WO2020182153A1 - 基于自适应语种进行语音识别的方法及相关装置 - Google Patents

基于自适应语种进行语音识别的方法及相关装置 Download PDF

Info

Publication number
WO2020182153A1
WO2020182153A1 PCT/CN2020/078806 CN2020078806W WO2020182153A1 WO 2020182153 A1 WO2020182153 A1 WO 2020182153A1 CN 2020078806 W CN2020078806 W CN 2020078806W WO 2020182153 A1 WO2020182153 A1 WO 2020182153A1
Authority
WO
WIPO (PCT)
Prior art keywords
language
output
phoneme
input
model
Prior art date
Application number
PCT/CN2020/078806
Other languages
English (en)
French (fr)
Inventor
苏丹
傅天晓
罗敏
陈祺
张宇露
罗琳
Original Assignee
腾讯科技（深圳）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 腾讯科技（深圳）有限公司 filed Critical 腾讯科技（深圳）有限公司
Publication of WO2020182153A1 publication Critical patent/WO2020182153A1/zh
Priority to US17/231,945 priority Critical patent/US20210233521A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00—Speech recognition
    • G10L15/005—Language recognition
    • 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
    • G10L15/02—Feature extraction for speech recognition; Selection of recognition unit
    • 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
    • G10L15/22—Procedures used during a speech recognition process, e.g. man-machine dialogue
    • 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
    • G10L17/00—Speaker identification or verification techniques
    • 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
    • G10L17/00—Speaker identification or verification techniques
    • G10L17/04—Training, enrolment or model building
    • 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
    • G10L15/18—Speech classification or search using natural language modelling
    • G10L15/183—Speech classification or search using natural language modelling using context dependencies, e.g. language models
    • G10L15/187—Phonemic context, e.g. pronunciation rules, phonotactical constraints or phoneme n-grams
    • 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
    • G10L15/28—Constructional details of speech recognition systems
    • G10L15/32—Multiple recognisers used in sequence or in parallel; Score combination systems therefor, e.g. voting systems
    • 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
    • G10L15/02—Feature extraction for speech recognition; Selection of recognition unit
    • G10L2015/025—Phonemes, fenemes or fenones being the recognition units

