A method and related apparatus for generating an automatic speech recognition model using unsupervised learning

Abstract

A method for generating an automatic speech recognition (ASR) model using unsupervised learning includes obtaining text information by a device. The method includes determining, by the device, a phoneme sequence group associated with the text information. The method includes obtaining speech waveform data by the device. The method includes determining, by the device, a phoneme boundary group associated with the speech waveform data. The method includes: based on determining the phoneme sequence group associated with the text information and based on determining the phoneme boundary group associated with the speech waveform data, generating the ASR model by the device using an output distribution matching (ODM) technique.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to U.S. Application No. 16 / 269,951, filed on February 7, 2019, the disclosure of which is incorporated herein by reference in its entirety. Technical Field

[0003] The present application relates to automatic speech recognition technology, and more particularly to a method and related apparatus for generating automatic speech recognition patterns using unsupervised learning. Background Art

[0004] Automatic speech recognition (ASR) includes widespread applications in smartphones, smart speakers, tablet computers, etc. Recently, the performance of ASR has improved, and the recognition accuracy in certain scenarios is comparable to that of humans.

[0005] ASR systems usually rely on supervised deep learning, which requires a large amount of manually labeled data to train deep neural network models. In order to obtain training data, humans need to listen, transcribe and annotate a large amount of data. In this way, training data acquisition is time-intensive, error-prone, expensive, inaccurate, inefficient, etc. In addition, for some resource-poor languages, obtaining the necessary amount of training data is a difficult or impossible task. Summary of the invention

[0006] According to some possible implementations, a method for generating an automatic speech recognition (ASR) model using unsupervised learning includes: obtaining text information; determining a phoneme sequence group associated with the text information; obtaining speech waveform data; determining a phoneme boundary group associated with the speech waveform data; and generating an ASR model based on the phoneme sequence group and the phoneme boundary group using an output distribution matching (ODM) technique.

[0007] According to some possible implementations, a device includes: at least one memory configured to store program code; and at least one processor configured to read the program code and execute the above-mentioned method for generating an automatic speech recognition (ASR) model using unsupervised learning according to the instructions of the program code.

[0008] According to some possible implementations, a device for generating an automatic speech recognition (ASR) model using unsupervised learning includes: a first obtaining unit, which is configured to obtain text information; a first determining unit, which is configured to determine a phoneme sequence group associated with the text information; a second obtaining unit, which is configured to obtain speech waveform data; a second determining unit, which is configured to determine a phoneme boundary group associated with the speech waveform data; and a generating unit, which is configured to generate an ASR model based on the phoneme sequence group and the phoneme boundary group using an output distribution matching (ODM) technique.

[0009] According to some possible implementations, a non-transitory computer-readable medium storing instructions includes one or more instructions that, when executed by one or more processors of a device, cause the one or more processors to perform the above-mentioned method for generating an automatic speech recognition (ASR) model using unsupervised learning.

[0010] Compared with the traditional method of generating ASR models based on supervised learning, the implementation of the present invention reduces the manual effort required for model generation through unsupervised learning, is able to develop ASR models for multiple languages, reduces human errors, reduces the time associated with model generation, and has other technical benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1A and Figure 1B is a diagram of an overview of an example implementation according to an embodiment;

[0012] Figure 2 is a diagram of an example environment in which the systems and / or methods described herein may be implemented according to another embodiment;

[0013] Figure 3 According to another embodiment Figure 2 diagrams of example components of one or more devices; and

[0014] Figure 4 is a flow chart of an example process for generating an automatic speech recognition (ASR) model using unsupervised learning and using output distribution matching (ODM) techniques according to another embodiment. DETAILED DESCRIPTION

[0015] ASR systems typically rely on supervised deep learning, which requires a large amount of manually labeled data to train deep neural network models. For example, ASR systems are typically trained using speech data of, for example, 1,000 hours, 4,000 hours, or 10,000 hours of actual speech. In order to obtain training data, humans need to listen to, transcribe, and annotate large amounts of data. In this way, training data acquisition is time-intensive, error-prone, expensive, inaccurate, inefficient, etc. In addition, for some resource-poor languages, obtaining the necessary amount of training data is a difficult or impossible task.

