speech signals containing multiple languages.
  • Fig. 1 is a schematic structural diagram of a computer system provided by an exemplary embodiment of the present disclosure.
  • Fig. 2 is a flowchart of a speech recognition method provided by an exemplary embodiment of the present disclosure.
  • Fig. 3 is a flowchart of a method for training a speech recognition model provided by an exemplary embodiment of the present disclosure.
  • Fig. 4 is a flow chart of the sub-steps of step S202 provided by an exemplary embodiment of the present disclosure.
  • Fig. 5 is a block diagram of a speech recognition device provided by an exemplary embodiment of the present disclosure.
  • Fig. 6 is a schematic structural diagram of an electronic device provided by an exemplary 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.”
  • FIG. 1 shows a schematic structural diagram of a computer system provided by an exemplary embodiment of the present disclosure.
  • the computer system includes a terminal 120 and a server 140 .
  • the terminal 120 and the server 140 are connected to each other through a wired or wireless network.
  • the terminal 120 may include at least one of a smart phone, a notebook computer, a desktop computer, a tablet computer, a smart speaker, and a smart robot.
  • Terminal 120 includes a display; the display may be used to display speech recognition results.
  • Terminal 120 includes a first memory and a first processor.
  • a first program is stored in the first memory; the above-mentioned first program is invoked and executed by the first processor to implement the voice recognition method provided by the present disclosure.
  • the first memory may include but not limited to the following: Random Access Memory (Random Access Memory, RAM), Read Only Memory (Read Only Memory, ROM), Programmable Read-Only Memory (Programmable Read-Only Memory, PROM), Erasable Programmable Read-Only Memory (EPROM), and Electric Erasable Programmable Read-Only Memory (EEPROM).
  • the first processor may be composed of one or more integrated circuit chips.
  • the first processor may be a general processor, such as a central processing unit (Central Processing Unit, CPU) or a network processor (Network Processor, NP).
  • the speech recognition model in the terminal may be trained by the terminal; or, trained by the server, and the terminal obtains it from the server.
  • Server 140 includes a second memory and a second processor.
  • a second program is stored in the second memory, and the above-mentioned second program is invoked by the second processor to implement the speech recognition method provided by the present disclosure.
  • the speech recognition model is stored in the second memory, and the speech recognition model is invoked by the second processor to implement the speech recognition method.
  • the second memory may include but not limited to the following: RAM, ROM, PROM, EPROM, EEPROM.
  • the second processor may be a general processor, such as CPU or NP.
  • the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or it can provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, intermediate Cloud servers for basic cloud computing services such as mail service, domain name service, security service, content delivery network (Content Delivery Network, CDN), and big data and artificial intelligence platforms.
  • the terminal may be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited thereto.
  • the terminal and the server may be connected directly or indirectly through wired or wireless communication, which is not limited in the present disclosure.
  • model pruning In order to improve the efficiency of pre-trained language models, various model compression methods have been proposed, including model pruning.
  • the present disclosure provides a voice recognition method provided in an exemplary embodiment, including: acquiring a target voice signal to be recognized that includes multiple languages, and recognizing the target voice signal through a voice recognition model that integrates sparse sub-networks of various languages Semantics; the sparse sub-network is obtained by pruning the parameters of the multilingual pre-training model.
  • the multilingual pre-training model is obtained from the training of speech signals containing multiple languages.
  • the multiple languages include some large languages, such as Chinese, English, etc., as well as some minor languages, such as French, Spanish, etc. This disclosure solves the problem of language interference in cross-lingual representation learning from the perspective of self-adaptation.
  • the entire multilingual pre-training model performs parameter pruning for different languages, and constructs a set of sparse sub-networks that share some parameters for training, thus endowing
  • the ability of the speech recognition model to specifically model different languages has been greatly improved in the process of cross-lingual representation learning for both large and small languages.
  • FIG. 2 is a flowchart of a voice recognition method provided by an exemplary embodiment of the present disclosure.
  • the method is executed by an electronic device, for example, by a terminal or a server in the computer system shown in FIG. 1 .
  • the speech recognition method shown in Fig. 2 comprises the following steps:
  • step S101 target speech signals to be recognized including multiple languages are acquired.
  • the multiple languages include some major languages, such as Chinese and English, and some small languages, such as French and Spanish.
  • step S102 the semantics of the target speech signal is recognized by a speech recognition model fused with sparse sub-networks of various languages.
  • the sparse sub-network is obtained by performing parameter pruning on a multilingual pre-training model trained on speech signals containing multiple languages.
  • FIG. 3 is a flowchart of a method for training a speech recognition model provided by an exemplary embodiment of the present disclosure.
  • the training method of this speech recognition model comprises the following steps:
  • step S201 speech signals containing multiple languages are acquired as training samples.
  • the multiple languages include some major languages, such as Chinese and English, and some small languages, such as French and Spanish.
  • the number of samples in large languages is far greater than the number of samples in small languages.
  • the number of samples in large languages may be 1 million, while the number of samples in small languages may only be tens of thousands.
  • the training samples are unsupervised speech data, that is, speech data without manual annotation.
  • step S202 a multilingual pre-training model is obtained through training according to training samples.
  • the multilingual pre-training model is used for speech recognition, and the multilingual pre-training model can recognize different languages respectively.
  • the Wav2vec 2.0 framework is used for cross-language speech representation learning, which mainly includes a feature extractor, a context network and a quantization module.
  • the feature extractor is composed of a multi-layer convolutional neural network
  • the context network is composed of a multi-layer Transformer layer, which is used for semantic learning of the speech signal output by the feature extractor, and outputs a representation vector with context information
  • the quantization module uses To provide the original unsupervised speech data for comparative learning, and quantify the speech signal output by the feature extractor.
  • additional diversity loss is added to promote the use of quantization modules in multilingual pre-training models and avoid the collapse of quantization modules.
  • step S202 includes sub-step S2021, sub-step S2022, sub-step S2023, sub-step S2024 and sub-step S2025, and the specific way to train the multilingual pre-training model will be described in detail in the sub-steps of step S202. Please refer to FIG. 4 .
  • FIG. 4 is a flowchart of substeps of step S202 shown in an exemplary embodiment of the present disclosure.
  • the speech signal is converted into a plurality of low-dimensional signal frames.
  • the speech signal before converting the speech signal into multiple low-dimensional signal frames, it is necessary to upsample the training samples of the language whose number of samples is lower than the first threshold, so as to expand the language samples lower than the first threshold in the sampled data.
  • the number of training samples for example, the number of samples in a small language may only be tens of thousands, so it is necessary to upsample the training samples in a small language.
  • Uniform sampling is performed on languages whose number of samples is higher than the second threshold. For example, the number of samples of a large language may be 1 million. At this time, it is only necessary to uniformly sample them to obtain sufficient sampling data.
  • each signal frame is a speech representation signal of fixed duration.
  • each signal frame may be about 25 ms long with a step of 20 ms.
  • any one frame among the plurality of signal frames is masked to obtain a masked speech signal.
  • any one or two frames in the 10 signal frames can be masked out. If a speech signal is divided into 100 frames in total, 100 frames can be masked out. Any 10 or 20 frames in the signal frame, and then get the masked speech signal.
  • sub-step S2023 input the masked speech signal into the initial multilingual pre-training model for semantic learning, so as to predict the masked signal frame.