Definitions

  • This application relates to the field of computer application technology, in particular to a speech recognition technology.
  • acoustic models are simultaneously trained to support speech recognition in multiple languages. In this way, because the pronunciations of different languages are directly confused with each other, it will greatly affect the accuracy of speech recognition in different languages, and may cause a worse user experience.
  • this application provides a method and related device for speech recognition based on adaptive languages.
  • a method for speech recognition based on an adaptive language includes: extracting phoneme features representing pronunciation phoneme information based on the acquired speech data; and inputting the phoneme features into pre-trained based on multilingual corpus
  • the language discrimination model obtains the language discrimination result of the speech data; the speech recognition result of the speech data is obtained based on the language acoustic model of the language corresponding to the language discrimination result.
  • an apparatus for adaptive language speech recognition including: an extraction module for extracting phoneme features representing pronunciation phoneme information based on the acquired speech data; a discrimination module for comparing all The phoneme feature input is preliminarily based on a language discrimination model trained on a multilingual corpus to obtain the language discrimination result of the speech data; and a recognition module for obtaining the speech data based on the language acoustic model of the language corresponding to the language discrimination result The result of speech recognition.
  • a voice recognition method based on artificial intelligence including:
  • a voice recognition device based on artificial intelligence including:
  • the acoustic model unit is configured to use the corpus of each language as input to train multiple language acoustic models, and the multiple language acoustic models correspond to different languages;
  • a discriminant model unit for extracting phoneme features of the corpus of each language by using the multiple language acoustic models, and training a language discriminant model based on the phoneme features;
  • the recognition module is configured to perform voice recognition on the collected voice data based on the multiple language acoustic models and the language discrimination model.
  • a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the method according to the above is implemented.
  • an electronic device including: a processor; and a memory, on which computer-readable instructions are stored, and when the computer-readable instructions are executed by the processor, the foregoing method.
  • a voice interaction device including: a collection unit for collecting user voice data; a processing unit for extracting phoneme features representing pronunciation phoneme information based on the voice data; Phoneme feature input is preliminarily based on a language discrimination model trained on a multilingual corpus to obtain the language discrimination result of the speech data; and the speech recognition result of the speech data is obtained based on the language acoustic model of the language corresponding to the language discrimination result; and The interaction unit is configured to present corresponding interaction content to the user based on the voice recognition result of the processing unit.
  • the language is distinguished by the extracted phoneme features, thereby switching to the language acoustic model of the language corresponding to the speech discrimination result, and the speech is obtained based on the language acoustic model of the language corresponding to the speech discrimination result Data speech recognition result.
  • it can automatically switch to different language acoustic models for speech recognition, avoiding the problem of low recognition rate caused by the confusion of different language pronunciations, thereby improving the processing efficiency while ensuring the accuracy of speech recognition.
  • redundant operations caused by the user's choice of language are saved, and processing efficiency and user experience are improved.
  • Figure 1 shows a schematic diagram of a system architecture to which embodiments of the present application can be applied
  • FIG. 2 shows a schematic structural diagram of a computer system suitable for implementing an electronic device according to an embodiment of the present application
  • Figure 3 shows a flow chart of a method for speech recognition based on adaptive language
  • Figure 4 shows an example of a language acoustic model based on a neural network
  • Figure 5 shows a flow chart of a model training method for adaptive language for speech recognition
  • FIG. 6 shows a schematic diagram of a framework for training multiple language acoustic models based on step 510 of the embodiment in FIG. 5;
  • FIG. 7 shows a schematic diagram of a framework for training a mixed language acoustic model based on step 520 in the embodiment of FIG. 5;
  • Figure 8 shows a flow chart of a method for speech recognition based on adaptive language
  • Figure 9 shows a flow chart of a method for speech recognition based on adaptive language
  • Figure 10 shows a block diagram of a device for speech recognition based on adaptive language
  • Figure 11 shows a block diagram of a device for speech recognition based on adaptive language
  • Fig. 12 shows a block diagram of a voice interaction device.
  • AI Artificial Intelligence
  • digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results.
  • artificial intelligence is a comprehensive technology of computer science, which attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a similar way to human intelligence.
  • Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making.
  • Artificial intelligence technology is a comprehensive discipline, covering a wide range of fields, including both hardware-level technology and software-level technology.
  • Basic artificial intelligence technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, and mechatronics.
  • Artificial intelligence software technology mainly includes computer vision technology, speech technology, natural language processing technology, and machine learning/deep learning.
  • the key technologies of speech technology include automatic speech recognition technology (Automatic Speech Recognition, ASR), speech synthesis technology (Text-To-Speech, TTS) and voiceprint recognition technology. Enabling computers to be able to listen, see, speak, and feel is the future development direction of human-computer interaction, and voice has become one of the most promising human-computer interaction methods in the future.
  • ASR Automatic Speech Recognition
  • TTS speech synthesis technology
  • voiceprint recognition technology Enabling computers to be able to listen, see, speak, and feel is the future development direction of human-computer interaction, and voice has become one of the most promising human-computer interaction methods in the future.
  • Machine Learning is a multi-field interdisciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other subjects. Specializing in the study of how computers simulate or realize human learning behaviors in order to acquire new knowledge or skills, and reorganize the existing knowledge structure to continuously improve its own performance.
  • Machine learning is the core of artificial intelligence, the fundamental way to make computers intelligent, and its applications cover all fields of artificial intelligence.
  • Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning techniques.
  • artificial intelligence technology has been researched and applied in many fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned driving, autonomous driving, drones , Robotics, intelligent medical care, intelligent customer service, etc., I believe that with the development of technology, artificial intelligence technology will be applied in more fields and play more and more important values.
  • the solutions provided in the embodiments of the present application may be executed by an electronic device with a voice recognition function based on an adaptive language, and the electronic device may be a terminal device or a server.
  • FIG. 1 takes the voice processing device as a server as an example, and shows a schematic diagram of an exemplary system architecture 100 to which the embodiments of the present application can be applied.
  • the system architecture 100 may include one or more of terminal devices 101, 102, 103, a network 104 and a server 105.
  • the network 104 is used to provide a medium for communication links between the terminal devices 101, 102, 103 and the server 105.
  • the network 104 may include various connection types, such as wired, wireless communication links, or fiber optic cables.
  • the numbers of terminal devices, networks, and servers in FIG. 1 are merely illustrative. According to implementation needs, there can be any number of terminal devices, networks and servers.
  • the server 105 may be a server cluster composed of multiple servers.
  • the user can use the terminal devices 101, 102, 103 to interact with the server 105 through the network 104 to receive or send messages and so on.
  • the terminal devices 101, 102, and 103 may be various electronic devices with display screens, including but not limited to smart phones, tablet computers, portable computers, desktop computers, and so on.
  • the server 105 may be a server that provides various services.
  • the terminal device 101 may collect the user's voice through a voice collection module, and convert it into a digital data form (ie, voice data) and send it to the server 105 through the network 104.
  • the server 105 can extract phoneme features representing pronunciation phoneme information based on the acquired voice data.
  • the phoneme features can reflect the pronunciation phoneme features of different languages, and input the phoneme features into the language discrimination model trained in advance based on the multilingual corpus.
  • the voice recognition result of the voice data is obtained based on the language acoustic model of the language corresponding to the language discrimination result, and returned to the terminal device 101 via the network 104.
  • the method for performing speech recognition based on the adaptive language provided by the embodiments of the present application may be executed by the server 105. Accordingly, the apparatus for performing voice recognition based on the adaptive language may be set in the server 105. In other embodiments, some terminal devices may have similar functions as the server to perform this method. Therefore, the method provided in the embodiments of the present application is not strictly limited to be executed on the server side, that is, the electronic device with the function of performing voice recognition based on the adaptive language may be a terminal device.
  • Fig. 2 shows a schematic structural diagram of a computer system suitable for implementing an electronic device according to an embodiment of the present application.
  • the computer system 200 includes a central processing unit (CPU) 201, which can be based on a program stored in a read only memory (ROM) 202 or a program loaded from a storage part 208 into a random access memory (RAM) 203 And perform various appropriate actions and processing.
  • CPU 201 central processing unit
  • ROM read only memory
  • RAM random access memory
  • various programs and data required for system operation are also stored.
  • the CPU 201, ROM 202, and RAM 203 are connected to each other through a bus 204.
  • An input/output (I/O) interface 205 is also connected to the bus 204.
  • the following components are connected to the I/O interface 205: an input part 206 including a keyboard, a mouse, etc.; an output part 207 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker; a storage part 208 including a hard disk, etc. ; And a communication section 209 including a network interface card such as a LAN card, a modem, etc.
  • the communication section 209 performs communication processing via a network such as the Internet.
  • the drive 210 is also connected to the I/O interface 205 as needed.
  • a removable medium 211 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 210 as needed, so that the computer program read from it is installed into the storage section 208 as needed.
  • the process described below with reference to the flowchart can be implemented as a computer software program.
  • the embodiments of the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program contains program code for executing the method shown in the flowchart.
  • the computer program may be downloaded and installed from the network through the communication part 209, and/or installed from the removable medium 211.
  • the central processing unit (CPU) 201 various functions defined in the embodiments of the present application are executed.
  • the computer-readable medium shown in this application may be a computer-readable signal medium or a computer-readable storage medium or any combination of the two.
  • the computer-readable storage medium may be, for example, but is not limited to an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable Programmable read only memory (EPROM or flash memory), optical fiber, portable compact disk read only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.
  • the computer-readable storage medium may be any tangible medium that contains or stores a program, and the program may be used by or in combination with an instruction execution system, apparatus, or device.
  • a computer-readable signal medium may include a data signal propagated in a baseband or as a part of a carrier wave, and a computer-readable program code is carried therein. This propagated data signal can take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing.
  • the computer-readable signal medium may also be any computer-readable medium other than the computer-readable storage medium.
  • the computer-readable medium may send, propagate or transmit the program for use by or in combination with the instruction execution system, apparatus, or device .
  • the program code contained on the computer-readable medium can be transmitted by any suitable medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.
  • this application also provides a computer-readable medium.
  • the computer-readable medium may be included in the electronic device described in the above-mentioned embodiments; or it may exist alone without being assembled into the electronic device. in.
  • the above-mentioned computer-readable medium carries one or more programs.
  • the electronic device realizes the method described in the following embodiments. For example, the electronic device can implement the steps shown in FIGS. 3 to 9.
  • Fig. 3 is a flow chart showing a method for speech recognition based on an adaptive language type according to an exemplary embodiment. As shown in FIG. 3, the method for speech recognition based on adaptive language can be executed by any electronic device, and can include the following steps 310-330.
  • step 310 phoneme features representing pronunciation phoneme information are extracted based on the acquired speech data.
  • the voice data here refers to the digital format data obtained after the voice collection module of the electronic device collects and processes the user's voice.
  • Electronic devices include, but are not limited to, devices such as smart phones, tablet computers, personal computers, notebook computers, and the like.
  • Voice collection modules include, for example, components such as microphones and sound cards of these electronic devices.
  • the electronic device may use a preset feature extraction algorithm to process the aforementioned voice data to obtain corresponding phoneme features.
  • the phoneme feature reflects the pronunciation phoneme information, and the language of the speech data can be identified through the phoneme feature.
  • the phoneme feature can be, for example, the bottleneck bottleneck feature extracted from the bottleneck layer of the language acoustic model.
  • step 320 the phoneme feature is input to the language discrimination model obtained by training based on the multilingual corpus in advance to obtain the language discrimination result of the speech data.
  • One of the ways to realize multilingual speech recognition in related technologies is to manually select the language by the user so that the speech recognition product can switch to the corresponding language acoustic model for recognition, but this increases the user’s operational burden and reduces Improve the user experience and the processing efficiency of speech recognition.
  • the embodiment of the present application can automatically determine the language to which the voice data belongs based on the extracted phoneme features to obtain the language discrimination result of the voice data, so as to automatically switch to the language acoustic recognition model corresponding to the language discrimination result to recognize the voice data. .
  • Phoneme is the smallest phonetic unit divided according to the natural attributes of speech. From the perspective of acoustic properties, phoneme is the smallest phonetic unit divided from the perspective of sound quality. From the physiological point of view, a pronunciation action forms a phoneme. For example, [ma] contains [m][a] two pronunciation actions, which are two phonemes. The sounds produced by the same pronunciation action are the same phoneme, and the sounds produced by different pronunciation actions are different phonemes. For example, in [ma-mi], two [m] pronunciation actions are the same, they are the same phoneme, and [a][i] pronunciation actions are different, they are different phonemes.
  • Phonemes are generally divided into two categories: vowels and consonants. Different languages can be divided into different pronunciation phonemes. Take Mandarin Chinese as an example, including 22 consonants and 10 vowels; while the English International Phonetic Alphabet has 48 phonemes, including 20 vowel phonemes and 28 consonant phonemes.
  • the embodiment of the present application extracts phoneme features representing pronunciation phoneme information in the speech data, and inputs a language discrimination model pre-trained based on a multilingual corpus to realize the discrimination of the language of the speech data.
  • a language discrimination model pre-trained based on a multilingual corpus to realize the discrimination of the language of the speech data.
  • step 330 the voice recognition result of the voice data is obtained based on the language acoustic model of the language corresponding to the language discrimination result.
  • one of the ways in related technologies to realize multilingual speech recognition is to train an acoustic model to support speech recognition in multiple languages by including a set of multilingual mixed acoustic pronunciation units. Because the pronunciation of different languages is directly confused with each other, it will greatly affect the accuracy of speech recognition in different languages.
  • the embodiment of the present application is based on the language discrimination result in step 320, and obtains the speech recognition result output by the language acoustic model of the corresponding language according to the discriminated language.
  • the language acoustic model here is used to detect words matching the pronunciation from the voice data, and then determine the subsequent response mode, for example, recognize the corresponding voice command to interact with the smart device.
  • the speech acoustic model can be obtained by training the initial model on the known corpus and its speech features.
  • the initial model can be obtained through CNN (Convolutional Neural Network, Convolutional Neural Network), DNN (Deep Neural Network, Deep Neural Network) ) And other neural networks or their combined networks.
  • Figure 4 shows an example of a speech acoustic model based on a neural network.
  • the model includes a first input layer 410, a first hidden layer 420, and a first output layer 430.
  • the first input layer 410 is used to receive the input of phoneme features, where the number of nodes depends on the number of phone features;
  • the first hidden layer 420 is used to process the input phoneme features, which may include a multi-layer network (multiple sub-hidden layers).
  • each layer of the network can include multiple nodes. The number of network layers and nodes can be determined according to the size of the training corpus and the computing power of the device.
  • the first output layer 430 includes multiple output nodes, which may correspond to different phonemes, for example.
  • the output probability of each node is calculated, which represents the probability that the input phoneme feature belongs to the phoneme corresponding to each node.
  • the phoneme features extracted based on the known corpus are input into the first input layer 410, and the probability results of the output nodes in the first output layer 430 are calculated, so that
  • the objective loss function (for example, the softmax function) is minimized as the objective, and the first hidden layer 420 that can accurately express the phoneme characteristics of the input corpus can be trained to obtain a trained language acoustic model.
  • the speech acoustic model For training of the speech acoustic model, refer to the embodiments shown in FIG. 5 and FIG. 6.
  • the embodiment of the present application trains multiple language acoustic models by language.
  • the speech recognition result is output based on the language acoustic model of the corresponding language in step 330, which can avoid the low recognition rate due to the confusion of the pronunciation of different languages when the mixed language acoustic model is used. Therefore, while improving the processing efficiency, the accuracy of speech recognition is ensured.
  • the language is distinguished by the extracted phoneme features, so as to switch to the language acoustic model of the language corresponding to the speech discrimination result, and the language acoustic model of the language corresponding to the speech discrimination result is obtained Voice recognition result of voice data.
  • it can automatically switch to different language acoustic models for speech recognition, avoiding the problem of low recognition rate caused by the confusion of different language pronunciations, thereby improving the processing efficiency while ensuring the accuracy of speech recognition.
  • redundant operations caused by the user's choice of language are saved, and processing efficiency and user experience are improved.
  • Fig. 5 is a flow chart showing a model training method based on adaptive language for speech recognition according to an exemplary embodiment.
  • the model training method for speech recognition based on adaptive language can be executed by any electronic device, and can include the following steps 510-540.
  • step 510 the corpus of each language is used as input to train multiple language acoustic models, and the multiple language acoustic models correspond to different languages.
  • the corpus here includes phoneme feature sets obtained by respectively preprocessing and feature extraction of user speech belonging to each language.
  • the preprocessing includes, but is not limited to, sampling and quantization of user voice and endpoint detection (Voice Activity Detection, VAD), etc.
  • VAD Voice Activity Detection
  • VAD here refers to detecting the presence or absence of speech in a noisy environment, which can be used in systems such as speech coding to reduce the speech coding rate, save communication bandwidth, and improve the recognition rate.
  • VAD can be performed on various sounds acquired in the environment through a collection device such as a built-in or external microphone, and the user's voice can be detected from it.
  • the electronic device detects the user's voice, it further determines the starting point of the voice, and then starts to collect the voice in the environment and form voice data in digital form.
  • the high-frequency resolution of the voice data is improved, the voice data becomes smoother, and the subsequent processing of the voice data is convenient.
  • the feature extraction includes, but is not limited to, for example, removing redundant parts in the voice data to extract parameters that can represent the essential features of the voice data.
  • a neural network similar to that shown in Figure 4 can be used for training to obtain a language acoustic model of the corresponding language.
  • FIG. 6 shows a schematic diagram of the framework of training multiple language acoustic models based on step 510.
  • the extracted phoneme feature set is input to the first input layer 611; the output value output to the first hidden layer 612 is calculated in the first input layer 611 based on the preset first weight matrix ;
  • the first hidden layer 612 includes multiple sub-hidden layers, and each sub-hidden layer receives the output value of the next sub-hidden layer (or the adjacent first input layer 611), and uses the weight matrix of this layer (such as the second weight matrix) Perform weighted calculation and output the result to the previous sub-hidden layer (or the adjacent first output layer 613); the first output layer 613 includes a plurality of output elements.
  • each output element of the first output layer 613 corresponds to a phoneme included in language 1.
  • the first output layer 613 receives the output value of the adjacent sub-hidden layer, and uses the weight matrix of this layer (such as the third weight matrix) to perform weighting calculation, and uses the loss function to calculate the output probability (such as the first output Probability), the first output probability represents the probability that the input phoneme feature belongs to the pronunciation phoneme to which each output element belongs.
  • the loss function here may include a softmax function, for example.
  • the target probability value (for example, 0 or 1) belonging to each pronunciation phoneme can be determined in advance, and the first input layer 611, the first hidden layer 612, and the first input layer 611 can be continuously adjusted through the above training process.
  • the weight matrix of the first output layer 613 (respectively the first weight matrix, the second weight matrix and the third weight matrix) finally meets the preset conditions (for example, the training reaches the preset number of iterations or the error with the target probability falls within the preset Set range) language acoustic model 610.
  • the corpus of the corresponding language can be input in the same way, and the language acoustic model 620-6N0 of the corresponding language can be trained to obtain the speech recognition result of the corresponding language in step 330.
  • the corresponding voice recognition result can be output based on the language acoustic model 620.
  • the speech recognition result here may include, for example, which pronunciation phoneme of language 2 the input speech belongs to, and the text information corresponding to the speech can be obtained after further processing, which will not be repeated here.
  • step 520 a mixed language corpus including multiple languages is used as input to train a mixed language acoustic model supporting multiple languages.
  • a mixed language acoustic model supporting multiple languages at the same time is also trained. Different from using a mixed-language acoustic model in the related art to perform speech recognition on user speech that may include multiple languages, in the embodiment of the present application, the mixed-language acoustic model is used to support the language discrimination in step 320 above.
  • the training process of the mixed language acoustic model is similar to the language acoustic model of each language in step 510.
  • the following is only a brief introduction, and the repetitive parts will not be repeated.
  • the mixed corpus here includes, for example, a phoneme feature set obtained by preprocessing and feature extraction of user speech in various languages.
  • a neural network similar to that shown in Figure 4 can also be used for training to obtain the final mixed language acoustic model.
  • FIG. 7 shows a schematic diagram of the framework of training a mixed language acoustic model based on step 520.
  • the extracted phoneme feature set is input to the input layer 710; the output value output to the hidden layer 720 is calculated based on the preset weight matrix in the input layer 710; the hidden layer 720 includes multiple sub-hidden Layer, each sub-hidden layer receives the output value of the next sub-hidden layer (or the adjacent input layer 710), uses the weight matrix of this layer to perform weighting calculations and outputs the result to the previous sub-hidden layer (or adjacent output layer) );
  • the output layer 730 includes multiple output modules, and each output module includes multiple output elements.
  • each output module 731 to 73N of the output layer 730 corresponds to a language 1-N, and each output element in an output module corresponds to a pronunciation phoneme included in the language.
  • the output layer 730 receives the output value of the adjacent sub-hidden layer, and uses the weight matrix of this layer to perform weighting calculation.
  • the output probability is calculated based on the result of the weighted calculation using the loss function.
  • the output probability indicates that the input voice feature belongs to each output module
  • the loss function here may include a softmax function, for example.
  • the target probability value (for example, 0 or 1) belonging to each pronunciation phoneme can be determined in advance, and the input layer 710, hidden layer 720 and output layer 730 can be continuously adjusted through the above training process Finally, a mixed language acoustic model 700 that satisfies the preset condition (for example, the training reaches the preset number of iterations or the error with the target probability value falls within the preset range) is finally obtained.