[0016] Some implementations herein provide techniques for generating an ASR model that maps a speech waveform to a text transcription. The ASR model can be a function that maps an acoustic speech signal to a sequence of words. In some implementations, the ASR model initially maps the speech signal to a sequence of intermediate phonemes (e.g., pronunciation units), and then maps the intermediate phoneme sequence to a sequence of words.

[0017] Some implementations herein provide ASR model generation techniques that do not require manually labeled data (e.g., model generation that can use unsupervised learning). In this way, some implementations herein can improve model generation, can improve the accuracy of the generated model, reduce the need for manual work, reduce manual errors associated with supervised learning, improve the efficiency of model generation, and have other technical benefits.

[0018] In addition, some implementations herein provide the following technical benefits related to unsupervised training of ASR models. First, some implementations herein can identify phoneme boundaries in continuous speech waveforms through iterative identification and refinement of phoneme boundaries. Second, some implementations herein can map speech waveforms into phonemes without the need for labeled data.

[0019] Figure 1A and Figure 1B FIG is a diagram that provides an overview of the embodiments described herein. Figure 1A As shown, and as indicated by reference numeral 101, a platform (e.g., a server) can obtain text information. In addition, and as indicated by reference numeral 102, the platform can obtain dictionary information. The dictionary information may include a data structure that maps a word sequence and a phoneme sequence. As shown by reference numeral 103, the platform can use the dictionary information to determine a phoneme sequence associated with the text information. Further, and as shown by reference numeral 104, the platform can generate a language model based on determining the phoneme sequence associated with the text information. That is, the platform can use the phoneme sequence to train the language model.

[0020] like Figure 1B As shown, and as shown by reference numeral 105, the platform can obtain Figure 1A In addition, and as indicated by reference numeral 106, the platform can obtain speech waveform data. Figure 1B The speech waveform data shown depicts phonemes and word sequences. In reality, the speech waveform data may be unlabeled and unannotated speech waveform data. In other words, the phonemes and word sequences are for illustration purposes only. As shown by reference numeral 107, the platform can generate an ASR model using the speech waveform data and the language model through unsupervised learning. The platform can use the same Figure 4 The associated techniques are described in more detail to generate ASR models.

[0021] In this way, some implementations herein can generate ASR models using unsupervised learning. In other words, some implementations herein can generate ASR models without the need for manually labeled training data.

[0022] Figure 2 2 is a diagram of an example environment 200 that may be used to implement the systems and / or methods described herein. Figure 2 As shown, environment 200 may include user device 210, platform 220, and network 230. The various devices of environment 200 may be interconnected via wired connections, wireless connections, or a combination of wired connections and wireless connections.

[0023] User device 210 includes one or more devices capable of receiving, generating, storing, processing, and / or providing information associated with platform 220. For example, user device 210 may include a computing device (e.g., a desktop computer, a laptop computer, a tablet computer, a handheld computer, a smart speaker, a server, etc.), a mobile phone (e.g., a smart phone, a wireless phone, etc.), a wearable device (e.g., a pair of smart glasses or a smart watch), or the like. In some implementations, user device 210 can receive information from platform 220 and / or send information to platform 220.

[0024] The platform 220 includes one or more devices capable of generating an ASR model, as described elsewhere herein. In some implementations, the platform 220 may include a cloud server or cloud server group. In some implementations, the platform 220 may be designed to be modular so that certain software components may be swapped in or out according to specific needs. In this way, the platform 220 may be easily and / or quickly reconfigured for different purposes.

[0025] In some implementations, as shown, the platform 220 can be hosted in a cloud computing environment 222. It is worth noting that, although the implementations described herein describe the platform 220 as being hosted in a cloud computing environment 222, in some implementations, the platform 220 is not cloud-based (i.e., can be implemented in a non-cloud computing environment) or can be partially cloud-based.