  • the quantization module receives the feature extractor The output speech signal is unmasked, so the predicted masked signal frame can be compared with the unsupervised speech data provided by the quantization module.
  • sub-step S2024 when the predicted masked signal frame is consistent with the actual masked signal frame, it is determined that the prediction is correct and the parameters of the initial multilingual pre-trained model are updated.
  • the quantization module When the predicted masked signal frame is consistent with the actual masked signal frame output by the quantization module, it is determined that the signal frame predicted by the context network is correct, and the parameters of the initial multilingual pre-training model are updated.
  • sub-step S2025 the step of updating the parameters of the initial multilingual pre-training model is repeated to obtain a multilingual pre-training model.
  • the masked signal frame is not repeated each time, and then perform semantic learning on the masked speech signal through the context network to predict the masked signal frame, and predict the masked signal frame and
  • the actual masked signal frame output by the quantization module is compared, and when the comparison is consistent, it is determined that the signal frame predicted by the context network is correct, and the parameters of the initial multilingual pre-training model are updated to obtain a multilingual pre-training model.
  • substeps S2022-S2024 can be repeatedly executed for a predetermined number of times, which can be a predetermined ratio of the number of signal frames of the voice signal, such as 50% of the frames of 10 signals, which is 5 times, and the predetermined number of times can be based on artificial The experience is obtained, or obtained according to other feasible methods, so I won’t repeat them here.
  • a predetermined number of times which can be a predetermined ratio of the number of signal frames of the voice signal, such as 50% of the frames of 10 signals, which is 5 times, and the predetermined number of times can be based on artificial The experience is obtained, or obtained according to other feasible methods, so I won’t repeat them here.
  • step S203 parameter pruning is performed on the multilingual pre-trained model corresponding to multiple languages to obtain a sparse sub-network corresponding to each language.
  • the parameter pruning process mainly includes the following two methods, the method based on the lottery hypothesis and the method based on Taylor expansion. Any one of the pruning methods may be used in the present disclosure, and the two pruning methods will be described in detail below.
  • the step of performing parameter pruning on the multilingual pre-training model corresponding to the multiple languages based on the lottery ticket hypothesis method includes: using the speech signals of each language as a training sample to train the multilingual pre-training model respectively, where the multilingual
  • the pre-training model refers to the multilingual pre-training model trained in step S202, and the training convergence condition here is also the same as in step S202; then obtain all the parameters of the multilingual pre-training model corresponding to each language; form parameters according to these parameters Matrix; then construct a masking matrix corresponding to the parameter matrix, the length and width of the masking matrix are consistent with the parameter matrix; then obtain the absolute value of each parameter in the parameter matrix; cut the parameters of a predetermined ratio according to the size of the absolute value, for example, According to the size of the absolute value, the parameters of the predetermined ratio are clipped from small to large, and the predetermined ratio can be 20%, 30% or 50%, etc., and there is no limit here; finally, the masking state of the corresponding position of the clipped parameter in the
  • the steps of performing parameter pruning on the multilingual pre-training model corresponding to multiple languages based on the Taylor expansion method include: using speech signals of each language as training samples to train the multilingual pre-training model respectively.
  • the multilingual pre-training model here is Refers to the multilingual pre-training model trained in step S202.
  • the training convergence condition is the same as in step S202; all parameters of the multilingual pre-training model corresponding to each language are obtained;
  • the loss value caused to the multilingual pre-training model after each parameter is clipped, in one embodiment, the formula for predicting the loss value caused to the multilingual pre-training model after each parameter is clipped includes:
  • g is the gradient of the parameter
  • w is the weight of the parameter
  • is an operator for taking the absolute value; then cut the parameters of a predetermined proportion according to the size of the loss value, for example, it can be cut according to the size of the loss value from small to large
  • the predetermined ratio may be 20%, 30% or 50%, etc., which is not limited here.
  • the multilingual pre-training model obtained after parameter pruning can speed up inference and calculation, meet the minimum delay limit, and reduce memory consumption. It is easier to deploy on the terminal side, such as mobile phones, and facilitates model training and fine-tuning.
  • the speech signals of each language are used as training samples, and the multilingual pre-training models are trained separately to obtain the sparse sub-network corresponding to each language.
  • step S204 multilingual adaptive pre-training is performed on each sparse sub-network with the corresponding language to update the parameters of each sparse sub-network, so as to obtain shared parameters and exclusive parameters between each sparse sub-network.
  • each training batch is only composed of training samples in one language, where batch refers to: use a small part of the training samples to perform a backpropagation parameter update on the model weight, this small part A sample is called a "batch of data".
  • the sparse subnetwork corresponding to this language is used for forward propagation, and the loss of the sparse subnetwork is calculated, and only the parameters corresponding to the sparse subnetwork are updated during backpropagation.
  • the final sparse sub-network automatically assigns shared parameters and exclusive parameters between different languages within the network, so as to achieve the effect of adaptive training.
  • step S205 the speech recognition model of the sparse sub-network integrating various languages is obtained according to the shared parameters and the exclusive parameters.
  • the speech recognition model based on the sparse shared subnetwork cross-language speech representation proposed by the present disclosure can surpass the baseline cross-language speech representation learning method in both large and small languages.
  • the proposed speech recognition method On the public shared voice (Common Voice) data set, the proposed speech recognition method has a relative 9.8% average phoneme error rate reduction in the 100M model compared to the baseline system, and a 7.4% phoneme error rate reduction in the 300M model. Moreover, this method can greatly alleviate the language interference problem suffered by large languages. On the 100M model and the 300M model, the relative phoneme error rates of large languages are reduced by 17.8% and 16.7%.
  • the speech recognition method includes acquiring the target speech signal to be recognized including multiple languages, and recognizing the semantics of the target speech signal through a speech recognition model that integrates sparse sub-networks of various languages, the sparse sub-network
  • the network is obtained by performing parameter pruning on a multilingual pre-training model trained on speech signals containing multiple languages.
  • the present disclosure addresses cross-lingual representation from an adaptive perspective For the problem of language interference in learning, the entire multilingual pre-training model is pruned separately for different languages, and a set of sparse sub-networks sharing some parameters are constructed for training, thus endowing the speech recognition model with specific modeling for different languages. In the process of cross-lingual representation learning, both large and small languages have been greatly improved.
  • Fig. 5 is a block diagram of a speech recognition device according to an exemplary embodiment of the present disclosure.
  • the device 20 includes an acquisition module 201 and an identification module 203 .
  • the acquiring module 201 is configured to acquire target speech signals to be recognized that include multiple languages;
  • the recognition module 203 is used to recognize the semantics of the target speech signal by fusing the speech recognition models of the sparse sub-networks of various languages; the sparse sub-network is obtained by performing parameter pruning on the multilingual pre-training model, so The multilingual pre-training model is trained according to speech signals containing the multiple languages.
  • the device 20 also includes a processing module 205 .
  • the processing module 205 is configured to acquire speech signals containing the multiple languages as training samples;
  • the multilingual pre-training model is used for speech recognition