  • the preset condition for example, the training reaches the preset number of iterations or the error with the target probability value falls within the preset range
  • the final trained mixed language acoustic model 700 is used to obtain the language discrimination result in step 320. For example, based on the phoneme features extracted from the current user's speech in step 310, the input mixed language acoustic model 700 can get the probability that it belongs to the pronunciation phoneme of each output element in each output module, and accordingly can get the phoneme that belongs to each language. Probability.
  • the essential difference between the embodiments of this application and related technologies is that for user speech that may include multiple languages, input the mixed language acoustic model trained here, and the result obtained is not used to output speech recognition results, but to Determine the language to which it belongs, and further output the speech recognition results based on the language acoustic model of the corresponding language.
  • step 530 the phoneme feature of the mixed corpus is extracted based on the mixed language acoustic model, and the language discrimination model is trained based on the phoneme feature.
  • step 540 the phoneme features of the corpus of each language are extracted from the multiple language acoustic models, and the language discrimination model is assisted based on the phoneme features.
  • steps 510 and 520 can be trained to obtain the language acoustic model and the mixed language acoustic model used in steps 330 and 320, respectively.
  • the application may further include steps 530 and 540 to train the language discrimination model based on the phoneme features extracted from the above-mentioned mixed language acoustic model and language acoustic model, so as to further improve the accuracy of the language discrimination in step 320.
  • steps 530 and 540 to train the language discrimination model based on the phoneme features extracted from the above-mentioned mixed language acoustic model and language acoustic model, so as to further improve the accuracy of the language discrimination in step 320.
  • steps 530 and 540 to train the language discrimination model based on the phoneme features extracted from the above-mentioned mixed language acoustic model and language acoustic model, so as to further improve the accuracy of the language discrimination in step 320.
  • steps 530 and 540 to train the language discrimination model based on the phoneme features extracted from the above-mentioned mixed language acoustic model and language acoustic model, so as to further improve the accuracy of
  • multiple language acoustic models and language discriminant models can be applied to the embodiment corresponding to FIG. 3, so as to obtain the collected data based on multiple language acoustic models and language discriminant models.
  • Voice data for voice recognition can be applied to the embodiment corresponding to FIG. 3, so as to obtain the collected data based on multiple language acoustic models and language discriminant models.
  • the language discrimination model includes a second input layer, a second hidden layer, and a second output layer.
  • the training process of the language discrimination model is similar to the language acoustic model of each language in step 510. The following is only a brief introduction and repetitions No longer.
  • the input of the language discrimination model may include phoneme features extracted through the language acoustic model and the mixed language acoustic model. Based on the phoneme features extracted by the bottleneck layer, a neural network similar to that shown in Figure 4 can also be used for training to obtain the final language discrimination model.
  • the extracted phoneme features are input to the second input layer 810; the second input layer 810 is based on the preset fourth weight Matrix calculation of the output value output to the second hidden layer 820; the second hidden layer 820 uses the weight matrix of this layer (such as the fifth weight matrix) to perform weighting calculation and output the result to the adjacent second output layer 830; second The output layer 830 includes a plurality of output elements.
  • each output element of the second output layer 830 corresponds to a language.
  • the second output layer 830 receives the output value of the adjacent sub-hidden layer, and uses the weight matrix of this layer (for example, the sixth weight matrix) to perform weighting calculation.
  • the output probability is calculated based on the result of the weighting calculation using the loss function, and the output probability represents The probability that the input phoneme feature belongs to the language of each output element.
  • the loss function here may include a softmax function, for example.
  • the target probability value (for example, 0 or 1) belonging to each language can be determined in advance, and the second input layer 810, the second hidden layer 820, and the second input layer can be continuously adjusted through the above training process.
  • the weight matrix of the output layer 830 (respectively the fourth weight matrix, the fifth weight matrix and the sixth weight matrix) finally meets the preset conditions (for example, the training reaches the preset number of iterations or the error with the target probability value falls within the preset Range) of the language discrimination model 800.
  • the finally trained language discrimination model 800 can also be used to obtain the language discrimination result in step 320. For example, based on the voice features extracted from the current user's voice in step 310, the phoneme features are input into the language discrimination model 800 to obtain the probability that they belong to the language of each output element.
  • the trained mixed language acoustic model or language discriminant model can be used to determine the language to which the user’s voice belongs, and then the speech recognition can be obtained based on the language acoustic model of the corresponding language.
  • Fig. 8 is a flow chart showing a method for speech recognition based on an adaptive language type according to another exemplary embodiment. As shown in FIG. 8, the method for speech recognition based on adaptive language can be executed by any electronic device, and can include the following steps 911-913.
  • step 911 phoneme features are extracted based on the acquired voice data.
  • step 310 For this step, refer to step 310 in the foregoing embodiment.
  • step 912 the phoneme characteristics are respectively input to multiple language acoustic models corresponding to different languages, and the language discrimination result of the speech data is obtained according to the phoneme characteristics.
  • step 913 among the speech recognition results returned from the multiple language acoustic models, the speech recognition result of the language acoustic model corresponding to the language discrimination result is selected for output.
  • the language discrimination model and multiple language acoustic models are simultaneously input based on the phoneme features extracted in step 911.
  • Multiple language acoustic models start processing based on the input voice data and output their respective speech recognition results; at the same time, the language discrimination model outputs the language discrimination results based on the input phoneme characteristics, and then selects the corresponding language from the multiple language acoustic models based on the language discrimination results The speech recognition result output of the speech acoustic model.
  • the language discrimination model here may be trained based on steps 530 and 540 of the embodiment shown in FIG. 5, for example, and multiple language acoustic models may be trained based on steps 510 and 520 of the embodiment shown in FIG. 5, for example.
  • the phoneme features extracted in step 911 are input into multiple language acoustic models 620-6N0.
  • the above-mentioned voice features are first input to the first input layer 611; the output value output to the first hidden layer 612 is calculated in the first input layer 611 based on the weight matrix obtained by training; the first hidden layer 612 includes multiple sub- Hidden layer, each sub-hidden layer receives the output value of the next sub-hidden layer (or the adjacent first input layer 611), uses the trained weight matrix of this layer to perform weighting calculation and outputs the result to the previous sub-hidden layer (or The adjacent first output layer 613); the first output layer 613 includes a plurality of output elements, and each output element corresponds to a pronunciation phoneme included in language 1.
  • the first output layer 613 receives the output value of the adjacent sub-hidden layer, and uses the trained weight matrix of this layer to perform weighting calculation.
  • the loss function is used to calculate the output probability based on the result of the weighting calculation.
  • the output probability indicates that the input voice feature belongs to The probability that each output element belongs to the pronunciation phoneme.
  • the phoneme features extracted in step 911 are also input into the language type discrimination model 800 at the same time.
  • the above-mentioned phoneme features are first input to the second input layer 810; in the second input layer 810, the output value to the second hidden layer 820 is calculated based on the weight matrix obtained by training; the second hidden layer 820 uses the trained weight matrix of this layer to perform Weighted calculation and output to the second output layer 830; the second output layer 830 includes a plurality of output elements, each output element corresponding to a language.
  • the second output layer 830 receives the output value of the second hidden layer 820, and uses the trained weight matrix of this layer to perform weighting calculation.
  • the output probability is calculated based on the result of the weighted calculation using the loss function.
  • the output probability indicates that the input phoneme feature belongs to The probability of each language.
  • step 913 the language to which the current user language belongs is determined based on the output element with the highest output probability in the language discrimination model 800, and based on this, one of the language acoustic models 620-6N0 can be selected to output the speech recognition result.
  • the text information corresponding to the speech can be obtained after further processing, which will not be repeated here.
  • the language is distinguished by the extracted phoneme features, and the speech recognition of multiple languages is performed at the same time, and then the speech data is obtained from the language acoustic model of the language corresponding to the speech discrimination result Speech recognition results.
  • it can automatically switch to different language acoustic models for speech recognition, avoiding the problem of low recognition rate caused by the confusion of different language pronunciations, thereby improving the processing efficiency while ensuring the accuracy of speech recognition.
  • redundant operations caused by the user's choice of language are saved, and processing efficiency and user experience are improved.
  • the speech recognition of different languages is performed at the same time as the speech discrimination, the processing speed of the speech recognition can be further improved.
  • Fig. 9 is a flowchart showing a method for self-adaptive language type speech recognition according to still another exemplary embodiment. As shown in Figure 9, the method for adaptive language speech recognition can be executed by any computing device, and can include the following steps 921-923.
  • step 921 phoneme features are extracted based on the acquired voice data.
  • step 310 For this step, refer to step 310 in the foregoing embodiment.
  • step 922 the language discrimination result of the speech data is obtained according to the phoneme characteristics.
  • step 923 from a plurality of language acoustic models respectively corresponding to different languages, the language acoustic model of the language corresponding to the language discrimination result is selected, and the phoneme features are input to obtain the speech recognition result.
  • the language discrimination model is first input, and the language discrimination model outputs the language discrimination result based on the input phoneme feature, and then according to the language discrimination result, the speech feature is input to the language of the language corresponding to the language discrimination result Acoustic model to obtain the corresponding speech recognition results.
  • the language discrimination model here may be trained based on steps 530 and 540 of the embodiment shown in FIG. 5, for example, and multiple language acoustic models may be trained based on steps 510 and 520 of the embodiment shown in FIG. 5, for example.
  • the phoneme features of the speech features extracted in step 921 are input to the language discrimination model 800; first, the second input layer 810 is calculated based on the weight matrix obtained by training to the second The output value of the hidden layer 820; the second hidden layer 820 uses the trained weight matrix of this layer to perform weighting calculation and outputs the result to the second output layer 830; the second output layer 830 includes a plurality of output elements, each output element Corresponds to a language.
  • the second output layer 830 receives the output value of the second hidden layer 820, and uses the trained weight matrix of this layer to perform weighting calculation.
  • the output probability is calculated based on the result of the weighted calculation using the loss function.
  • the output probability indicates that the input phoneme feature belongs to The probability of each language.
  • step 923 the language to which the current user language belongs is determined based on the output element with the highest output probability in the language discrimination model 800, and based on this, one of the language acoustic models 620-6N0 can be selected to input the aforementioned phoneme features.
  • the phoneme feature is input into the language acoustic model 610 corresponding to language type 1 in step 923 based on the speech discrimination result.
  • the above-mentioned phoneme features are first input to the first input layer 611; the output value output to the first hidden layer 612 is calculated in the first input layer 611 based on the weight matrix obtained by training; the first hidden layer 612 includes a plurality of sub-hidden layers, each of which is hidden The layer receives the output value of the next sub-hidden layer (or the adjacent first input layer 611), uses the trained weight matrix of this layer to perform weighting calculation and outputs the result to the previous sub-hidden layer (or the adjacent first output Layer 613); the first output layer 613 includes a plurality of output elements, and each output element corresponds to a pronunciation phoneme included in language 1.
  • the first output layer 613 receives the output value of the adjacent sub-hidden layer, and uses the trained weight matrix of this layer to perform weighting calculation.
  • the loss function is used to calculate the output probability based on the result of the weighting calculation.
  • the output probability indicates that the input voice feature belongs to The probability that each output element belongs to the pronunciation phoneme.
  • the text information corresponding to the speech can be obtained after further processing, which will not be repeated here.
  • the language is distinguished by the extracted phoneme features, and then the language acoustic model of the corresponding language is selected for speech recognition according to the speech discrimination result, and the speech recognition result is output to avoid pronunciation in different languages
  • the problem of low recognition rate caused by mutual confusion thereby improving the processing efficiency while ensuring the accuracy of speech recognition.
  • redundant operations caused by the user's choice of language are saved, and processing efficiency and user experience are improved.
  • the speech recognition by language is performed on the basis of speech discrimination, the processing amount of speech recognition can be saved and the processing efficiency can be improved.
  • Fig. 10 is a block diagram showing an apparatus for speech recognition based on an adaptive language type according to an exemplary embodiment.
  • the apparatus for adaptive language for speech recognition can be implemented by any electronic device, and can include the following extraction module 1010, discrimination module 1020, and recognition module 1030.
  • the extraction module 1010 is configured to extract phoneme features representing pronunciation phoneme information based on the acquired speech data.
  • the discrimination module 1020 is configured to input the phoneme feature into a language discrimination model trained based on a multilingual corpus in advance to obtain a language discrimination result of the speech data.
  • the recognition module 1030 is configured to obtain the voice recognition result of the voice data based on the language acoustic model of the language corresponding to the language discrimination result.
  • the extraction module 1010 is further configured to input the phoneme features into multiple language acoustic models corresponding to different languages; the recognition module 1030 is also configured to return speech recognition results from the multiple language acoustic models Select the speech recognition result output of the language acoustic model of the language corresponding to the language discrimination result.
  • the recognition module 1030 is further configured to select the language acoustic model of the language corresponding to the language discrimination result from a plurality of language acoustic models corresponding to different languages, and input the phoneme feature to obtain the speech recognition result.
  • the apparatus for performing speech recognition based on adaptive languages further includes a recognition model training module, which is configured to use the corpus of each language as input to train multiple language acoustic models. Acoustic models respectively correspond to different languages; and a discriminant model training module for extracting phoneme features of the corpus of each language by using the multiple language acoustic models, and training the language discriminant model based on the phoneme features.
  • a recognition model training module which is configured to use the corpus of each language as input to train multiple language acoustic models. Acoustic models respectively correspond to different languages; and a discriminant model training module for extracting phoneme features of the corpus of each language by using the multiple language acoustic models, and training the language discriminant model based on the phoneme features.
  • the apparatus for performing speech recognition based on adaptive languages it further includes a discriminant model training module for using a mixed corpus including multiple languages as input to train a mixed language supporting the multiple languages
  • the acoustic model extracts phoneme features of the mixed corpus based on the mixed language acoustic model, and trains the language discrimination model based on the phoneme features.
  • the speech acoustic model includes a first input layer, a first hidden layer and a first output layer
  • the recognition model training module is used for:
  • the output value of the adjacent first input layer or the previous sub-hidden layer is received, and the corresponding second weight matrix is used for weighting calculation, and the result is output To the adjacent first output layer or the next sub-hidden layer;
  • the first output layer including a plurality of first output elements receives the output values of the adjacent sub-hidden layers, uses the third weight matrix to perform weighting calculation, and obtains the first output probability based on the calculation result, the first output probability Represents the probability that the input phoneme feature belongs to the pronunciation phoneme to which each first output element belongs;
  • the difference between the first output probability and the target probability iteratively adjust the first weight matrix, the second weight matrix, and the third weight matrix to obtain the speech acoustic model meeting preset conditions .
  • the language discrimination model includes a second input layer, a second hidden layer and a second output layer
  • the discrimination model training module is used for:
  • a sixth weight matrix is used for weighting calculation, and a second output probability is obtained based on the calculation result.
  • the second output probability represents the probability that the input phoneme feature belongs to the language of each second output element ;
  • the phoneme feature includes a bottleneck feature.
  • the language is distinguished by the extracted phoneme features, thereby switching to the language acoustic model of the language corresponding to the speech discrimination result, and the speech is obtained based on the language acoustic model of the language corresponding to the speech discrimination result Data speech recognition result.
  • it can automatically switch to different language acoustic models for speech recognition, avoiding the problem of low recognition rate caused by the confusion of different language pronunciations, thereby improving the processing efficiency while ensuring the accuracy of speech recognition.
  • redundant operations caused by the user's choice of language are saved, and processing efficiency and user experience are improved.
  • Fig. 11 is a block diagram showing an apparatus for speech recognition based on an adaptive language type according to another exemplary embodiment.
  • the apparatus for speech recognition based on adaptive language can be implemented by any electronic device.
  • the discrimination module 1020 includes a discrimination model unit 1021
  • the recognition module 1030 includes multiple acoustic model units. 1031.
  • the acoustic model unit 1031 uses the corpus of each language as an input to train multiple language acoustic models, and the multiple language acoustic models correspond to different languages.
  • the recognition module 1030 also uses a mixed corpus including multiple languages as input to train a mixed language acoustic model supporting multiple languages.
  • the discrimination model unit 1021 extracts phoneme features of the mixed corpus based on the above-mentioned mixed language acoustic model, and trains a language discrimination model based on the phoneme characteristics. In one embodiment, the discriminant model unit 1021 further extracts phoneme features of the corpus of each language from multiple language acoustic models, and trains the language discriminant model based on the phoneme features.
  • the recognition module 1030 is configured to perform voice recognition on the collected voice data based on the multiple language acoustic models and the language discrimination model.
  • the extraction module 1010 extracts phoneme features in the phoneme features based on the mixed language acoustic model trained by the recognition module 1030, and inputs them to the language discrimination model trained by the discriminant model unit 1021 to obtain a language discrimination result.
  • the aforementioned phoneme feature includes a bottleneck feature.
  • the extraction module 1010 sends the above-mentioned speech features to the multiple language acoustic models trained by the multiple acoustic model units 1031, and at the same time sends them to the language discrimination model trained by the discrimination model unit 1021 to obtain the language discrimination result;
  • the module 1030 selects and outputs the speech recognition result of the language acoustic model corresponding to the language of the speech recognition result from the speech recognition results returned by the multiple acoustic model units 1031 according to the language discrimination result.
  • the extraction module 1010 sends the above-mentioned speech features to the language discrimination model trained by the discrimination model unit 1021 to obtain the language discrimination result; and then the recognition module 1030 selects from a plurality of acoustic model units 1031 according to the language discrimination result The speech acoustic model of the language corresponding to the speech discrimination result, and input the aforementioned speech features to obtain the speech recognition result.
  • the language is distinguished by the extracted speech features and the speech recognition of multiple languages is performed at the same time, thereby switching to the language acoustic model of the language corresponding to the speech discrimination result, based on the speech discrimination
  • the speech acoustic model of the corresponding language obtains the speech recognition result of the speech data.
  • it can automatically switch to different language acoustic models for speech recognition, avoiding the problem of low recognition rate caused by the confusion of different language pronunciations, thereby improving the processing efficiency while ensuring the accuracy of speech recognition.
  • redundant operations caused by the user's choice of language are saved, and processing efficiency and user experience are improved.
  • Fig. 12 is a block diagram showing a voice interaction device according to an exemplary embodiment.
  • the voice interaction device of this embodiment includes a collection unit 1201, a processing unit 1202, and an interaction unit 1203.
  • the voice interaction device here may include, but is not limited to, any device among mobile terminals, smart speakers, smart TVs, and smart wearable devices, for example.
  • the collection unit 1201 is used to collect the user's voice data.
  • the collection unit here may include, for example, a microphone, a voice collection chip, etc., through preliminary processing of the voice data, to obtain digitized voice data.
  • the collection unit 1201 is not limited to only collecting the user's voice data. For example, it can also collect all environmental sounds, and obtain the data belonging to the voice through preliminary analysis and screening.
  • the processing unit 1202 is configured to extract phoneme features representing pronunciation phoneme information based on the speech data; input the phoneme features into a language discrimination model pre-trained on a multilingual corpus to obtain the language type of the speech data Discriminating result; and obtaining the speech recognition result of the speech data based on the language acoustic model of the language corresponding to the language discriminating result.
  • a language discrimination model pre-trained on a multilingual corpus to obtain the language type of the speech data Discriminating result
  • obtaining the speech recognition result of the speech data based on the language acoustic model of the language corresponding to the language discriminating result.
  • the interaction unit 1203 is configured to present corresponding interaction content to the user based on the voice recognition result of the processing unit 1202.
  • the interaction unit 1203 here may include, for example, any combination of an image output device and a voice output device.
  • the interaction unit 1203 may directly display the recognized voice content in the form of text.
  • the interaction unit 1203 may also generate reply content based on the voice recognition result, and present the reply content to the user in the form of text or voice.
  • the processing unit 1202 recognizes the user's voice data as an operation instruction
  • the interaction unit 1203 may also present the execution result of the operation instruction to the user.
  • the interaction unit 1203 may also be used to display the language discrimination result of the processing unit 1202 for the user to perform confirmation or modification selection operations. Then, based on the received user selection operation, the interaction unit 1203 may notify the processing unit 1202 to adjust the language discrimination result, and obtain the adjusted speech recognition result from the processing unit 1202.
  • the language type is distinguished based on the extracted speech features, and the speech recognition result of the speech data is obtained based on the language acoustic model of the language corresponding to the speech judgment result.
  • it can automatically switch to different language acoustic models for speech recognition, avoiding the problem of low recognition rate caused by the confusion of different language pronunciations, so as to improve the processing efficiency while ensuring the accuracy of speech recognition and saving
  • the redundant operations caused by the user's language selection are improved, and the processing efficiency and user experience are improved.
  • automatic sentence-level recognition can be realized.
  • modules or units of the device for action execution are mentioned in the above detailed description, this division is not mandatory.
  • the features and functions of two or more modules or units described above may be embodied in one module or unit.
  • the features and functions of a module or unit described above can be further divided into multiple modules or units to be embodied.
  • a component displayed as a module or unit may or may not be a physical unit, that is, it may be located in one place, or may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the objectives of the solutions of the present disclosure.