[0026] Cloud computing environment 222 includes an environment hosting platform 220. Cloud computing environment 222 can provide computing, software, data access, storage, etc. services that do not require end users (e.g., user devices 210) to be aware of the physical location and configuration of one or more systems and / or one or more devices hosting platform 220. As shown, cloud computing environment 222 can include a computing resource group 224 (collectively referred to as "computing resources 224" and individually referred to as "computing resource 224").

[0027] The computing resources 224 include one or more personal computers, workstation computers, server devices, or other types of computing and / or communication devices. In some implementations, the computing resources 224 can host the platform 220. Cloud resources can include computing instances executed in the computing resources 224, storage devices provided in the computing resources 224, data transmission devices provided by the computing resources 224, etc. In some implementations, the computing resources 224 can communicate with other computing resources 224 via wired connections, wireless connections, or a combination of wired and wireless connections.

[0028] Further Figure 2 As shown, computing resources 224 include cloud resource groups, such as one or more applications ("APP") 224-1, one or more virtual machines ("VM") 224-2, virtualized storage devices ("VS") 224-3, one or more hypervisors ("HYP") 224-4, and the like.

[0029] The applications 224-1 include one or more software applications that can be provided to or accessed by the user device 210 and / or the sensor device 220. The applications 224-1 can eliminate the need for installing and executing software applications on the user device 210. For example, the applications 224-1 can include software associated with the platform 220 and / or any other software that can be provided via the cloud computing environment 222. In some implementations, one application 224-1 can send information to or receive information from one or more other applications 224-1 via the virtual machine 224-2.

[0030] Virtual machine 224-2 includes a software implementation of a machine (e.g., a computer) that executes programs like a physical machine. Virtual machine 224-2 can be a system virtual machine or a process virtual machine, depending on the use and correspondence of virtual machine 224-2 to any real machine. System virtual machines can provide a complete system platform that executes support for a complete operating system ("OS"). Process virtual machines can execute a single program and can support a single process. In some implementations, virtual machine 224-2 can execute on behalf of a user (e.g., user device 210) and can manage the infrastructure of cloud computing environment 222, such as data management, synchronization, or long-term data transfer.

[0031] The virtualized storage device 224-3 includes one or more storage systems and / or one or more devices using virtualization technology within the device or storage system of the computing resource 224. In some implementations, within the context of the storage system, the types of virtualization may include block virtualization and file virtualization. Block virtualization may refer to the extraction (or separation) of logical storage from physical storage, so that the storage system can be accessed without considering physical storage or heterogeneous structures. Separation may allow administrators of the storage system flexibility in terms of administrator management of storage for end users. File virtualization may eliminate the dependency between data accessed at the file level and the location of the physical storage file. This may enable optimization of storage usage, server consolidation, and / or performance of non-disruptive file migration.

[0032] Hypervisor 224-4 can provide hardware virtualization technology that allows multiple operating systems (e.g., "guest operating systems") to execute simultaneously on a host computer such as computing resource 224. Hypervisor 224-4 can present a virtual operating platform to the guest operating systems and can manage the execution of the guest operating systems. Multiple instances of various operating systems can share virtualized hardware resources.

[0033] The network 230 includes one or more wired networks and / or wireless networks. For example, the network 230 may include a cellular network (e.g., a fifth generation (5G) network, a long term evolution (LTE) network, a third generation (3G) network, a code division multiple access (CDMA) network, etc.), a public land mobile network (PLMN), a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), a telephone network (e.g., a public switched telephone network (PSTN)), a private network, an ad hoc network, an intranet, the Internet, a fiber-based network, etc., and / or a combination of these or other types of networks.