  • the processing module 205 is also configured to convert the speech signal into a plurality of low-dimensional signal frames; the signal frames are speech representation signals of fixed duration;
  • the processing module 205 is further configured to upsample the training samples of the language whose number of samples is lower than the first threshold, so as to expand the number of training samples of the language lower than the first threshold in the sampled data;
  • Uniform sampling is performed on languages whose sample size is higher than the second threshold.
  • the processing module 205 is further configured to perform parameter pruning processing on the multilingual pre-training model corresponding to the multiple languages based on the lottery ticket assumption, to obtain a sparse sub-network corresponding to each language;
  • processing module 205 is further configured to use the speech signals of various languages as training samples to train the multilingual pre-training models respectively;
  • processing module 205 is further configured to use the speech signals of various languages as training samples to train the multilingual pre-training models respectively;
  • FIG. 6 it shows a schematic structural diagram of an electronic device (such as the terminal device or server in FIG. 1 ) 600 suitable for implementing the embodiments 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, personal digital assistant (Personal Digital Assistant, PDA), PAD (tablet computer), portable multimedia player (Portable Multimedia Player , PMP), mobile terminals such as vehicle-mounted terminals (such as vehicle-mounted navigation terminals), and fixed terminals such as digital TVs, desktop computers, and the like.
  • PDA Personal Digital Assistant
  • PAD tablet computer
  • PMP portable multimedia player
  • mobile terminals such as vehicle-mounted terminals (such as vehicle-mounted navigation terminals)
  • 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 processor (such as a central processing unit, a graphics processing unit, etc.) (RAM) 603 to execute various appropriate actions and processing.
  • a processor such as a central processing unit, a graphics processing unit, etc.
  • RAM random access memory
  • various programs and data necessary for the operation of the electronic device 600 are also stored.
  • the processor 601, ROM 602, and RAM 603 are connected to each other through a bus 604.
  • An input/output (Input/Output, I/O) interface 605 is also connected to the bus 604 .
  • an input device 606 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; including, for example, a liquid crystal display (Liquid Crystal Display, LCD) , an output device 607 such as a speaker, a vibrator, etc.; a memory 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 memory 608, or from ROM 602.
  • the processor 601 When the computer program is executed by the processor 601, the above functions defined in the methods of the embodiments of the present disclosure are executed.
  • the computer-readable medium mentioned above in the present disclosure may be a computer-readable signal medium or a computer-readable storage medium or any combination of the two.
  • a computer readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, 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 (Compact Disc 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. This propagated data signal can take many forms, Including but not limited to electromagnetic signals, optical signals or any suitable combination of the above.
  • a computer-readable signal medium may also be any computer-readable storage medium other than a computer-readable storage medium that can transmit, propagate, or transmit information for use by or in connection with an instruction execution system, apparatus, or device. program.
  • the program code contained on the computer readable medium can be transmitted by any appropriate medium, including but not limited to: electric wire, optical cable, radio frequency (Radio Frequency, RF), etc., or any suitable combination of the above.
  • the terminal and the server can communicate using any currently known or future developed network protocols such as Hypertext Transfer Protocol (HyperText Transfer Protocol, HTTP), and can communicate with digital data in any form or medium (eg, communication network) interconnections.
  • network protocols such as Hypertext Transfer Protocol (HyperText Transfer Protocol, HTTP)
  • Examples of communication networks include local area networks (Local Area Network, LAN), wide area networks (Wide Area Network, WAN), internetworks (e.g., the Internet), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently existing networks that are known or developed in the future.
  • 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: acquires target speech signals to be recognized including multiple languages;
  • the speech recognition model of the sparse sub-network recognizes the semantics of the target speech signal;
  • the sparse sub-network is obtained by performing parameter pruning on the multilingual pre-training model, and the multilingual pre-training model is based on the trained with speech signals from different languages.
  • 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.
  • 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.
  • exemplary types of hardware logic components include: Field Programmable Gate Arrays (Field Programmable Gate Arrays, FPGAs), Application Specific Integrated Circuits (ASICs), Application Specific Standard Products (Application Specific Standard Product, ASSP), System on a Chip (System on a Chip, SOC), Complex Programmable Logic Device (Complex Programmable Logic Device, CPLD) and so on.
  • a machine-readable storage 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 storage medium may be a machine-readable signal medium or a machine-readable storage medium.
  • a machine-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, 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 voice recognition method, including: acquiring a target voice signal to be recognized that includes multiple languages;
  • the sparse sub-network is obtained by performing parameter pruning on a multilingual pre-training model, and the multilingual pre-training model is obtained by training according to speech signals containing the multiple languages.
  • Example 2 provides the method of Example 1, and the training method of the speech recognition model includes the following steps:
  • the multilingual pre-training model is used for speech recognition
  • Example 3 provides the method of Example 2, and the step of obtaining the multilingual pre-trained model according to the training sample training includes:
  • the signal frame is a speech representation signal of fixed duration
  • Example 4 provides the method of Example 3, and the step of obtaining a multilingual pre-trained model according to the training sample training further includes:
  • Uniform sampling is performed on languages whose sample size is higher than the second threshold.
  • Example 5 provides the method of Example 2, performing parameter pruning processing on the multilingual pre-training model corresponding to the multiple languages to obtain the sparse subclasses corresponding to each language
  • the steps of the network include:
  • Example 6 provides the method of Example 5, and the step of performing parameter pruning processing on the multilingual pre-training model corresponding to the multiple languages based on the lottery ticket assumption method includes:
  • Example 7 provides the method of Example 5, and the step of performing parameter pruning on the multilingual pre-training model corresponding to the multiple languages based on the Taylor expansion method includes:
  • Example 8 provides the method of Example 7, and the formula for predicting the loss value of the multilingual pre-training model after each parameter is clipped includes:
  • g is the gradient of the parameter
  • w is the weight of the parameter
  • Example 9 provides a speech recognition device, including: an acquisition module, configured to acquire a target speech signal to be recognized including multiple languages;
  • a recognition module configured to recognize the semantics of the target speech signal by merging the speech recognition model of the sparse sub-network of various languages
  • the sparse sub-network is obtained by performing parameter pruning on a multilingual pre-training model, and the multilingual pre-training model is obtained by training according to speech signals containing the multiple languages.
  • Example 10 provides a computer-readable medium on which a computer program is stored, and when the computer program is executed by a processor, the steps of the aforementioned speech recognition method are implemented.
  • Example 11 provides an electronic device, including: a memory on which a computer program is stored; a processor configured to execute the computer program in the memory to implement the aforementioned The steps of the speech recognition method.
  • Example 12 provides a computer program product, the computer program product includes a computer program carried on a computer-readable medium, and the program code included in the computer program can be used to implement the aforementioned Steps of speech recognition method.
  • Example 13 provides a computer program, which implements the steps of the aforementioned speech recognition method when the computer program is executed by a processor.