Landscapes

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

Abstract

一种用于自适应语种进行语音识别的方法，包括：基于获取的语音数据提取表示发音音素信息的音素特征（310）；将音素特征输入预先基于多语种语料训练得到的语种判别模型，得到语音数据的语种判别结果（320）；以及基于语种判别结果所对应语种的语言声学模型获取语音数据的语音识别结果（330）。该方法根据不同的语种可以自动切换到不同的语言声学模型进行语音识别，避免不同语言发音相互混淆导致的识别率较低的问题，并且在提高处理效率的同时，保证了语音识别的准确率。

Description

基于自适应语种进行语音识别的方法及相关装置
本申请要求于2019年3月11日提交中国专利局、申请号201910182266.5、申请名称为“用于自适应语种进行语音识别的方法、装置及设备”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及计算机应用技术领域，特别涉及一种语音识别技术。
背景技术
随着语音识别技术的飞速发展，目前的语音识别准确度已达到实际应用的水平，从而成为人机交互的重要接口之一，被广泛应用于各类场景，例如语音输入、语音搜索、语音翻译、智能家居等等。同时，使用语音识别技术的用户也越来越多，这些用户可能来自不同的国家，使用不同的语种，因此要求语音识别模型可以支持不同语种的语音识别。
目前，通过包括多语种的混合声学发音单元集合，同时训练声学模型以支持多种语言的语音识别。这种方式由于不同语言发音直接相互混淆，会大大影响不同语言的语音识别准确率，反而可能造成用户使用体验更差。
发明内容
针对相关技术中多语种语音识别存在的识别准确率较低的问题，本申请提供一种基于自适应语种进行语音识别的方法及相关装置。
根据本申请的实施例，提供一种基于自适应语种进行语音识别的方法，包括：基于获取的语音数据提取表示发音音素信息的音素特征；将所述音素特征输入预先基于多语种语料训练得到的语种判别模型，得到所述语音数据的语种判别结果；基于所述语种判别结果所对应语种的语言声学模型获取所述语音数据的语音识别结果。
根据本申请的实施例，还提供一种用于自适应语种进行语音识别的装置，包括：提取模块，用于基于获取的语音数据提取表示发音音素信息的音素特征；判别模块，用于将所述音素特征输入预先基于多语种语料训练得到的语种判别模型，得到所述语音数据的语种判别结果；以及识别模块，用于基于所述语种判别结果所对应语种的语言声学模型获取所述语音数据的语音识别结果。
根据本申请的实施例，还提供基于人工智能的语音识别方法，包括：
分别使用各个语种的语料作为输入，训练多个语言声学模型，所述多个语言声学模型分别对应不同语种；
利用所述多个语言声学模型分别提取所述各个语种的语料的音素特征，并基于所述音素特征训练语种判别模型；
基于所述多个语言声学模型和所述语种判别模型对采集的语音数据进行语音识别。
根据本申请的实施例，提供一种基于人工智能的语音识别装置，包括：
声学模型单元，用于分别使用各个语种的语料作为输入，训练多个语言声学模型，所述多个语言声学模型分别对应不同语种；
判别模型单元，用于利用所述多个语言声学模型分别提取所述各个语种的语料的音素特征，并基于所述音素特征训练语种判别模型；
识别模块，用于基于所述多个语言声学模型和所述语种判别模型对采集的语音数据进行语音识别。
根据本申请的实施例，提供一种计算机可读存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时实现根据以上所述的方法。
根据本申请的实施例，提供一种电子设备，包括：处理器；以及存储器，所述存储器上存储有计算机可读指令，所述计算机可读指令被所述处理器执行时实现以上所述的方法。
根据本申请的实施例，还提供一种语音交互设备，包括：采集单元，用于采集用户的语音数据；处理单元，用于基于所述语音数据提取表示发音音素信息的音素特征；将所述音素特征输入预先基于多语种语料训练得到的语种判别模型，得到所述语音数据的语种判别结果；并基于所述语种判别结果所对应语种的语言声学模型获取所述语音数据的语音识别结果；以及交互单元，用于基于所述处理单元的语音识别结果向所述用户呈现相应的交互内容。
基于上述实施例基于自适应语种进行语音识别的方案，通过提取的音素特征进行语种判别，从而切换到语音判别结果所对应语种的语言声学模型，基于语音判别结果所对应语种的语言声学模型获取语音数据的语音识别结果。根据不同的语种可以自动切换到不同的语言声学模型进行语音识别，避免不同语言发音相互混淆导致的识别率较低的问题，从而在提高处理效率的同时，保证了语音识别的准确率。同时节省了由用户选择语种带来的多余操作，提高了处理效率和用户体验。
应当理解的是，以上的一般描述和后文的细节描述仅是示例性的，并不能限制本申请。
附图说明
此处的附图被并入说明书中并构成本说明书的一部分，示出了符合本申请的实施例，并于说明书一起用于解释本申请的原理。
图1示出可以应用本申请实施例的系统架构示意图；
图2示出适于用来实现本申请实施例的电子设备的计算机系统的结构示意图；
图3示出了一种基于自适应语种进行语音识别的方法流程图；
图4示出基于神经网络的语言声学模型示例；
图5示出了一种用于自适应语种进行语音识别的模型训练方法流程图；
图6示出基于图5实施例步骤510训练多个语言声学模型的框架示意图；
图7示出基于图5实施例步骤520训练混合语言声学模型的框架示意图；
图8示出了一种基于自适应语种进行语音识别的方法的流程图；
图9示出了一种基于自适应语种进行语音识别的方法的流程图；
图10示出了一种基于自适应语种进行语音识别的装置的框图；
图11示出了一种基于自适应语种进行语音识别的装置的框图；
图12示出了一种语音交互设备的框图。
具体实施方式
本部分将参考附图更全面地描述示例实施方式。然而，示例实施方式能够以多种形式实施，且不应被理解为限于在此阐述的范例；相反，提供这些实施方式使得本申请将更加全面和完整，并将示例实施方式的构思全面地传达给本领域的技术人员。
此外，所描述的特征、结构或特性可以以任何合适的方式结合在一个或更多实施例中。在下面的描述中，提供许多具体细节从而给出对本申请的实施例的充分理解。然而，本领域技术人员将意识到，可以实践本申请的技术方案而没有特定细节中的一个或更多，或者可以采用其它的方法、组元、装置、步骤等。在其它情况下，不详细示出或描述公知方法、装置、实现或者操作以避免模糊本申请的各方面。
附图中所示的方框图仅仅是功能实体，不一定必须与物理上独立的实体相对应。即，可以采用软件形式来实现这些功能实体，或在一个或多个硬件模块或集成电路中实现这些功能实体，或在不同网络和/或处理器装置和/或微控制器装置中实现这些功能实体。
附图中所示的流程图仅是示例性说明，不是必须包括所有的内容和操作/步骤，也不是必须按所描述的顺序执行。例如，有的操作/步骤还可以分解，而有的操作/步骤可以合并或部分合并，因此实际执行的顺序有可能根据实际情况改变。
需要强调的是，本申请实施例所提供的基于自适应语种的进行语音识别的方法可以是基于人工智能实现的。人工智能(Artificial Intelligence，AI)是利用数字计算机或者数字计算机控制的机器模拟、延伸和扩展人的智能，感知环境、获取知识并使用知识获得最佳结果的理论、方法、技术及应用系统。换句话说，人工智能是计算机科学的一个综合技术，它企图了解智能的实质，并生产出一种新的能以人类智能相似的方式做出反应的智能机器。 人工智能也就是研究各种智能机器的设计原理与实现方法，使机器具有感知、推理与决策的功能。
人工智能技术是一门综合学科，涉及领域广泛，既有硬件层面的技术也有软件层面的技术。人工智能基础技术一般包括如传感器、专用人工智能芯片、云计算、分布式存储、大数据处理技术、操作/交互系统、机电一体化等技术。人工智能软件技术主要包括计算机视觉技术、语音技术、自然语言处理技术以及机器学习/深度学习等几大方向。
其中，语音技术(Speech Technology)的关键技术有自动语音识别技术(Automatic Speech Recognition，ASR)和语音合成技术(Text-To-Speech，TTS)以及声纹识别技术。让计算机能听、能看、能说、能感觉，是未来人机交互的发展方向，其中语音成为未来最被看好的人机交互方式之一。
机器学习(Machine Learning，ML)是一门多领域交叉学科，涉及概率论、统计学、逼近论、凸分析、算法复杂度理论等多门学科。专门研究计算机怎样模拟或实现人类的学习行为，以获取新的知识或技能，重新组织已有的知识结构使之不断改善自身的性能。机器学习是人工智能的核心，是使计算机具有智能的根本途径，其应用遍及人工智能的各个领域。机器学习和深度学习通常包括人工神经网络、置信网络、强化学习、迁移学习、归纳学习、式教学习等技术。
随着人工智能技术研究和进步，人工智能技术在多个领域展开研究和应用，例如常见的智能家居、智能穿戴设备、虚拟助理、智能音箱、智能营销、无人驾驶、自动驾驶、无人机、机器人、智能医疗、智能客服等，相信随着技术的发展，人工智能技术将在更多的领域得到应用，并发挥越来越重要的价值。
本申请实施例提供的方案涉及人工智能的语音、机器学习/深度学习等技术，具体通过如下实施例进行说明。
本申请实施例提供的方案可以由具有基于自适应语种进行语音识别功能的电子设备执行，该电子设备可以是终端设备，也可以是服务器。
以下对本申请实施例的技术方案的实现细节进行详细阐述。
图1以语音处理设备是服务器为例，示出了可以应用本申请实施例的示例性系统架构100的示意图。
如图1所示，系统架构100可以包括终端设备101、102、103中的一种或多种，网络104和服务器105。网络104用以在终端设备101、102、103和服务器105之间提供通信链路的介质。网络104可以包括各种连接类型，例如有线、无线通信链路或者光纤电缆等等。
应该理解，图1中的终端设备、网络和服务器的数目仅仅是示意性的。根据实现需要，可以具有任意数目的终端设备、网络和服务器。比如服务器105可以是多个服务器组成的服务器集群等。
用户可以使用终端设备101、102、103通过网络104与服务器105交互，以接收或发送消息等。终端设备101、102、103可以是具有显示屏的各种电子设备，包括但不限于智能手机、平板电脑、便携式计算机和台式计算机等等。
服务器105可以是提供各种服务的服务器。例如，终端设备101可以通过语音采集模块采集用户的语音，并转化为数字化的数据形式(即语音数据)通过网络104发送至服务器105。接续，服务器105可以基于获取的语音数据提取表示发音音素信息的音素特征，音素特征可以体现出不同语种的发音音素特点，将所述音素特征输入预先基于多语种语料训练得到的语种判别模型，可以得到所述语音数据的语种判别结果。然后，基于语种判别结果所对应语种的语言声学模型获取所述语音数据的语音识别结果，并通过网络104返回给终端设备101。
在一些实施例中，本申请实施例所提供的基于自适应语种进行语音识别的方法可以由服务器105执行，相应地，基于自适应语种进行语音识别的装置可以设置于服务器105中。在另一些实施例中，某些终端设备可以和服务器具有相似的功能从而执行本方法。因此，本申请实施例所提供的方法不严格限定在服务器端执行，即具有基于自适应语种进行语音识别功能的电子设备可以是终端设备。
图2示出了适于用来实现本申请实施例的电子设备的计算机系统的结构示意图。
需要说明的是，图2示出的电子设备的计算机系统200仅是一个示例，不应对本申请实施例的功能和使用范围带来任何限制。
如图2所示，计算机系统200包括中央处理单元(CPU)201，其可以根据存储在只读存储器(ROM)202中的程序或者从存储部分208加载到随机访问存储器(RAM)203中的程序而执行各种适当的动作和处理。在RAM 203中，还存储有系统操作所需的各种程序和数据。CPU 201、ROM 202以及RAM 203通过总线204彼此相连。输入/输出(I/O)接口205也连接至总线204。
以下部件连接至I/O接口205：包括键盘、鼠标等的输入部分206；包括诸如阴极射线管(CRT)、液晶显示器(LCD)等以及扬声器等的输出部分207；包括硬盘等的存储部分208；以及包括诸如LAN卡、调制解调器等的网络接口卡的通信部分209。通信部分209经由诸如因特网的网络执行通信处理。驱动器210也根据需要连接至I/O接口205。可拆卸介质211，诸如磁盘、光盘、磁光盘、半导体存储器等等，根据需要安装在驱动器210上，以便于从其上读出的计算机程序根据需要被安装入存储部分208。
特别地，根据本申请的实施例，下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，本申请的实施例包括一种计算机程序产品，其包括承载在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部分209从网络上被下载和安装，和/或从可拆卸介质211 被安装。在该计算机程序被中央处理单元(CPU)201执行时，执行本申请实施例中限定的各种功能。
需要说明的是，本申请所示的计算机可读介质可以是计算机可读信号介质或者计算机可读存储介质或者是上述两者的任意组合。计算机可读存储介质例如可以是，但不限于电、磁、光、电磁、红外线、或半导体的系统、装置或器件，或者任意以上的组合。计算机可读存储介质的更具体的例子可以包括但不限于：具有一个或多个导线的电连接、便携式计算机磁盘、硬盘、随机访问存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、光纤、便携式紧凑磁盘只读存储器(CD-ROM)、光存储器件、磁存储器件、或者上述的任意合适的组合。在本申请中，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。而在本申请中，计算机可读的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。计算机可读的信号介质还可以是计算机可读存储介质以外的任何计算机可读介质，该计算机可读介质可以发送、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：无线、电线、光缆、RF等等，或者上述的任意合适的组合。
作为另一方面，本申请还提供了一种计算机可读介质，该计算机可读介质可以是上述实施例中描述的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一个该电子设备执行时，使得该电子设备实现如下述实施例中所述的方法。例如，所述的电子设备可以实现如图3至图9所示的各个步骤。
图3是根据一示例性实施例示出的一种基于自适应语种进行语音识别的方法的流程图。如图3所示，该基于自适应语种进行语音识别的方法可以由任意电子设备执行，可包括以下步骤310-330。
在步骤310中，基于获取的语音数据提取表示发音音素信息的音素特征。