[0034] Figure 2 The number and arrangement of devices and networks shown are provided as examples. Figure 2In addition to the devices and / or networks shown, there may be additional devices and / or networks, fewer devices and / or networks, different devices and / or networks, or differently arranged devices and / or networks. Figure 2 Two or more of the devices shown may be implemented in a single device, or Figure 2 The single device shown may be implemented as multiple distributed devices. Additionally or alternatively, a device group (eg, one or more devices) of environment 200 may perform one or more functions described as being performed by another device group of environment 200.

[0035] Figure 3 300 is a diagram of example components of device 300. Device 300 may correspond to user device 210 and / or platform 220. Figure 3 As shown, device 300 may include a bus 310 , a processor 320 , a memory 330 , a storage component 340 , an input component 350 , an output component 360 , and a communication interface 370 .

[0036] The bus 310 includes components that allow communication between components of the device 300. The processor 320 is implemented in hardware, firmware, or a combination of hardware and software. The processor 320 is a central processing unit (CPU), a graphics processing unit (GPU), an accelerated processing unit (APU), a microprocessor, a microcontroller, a digital signal processor (DSP), a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), or another type of processing component. In some implementations, the processor 320 includes one or more processors that can be programmed to perform functions. The memory 330 includes a random access memory (RAM), a read-only memory (ROM), and / or another type of dynamic or static storage device (e.g., flash memory, magnetic memory, and / or optical memory) that stores information and / or instructions for use by the processor 320.

[0037] Storage component 340 stores information and / or software related to the operation and use of device 300. For example, storage component 340 may include a hard disk (e.g., a magnetic disk, an optical disk, a magneto-optical disk, and / or a solid-state disk), a compact disk (CD), a digital versatile disk (DVD), a floppy disk, a cassette, a magnetic tape, and / or another type of non-transitory computer-readable medium and a corresponding drive.

[0038] Input components 350 include components that allow device 300 to receive information, for example, via user input (e.g., a touch screen display, a keyboard, a keypad, a mouse, buttons, switches, and / or a microphone). Additionally or alternatively, input components 350 may include sensors for sensing information (e.g., a global positioning system (GPS) component, an accelerometer, a gyroscope, and / or an actuator). Output components 360 include components that provide output information from device 300 (e.g., a display, a speaker, and / or one or more light emitting diodes (LEDs)).

[0039] The communication interface 370 includes transceiver-like components (e.g., a transceiver and / or a separate receiver and transmitter) that enable the device 300 to communicate with other devices, such as via a wired connection, a wireless connection, or a combination of a wired connection and a wireless connection. The communication interface 370 may allow the device 300 to receive information from another device and / or provide information to another device. For example, the communication interface 370 may include an Ethernet interface, an optical interface, a coaxial interface, an infrared interface, a radio frequency (RF) interface, a universal serial bus (USB) interface, a Wi-Fi interface, a cellular network interface, etc.

[0040] Device 300 may perform one or more processes described herein. Device 300 may perform these processes in response to processor 320 executing software instructions stored by non-transitory computer-readable media such as memory 330 and / or storage component 340. Computer-readable media are defined herein as non-transitory memory devices. Memory devices include memory space within a single physical storage device or memory space distributed across multiple physical storage devices.

[0041] The software instructions may be read into the memory 330 and / or storage component 340 from another computer-readable medium or from another device via the communication interface 370. When executed, the software instructions stored in the memory 330 and / or storage component 340 may cause the processor 320 to perform one or more processes described herein. Additionally or alternatively, hardwired circuitry may be used in place of or in combination with software instructions to perform one or more processes described herein. Therefore, the implementations described herein are not limited to any particular combination of hardware circuitry and software.

[0042] Figure 3 The number and arrangement of components shown are provided as examples. Figure 3 The components shown, device 300 may include additional components, fewer components, different components, or differently arranged components. Additionally or alternatively, a component group (e.g., one or more components) of device 300 may perform one or more functions described as being performed by another component group of device 300.

[0043] Figure 4 is a flow chart of an example process 400 for generating an automatic speech recognition (ASR) model using unsupervised learning and using output distribution matching (ODM) techniques. In some implementations, Figure 4 One or more processing blocks of may be executed by platform 220. In some implementations, Figure 4 One or more processing blocks of may be performed by another device or group of devices, such as user device 210, separate from or including platform 220.