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

一种语音识别方法、语音识别装置、计算机可读介质、电子设备、计算机程序产品以及计算机程序，该方法包括：获取包含多种语言的待识别的目标语音信号（S101），通过融合各种语言的稀疏子网络的语音识别模型识别目标语音信号的语义（S102），该稀疏子网络是对多语言预训练模型进行参数剪枝处理得到的，该多语言预训练模型是根据包含多种语言的语音信号训练得到的。

Description

语音识别方法、装置、计算机可读介质及电子设备
本申请要求于2022年3月3日提交中国专利局、申请号为202210204891.7、申请名称为“语音识别方法、装置、计算机可读介质及电子设备”的中国专利申请的优先权，其全部内容通过引用结合在本文中。
技术领域
本公开涉及计算机技术邻域，具体地，涉及一种语音识别方法、语音识别装置、计算机可读介质、电子设备、计算机程序产品以及计算机程序。
背景技术
目前的跨语言表征学习方法并没有考虑到不同语言之间发音的多样性，仍沿用了单语表征学习的模型结构，模型不存在专门用于建模特定语言特性的模块，因此往往会面临语言之间相互干扰的问题。这个问题在语言数量变多、无监督数据增加时会更加严重，当这个多语言预训练模型用于下游多语言语音识别任务时，会导致大语种语言(如中文、英文)的识别性能大幅下降，相比单语言预训练模型有较大的差距。
发明内容
提供该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方法，包括：获取包含多种语言的待识别的目标语音信号；通过融合各种语言的稀疏子网络的语音识别模型识别所述目标语音信号的语义；所述稀疏子网络是对多语言预训练模型进行参数剪枝处理得到的，所述多语言预训练模型是根据包含所述多种语言的语音信号训练得到的。
第二方面，本公开提供一种语音识别装置，包括：获取模块，用于获取包含多种语言的待识别的目标语音信号；识别模块，用于通过融合各种语言的稀疏子网络的语音识别模 型识别所述目标语音信号的语义；所述稀疏子网络是对多语言预训练模型进行参数剪枝处理得到的，所述多语言预训练模型是根据包含所述多种语言的语音信号训练得到的。
第三方面，本公开提供一种计算机可读介质，其上存储有计算机程序，该计算机程序被处理器执行时实现前述的语音识别方法的步骤。
第四方面，本公开提供一种电子设备，包括：存储器，其上存储有计算机程序；处理器，用于执行所述存储器中的所述计算机程序，以实现前述的语音识别方法的步骤。
第五方面，本公开实施例还提供一种计算机程序产品，该计算机程序产品包括承载在计算机可读介质上的计算机程序，所述计算机程序包括的程序代码可用于实现上述第一方面中的步骤。
第六方面，本公开实施例还提供一种计算机程序，该计算机程序被处理器执行时实现上述第一方面中的步骤。
通过上述技术方案，获取包含多种语言的待识别的目标语音信号，通过融合各种语言的稀疏子网络的语音识别模型识别目标语音信号的语义，该稀疏子网络是对多语言预训练模型进行参数剪枝处理得到的，该多语言预训练模型是根据包含多种语言的语音信号训练得到的。本公开从自适应的角度解决跨语言表征学习的语言干扰问题，将整个多语言预训练模型对不同语言分别进行参数剪枝处理，构造出一组共享部分参数的稀疏子网络进行训练，从而赋予了语音识别模型针对不同语言特异性建模的能力，在跨语言表征学习过程中，对大语种和小语种均有大幅的改善。
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是本公开一个示例性实施例提供的计算机系统的结构示意图。
图2是本公开一个示例性实施例提供的语音识别方法的流程图。
图3为本公开一个示例性实施例提供的语音识别模型的训练方法的流程图。
图4是本公开一个示例性实施例提供的步骤S202的子步骤的流程图。
图5是本公开一个示例性实施例提供的语音识别装置框图。
图6是本公开一个示例性实施例提供的电子设备的结构示意图。
附图标记说明
120-终端；140-服务器；20-语音识别装置；201-获取模块；203-识别模块；205-
处理模块；600-电子设备；601-处理器；602-ROM；603-RAM；604-总线；605-I/O接口；606-输入装置；607-输出装置；608-存储器；609-通信装置。
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”。
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示出了本公开一个示例性实施例提供的计算机系统的结构示意图，该计算机系统包括终端120和服务器140。
终端120与服务器140之间通过有线或者无线网络相互连接。
终端120可以包括智能手机、笔记本电脑、台式电脑、平板电脑、智能音箱、智能机器人中的至少一种。
终端120包括显示器；显示器可以用于显示语音识别结果。
终端120包括第一存储器和第一处理器。第一存储器中存储有第一程序；上述第一程序被第一处理器调用执行以实现本公开提供的语音识别方法。第一存储器可以包括但不限于以下几种：随机存取存储器(Random Access Memory，RAM)、只读存储器(Read Only Memory，ROM)、可编程只读存储器(Programmable Read-Only Memory， PROM)、可擦除只读存储器(Erasable Programmable Read-Only Memory，EPROM)、以及电可擦除只读存储器(Electric Erasable Programmable Read-Only Memory，EEPROM)。
第一处理器可以是一个或者多个集成电路芯片组成。可选地，第一处理器可以是通用处理器，比如，中央处理器(Central Processing Unit，CPU)或者网络处理器(Network Processor，NP)。示例性的，终端中的语音识别模型可以是由终端训练得到的；或，由服务器训练得到，终端从服务器获取。
服务器140包括第二存储器和第二处理器。第二存储器中存储有第二程序，上述第二程序被第二处理器调用来实现本公开提供的语音识别方法。示例性的，第二存储器中存储有语音识别模型，语音识别模型被第二处理器调用以实现语音识别方法。可选地，第二存储器可以包括但不限于以下几种：RAM、ROM、PROM、EPROM、EEPROM。可选地，第二处理器可以是通用处理器，比如，CPU或者NP。
服务器可以是独立的物理服务器，也可以是多个物理服务器构成的服务器集群或者分布式系统，还可以是提供云服务、云数据库、云计算、云函数、云存储、网络服务、云通信、中间件服务、域名服务、安全服务、内容分发网络(Content Delivery Network，CDN)、以及大数据和人工智能平台等基础云计算服务的云服务器。终端可以是智能手机、平板电脑、笔记本电脑、台式计算机、智能音箱、智能手表等，但并不局限于此。终端以及服务器可以通过有线或无线通信方式进行直接或间接地连接，本公开在此不做限制。
近年来，预训练语言模型迅速发展，预训练语言模型参数量也不断增加，导致计算成本也越来越高。为了提高预训练语言模型的效率，各种各样的模型压缩方法被提出，其中就包括模型剪枝。
基于此，本公开提供了一个示例性实施例提供的语音识别方法，包括：获取包含多种语言的待识别的目标语音信号，通过融合各种语言的稀疏子网络的语音识别模型识别目标语音信号的语义；该稀疏子网络是对多语言预训练模型进行参数剪枝处理得到的，多语言预训练模型是根据包含多种语言的语音信号训练得到的，该多种语言包括一些大语种，如中文、英文等，以及包括一些小语种，如法语、西班牙语等。本公开从自适应的角度解决跨语言表征学习的语言干扰问题，将整个多语言预训练模型对不同语言分别进行参数剪枝处理，构造出一组共享部分参数的稀疏子网络进行训练，从而赋予了语音识别模型针对不同语言特异性建模的能力，在跨语言表征学习过程中，对大语种和小语种均有大幅的改善。
需要说明的是，下面将对本实施例提供的语音识别方法进行详细说明，此处未提及之处可以参考下面图2的描述，在此不再赘述。
请参阅图2，图2为本公开一个示例性实施例提供的语音识别方法的流程图。该方法由电子设备来执行，例如，由图1所示的计算机系统中的终端或服务器来执行。图2所示的语音识别方法包括以下步骤：
在步骤S101中，获取包含多种语言的待识别的目标语音信号。
需要说明的是，该多种语言包括一些大语种，如中文、英文等，以及包括一些小语种，如法语、西班牙语等。
在步骤S102中，通过融合各种语言的稀疏子网络的语音识别模型识别目标语音信号的语义。
该稀疏子网络是对多语言预训练模型进行参数剪枝处理得到的，该多语言预训练模型是根据包含多种语言的语音信号训练得到的。
请参阅图3，图3为本公开一个示例性实施例提供的语音识别模型的训练方法的流程图。该语音识别模型的训练方法包括以下步骤：
在步骤S201中，获取包含多种语言的语音信号作为训练样本。
需要说明的是，该多种语言包括一些大语种，如中文、英文等，以及包括一些小语种，如法语、西班牙语等。其中大语种的样本数量远远大于小语种的样本数量，例如大语种的样本数量可以是100万条，而小语种的样本数量可能仅有几万条。并且训练样本为无监督语音数据，即没有进行人工标注的语音数据。
在步骤S202中，根据训练样本训练得到多语言预训练模型。
需要说明的是，多语言预训练模型用于语音识别，并且该多语言预训练模型可以对不同语言分别进行识别。