这里的语音数据是指由电子设备的语音采集模块对用户语音进行采集、处理后得到的数字化格式数据。电子设备例如包括但不限于智能手机、平板电脑、个人计算机、笔记本电脑等设备，语音采集模块例如包括这些电子设备的麦克风和声卡等部件。
在一个实施例中，电子设备可使用预设的特征提取算法处理上述语音数据，以得到相应的音素特征。音素特征体现的是发音音素信息，通过音素特征可以识别出语音数据的语种，在一种可能的实现方式中，音素特征例如可以是基于语言声学模型的bottleneck层提取的瓶颈bottleneck特征。
接续图3所示，在步骤320中，将音素特征输入预先基于多语种语料训练得到的语种判别模型，得到语音数据的语种判别结果。
相关技术中实现多语种语音识别的其中一种方式是，通过用户手动选择语种，以便语音识别产品能够切换到相应的语言声学模型进行识别，但这样一来就增加了用户的操作负担，也降低了用户体验和语音识别的处理效率。
本申请的实施例基于提取的音素特征可以自动判别语音数据所属的语种，以得到所述语音数据的语种判别结果，以便后续自动切换到语种判别结果所对应的语言声学识别模型对语音数据进行识别。
音素是根据语音的自然属性划分出来的最小语音单位。从声学性质来看，音素是从音质角度划分出来的最小语音单位。从生理性质来看，一个发音动作形成一个音素。如〔ma〕包含〔m〕〔a〕两个发音动作，是两个音素。相同发音动作发出的音就是同一音素，不同发音动作发出的音就是不同音素。如〔ma-mi〕中，两个〔m〕发音动作相同，是相同音素，〔a〕〔i〕发音动作不同，是不同音素。
音素一般分为元音和辅音两大类，不同的语种可划分出不同的发音音素。以汉语普通话为例，包括22个辅音和10个元音；而英语国际音标共有48个音素，其中元音音素20个，辅音音素28个。
相应的，本申请的实施例通过提取语音数据中表示发音音素信息的音素特征，输入预先基于多语种语料训练得到的语种判别模型，可实现对语音数据所属语种的判别。语种判别模型的训练可参见图5所示实施例。
接续图3所示，在步骤330中，基于语种判别结果所对应语种的语言声学模型获取语音数据的语音识别结果。
如背景技术部分所述，相关技术实现多语种语音识别的其中一种方式是，通过包括多语种的混合声学发音单元集合，同时训练声学模型以支持多种语言的语音识别。由于不同语言发音直接相互混淆，会大大影响不同语言的语音识别准确率。
本申请的实施例基于步骤320的语种判别结果，按照判别出的语种获取相应语种的语言声学模型输出的语音识别结果。
这里的语言声学模型用于从语音数据中检测出与发音相匹配的文字，进而确定后续的响应方式，例如识别为相应的语音指令与智能设备进行人机交互。在一个实施例中，语言声学模型可通过已知语料及其语音特征对初始模型进行训练得到，初始模型可通过CNN(Convolutional Neural Network，卷积神经网络)、DNN(Deep Neural Network，深度神经网络)等神经网络或者它们的组合网络来实现。
图4示出基于神经网络的语言声学模型示例。如图4所示，该模型包括第一输入层410、第一隐藏层420和第一输出层430。第一输入层410用于接收音素特征的输入，其中节点 的数量取决于音素特征的数量；第一隐藏层420用于对输入的音素特征进行处理，其可以包括多层网络(多个子隐层)，每层网络又可包括多个节点，网络层数和节点数可根据训练语料的规模和设备的计算能力来确定，层数和节点越多，可视为训练得到的模型精度越高；第一输出层430包括多个输出节点，例如可分别对应不同的音素。在第一输出层430会计算各个节点的输出概率，表示输入的音素特征属于各个节点所对应音素的概率。由此，基于参数经过初始化(例如随机初始化)的第一隐藏层420，在第一输入层410输入基于已知语料提取的音素特征，计算第一输出层430中输出节点的概率结果，以使目标损失函数(例如softmax函数)最小化为目标，可训练得到能够准确表达出输入语料音素特征的第一隐藏层420，从而得到训练好的语言声学模型。语言声学模型的训练可参见图5和图6所示实施例。
需要说明的是，图4中所示出的网络层数和每层中的节点个数均为示例，本申请的实施例并不仅限于此。
如上所述，为了按照步骤320判别出的语种获取相应语种的语言声学模型输出的语音识别结果，本申请的实施例是分语种来训练多个语言声学模型。在步骤320判别出语音数据所属的语种时，步骤330中便基于相应语种的语言声学模型输出语音识别结果，如此能够避免使用混合语言声学模型时，由于不同语言发音相互混淆导致的识别率较低的问题，从而在提高处理效率的同时，保证了语音识别的准确率。
基于上述实施例用于自适应语种进行语音识别的方案，通过提取的音素特征进行语种判别，从而切换到语音判别结果所对应语种的语言声学模型，基于语音判别结果所对应语种的语言声学模型获取语音数据的语音识别结果。根据不同的语种可以自动切换到不同的语言声学模型进行语音识别，避免不同语言发音相互混淆导致的识别率较低的问题，从而在提高处理效率的同时，保证了语音识别的准确率。同时节省了由用户选择语种带来的多余操作，提高了处理效率和用户体验。
图5是根据一示例性实施例示出的一种基于自适应语种进行语音识别的模型训练方法流程图。如图5所示，该基于自适应语种进行语音识别的模型训练方法可以由任意电子设备执行，可包括以下步骤510-540。
在步骤510中，分别使用各个语种的语料作为输入，训练多个语言声学模型，所述多个语言声学模型分别对应不同语种。
在一个实施例中，这里的语料包括分别对属于各个语种的用户语音进行预处理和特征提取所得到的音素特征集合。
所述的预处理例如包括但不限于对用户语音的采样量化和端点检测(Voice Activity Detection，VAD)等等。这里的VAD是指在噪声环境中检测语音的存在与否，可用于语音编码等系统中，起到降低语音编码速率、节省通信带宽、提高识别率等作用。当上述电子计算设备开启语音采集功能时，可通过内置或外置的麦克风等采集设备对环境中获取的各 种声音进行VAD，从其中检测出用户语音。接续，当电子设备检测到用户语音后，进一步确定语音的起始点，进而开始采集环境中的语音，并形成数字形式的语音数据。
经过对采集语音的预处理，提高了语音数据的高频分辨率，使得语音数据变得更加平滑，方便了语音数据的后续处理。
所述的特征提取例如包括但不限于去掉语音数据中的冗余部分，以提取出能够代表语音数据本质特征的参数。
基于得到的音素特征集合，可使用类似图4所示的神经网络进行训练，得到相应语种的语言声学模型。
图6示出基于步骤510训练多个语言声学模型的框架示意图。如图6所示，针对语种1，将提取得到的音素特征集合输入至第一输入层611；在第一输入层611基于预设的第一权重矩阵计算向第一隐藏层612输出的输出值；第一隐藏层612包括多个子隐层，每个子隐层接收下一个子隐层(或相邻的第一输入层611)的输出值，使用本层的权重矩阵(例如第二权重矩阵)进行加权计算并将结果输出至上一个子隐层(或相邻的第一输出层613)；第一输出层613包括多个输出元。
在一个实施例中，第一输出层613的每个输出元对应于语种1所包括的一个发音音素。第一输出层613接收相邻子隐层的输出值，并使用本层的权重矩阵(例如第三权重矩阵)进行加权计算，在利用损失函数基于加权计算的结果计算输出概率(例如第一输出概率)，该第一输出概率表示输入的音素特征属于每个输出元所属发音音素的概率。这里的损失函数例如可包括softmax函数。
基于已知语料提取的音素特征集合，可预先确定其属于每个发音音素的目标概率值(例如为0或1)，通过上述训练过程可不断调整第一输入层611、第一隐藏层612和第一输出层613的权重矩阵(分别为第一权重矩阵、第二权重矩阵和第三权重矩阵)，最终得到满足预设条件(例如训练达到预设迭代次数或与目标概率的误差落入预设范围)的语言声学模型610。
接续如图6所示，针对语种2至语种N，可按照同样方式输入相应语种的语料，训练得到相应语种的语言声学模型620-6N0，用于在步骤330中获取相应语种的语音识别结果。例如，在步骤320的语种判别结果指示当前用户语音属于语种2后，步骤330中便可基于语言声学模型620输出相应的语音识别结果。如上所述，这里的语音识别结果例如可包括输入的语音属于语种2中的哪个发音音素，后续经过进一步处理便可得到语音对应的文本信息，此处不加以赘述。
在步骤520中，使用包括多个语种的混合语料作为输入，训练支持多个语种的混合语言声学模型。
本申请的实施例中，除了通过步骤510分别训练对应于多个语种的多个语言声学模型之外，还训练同时支持多个语种的混合语言声学模型。与相关技术中使用混合语言声学模 型对可能包括多语种的用户语音进行语音识别不同，在本申请的实施例中，该混合语言声学模型用于支持上述步骤320中的语种判别。
混合语言声学模型的训练过程与步骤510中各语种的语言声学模型类似，以下仅简单介绍，重复之处不再赘述。
在一个实施例中，这里的混合语料例如包括对具有各个语种的用户语音进行预处理和特征提取所得到的音素特征集合。
基于得到的音素特征集合，同样可使用类似图4所示的神经网络进行训练，得到最终的混合语言声学模型。
图7示出基于步骤520训练混合语言声学模型的框架示意图。如图7所示，针对混合预料，将提取得到的音素特征集合输入至输入层710；在输入层710基于预设的权重矩阵计算向隐藏层720输出的输出值；隐藏层720包括多个子隐层，每个子隐层接收下一个子隐层(或相邻的输入层710)的输出值，使用本层的权重矩阵进行加权计算并将结果输出至上一个子隐层(或相邻的输出层)；输出层730包括多个输出模块，每个输出模块又包括多个输出元。
在一个实施例中，输出层730的每个输出模块731至73N分别对应一个语种1-N，而一个输出模块中的每个输出元又对应于该语种所包括的一个发音音素。输出层730接收相邻子隐层的输出值，并使用本层的权重矩阵进行加权计算，在利用损失函数基于加权计算的结果计算输出概率，该输出概率表示输入的语音特征属于每个输出模块中每个输出元所属发音音素的概率。这里的损失函数例如可包括softmax函数。
基于已知混合语料提取的音素特征集合，可预先确定其属于每个发音音素的目标概率值(例如为0或1)，通过上述训练过程可不断调整输入层710、隐藏层720和输出层730的权重矩阵，最终得到满足预设条件(例如训练达到预设迭代次数或与目标概率值的误差落入预设范围)的混合语言声学模型700。
最终训练得到的混合语言声学模型700，用于在步骤320中获取语种判别结果。例如，基于步骤310对当前用户语音提取的音素特征，输入混合语言声学模型700便可得到其属于每个输出模块中每个输出元所属发音音素的概率，相应便可得到其属于每个语种的概率。如上所述，本申请实施例与相关技术的实质不同在于，针对可能包括多语种的用户语音，输入此处训练的混合语言声学模型，得到的结果并非用于输出语音识别结果，而是用于确定其所属的语种，从而进一步基于相应语种的语言声学模型输出语音识别结果。
在步骤530中，基于混合语言声学模型提取混合语料的音素特征，并基于音素特征训练语种判别模型。
在步骤540中，从多个语言声学模型中，分别提取各个语种的语料的音素特征，并基于音素特征辅助训练语种判别模型。
如上所述，步骤510和520分别可训练得到用于步骤330和320的语言声学模型和混合语言声学模型。
在一个实施例中，本申请还可包括步骤530和540，基于上述混合语言声学模型和语言声学模型中提取的音素特征训练语种判别模型，以进一步提高步骤320中语种判别的准确率。然而本领域技术人员可以理解，本申请的其他实施例中可不包括步骤520，而仅从步骤510的语言声学模型中提取音素特征来进行后续的语种判别模型训练。
在训练得到多个语言声学模型和语种判别模型后，可以将多个语言声学模型和语种判别模型应用于图3所对应的实施例中，从而基于多个语言声学模型和语种判别模型对采集的语音数据进行语音识别。
需要说明的是，语种判别模型包括第二输入层、第二隐藏层和第二输出层，语种判别模型的训练过程与步骤510中各语种的语言声学模型类似，以下仅简单介绍，重复之处不再赘述。
在一个实施例中，语种判别模型的输入可以包括经过语言声学模型和混合语言声学模型提取到的音素特征。基于bottleneck层提取的音素特征，同样可使用类似图4所示的神经网络进行训练，得到最终的语种判别模型。
结合图6和图7所示，基于语言声学模型610-6N0和混合语言声学模型700，将提取得到的音素特征输入至第二输入层810；在第二输入层810基于预设的第四权重矩阵计算向第二隐藏层820输出的输出值；第二隐藏层820使用本层的权重矩阵(例如第五权重矩阵)进行加权计算并将结果输出至相邻的第二输出层830；第二输出层830包括多个输出元。
在一个实施例中，第二输出层830的每个输出元对应于一个语种。第二输出层830接收相邻子隐层的输出值，并使用本层的权重矩阵(例如第六权重矩阵)进行加权计算，在利用损失函数基于加权计算的结果计算输出概率，该输出概率表示输入的音素特征属于每个输出元所属语种的概率。这里的损失函数例如可包括softmax函数。
基于已知语料提取的音素特征，可预先确定其属于每个语种的目标概率值(例如为0或1)，通过上述训练过程可不断调整第二输入层810、第二隐藏层820和第二输出层830的权重矩阵(分别为第四权重矩阵、第五权重矩阵和第六权重矩阵)，最终得到满足预设条件(例如训练达到预设迭代次数或与目标概率值的误差落入预设范围)的语种判别模型800。
最终训练得到的语种判别模型800，也可用于在步骤320中获取语种判别结果。例如，基于步骤310对当前用户语音提取的语音特征，将其中的音素特征输入语种判别模型800便可得到其属于每个输出元所属语种的概率。
基于上述实施例用于自适应语种进行语音识别的模型训练方法，训练得到的混合语言声学模型或者语种判别模型可用于确定用户语音所属的语种，进而便可基于相应语种的语 言声学模型获取语音识别结果，既能保证分语种进行识别的准确率，同时节省了由用户选择语种带来的多余操作，提高了处理效率和用户体验。
图8是根据另一示例性实施例示出的一种基于自适应语种进行语音识别的方法的流程图。如图8所示，该基于自适应语种进行语音识别的方法可以由任意电子设备执行，可包括以下步骤911-913。
在步骤911中，基于获取的语音数据提取音素特征。
该步骤可参考上述实施例的步骤310。
在步骤912中，将音素特征分别输入至对应于不同语种的多个语言声学模型，并根据音素特征得到语音数据的语种判别结果。
在步骤913中，从多个语言声学模型返回的语音识别结果中，选择语种判别结果所对应语种的语言声学模型的语音识别结果输出。
在该实施例中，基于步骤911提取的音素特征同时输入语种判别模型和多个语言声学模型。多个语言声学模型基于输入的语音数据开始处理，输出各自的语音识别结果；同时语种判别模型基于输入的音素特征输出语种判别结果，进而根据语种判别结果从多个语言声学模型中选择相应语种的语言声学模型的语音识别结果输出。
这里的语种判别模型例如可基于图5所示实施例的步骤530和540训练得到，多个语言声学模型例如可基于图5所示实施例的步骤510和520训练得到。