[0044] like Figure 4 As shown, processing 400 may include obtaining text information (block 410). For example, platform 220 may receive text information from user device 210, and obtain text information based on receiving text information from user device 210. Additionally or alternatively, platform 220 may implement data extraction techniques (e.g., web mining techniques, web crawling techniques, information retrieval techniques, etc.), and obtain text information based on implementing data extraction techniques. Additionally or alternatively, platform 220 may retrieve text information from a memory location. The text information may include thousands, millions, billions, etc. of data sets. In this way, platform 220 may implement big data techniques to process text information.

[0045] like Figure 4 As further shown, the process 400 may include determining a group of phoneme sequences associated with the text information (block 420). For example, the platform 220 may process the text information and determine a phoneme sequence associated with the text information. Although the implementations herein are used in conjunction with phonemes, it should be understood that the implementations herein are also applicable to other units, such as words, sentences, n-grams, characters, etc.

[0046] In some implementations, the platform 220 can use a lexicon to determine a group of phoneme sequences associated with the text information. For example, a lexicon can refer to a data structure that maps phonemes and word sequences.

[0047] In some implementations, the platform 220 can generate a language model based on determining a group of phoneme sequences associated with the text information. In this way, and as described elsewhere herein, the platform 220 can use the language model when generating an ASR model.

[0048] like Figure 4As further shown, the process 400 may include obtaining speech waveform data (block 430). For example, the platform 220 may receive the speech waveform data from the user device 210. Additionally or alternatively, the platform 220 may obtain the speech waveform data in a similar manner and / or using similar techniques associated with the text information described above. The speech waveform data may constitute unlabeled data. In other words, the speech waveform data may not have been pre-processed by an operator.

[0049] like Figure 4 As further shown, process 400 may include determining phoneme boundary groups associated with the speech waveform data (block 440). For example, platform 220 may determine phoneme boundary groups that divide individual phonemes associated with the speech waveform data.

[0050] In some implementations, and as described elsewhere herein, the platform 220 can iteratively determine phoneme boundaries associated with the speech waveform data. For example, the platform 220 can determine an initial set of phoneme boundaries associated with the speech waveform data, and can iteratively determine updated and / or refined phoneme boundaries after determining the initial set of phoneme boundaries.

[0051] In some implementations, the platform 220 may determine the phoneme boundaries associated with the speech waveform data by training a long short-term memory (LSTM) recurrent neural network (RNN). In addition, the platform 220 may use the activation signal of the forget gate of the LSTM RNN to estimate the phoneme boundaries. Alternatively, the platform 220 may train a gated recurrent unit (GRU) RNN and use the activation signal of the update gate of the GRURNN to estimate the phoneme boundary group.

[0052] like Figure 4 As further shown, the process 400 may include generating an automatic speech recognition (ASR) model using unsupervised learning and using an output distribution matching (ODF) technique based on determining a group of phoneme sequences associated with the text information and a group of phoneme boundaries associated with the speech waveform data (block 450). For example, the platform 220 may generate the ASR model using unsupervised learning.

[0053] In some implementations, the platform 220 may generate an ASR model based on the following objective function:

[0054]

[0055] Referring to the above objective function, “L ODM " can represent the empirical output distribution matching (ODM) cost function. For example, the empirical ODM cost function can be expressed as:

[0056]

[0057] The empirical ODM cost function compares the frequencies of n consecutive letters (e.g., n-grams, where n=1, 2, 3, 6, etc.). For example, the empirical ODM cost function compares the frequencies of tri-grams (e.g., n=3) such as "qui", "uic", "ick", etc.

[0058] Referring to the empirical ODM cost function, “Z” can represent a certain n-gram, “P LM “(z)” may represent the probability of a particular n-gram “Z” (eg, the probability of a tri-gram “qui”), and “LM” may represent a language model (eg, as described above in connection with block 420).