在训练阶段沿用了Wav2vec 2.0框架进行跨语言语音表征学习，其主要包括了特征提取器、上下文网络和一个量化模块。该特征提取器由多层卷积神经网络构成；该上下文网络由多层的Transformer层组成，用于对特征提取器的输出的语音信号进行语义学习，输出具有上下文信息的表征向量；量化模块用于提供原始的无监督语音数据，用于对比学习，将特征提取器的输出的语音信号进行量化。为了语音表征学习的稳定，额外添加了多样性损失来促进多语言预训练模型对量化模块的使用，避免量化模块的坍塌现象。
需要说明的是，步骤S202包括子步骤S2021、子步骤S2022、子步骤S2023、子步骤S2024及子步骤S2025，训练得到多语言预训练模型的具体方式将在步骤S202的子步骤中进行详细描述。请参阅图4，图4是本公开一个示例性实施例示出的步骤S202的子步骤的流程图。
在子步骤S2021中，将语音信号转化成多个低维的信号帧。
需要说明的是，将语音信号转化成多个低维的信号帧之前还需要对对样本数量低于第一阈值的语言的训练样本进行上采样，以扩大采样数据中低于第一阈值的语言的训练样本数量；例如小语种的样本数量可能仅有几万条，因此需要对小语种的训练样本进行上采样。对样本数量高于第二阈值的语种进行均匀采样，例如大语种的样本数量可以是100万条，此时仅需要对其进行均匀采样即可得到足够的采样数据。
通过前述提到的特征提取器将输入的语音信号转化成多个低维的信号帧，其中每一个信号帧为固定时长的语音表征信号。示例性的，每一个信号帧可以是约25ms长，步幅20ms。
在子步骤S2022中，掩蔽掉多个信号帧中的任意一帧，得到被掩蔽的语音信号。
示例性的，某个语音信号总共被划分为10帧，那么可以掩蔽掉10个信号帧中的任意一帧或两帧，如果某个语音信号总共被划分为100帧，那么可以掩蔽掉100个信号帧中的任意10帧或20帧，然后得到被掩蔽的语音信号。
在子步骤S2023中，将被掩蔽的语音信号输入初始多语言预训练模型进行语义学习，以预测被掩蔽的信号帧。
通过前述提到的上下文网络，对被掩蔽的语音信号进行语义学习，根据语义学习的结果重建出被掩蔽的信号帧，以预测出被掩蔽的信号帧，与此同时，量化模块接收特征提取器输出的语音信号是未受掩蔽的，因此可以将预测出的被掩蔽的信号帧与量化模块用于提供的无监督语音数据进行对比。
在子步骤S2024中，当预测的被掩蔽的信号帧与实际被掩蔽的信号帧一致时，确定预测正确并更新初始多语言预训练模型的参数。
当预测的被掩蔽的信号帧与量化模块输出的实际被掩蔽的信号帧一致时，确定上下文网络预测的信号帧是正确的，并更新初始多语言预训练模型的参数。
在子步骤S2025中，重复执行更新初始多语言预训练模型的参数的步骤，以得到多语言预训练模型。
重复执行子步骤S2022-S2024，每一次掩蔽的信号帧都不重复，然后通过上下文网络对被掩蔽的语音信号进行语义学习，以预测出被掩蔽的信号帧，将预测出被掩蔽的信号帧与量化模块输出的实际被掩蔽的信号帧进行比对，当比对一致时确定上下文网络预测的信号帧是正确的，并更新初始多语言预训练模型的参数，以得到多语言预训练模型。
示例性的，可以重复执行子步骤S2022-S2024预定次数，该预定次数可以是语音信号的信号帧数量的预定比例，如10个信号的帧的50％，就是5次，该预定次数可以基于人为经验取得，或者根据其他可行的方法取得，在此不再赘述。
在步骤S203中，将多语言预训练模型对应多种语言分别进行参数剪枝处理，得到每种语言对应的稀疏子网络。
需要说明的是，参数剪枝处理主要包括以下两种方式，基于彩票假设方式和基于泰勒展开方式。在本公开中可以采用其中任意一种剪枝方式，下面将详细阐述这两种剪枝方式。
基于彩票假设方式将所述多语言预训练模型对应所述多种语言分别进行参数剪枝处理的步骤包括：将各个语种的语音信号作为训练样本，分别训练多语言预训练模型，这里的多语言预训练模型是指步骤S202中训练得到的多语言预训练模型，此外这里训练收敛条件也与步骤S202中一样；然后获取每种语言对应的多语言预训练模型的所有参数；根据这些参数构成参数矩阵；然后构建与参数矩阵对应的掩蔽矩阵，该掩蔽矩阵的长、宽与参数矩阵一致；然后获取参数矩阵中每个参数的绝对值；根据该绝对值的大小裁剪预定比例的参数，例如可以按照绝对值的大小从小到大裁剪预定比例的参数，该预定比例可以是20％、30％或50％等，在此不做限制；最后将被裁剪的参数在掩蔽矩阵中对应位置的掩蔽状态置为第一值，其余位置的掩蔽状态置为第二值，在一种实施方式中第一值为0，第二值为1。
基于泰勒展开方式将多语言预训练模型对应多种语言分别进行参数剪枝处理的步骤包括：将各个语种的语音信号作为训练样本，分别训练多语言预训练模型，这里的多语言预训练模型是指步骤S202中训练得到的多语言预训练模型，此外这里训练收敛条件也与步骤S202中一样；获取每种语言对应的多语言预训练模型的所有参数；通过一阶泰勒展开参数后，预测每个参数被裁剪后对多语言预训练模型造成的损失值，在一种实施方式中，预测每个参数被裁剪后对多语言预训练模型造成的损失值的公式包括：
|g2w2|
其中，g为所述参数的梯度，w为所述参数的权重，|为取绝对值的运算符；然后根据损失值的大小裁剪预定比例的参数，例如可以按照损失值的大小从小到大裁剪预定比例的参数，该预定比例可以是20％、30％或50％等，在此不做限制。
参数剪枝处理后得到的多语言预训练模型，可以加快推理计算速度，满足最低延迟限制，可以减少所消耗的内存，对于终端侧更易于部署，比如手机，更便于模型的训练和微调。
通过上述两种剪枝方式中的任意一种，将各个语种的语音信号作为训练样本，分别训练多语言预训练模型，得到每种语言对应的稀疏子网络。
在步骤S204中，通过对应的语言对各个稀疏子网络进行多语言自适应预训练来更新各个稀疏子网络的参数，以得到各个稀疏子网络之间的共享参数和独享参数。
在训练过程中，每个训练batch都只由一种语言的训练样本组成，其中batch是指：使用训练样本中的一小部分样本对模型权重进行一次反向传播的参数更新，这一小部分样本被称为“一批数据”。
而对每种语言的输入样本数据，只使用这个语言所对应的稀疏子网络进行前向传播，并计算稀疏子网络损失，反向传播时也只更新这个稀疏子网络对应的参数。通过这种方式，最终的稀疏子网络在网络内部自动分配了不同语言之间共享参数和独享参数，从而达到自适应训练的效果。
在步骤S205中，根据共享参数和独享参数得到融合各种语言的稀疏子网络的语音识别模型。
本公开提出的基于稀疏共享子网络跨语言语音表征的语音识别模型可以在大语种和小语种上均超过基线跨语言语音表征学习方法。在公开的共享语音(Common Voice)数据集合上，所提出的语音识别方法在100M模型相比基线系统有相对9.8％平均音素错误率下降，在300M模型上有7.4％音素错误率下降。而且这种方法可以大幅缓解大语种所受到的语言干扰问题，在100M模型和300M模型上对大语种有相对17.8％和16.7％的音素错误率下降。
综上所述，本公开提供的语音识别方法，包括获取包含多种语言的待识别的目标语音信号，通过融合各种语言的稀疏子网络的语音识别模型识别目标语音信号的语义，该稀疏子网络是对多语言预训练模型进行参数剪枝处理得到的，该多语言预训练模型是根据包含多种语言的语音信号训练得到的。本公开从自适应的角度解决跨语言表征 学习的语言干扰问题，将整个多语言预训练模型对不同语言分别进行参数剪枝处理，构造出一组共享部分参数的稀疏子网络进行训练，从而赋予了语音识别模型针对不同语言特异性建模的能力，在跨语言表征学习过程中，对大语种和小语种均有大幅的改善。
图5是本公开一个示例性实施例示出的一种语音识别装置框图。参照图5，该装置20包括获取模块201和识别模块203。
该获取模块201，用于获取包含多种语言的待识别的目标语音信号；
该识别模块203，用于通过融合各种语言的稀疏子网络的语音识别模型识别所述目标语音信号的语义；所述稀疏子网络是对多语言预训练模型进行参数剪枝处理得到的，所述多语言预训练模型是根据包含所述多种语言的语音信号训练得到的。
该装置20还包括处理模块205。
可选地，该处理模块205，用于获取包含所述多种语言的语音信号作为训练样本；
根据所述训练样本训练得到所述多语言预训练模型；所述多语言预训练模型用于语音识别；
将所述多语言预训练模型对应所述多种语言分别进行参数剪枝处理，得到每种语言对应的稀疏子网络；
通过对应的语言对各个所述稀疏子网络进行多语言自适应预训练来更新各个所述稀疏子网络的参数，以得到各个所述稀疏子网络之间的共享参数和独享参数；
根据所述共享参数和所述独享参数得到融合各种语言的稀疏子网络的语音识别模型。
可选地，该处理模块205，还用于将所述语音信号转化成多个低维的信号帧；所述信号帧为固定时长的语音表征信号；
掩蔽掉多个所述信号帧中的任意一帧，得到被掩蔽的语音信号；
将所述被掩蔽的语音信号输入初始多语言预训练模型进行语义学习，以预测被掩蔽的信号帧；
当预测的被掩蔽的信号帧与实际被掩蔽的信号帧一致时，确定预测正确并更新初始多语言预训练模型的参数；
重复执行所述更新初始多语言预训练模型的参数的步骤，以得到所述多语言预训练模型。
可选地，该处理模块205，还用于对样本数量低于第一阈值的语言的训练样本进行上采样，以扩大采样数据中所述低于第一阈值的语言的训练样本数量；
对样本数量高于第二阈值的语种进行均匀采样。
可选地，该处理模块205，还用于基于彩票假设方式将所述多语言预训练模型对应所述多种语言分别进行参数剪枝处理，得到每种语言对应的稀疏子网络；
或基于泰勒展开方式将所述多语言预训练模型对应所述多种语言分别进行参数剪枝处理，得到每种语言对应的稀疏子网络。
可选地，该处理模块205，还用于将各个语种的所述语音信号作为训练样本，分别训练所述多语言预训练模型；