结合图6所示多个语言声学模型的框架示意图，此处步骤911提取的音素特征分别输入多个语种的语言声学模型620-6N0。以模型610为例，上述语音特征首先输入至第一输入层611；在第一输入层611基于训练得到的权重矩阵计算向第一隐藏层612输出的输出值；第一隐藏层612包括多个子隐层，每个子隐层接收下一个子隐层(或相邻的第一输入层611)的输出值，使用本层经过训练的权重矩阵进行加权计算并将结果输出至上一个子隐层(或相邻的第一输出层613)；第一输出层613包括多个输出元，每个输出元对应于语种1所包括的一个发音音素。第一输出层613接收相邻子隐层的输出值，并使用本层经过训练的权重矩阵进行加权计算，在利用损失函数基于加权计算的结果计算输出概率，该输出概率表示输入的语音特征属于每个输出元所属发音音素的概率。
结合图6和图7所示语种判别模型800的框架示意图，此处步骤911提取的音素特征还同时输入至语种判别模型800。上述音素特征首先输入至第二输入层810；在第二输入层810基于训练得到的权重矩阵计算向第二隐藏层820输出的输出值；第二隐藏层820使用本层经过训练的权重矩阵进行加权计算并将结果输出至第二输出层830；第二输出层830包括多个输出元，每个输出元对应于一个语种。第二输出层830接收第二隐藏层820的输 出值，并使用本层经过训练的权重矩阵进行加权计算，在利用损失函数基于加权计算的结果计算输出概率，该输出概率表示输入的音素特征属于每个语种的概率。
最后，在步骤913中，基于语种判别模型800中输出概率最大的输出元确定当前用户语言所属的语种，据此可选择语言声学模型620-6N0中的一个输出语音识别结果。对于语言声学模型输出的语音识别结果，后续经过进一步处理便可得到语音对应的文本信息，此处不加以赘述。
基于上述实施例用于自适应语种进行语音识别的方案，通过提取的音素特征进行语种判别，并同时进行多个语种的语音识别，进而从语音判别结果所对应语种的语言声学模型获取语音数据的语音识别结果。根据不同的语种可以自动切换到不同的语言声学模型进行语音识别，避免不同语言发音相互混淆导致的识别率较低的问题，从而在提高处理效率的同时，保证了语音识别的准确率。同时节省了由用户选择语种带来的多余操作，提高了处理效率和用户体验。并且，由于在语音判别的同时进行分语种的语音识别，可进一步提高语音识别的处理速度。
图9是根据再一示例性实施例示出的一种用于自适应语种进行语音识别的方法的流程图。如图9所示，该用于自适应语种进行语音识别的方法可以由任意计算设备执行，可包括以下步骤921-923。
在步骤921中，基于获取的语音数据提取音素特征。
该步骤可参考上述实施例的步骤310。
在步骤922中，根据音素特征得到语音数据的语种判别结果。
在步骤923中，从分别对应于不同语种的多个语言声学模型中，选择语种判别结果所对应语种的语言声学模型，输入音素特征以获取语音识别结果。
在该实施例中，基于步骤911提取的音素特征先输入语种判别模型，语种判别模型基于输入的音素特征输出语种判别结果，进而根据语种判别结果使语音特征输入至语种判别结果所对应语种的语言声学模型，以获取相应的语音识别结果。
这里的语种判别模型例如可基于图5所示实施例的步骤530和540训练得到，多个语言声学模型例如可基于图5所示实施例的步骤510和520训练得到。
结合图6所示多个语言声学模型的框架示意图，此处步骤921提取的语音特征中的音素特征输入至语种判别模型800；首先在第二输入层810基于训练得到的权重矩阵计算向第二隐藏层820输出的输出值；第二隐藏层820使用本层经过训练的权重矩阵进行加权计算并将结果输出至第二输出层830；第二输出层830包括多个输出元，每个输出元对应于一个语种。第二输出层830接收第二隐藏层820的输出值，并使用本层经过训练的权重矩 阵进行加权计算，在利用损失函数基于加权计算的结果计算输出概率，该输出概率表示输入的音素特征属于每个语种的概率。
在步骤923中，基于语种判别模型800中输出概率最大的输出元确定当前用户语言所属的语种，据此可选择语言声学模型620-6N0中的一个来输入上述音素特征。
以步骤922的语音判别结果为语种1为例，步骤923中基于该语音判别结果将音素特征输入至语种1对应的语言声学模型610。上述音素特征首先输入至第一输入层611；在第一输入层611基于训练得到的权重矩阵计算向第一隐藏层612输出的输出值；第一隐藏层612包括多个子隐层，每个子隐层接收下一个子隐层(或相邻的第一输入层611)的输出值，使用本层经过训练的权重矩阵进行加权计算并将结果输出至上一个子隐层(或相邻的第一输出层613)；第一输出层613包括多个输出元，每个输出元对应于语种1所包括的一个发音音素。第一输出层613接收相邻子隐层的输出值，并使用本层经过训练的权重矩阵进行加权计算，在利用损失函数基于加权计算的结果计算输出概率，该输出概率表示输入的语音特征属于每个输出元所属发音音素的概率。
对于语言声学模型输出的语音识别结果，后续经过进一步处理便可得到语音对应的文本信息，此处不加以赘述。
基于上述实施例用于自适应语种进行语音识别的方案，通过提取的音素特征进行语种判别，进而根据语音判别结果选择相应语种的语言声学模型进行语音识别，并输出语音识别结果，避免不同语言发音相互混淆导致的识别率较低的问题，从而在提高处理效率的同时，保证了语音识别的准确率。同时节省了由用户选择语种带来的多余操作，提高了处理效率和用户体验。并且，由于在语音判别的基础上进行分语种的语音识别，可节省语音识别的处理量，提高处理效率。
下述为本申请装置实施例，可以用于执行本申请上述对话模型的更新训练方法的实施例。对于本申请装置实施例中未披露的细节，请参照本申请对话模型的更新训练方法实施例。
图10是根据一示例性实施例示出的一种基于自适应语种进行语音识别的装置的框图。如图10所示，该用于自适应语种进行语音识别的装置可以由任意电子设备实现，可包括以下提取模块1010、判别模块1020和识别模块1030。
提取模块1010用于基于获取的语音数据提取表示发音音素信息的音素特征。
判别模块1020用于将所述音素特征输入预先基于多语种语料训练得到的语种判别模型，得到所述语音数据的语种判别结果。
识别模块1030用于基于所述语种判别结果所对应语种的语言声学模型获取所述语音数据的语音识别结果。
在一个实施例中，提取模块1010还用于将所述音素特征分别输入至对应于不同语种的多个语言声学模型；识别模块1030还用于从所述多个语言声学模型返回的语音识别结果中，选择所述语种判别结果所对应语种的语言声学模型的语音识别结果输出。
在一个实施例中，识别模块1030还用于从对应于不同语种的多个语言声学模型中，选择所述语种判别结果所对应语种的语言声学模型，输入所述音素特征以获取所述语音识别结果。
在一个实施例中，基于上述基于自适应语种进行语音识别的装置实施例，还包括识别模型训练模块，用于分别使用各个语种的语料作为输入，训练多个语言声学模型，所述多个语言声学模型分别对应不同语种；以及判别模型训练模块，用于利用所述多个语言声学模型分别提取所述各个语种的语料的音素特征，并基于所述音素特征训练所述语种判别模型。
在一个实施例中，基于上述基于自适应语种进行语音识别的装置实施例，还包括判别模型训练模块，用于使用包括多个语种的混合语料作为输入，训练支持所述多个语种的混合语言声学模型，基于所述混合语言声学模型提取所述混合语料的音素特征，并基于所述音素特征训练所述语种判别模型。
在一个实施例中，所述语言声学模型包括第一输入层、第一隐藏层和第一输出层，所述识别模型训练模块，用于：
对于与所述多个语种中每个语种对应的语料，将提取得到的音素特征输入至所述第一输入层；
在所述第一输入层基于预设的第一权重矩阵计算向所述第一隐藏层输出的输出值；
在所述第一隐藏层包括的多个子隐层中，接收相邻的所述第一输入层或前一个子隐层的输出值，使用相应的第二权重矩阵进行加权计算，并将结果输出至相邻的所述第一输出层或后一个子隐层；
在包括多个第一输出元的所述第一输出层接收相邻子隐层的输出值，使用第三权重矩阵进行加权计算，并基于计算结果获取第一输出概率，所述第一输出概率表示输入的所述音素特征属于每个第一输出元所属发音音素的概率；
根据所述第一输出概率与目标概率之间的差值，迭代调整所述第一权重矩阵、所述第二权重矩阵和所述第三权重矩阵，得到满足预设条件的所述语言声学模型。
在一个实施例中，所述语种判别模型包括第二输入层、第二隐藏层和第二输出层，所述判别模型训练模块，用于：
在训练得到的所述多个语言声学模型中，基于所述第一隐藏层的输出值提取所述音素特征，输入至所述第二输入层；
在所述第二输入层基于预设的第四权重矩阵计算向所述第二隐藏层输出的输出值；
在所述第二隐藏层使用相应的第五权重矩阵进行加权计算，并将结果输出至所述第二输出层包括的多个第二输出元；
在所述第二输出层使用第六权重矩阵进行加权计算，并基于计算结果获取第二输出概率，所述第二输出概率表示输入的所述音素特征属于每个第二输出元所属语种的概率；
根据所述第二输出概率与所述目标概率之间的差值，迭代调整所述第四权重矩阵、所述第五权重矩阵和所述第六权重矩阵，得到满足预设条件的所述语种判别模型。
在一个实施例中，所述音素特征包括瓶颈bottleneck特征。
基于上述实施例基于自适应语种进行语音识别的方案，通过提取的音素特征进行语种判别，从而切换到语音判别结果所对应语种的语言声学模型，基于语音判别结果所对应语种的语言声学模型获取语音数据的语音识别结果。根据不同的语种可以自动切换到不同的语言声学模型进行语音识别，避免不同语言发音相互混淆导致的识别率较低的问题，从而在提高处理效率的同时，保证了语音识别的准确率。同时节省了由用户选择语种带来的多余操作，提高了处理效率和用户体验。
图11是根据另一示例性实施例示出的一种基于自适应语种进行语音识别的装置的框图。如图11所示，该基于自适应语种进行语音识别的装置可以由任意电子设备实现，在图10实施例的基础上，判别模块1020包括判别模型单元1021，识别模块1030包括多个声学模型单元1031。
声学模型单元1031分别使用各个语种的语料作为输入，训练多个语言声学模型，所述多个语言声学模型分别对应不同语种。在一个实施例中，识别模块1030还使用包括多个语种的混合语料作为输入，训练支持多个语种的混合语言声学模型。
判别模型单元1021基于上述混合语言声学模型提取混合语料的音素特征，并基于音素特征训练语种判别模型。在一个实施例中，判别模型单元1021还从多个语言声学模型中，分别提取各个语种的语料的音素特征，并基于音素特征训练语种判别模型。
识别模块1030用于基于所述多个语言声学模型和所述语种判别模型对采集的语音数据进行语音识别。
在一个实施例中，提取模块1010基于识别模块1030训练得到的混合语言声学模型提取音素特征中的音素特征，以输入至判别模型单元1021训练得到的语种判别模型得到语种判别结果。
在一个实施例中，上述的音素特征包括bottleneck特征。
在一个实施例中，提取模块1010将上述语音特征分别发送至多个声学模型单元1031训练得到的多个语言声学模型，同时发送至判别模型单元1021训练得到的语种判别模型得 到语种判别结果；进而识别模块1030根据语种判别结果，从多个声学模型单元1031返回的语音识别结果中，选择语音判别结果所对应语种的语言声学模型的语音识别结果输出。
在另一个实施例中，提取模块1010将上述语音特征发送至判别模型单元1021训练得到的语种判别模型得到语种判别结果；进而识别模块1030根据语种判别结果，从多个声学模型单元1031中，选择语音判别结果所对应语种的语言声学模型，输入上述语音特征以获取语音识别结果。
基于上述实施例用于自适应语种进行语音识别的方案，通过提取的语音特征进行语种判别并同时进行多个语种的语音识别，从而切换到语音判别结果所对应语种的语言声学模型，基于语音判别结果所对应语种的语言声学模型获取语音数据的语音识别结果。根据不同的语种可以自动切换到不同的语言声学模型进行语音识别，避免不同语言发音相互混淆导致的识别率较低的问题，从而在提高处理效率的同时，保证了语音识别的准确率。同时节省了由用户选择语种带来的多余操作，提高了处理效率和用户体验。
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。
图12是根据一示例性实施例示出的一种语音交互设备的框图。如图12所示，本实施例的语音交互设备包括采集单元1201、处理单元1202和交互单元1203。这里的语音交互设备例如可包括但不限于移动终端、智能音箱、智能电视、智能穿戴设备中的任意设备。
在一个实施例中，采集单元1201用于采集用户的语音数据。这里的采集单元例如可包括麦克风、语音采集芯片等，通过对语音数据进行初步处理，得到数字化的语音数据。另外，采集单元1201不限于仅采集用户的语音数据，例如还可采集所有环境声音，并通过初步分析和筛选得到其中属于语音的数据。
在一个实施例中，处理单元1202用于基于所述语音数据提取表示发音音素信息的音素特征；将所述音素特征输入预先基于多语种语料训练得到的语种判别模型，得到所述语音数据的语种判别结果；并基于所述语种判别结果所对应语种的语言声学模型获取所述语音数据的语音识别结果。这里处理单元1202的处理可参见上述方法实施例的内容，此处不再赘述。
在一个实施例中，交互单元1203用于基于处理单元1202的语音识别结果向用户呈现相应的交互内容。这里的交互单元1203例如可包括图像输出装置和语音输出装置的任意组合。作为示例，基于处理单元1202输出的语音识别结果，交互单元1203可以以文字的形式直接展示识别出的语音内容。作为另一示例，交互单元1203还可以基于语音识别结果生成回复内容，并以文字或语音的形式将回复内容呈现给与用户。再例如，如果处理单元1202将用户的语音数据识别为操作指令，交互单元1203还可以向用户呈现操作指令的执行结果。
在另一个实施例中，作为对处理单元1202对语种判别的补充，交互单元1203还可用于展示处理单元1202的语种判别结果，供用户执行确认或修改的选择操作。接续，基于接 收的用户选择操作，交互单元1203可通知处理单元1202调整语种判别结果，并从处理单元1202获取调整后的语音识别结果。
基于上述实施例的语音交互设备，通过提取的语音特征进行语种判别，并基于语音判别结果所对应语种的语言声学模型获取语音数据的语音识别结果。根据不同的语种可以自动切换到不同的语言声学模型进行语音识别，避免不同语言发音相互混淆导致的识别率较低的问题，从而在提高处理效率的同时，保证了语音识别的准确率，同时节省了由用户选择语种带来的多余操作，提高了处理效率和用户体验。在同时接收多个用户的多语种语音输入时，能够实现句子级别的自动识别。
应当注意，尽管在上文详细描述中提及了用于动作执行的设备的若干模块或者单元，但是这种划分并非强制性的。实际上，根据本公开的实施方式，上文描述的两个或更多模块或者单元的特征和功能可以在一个模块或者单元中具体化。反之，上文描述的一个模块或者单元的特征和功能可以进一步划分为由多个模块或者单元来具体化。作为模块或单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部模块来实现本公开方案的目的。
应当理解的是，本申请并不局限于上面已经描述并在附图中示出的精确结构，并且可以在不脱离其范围进行各种修改和改变。本申请的范围仅由所附的权利要求来限制。