[0059] The empirical prediction probability equation can be expressed as follows:

[0060]

[0061] For example, the empirical prediction probability function shown above can represent the empirical prediction probability for n-gram “Z”, “x bi+εi " can represent a randomly sampled frame in the ith phoneme segment, "b i " can represent the starting boundary of the ith phoneme segment, "ε i " may denote a random variable indexing a frame within the ith segment, and "j" may denote I+n-1 for an n-gram.

[0062] In some implementations, the speech data waveform may be represented as a sequence of segments corresponding to different phonemes. Thus, inputting "x bi+εi… x bj+εj " is sampling from the i-th segment to the j-th segment.

[0063] In some implementations, the outputs corresponding to frames in the same segment should be the same. That is, for the ith phoneme segment, for “x bi+εi "The predicted output phoneme for any random "ε i " should be the same. In this way, the following additional term, denoted as the phoneme smoothing cost function, can be added to the empirical ODM cost function:

[0064]

[0065] Referring to the phoneme smoothing cost function shown above, “PS” may stand for “phoneme smoothing.” The phoneme smoothing cost function may operate, for example, to penalize differences between predicted outputs between different frames.

[0066] In this manner, and with reference to the objective function shown above, the objective function can be expressed as the empirical ODM cost function “L ODM ” and the phoneme smoothing cost function “L PS ". " may represent a regularization parameter that controls the trade-off between the empirical ODM cost function and the phoneme smoothing function.

[0067] After the initial phoneme boundaries are determined, the platform 220 may initiate the optimization of the objective function shown above to train the ASR model. For example, the platform 220 may initialize the phoneme boundaries and may θ Initialize the model parameters (“θ”) in (e.g., the weights of a deep neural network).

[0068] like Figure 4 As further shown, process 400 may include determining whether a training condition is satisfied (block 460). For example, platform 220 may determine whether a training condition is satisfied based on training an ASR model.

[0069] In some implementations, the training condition may refer to a convergence or stopping criterion. As an example, the training condition may refer to a number of iterations that meets a threshold, a number of phoneme boundary refinements that meets a threshold, an optimization parameter that meets a threshold, a time frame that meets a threshold, an accuracy that meets a threshold, etc.

[0070] like Figure 4 As further shown, if the training condition is not met (block 460 —No), the process 400 may include refining the phoneme boundaries (block 470 ). For example, the platform 220 may refine the phoneme boundaries after initializing the phoneme boundaries, as described above in connection with block 440 .

[0071] After the initial phoneme boundaries are determined, the platform 220 may train the ASR model as described above. The platform 220 may determine updated P based on the training of the ASR model. θ (y t |x t ). Based on the determination of the updated P θ (y t |x t ), the platform 220 can further refine the estimate of the phoneme boundaries using the following method. First, the platform 220 can determine the most likely phoneme sequence Y for a given input X using the following equation:

[0072]

[0073] In addition, the following conditions may apply to the above equation:

[0074]

[0075] Referring to the above content, P(b i ≠t) may represent the probability that time t does not correspond to a boundary (eg, no phoneme switch). Using the above two equations, platform 220 may perform a standard beam search method to determine Y 1:n Approximate sequence of .

[0076] Additionally, the platform 220 may determine the refinement boundary by determining the time step t, where Y t ≠Y t-1 The platform 220 may determine a set of refined boundaries and utilize the refined boundaries associated with the above-mentioned objective function. In this manner, the platform 220 may iteratively determine the refined phoneme boundaries and implement the refined phoneme boundaries in association with the objective function. The platform 220 may iteratively perform the above operations until the training condition is met.

[0077] like Figure 4 As further shown, if the training conditions are met (block 460-yes), process 400 may include performing post-processing (block 480). For example, platform 220 may determine that the training conditions are met (e.g., using one or more of the above metrics) and may perform post-processing of the ASR model.