获取每种语言对应的所述多语言预训练模型的所有参数；
根据所述参数构成参数矩阵；
构建与所述参数矩阵对应的掩蔽矩阵；
获取所述参数矩阵中每个参数的绝对值；
根据所述绝对值的大小裁剪预定比例的所述参数；
将被裁剪的所述参数在所述掩蔽矩阵中对应位置的掩蔽状态置为第一值，其余位置的所述掩蔽状态置为第二值。
可选地，该处理模块205，还用于将各个语种的所述语音信号作为训练样本，分别训练所述多语言预训练模型；
获取每种语言对应的所述多语言预训练模型的所有参数；
通过一阶泰勒展开所述参数后，预测每个所述参数被裁剪后对所述多语言预训练模型造成的损失值；
根据所述损失值的大小裁剪预定比例的所述参数。
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(例如图1中的终端设备或服务器)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、个人数字助理(Personal Digital Assistant，PDA)、PAD(平板电脑)、便携式多媒体播放器(Portable Multimedia Player，PMP)、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可以包括处理器(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器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。处理器601、ROM 602以及RAM 603通过总线604彼此相连。输入/输出(Input/Output，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(Liquid Crystal Display，LCD)、扬声器、振动器等的输出装置607；包括例如磁带、硬盘等的存储器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，或者从存储器608被安装，或者从ROM 602被安装。在该计算机程序被处理器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(Compact Disc Read Only Memory，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存储介质，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、射频(Radio Frequency，RF)等等，或者上述的任意合适的组合。
在一些实施方式中，终端、服务器可以利用诸如超文本传输协议(HyperText Transfer Protocol，HTTP)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(Local Area Network，LAN)，广域网(Wide Area Network，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：获取包含多种语言的待识别的目标语音信号；通过融合各种语言的稀疏子网络的语音识别模型识别所述目标语音信号的语义；所述稀疏子网络是对多语言预训练模型进行参数剪枝处理得到的，所述多语言预训练模型是根据包含所述多种语言的语音信号训练得到的。
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。
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(Field Programmable Gate Array，FPGA)、专用集成电路(Application Specific Integrated Circuit，ASIC)、专用标准产品(Application Specific Standard Product，ASSP)、片上系统(System on a Chip，SOC)、复杂可编程逻辑设备(Complex Programmable Logic Device，CPLD)等等。
在本公开的上下文中，机器可读存储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。机器可读存储介质可以是机器可读信号介质或机器可读储存介质。机器可读存储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方法，包括：获取包含多种语言的待识别的目标语音信号；
通过融合各种语言的稀疏子网络的语音识别模型识别所述目标语音信号的语义；
所述稀疏子网络是对多语言预训练模型进行参数剪枝处理得到的，所述多语言预训练模型是根据包含所述多种语言的语音信号训练得到的。
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，所述语音识别模型的训练方法包括以下步骤：
获取包含所述多种语言的语音信号作为训练样本；
根据所述训练样本训练得到所述多语言预训练模型；所述多语言预训练模型用于语音识别；
将所述多语言预训练模型对应所述多种语言分别进行参数剪枝处理，得到每种语言对应的稀疏子网络；
通过对应的语言对各个所述稀疏子网络进行多语言自适应预训练来更新各个所述稀疏子网络的参数，以得到各个所述稀疏子网络之间的共享参数和独享参数；
根据所述共享参数和所述独享参数得到融合各种语言的稀疏子网络的语音识别模型。
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，所述根据所述训练样本训练得到所述多语言预训练模型的步骤包括：
将所述语音信号转化成多个低维的信号帧；所述信号帧为固定时长的语音表征信号；
掩蔽掉多个所述信号帧中的任意一帧，得到被掩蔽的语音信号；
将所述被掩蔽的语音信号输入初始多语言预训练模型进行语义学习，以预测被掩蔽的信号帧；
当预测的被掩蔽的信号帧与实际被掩蔽的信号帧一致时，确定预测正确并更新初始多语言预训练模型的参数；
重复执行所述更新初始多语言预训练模型的参数的步骤，以得到所述多语言预训练模型。
根据本公开的一个或多个实施例，示例4提供了示例3的方法，所述根据所述训练样本训练得到多语言预训练模型的步骤还包括：
对样本数量低于第一阈值的语言的训练样本进行上采样，以扩大采样数据中所述低于第一阈值的语言的训练样本数量；
对样本数量高于第二阈值的语种进行均匀采样。
根据本公开的一个或多个实施例，示例5提供了示例2的方法，所述将所述多语言预训练模型对应所述多种语言分别进行参数剪枝处理得到每种语言对应的稀疏子网络的步骤包括：
基于彩票假设方式将所述多语言预训练模型对应所述多种语言分别进行参数剪枝处理，得到每种语言对应的稀疏子网络；
或基于泰勒展开方式将所述多语言预训练模型对应所述多种语言分别进行参数剪枝处理，得到每种语言对应的稀疏子网络。
根据本公开的一个或多个实施例，示例6提供了示例5的方法，所述基于彩票假设方式将所述多语言预训练模型对应所述多种语言分别进行参数剪枝处理的步骤包括：
将各个语种的所述语音信号作为训练样本，分别训练所述多语言预训练模型；
获取每种语言对应的所述多语言预训练模型的所有参数；
根据所述参数构成参数矩阵；
构建与所述参数矩阵对应的掩蔽矩阵；
获取所述参数矩阵中每个参数的绝对值；
根据所述绝对值的大小裁剪预定比例的所述参数；
将被裁剪的所述参数在所述掩蔽矩阵中对应位置的掩蔽状态置为第一值，其余位置的所述掩蔽状态置为第二值。
根据本公开的一个或多个实施例，示例7提供了示例5的方法，所述基于泰勒展开方式将所述多语言预训练模型对应所述多种语言分别进行参数剪枝处理的步骤包括：
将各个语种的所述语音信号作为训练样本，分别训练所述多语言预训练模型；
获取每种语言对应的所述多语言预训练模型的所有参数；
通过一阶泰勒展开所述参数后，预测每个所述参数被裁剪后对所述多语言预训练模型造成的损失值；
根据所述损失值的大小裁剪预定比例的所述参数。
根据本公开的一个或多个实施例，示例8提供了示例7的方法，所述预测每个所述参数被裁剪后对所述多语言预训练模型造成的损失值的公式包括：
|g2w2|
其中，g为所述参数的梯度，w为所述参数的权重。
根据本公开的一个或多个实施例，示例9提供了一种语音识别装置，包括：获取模块，用于获取包含多种语言的待识别的目标语音信号；
识别模块，用于通过融合各种语言的稀疏子网络的语音识别模型识别所述目标语音信号的语义；
所述稀疏子网络是对多语言预训练模型进行参数剪枝处理得到的，所述多语言预训练模型是根据包含所述多种语言的语音信号训练得到的。
根据本公开的一个或多个实施例，示例10提供了一种计算机可读介质，其上存储有计算机程序，该计算机程序被处理器执行时实现前述的语音识别方法的步骤。
根据本公开的一个或多个实施例，示例11提供了一种电子设备，包括：存储器，其上存储有计算机程序；处理器，用于执行所述存储器中的所述计算机程序，以实现前述的语音识别方法的步骤。
根据本公开的一个或多个实施例，示例12提供了一种计算机程序产品，该计算机程序产品包括承载在计算机可读介质上的计算机程序，所述计算机程序包括的程序代码可用于实现前述的语音识别方法的步骤。
根据本公开的一个或多个实施例，示例13提供了一种计算机程序，该计算机程序被处理器执行时实现前述的语音识别方法的步骤。
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 (12)