Claims (19)

  1. 一种基于自适应语种进行语音识别的方法，包括：
    基于获取的语音数据提取表示发音音素信息的音素特征；
    将所述音素特征输入预先基于多语种语料训练得到的语种判别模型，得到所述语音数据的语种判别结果；
    基于所述语种判别结果所对应语种的语言声学模型获取所述语音数据的语音识别结果。
  2. 如权利要求1所述的方法，所述基于所述语种判别结果所对应语种的语言声学模型获取所述语音数据的语音识别结果，包括：
    将所述音素特征分别输入至对应于不同语种的多个语言声学模型；
    从所述多个语言声学模型返回的语音识别结果中，选择所述语种判别结果所对应语种的语言声学模型的语音识别结果输出。
  3. 如权利要求1所述的方法，所述基于所述语种判别结果所对应语种的语言声学模型获取所述语音数据的语音识别结果，包括：
    从对应于不同语种的多个语言声学模型中，选择所述语种判别结果所对应语种的语言声学模型，输入所述音素特征以获取所述语音识别结果。
  4. 如权利要求1所述的方法，所述方法还包括：
    分别使用各个语种的语料作为输入，训练多个语言声学模型，所述多个语言声学模型分别对应不同语种；
    利用所述多个语言声学模型分别提取所述各个语种的语料的音素特征；
    基于所述音素特征训练所述语种判别模型。
  5. 如权利要求1或4所述的方法，所述方法还包括：
    使用包括多个语种的混合语料作为输入，训练支持所述多个语种的混合语言声学模型；
    基于所述混合语言声学模型提取所述混合语料的音素特征；
    基于所述音素特征训练所述语种判别模型。
  6. 如权利要求4所述的方法，所述语言声学模型包括第一输入层、第一隐藏层和第一输出层，所述训练多个语言声学模型，包括：
    对于与所述多个语种中每个语种对应的语料，将提取得到的音素特征输入至所述第一输入层；
    在所述第一输入层基于预设的第一权重矩阵计算向所述第一隐藏层输出的输出值；
    在所述第一隐藏层包括的多个子隐层中，接收相邻的所述第一输入层或前一个子隐层的输出值，使用相应的第二权重矩阵进行加权计算，并将结果输出至相邻的所述第一输出层或后一个子隐层；
    在包括多个第一输出元的所述第一输出层接收相邻子隐层的输出值，使用第三权重矩阵进行加权计算，并基于计算结果获取第一输出概率，所述第一输出概率表示输入的所述音素特征属于每个第一输出元所属发音音素的概率；
    根据所述第一输出概率与目标概率之间的差值，迭代调整所述第一权重矩阵、所述第二权重矩阵和所述第三权重矩阵，得到满足预设条件的所述语言声学模型。
  7. 如权利要求6所述的方法，所述语种判别模型包括第二输入层、第二隐藏层和第二输出层，所述基于所述音素特征训练所述语种判别模型，包括：
    在训练得到的所述多个语言声学模型中，基于所述第一隐藏层的输出值提取所述音素特征，输入至所述第二输入层；
    在所述第二输入层基于预设的第四权重矩阵计算向所述第二隐藏层输出的输出值；
    在所述第二隐藏层使用相应的第五权重矩阵进行加权计算，并将结果输出至所述第二输出层包括的多个第二输出元；
    在所述第二输出层使用第六权重矩阵进行加权计算，并基于计算结果获取第二输出概率，所述第二输出概率表示输入的所述音素特征属于每个第二输出元所属语种的概率；
    根据所述第二输出概率与所述目标概率之间的差值，迭代调整所述第四权重矩阵、所述第五权重矩阵和所述第六权重矩阵，得到满足预设条件的所述语种判别模型。
  8. 如权利要求1至4、6、7任一项所述的方法，所述音素特征包括瓶颈bottleneck特征。
  9. 一种用于自适应语种进行语音识别的装置，包括：
    提取模块，用于基于获取的语音数据提取表示发音音素信息的音素特征；
    判别模块，用于将所述音素特征输入预先基于多语种语料训练得到的语种判别模型，得到所述语音数据的语种判别结果；
    识别模块，用于基于所述语种判别结果所对应语种的语言声学模型获取所述语音数据的语音识别结果。
  10. 一种基于人工智能的语音识别方法，包括：
    分别使用各个语种的语料作为输入，训练多个语言声学模型，所述多个语言声学模型分别对应不同语种；
    利用所述多个语言声学模型分别提取所述各个语种的语料的音素特征，并基于所述音素特征训练语种判别模型；
    基于所述多个语言声学模型和所述语种判别模型对采集的语音数据进行语音识别。
  11. 如权利要求10所述的方法，所述基于所述多个语言声学模型和所述语种判别模型对采集的语音数据进行语音识别，包括：
    基于所述语音数据提取表示发音音素信息的音素特征；
    将所述音素特征输入至所述语种判别模型，得到所述语音数据的语种判别结果；
    基于所述语种判别结果所对应语种的语言声学模型获取所述语音数据的语音识别结果。
  12. 一种基于人工智能的语音识别装置，包括：
    声学模型单元，用于分别使用各个语种的语料作为输入，训练多个语言声学模型，所述多个语言声学模型分别对应不同语种；
    判别模型单元，用于利用所述多个语言声学模型分别提取所述各个语种的语料的音素特征，并基于所述音素特征训练语种判别模型；
    识别模块，用于基于所述多个语言声学模型和所述语种判别模型对采集的语音数据进行语音识别。
  13. 一种计算机可读存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时实现根据权利要求1至8或10-11中任一项所述的方法。
  14. 一种电子设备，包括：
    处理器；以及
    存储器，所述存储器上存储有计算机可读指令，所述计算机可读指令被所述处理器执行时实现根据权利要求1至8或10-11中任一项所述方法。
  15. 一种语音交互设备，包括：
    采集单元，用于采集用户的语音数据；
    处理单元，用于基于所述语音数据提取表示发音音素信息的音素特征；将所述音素特征输入预先基于多语种语料训练得到的语种判别模型，得到所述语音数据的语种判别结果；并基于所述语种判别结果所对应语种的语言声学模型获取所述语音数据的语音识别结果；
    交互单元，用于基于所述处理单元的语音识别结果向所述用户呈现相应的交互内容。
  16. 如权利要求15所述的语音交互设备，所述交互单元还用于展示所述处理单元的语种判别结果，基于接收的用户选择操作通知所述处理单元调整所述语种判别结果，并从所述处理单元获取调整后的语音识别结果。
  17. 如权利要求15或16所述的语音交互设备，所述语音交互设备包括移动终端、智能音箱、智能电视、智能穿戴设备中的任意设备，所述交互单元包括图像输出装置和语音输出装置的任意组合。
  18. 如权利要求15所述的语音交互设备，所述交互单元具体用于基于所述语音识别结果，以文字的形式直接展示识别出的语音内容。
  19. 如权利要求15所述的语音交互设备，所述交互单元还用于：
    基于所述语音识别结果生成回复内容，并以文字和/或语音的形式向所述用户呈现所述回复内容；和/或
    在基于所述语音识别结果确定所述语音数据为操作指令时，向所述用户呈现所述操作指令的执行结果。
PCT/CN2020/078806 2019-03-11 2020-03-11 基于自适应语种进行语音识别的方法及相关装置 WO2020182153A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/231,945 US20210233521A1 (en) 2019-03-11 2021-04-15 Method for speech recognition based on language adaptivity and related apparatus