[0078] In some implementations, the platform 220 may apply a set of speech post-processing techniques to improve model performance. For example, the platform 220 may apply the training input to the training model P θ To obtain a predicted output (e.g., a predicted label). In addition, the platform 220 can use the predicted output in combination with the training input to train a hidden Markov model (HMM)-Gaussian mixture model (GMM). Additionally or alternatively, the platform 220 can implement any number of other post-processing techniques such as speaker adaptation to further improve the performance of the ASR model.

[0079] In this way, some implementations herein allow for completely unsupervised training of ASR models. That is, segmentation (e.g., boundary estimation) and neural network training do not require labeled data. For example, some implementations herein allow for model generation without supervised boundary information.

[0080] Compared to supervised learning, some implementations in this article improve model accuracy, reduce the human effort required for model generation, allow ASR models to be developed for multiple languages, reduce human errors, reduce the time associated with model generation, and have other technical benefits.

[0081] Although the implementations herein describe phoneme sequences, it should be understood that other implementations include word sequences, character sequences, etc. as intermediate sequences. In other words, other implementations include direct mapping between speech waveforms and word sequences and / or character sequences.

[0082] Furthermore, in some implementations, Mel-Frequency Cepstral Coefficients (MFCCs) are used as input. It should be appreciated that other implementations include other features, such as filter bank inputs, etc.

[0083] although Figure 4 Example blocks of process 400 are shown, but in some implementations, compared to Figure 4 In addition to the blocks depicted in , process 400 may include additional blocks, fewer blocks, different blocks, or differently arranged blocks. Additionally or alternatively, two or more of the blocks of process 400 may be executed in parallel.

[0084] The foregoing disclosure provides illustration and description, but is not intended to be exhaustive or to limit the implementations to the precise forms disclosed.Modifications and variations are possible in light of the above disclosure or may be acquired from practice of the implementations.

[0085] As used herein, the term component is intended to be broadly interpreted as hardware, firmware, or a combination of hardware and software.

[0086] It will be apparent that the systems and / or methods described herein can be implemented in various forms of hardware, firmware, or a combination of hardware and software. The actual dedicated control hardware or software code used to implement these systems and / or methods does not limit the implementation. Therefore, the operation and behavior of the systems and / or methods are described herein without reference to specific software code - it should be understood that software and hardware can be designed to implement the systems and / or methods based on the description herein.

[0087] Even though particular combinations of features are recited in the claims and / or disclosed in the specification, these combinations are not intended to limit the disclosure of possible implementations. In fact, many of these features may be combined in ways not specifically recited in the claims and / or disclosed in the specification. Although each of the listed dependent claims may be directly dependent on only one claim, the disclosure of possible implementations includes each dependent claim in combination with every other claim in the claim set.

[0088] Any element, action or instruction used in this article should not be interpreted as a key or necessary element, action or instruction, unless clearly described in this way. In addition, as used in this article, "a kind of" and "one" are intended to include one or more items, and can be used interchangeably with "one or more". In addition, as used in this article, the term "group (set)" is intended to include one or more items (for example, related items, irrelevant items, the combination of related items and irrelevant items, etc.), and can be used interchangeably with "one or more". In the case of meaning only one item, the term "one" or similar language is used. In addition, as used in this article, the term "has (has)", "have (have)", "having (having)" etc. are intended to be open terms. In addition, the phrase "based on" is intended to represent "based at least in part on", unless otherwise clearly stated.

Claims

1. A method for generating an automatic speech recognition (ASR) model using unsupervised learning, characterized in that The method comprises: Obtaining text information from the device; determining, by the device, a group of phoneme sequences associated with the text information; Acquiring speech waveform data by the device; Determining, by the device, a phoneme boundary group associated with the speech waveform data using an activation signal of a forget gate of a long short-term memory (LSTM) recurrent neural network (RNN); and The device generates an ASR model using a target function based on the phoneme sequence group and the phoneme boundary group; The device determines whether a training condition is met based on the generated ASR model, wherein the training condition includes a number of phoneme boundary refinements that meets a threshold; The device performs post-processing on the ASR model based on the number of phoneme boundary refinements that meet the threshold; the objective function is L ODM represents the empirical output distribution matching (ODM) cost function, and the objective function is expressed as the empirical ODM cost function L ODM and the phoneme smoothing cost function L PS , λ represents the regularization parameter that controls the trade-off between the empirical ODM cost function and the phoneme smoothing function.