  1. 一种语音识别方法，包括：
    获取包含多种语言的待识别的目标语音信号；
    通过融合各种语言的稀疏子网络的语音识别模型识别所述目标语音信号的语义；
    所述稀疏子网络是对多语言预训练模型进行参数剪枝处理得到的，所述多语言预训练模型是根据包含所述多种语言的语音信号训练得到的。
  2. 根据权利要求1所述的方法，其中，所述语音识别模型的训练方法包括以下步骤：
    获取包含所述多种语言的语音信号作为训练样本；
    根据所述训练样本训练得到所述多语言预训练模型；所述多语言预训练模型用于语音识别；
    将所述多语言预训练模型对应所述多种语言分别进行参数剪枝处理，得到每种语言对应的稀疏子网络；
    通过对应的语言对各个所述稀疏子网络进行多语言自适应预训练来更新各个所述稀疏子网络的参数，以得到各个所述稀疏子网络之间的共享参数和独享参数；
    根据所述共享参数和所述独享参数得到融合各种语言的稀疏子网络的语音识别模型。
  3. 根据权利要求2所述的方法，其中，所述根据所述训练样本训练得到所述多语言预训练模型的步骤包括：
    将所述语音信号转化成多个低维的信号帧；所述信号帧为固定时长的语音表征信号；
    掩蔽掉多个所述信号帧中的任意一帧，得到被掩蔽的语音信号；
    将所述被掩蔽的语音信号输入初始多语言预训练模型进行语义学习，以预测被掩蔽的信号帧；当预测的被掩蔽的信号帧与实际被掩蔽的信号帧一致时，确定预测正确并更新初始多语言预训练模型的参数；
    重复执行所述更新初始多语言预训练模型的参数的步骤，以得到所述多语言预训练模型。
  4. 根据权利要求3所述的方法，其中，所述根据所述训练样本训练得到所述多语言预训练模型的步骤还包括：
    对样本数量低于第一阈值的语言的训练样本进行上采样，以扩大采样数据中所述低于第一阈值的语言的训练样本数量；
    对样本数量高于第二阈值的语种进行均匀采样。
  5. 根据权利要求2-4中任一项所述的方法，其中，所述将所述多语言预训练模型对应所述多种语言分别进行参数剪枝处理，得到每种语言对应的稀疏子网络的步骤包括：
    基于彩票假设方式将所述多语言预训练模型对应所述多种语言分别进行参数剪枝处理，得到每种语言对应的稀疏子网络；
    或基于泰勒展开方式将所述多语言预训练模型对应所述多种语言分别进行参数剪枝处理，得到每种语言对应的稀疏子网络。
  6. 根据权利要求5所述的方法，其中，所述基于彩票假设方式将所述多语言预训练模型对应所述多种语言分别进行参数剪枝处理的步骤包括：
    将各个语种的所述语音信号作为训练样本，分别训练所述多语言预训练模型；
    获取每种语言对应的所述多语言预训练模型的所有参数；
    根据所述参数构成参数矩阵；
    构建与所述参数矩阵对应的掩蔽矩阵；
    获取所述参数矩阵中每个参数的绝对值；
    根据所述绝对值的大小裁剪预定比例的所述参数；
    将被裁剪的所述参数在所述掩蔽矩阵中对应位置的掩蔽状态置为第一值，其余位置的所述掩蔽状态置为第二值。
  7. 根据权利要求5所述的方法，其中，所述基于泰勒展开方式将所述多语言预训练模型对应所述多种语言分别进行参数剪枝处理的步骤包括：
    将各个语种的所述语音信号作为训练样本，分别训练所述多语言预训练模型；
    获取每种语言对应的所述多语言预训练模型的所有参数；
    通过一阶泰勒展开所述参数后，预测每个所述参数被裁剪后对所述多语言预训练模型造成的损失值；
    根据所述损失值的大小裁剪预定比例的所述参数。
  8. 一种语音识别装置，包括：
    获取模块，用于获取包含多种语言的待识别的目标语音信号；
    识别模块，用于通过融合各种语言的稀疏子网络的语音识别模型识别所述目标语音信号的语义；
    所述稀疏子网络是对多语言预训练模型进行参数剪枝处理得到的，所述多语言预训练模型是根据包含所述多种语言的语音信号训练得到的。
  9. 一种计算机可读介质，其上存储有计算机程序，其中，所述计算机程序被处理器执行时实现权利要求1-7中任一项所述方法的步骤。
  10. 一种电子设备，包括：
    存储器，其上存储有计算机程序；
    处理器，用于执行所述存储器中的所述计算机程序，以实现权利要求1-7中任一项所述方法的步骤。
  11. 一种计算机程序产品，所述计算机程序产品包括承载在计算机可读介质上的计算机程序，所述计算机程序包括的程序代码用于实现权利要求1-7中任一项所述方法的步骤。
  12. 一种计算机程序，所述计算机程序被处理器执行时实现权利要求1-7中任一项所述方法的步骤。
PCT/CN2023/079156 2022-03-03 2023-03-01 语音识别方法、装置、计算机可读介质及电子设备 WO2023165538A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210204891.7 2022-03-03
CN202210204891.7A CN114582329A (zh) 2022-03-03 2022-03-03 语音识别方法、装置、计算机可读介质及电子设备