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910182266.5 2019-03-11
CN201910182266.5A CN109817213B (zh) 2019-03-11 2019-03-11 用于自适应语种进行语音识别的方法、装置及设备

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/231,945 Continuation US20210233521A1 (en) 2019-03-11 2021-04-15 Method for speech recognition based on language adaptivity and related apparatus

Publications (1)

Publication Number Publication Date
WO2020182153A1 true WO2020182153A1 (zh) 2020-09-17

Family

ID=66608665

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/078806 WO2020182153A1 (zh) 2019-03-11 2020-03-11 基于自适应语种进行语音识别的方法及相关装置

Country Status (3)

Country Link
US (1) US20210233521A1 (zh)
CN (2) CN109817213B (zh)
WO (1) WO2020182153A1 (zh)

Families Citing this family (35)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109817213B (zh) * 2019-03-11 2024-01-23 腾讯科技（深圳）有限公司 用于自适应语种进行语音识别的方法、装置及设备
CN110349564B (zh) * 2019-07-22 2021-09-24 思必驰科技股份有限公司 一种跨语言语音识别方法和装置
CN110491375B (zh) * 2019-08-22 2022-02-01 新疆科大讯飞信息科技有限责任公司 一种目标语种检测的方法和装置
CN110600033B (zh) * 2019-08-26 2022-04-05 北京大米科技有限公司 学习情况的评估方法、装置、存储介质及电子设备
CN110648654A (zh) * 2019-10-09 2020-01-03 国家电网有限公司客户服务中心 一种引入语言向量的语音识别增强方法和装置
CN110970014B (zh) * 2019-10-31 2023-12-15 阿里巴巴集团控股有限公司 语音转换、文件生成、播音、语音处理方法、设备及介质
CN110930978A (zh) * 2019-11-08 2020-03-27 北京搜狗科技发展有限公司 一种语种识别方法、装置和用于语种识别的装置
CN112837674A (zh) * 2019-11-22 2021-05-25 阿里巴巴集团控股有限公司 语音识别方法、装置及相关系统和设备
CN111179938A (zh) * 2019-12-26 2020-05-19 安徽仁昊智能科技有限公司 一种基于人工智能的语音识别垃圾分类系统
CN111833865B (zh) * 2020-01-08 2024-05-24 北京嘀嘀无限科技发展有限公司 一种人机交互方法与终端、计算机可读存储介质
CN113205795A (zh) * 2020-01-15 2021-08-03 普天信息技术有限公司 多语种混说语音的语种识别方法及装置
CN111402861B (zh) * 2020-03-25 2022-11-15 思必驰科技股份有限公司 一种语音识别方法、装置、设备及存储介质
CN111488826B (zh) * 2020-04-10 2023-10-17 腾讯科技（深圳）有限公司 一种文本识别方法、装置、电子设备和存储介质
CN111613204B (zh) * 2020-04-29 2023-04-18 云知声智能科技股份有限公司 一种快速响应的神经语音合成系统及其方法
CN111627457A (zh) * 2020-05-13 2020-09-04 广州国音智能科技有限公司 语音分离方法、系统及计算机可读存储介质
CN111833845B (zh) * 2020-07-31 2023-11-24 平安科技（深圳）有限公司 多语种语音识别模型训练方法、装置、设备及存储介质
CN112017670B (zh) * 2020-08-13 2021-11-02 北京达佳互联信息技术有限公司 一种目标账户音频的识别方法、装置、设备及介质
CN112017630B (zh) * 2020-08-19 2022-04-01 北京字节跳动网络技术有限公司 一种语种识别方法、装置、电子设备及存储介质
CN112017648A (zh) * 2020-08-25 2020-12-01 北京声智科技有限公司 加权有限状态转换器构建方法、语音识别方法及装置
CN112185348B (zh) * 2020-10-19 2024-05-03 平安科技（深圳）有限公司 多语种语音识别方法、装置及电子设备
CN112259094A (zh) * 2020-10-20 2021-01-22 江苏峰鑫网络科技有限公司 一种语音客服机器人系统
CN112270923A (zh) * 2020-10-22 2021-01-26 江苏峰鑫网络科技有限公司 一种基于神经网络的语义识别系统
CN112002325B (zh) * 2020-10-27 2021-02-09 苏州思必驰信息科技有限公司 多语种语音交互方法和装置
CN112562640B (zh) * 2020-12-01 2024-04-12 北京声智科技有限公司 多语言语音识别方法、装置、系统及计算机可读存储介质
CN112652300B (zh) * 2020-12-24 2024-05-17 百果园技术(新加坡)有限公司 多方言语音识别方法、装置、设备和存储介质
CN112885330A (zh) * 2021-01-26 2021-06-01 北京云上曲率科技有限公司 一种基于低资源音频的语种识别方法及系统
CN113345418A (zh) * 2021-06-09 2021-09-03 中国科学技术大学 基于跨语种自训练的多语种模型训练方法
CN114038457B (zh) * 2021-11-04 2022-09-13 贝壳找房(北京)科技有限公司 用于语音唤醒的方法、电子设备、存储介质和程序
CN114398468A (zh) * 2021-12-09 2022-04-26 广东外语外贸大学 一种多语种识别方法和系统
CN114120972B (zh) * 2022-01-28 2022-04-12 科大讯飞华南有限公司 一种基于场景化的语音智能识别方法及系统
CN114724544B (zh) * 2022-04-13 2022-12-06 北京百度网讯科技有限公司 语音芯片、语音识别方法、装置、设备及智能汽车
CN114927135B (zh) * 2022-07-22 2022-12-13 广州小鹏汽车科技有限公司 语音交互方法、服务器及存储介质
CN117012184A (zh) * 2022-10-31 2023-11-07 腾讯科技（深圳）有限公司 一种语音识别方法和相关装置
CN115831094B (zh) * 2022-11-08 2023-08-15 北京数美时代科技有限公司 一种多语种语音识别方法、系统、存储介质和电子设备
CN116303697B (zh) * 2023-05-18 2023-08-08 深圳鹏锐信息技术股份有限公司 一种基于人工智能的模型展示系统

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20050197837A1 (en) * 2004-03-08 2005-09-08 Janne Suontausta Enhanced multilingual speech recognition system
CN103065622A (zh) * 2012-12-20 2013-04-24 中国科学院声学研究所 一种用于语种识别的语种模型的训练方法及系统
US20180137109A1 (en) * 2016-11-11 2018-05-17 The Charles Stark Draper Laboratory, Inc. Methodology for automatic multilingual speech recognition
CN108510976A (zh) * 2017-02-24 2018-09-07 芋头科技（杭州）有限公司 一种多语言混合语音识别方法
CN109192192A (zh) * 2018-08-10 2019-01-11 北京猎户星空科技有限公司 一种语种识别方法、装置、翻译机、介质和设备
CN110491382A (zh) * 2019-03-11 2019-11-22 腾讯科技（深圳）有限公司 基于人工智能的语音识别方法、装置及语音交互设备

Family Cites Families (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
ATE442641T1 (de) * 2004-07-22 2009-09-15 France Telecom Spracherkennungsverfahren und -system, das an die eigenschaften von nichtmuttersprachlern angepasst ist
JP5120826B2 (ja) * 2005-09-29 2013-01-16 独立行政法人産業技術総合研究所 発音診断装置、発音診断方法、記録媒体、及び、発音診断プログラム
US8046221B2 (en) * 2007-10-31 2011-10-25 At&T Intellectual Property Ii, L.P. Multi-state barge-in models for spoken dialog systems
CN101447184B (zh) * 2007-11-28 2011-07-27 中国科学院声学研究所 基于音素混淆的中英文双语语音识别方法
EP2736042A1 (en) * 2012-11-23 2014-05-28 Samsung Electronics Co., Ltd Apparatus and method for constructing multilingual acoustic model and computer readable recording medium for storing program for performing the method
CN103400577B (zh) * 2013-08-01 2015-09-16 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
CN104036774B (zh) * 2014-06-20 2018-03-06 国家计算机网络与信息安全管理中心 藏语方言识别方法及系统
CN104681036B (zh) * 2014-11-20 2018-09-25 苏州驰声信息科技有限公司 一种语言音频的检测系统及方法
CN106297773B (zh) * 2015-05-29 2019-11-19 中国科学院声学研究所 一种神经网络声学模型训练方法
KR102443087B1 (ko) * 2015-09-23 2022-09-14 삼성전자주식회사 전자 기기 및 그의 음성 인식 방법
CN107195296B (zh) * 2016-03-15 2021-05-04 阿里巴巴集团控股有限公司 一种语音识别方法、装置、终端及系统
CN105976812B (zh) * 2016-04-28 2019-04-26 腾讯科技（深圳）有限公司 一种语音识别方法及其设备
JP6615736B2 (ja) * 2016-11-30 2019-12-04 日本電信電話株式会社 音声言語識別装置、その方法、及びプログラム
CN108682417B (zh) * 2018-05-14 2020-05-19 中国科学院自动化研究所 语音识别中的小数据语音声学建模方法
CN108986791B (zh) * 2018-08-10 2021-01-05 南京航空航天大学 针对民航陆空通话领域的中英文语种语音识别方法及系统
CN109117484B (zh) * 2018-08-13 2019-08-06 北京帝派智能科技有限公司 一种语音翻译方法和语音翻译设备
CN109359307B (zh) * 2018-10-17 2023-07-07 深圳市沃特沃德股份有限公司 自动识别语种的翻译方法、装置及设备

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20050197837A1 (en) * 2004-03-08 2005-09-08 Janne Suontausta Enhanced multilingual speech recognition system
CN103065622A (zh) * 2012-12-20 2013-04-24 中国科学院声学研究所 一种用于语种识别的语种模型的训练方法及系统
US20180137109A1 (en) * 2016-11-11 2018-05-17 The Charles Stark Draper Laboratory, Inc. Methodology for automatic multilingual speech recognition
CN108510976A (zh) * 2017-02-24 2018-09-07 芋头科技（杭州）有限公司 一种多语言混合语音识别方法
CN109192192A (zh) * 2018-08-10 2019-01-11 北京猎户星空科技有限公司 一种语种识别方法、装置、翻译机、介质和设备
CN110491382A (zh) * 2019-03-11 2019-11-22 腾讯科技（深圳）有限公司 基于人工智能的语音识别方法、装置及语音交互设备

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
MA, ZHANYU ET AL.: "Language Identification with Deep Bottleneck Features", ARXIV:1809.08909V1 [CS.CL], 18 September 2018 (2018-09-18), DOI: 20200603130631X *
MARCOS, L.M. ET AL.: "Multi-Lingual Deep Neural Networks for Language Recognition", SLT 2016, 31 December 2016 (2016-12-31), DOI: 20200603130759A *

Also Published As

Publication number Publication date
CN109817213A (zh) 2019-05-28
CN109817213B (zh) 2024-01-23
CN110491382A (zh) 2019-11-22
US20210233521A1 (en) 2021-07-29
CN110491382B (zh) 2020-12-04

Similar Documents

Publication Publication Date Title
WO2020182153A1 (zh) 基于自适应语种进行语音识别的方法及相关装置
WO2021093449A1 (zh) 基于人工智能的唤醒词检测方法、装置、设备及介质
CN111933129B (zh) 音频处理方法、语言模型的训练方法、装置及计算机设备
JP6802005B2 (ja) 音声認識装置、音声認識方法及び音声認識システム
CN109036384B (zh) 语音识别方法和装置
CN113205817B (zh) 语音语义识别方法、系统、设备及介质
WO2020073944A1 (zh) 语音合成方法及设备
CN111312245B (zh) 一种语音应答方法、装置和存储介质
WO2022078146A1 (zh) 语音识别方法、装置、设备以及存储介质
CN110853618A (zh) 一种语种识别的方法、模型训练的方法、装置及设备
CN112214591B (zh) 一种对话预测的方法及装置
CN111243599B (zh) 语音识别模型构建方法、装置、介质及电子设备
CN111833845A (zh) 多语种语音识别模型训练方法、装置、设备及存储介质
CN113837299B (zh) 基于人工智能的网络训练方法及装置、电子设备
US20230127787A1 (en) Method and apparatus for converting voice timbre, method and apparatus for training model, device and medium
Sen et al. Speech processing and recognition system
CN110853669A (zh) 音频识别方法、装置及设备
CN112562723B (zh) 发音准确度确定方法、装置、存储介质和电子设备
CN112017690B (zh) 一种音频处理方法、装置、设备和介质
CN113393841A (zh) 语音识别模型的训练方法、装置、设备及存储介质
CN111968646A (zh) 一种语音识别方法及装置
KR20210051523A (ko) 도메인 자동 분류 대화 시스템
CN116978367A (zh) 语音识别方法、装置、电子设备和存储介质
CN113763925B (zh) 语音识别方法、装置、计算机设备及存储介质
KR102635031B1 (ko) 의미 단위 시각화를 기반으로 한 화자 분리 실시간 통역 서비스 제공 방법, 장치 및 시스템

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: 20768991

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20768991

Country of ref document: EP

Kind code of ref document: A1