2. The method according to claim 1, characterized in that The determining of the phoneme boundary groups associated with the speech waveform data includes: determining the phoneme boundary groups by the device using a long short-term memory (LSTM) recurrent neural network (RNN).

3. The method according to claim 1, characterized in that: The determining of the phoneme boundary groups associated with the speech waveform data includes: determining the phoneme boundary groups by the device using a gated recurrent unit (GRU) recurrent neural network (RNN).

4. The method according to claim 1, characterized in that: The method further comprises: Another phoneme boundary group associated with the speech waveform data is determined by the device based on the ASR model.

5. The method according to claim 1, characterized in that The method further comprises: Determining, by the device based on the ASR model, whether a training condition is met; and Another phoneme boundary group is selectively determined by the device based on the satisfaction of the training condition.

6. The method according to claim 1, characterized in that The method further comprises: identifying, by the device, a word sequence group associated with the text information; and Wherein, determining the phoneme sequence group associated with the text information includes: The phoneme sequence group is determined by the device based on the word sequence group.

7. The method according to claim 1, characterized in that The method further comprises: comparing, by the device, the group of n-gram frequency values ​​associated with the text information and the group of phoneme frequency values ​​associated with the speech waveform data; and The step of generating the ASR model by the device based on the phoneme sequence group and the phoneme boundary group using an objective function comprises: The device generates the ASR model according to the comparison result of the n-gram frequency value group and the phoneme frequency value group using the ODM technology.

8. The method according to claim 1, characterized in that The speech waveform data is unlabeled.

9. The method according to claim 1, characterized in that: The method further comprises: performing, by the device, a beam searching technique based on generating the ASR model; and A refined phoneme boundary group is generated by the device based on the execution result of the beam search technique.

10. The method according to claim 1, characterized in that The method further comprises: Using the text information to train a language model; and The step of generating the ASR model by the device based on the phoneme sequence group and the phoneme boundary group using an objective function comprises: The ASR model is generated by the device using the language model.

11. A device, characterized in that: The device comprises: at least one memory configured to store program code; and At least one processor is configured to read the program code and execute the method according to any one of claims 1 to 10 according to instructions of the program code.

12. An apparatus for generating an automatic speech recognition (ASR) model using unsupervised learning, characterized in that The device comprises: A first obtaining unit, configured to obtain text information from a device; A first determining unit configured to determine, by the device, a phoneme sequence group associated with the text information; A second obtaining unit, configured to obtain speech waveform data from the device; a second determination unit configured to determine, by the device, a phoneme boundary group associated with the speech waveform data using an activation signal of a forget gate of a long short-term memory (LSTM) recurrent neural network (RNN); and A generating unit configured to generate an ASR model by the device based on the phoneme sequence group and the phoneme boundary group using an objective function; the device determines whether a training condition is satisfied based on the generated ASR model, wherein the training condition includes a number of phoneme boundary refinements that satisfies a threshold; The device performs post-processing on the ASR model based on the number of phoneme boundary refinements that meet the threshold; the objective function is L ODM represents the empirical output distribution matching (ODM) cost function, and the objective function is expressed as the empirical ODM cost function L ODM and the phoneme smoothing cost function L PS , λ represents the regularization parameter that controls the trade-off between the empirical ODM cost function and the phoneme smoothing function.

13. A non-transitory computer-readable medium storing instructions, characterized in that: The instructions include one or more instructions that, when executed by one or more processors of a device, cause the one or more processors to perform the method according to any one of claims 1 to 10.

Images