Publications (1)

Publication Number Publication Date
WO2023165538A1 true WO2023165538A1 (zh) 2023-09-07

Family

ID=81776470

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/079156 WO2023165538A1 (zh) 2022-03-03 2023-03-01 语音识别方法、装置、计算机可读介质及电子设备

Country Status (2)

Country Link
CN (1) CN114582329A (zh)
WO (1) WO2023165538A1 (zh)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114582329A (zh) * 2022-03-03 2022-06-03 北京有竹居网络技术有限公司 语音识别方法、装置、计算机可读介质及电子设备
WO2024013905A1 (ja) * 2022-07-13 2024-01-18 日本電信電話株式会社 ニューラルネットワーク枝刈りシステム、ニューラルネットワーク枝刈り方法及びニューラルネットワーク枝刈りプログラム
CN115547334A (zh) * 2022-10-17 2022-12-30 上海城建职业学院 小学作文语音识别文本纠错系统及方法
CN115862601B (zh) * 2023-03-01 2023-05-02 贝壳找房(北京)科技有限公司 数据生成方法、电子设备及可读存储介质
CN116306601B (zh) * 2023-05-17 2023-09-08 上海蜜度信息技术有限公司 小语种纠错模型训练方法、纠错方法、系统、介质及设备
CN116776870B (zh) * 2023-08-26 2023-11-14 腾讯科技（深圳）有限公司 意图识别方法、装置、计算机设备及介质

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103400577A (zh) * 2013-08-01 2013-11-20 百度在线网络技术（北京）有限公司 多语种语音识别的声学模型建立方法和装置
CN108711420A (zh) * 2017-04-10 2018-10-26 北京猎户星空科技有限公司 多语言混杂模型建立、数据获取方法及装置、电子设备
CN110827805A (zh) * 2019-12-09 2020-02-21 苏州思必驰信息科技有限公司 语音识别模型训练方法、语音识别方法和装置
CN112489622A (zh) * 2019-08-23 2021-03-12 中国科学院声学研究所 一种多语言连续语音流语音内容识别方法及系统
US20210210077A1 (en) * 2020-01-03 2021-07-08 International Business Machines Corporation Cognitive analysis for speech recognition using multi-language vector representations
CN113924619A (zh) * 2019-05-28 2022-01-11 谷歌有限责任公司 通过流式端到端模型的大规模多语言语音识别
CN114582329A (zh) * 2022-03-03 2022-06-03 北京有竹居网络技术有限公司 语音识别方法、装置、计算机可读介质及电子设备

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103400577A (zh) * 2013-08-01 2013-11-20 百度在线网络技术（北京）有限公司 多语种语音识别的声学模型建立方法和装置
CN108711420A (zh) * 2017-04-10 2018-10-26 北京猎户星空科技有限公司 多语言混杂模型建立、数据获取方法及装置、电子设备
CN113924619A (zh) * 2019-05-28 2022-01-11 谷歌有限责任公司 通过流式端到端模型的大规模多语言语音识别
CN112489622A (zh) * 2019-08-23 2021-03-12 中国科学院声学研究所 一种多语言连续语音流语音内容识别方法及系统
CN110827805A (zh) * 2019-12-09 2020-02-21 苏州思必驰信息科技有限公司 语音识别模型训练方法、语音识别方法和装置
US20210210077A1 (en) * 2020-01-03 2021-07-08 International Business Machines Corporation Cognitive analysis for speech recognition using multi-language vector representations
CN114582329A (zh) * 2022-03-03 2022-06-03 北京有竹居网络技术有限公司 语音识别方法、装置、计算机可读介质及电子设备

Also Published As

Publication number Publication date
CN114582329A (zh) 2022-06-03

Similar Documents

Publication Publication Date Title
WO2023165538A1 (zh) 语音识别方法、装置、计算机可读介质及电子设备
JP7208952B2 (ja) 対話モデルを生成するための方法及び装置
WO2022156561A1 (zh) 一种自然语言处理方法以及装置
WO2023273611A1 (zh) 语音识别模型的训练方法、语音识别方法、装置、介质及设备
WO2023273985A1 (zh) 一种语音识别模型的训练方法、装置及设备
US20230306954A1 (en) Speech synthesis method, apparatus, readable medium and electronic device
CN109740167B (zh) 用于生成信息的方法和装置
US11783808B2 (en) Audio content recognition method and apparatus, and device and computer-readable medium
WO2024012251A1 (zh) 语义分割模型训练方法、装置、电子设备及存储介质
US11036996B2 (en) Method and apparatus for determining (raw) video materials for news
CN112259089A (zh) 语音识别方法及装置
WO2023273610A1 (zh) 语音识别方法、装置、介质及电子设备
US20240078385A1 (en) Method and apparatus for generating text
WO2024099342A1 (zh) 翻译方法、装置、可读介质及电子设备
US20230127787A1 (en) Method and apparatus for converting voice timbre, method and apparatus for training model, device and medium
CN115081616A (zh) 一种数据的去噪方法以及相关设备
CN116030792A (zh) 用于转换语音音色的方法、装置、电子设备和可读介质
WO2023211369A2 (zh) 语音识别模型的生成方法、识别方法、装置、介质及设备
CN113591490B (zh) 信息处理方法、装置和电子设备
CN113468857B (zh) 风格转换模型的训练方法、装置、电子设备以及存储介质
CN114550702A (zh) 一种语音识别方法和装置
CN111898338B (zh) 文本生成方法、装置和电子设备
WO2024001653A1 (zh) 特征提取方法、装置、存储介质及电子设备
WO2023179506A1 (zh) 韵律预测方法、装置、可读介质及电子设备
WO2023202543A1 (zh) 文字处理方法、装置、电子设备及存储介质

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: 23762936

Country of ref document: EP

Kind code of ref document: A1