WO2019080248A1 - 语音识别方法、装置、设备以及计算机可读存储介质 - Google Patents

语音识别方法、装置、设备以及计算机可读存储介质

Info

Publication number
WO2019080248A1
WO2019080248A1 PCT/CN2017/113230 CN2017113230W WO2019080248A1 WO 2019080248 A1 WO2019080248 A1 WO 2019080248A1 CN 2017113230 W CN2017113230 W CN 2017113230W WO 2019080248 A1 WO2019080248 A1 WO 2019080248A1
Authority
WO
WIPO (PCT)
Prior art keywords
information
speech
feature
information block
frame
Prior art date
Legal status (The legal status 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 status listed.)
Ceased
Application number
PCT/CN2017/113230
Other languages
English (en)
French (fr)
Inventor
梁浩
程宁
王健宗
肖京
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Ping An Technology Shenzhen Co Ltd
Original Assignee
Ping An Technology Shenzhen Co Ltd
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 Ping An Technology Shenzhen Co Ltd filed Critical Ping An Technology Shenzhen Co Ltd
Priority to SG11202001627XA priority Critical patent/SG11202001627XA/en
Priority to US16/642,371 priority patent/US11081103B2/en
Publication of WO2019080248A1 publication Critical patent/WO2019080248A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00—Speech recognition
    • G10L15/28—Constructional details of speech recognition 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
    • G10L15/26—Speech to text systems
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N7/00—Computing arrangements based on specific mathematical models
    • G06N7/01—Probabilistic graphical models, e.g. probabilistic networks
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00—Speech recognition
    • G10L15/02—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
    • G10L15/06—Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00—Speech recognition
    • G10L15/08—Speech classification or search
    • G10L15/16—Speech classification or search using artificial neural networks
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00—Speech recognition
    • G10L15/08—Speech classification or search
    • G10L15/18—Speech classification or search using natural language modelling
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Definitions

  • the present application relates to the field of computer technologies, and in particular, to a voice recognition method, apparatus, device, and computer readable storage medium.
  • Speech recognition technology also known as Automatic Speech Recognition (Automatic Speech Recognition, ASR), whose goal is to make the machine turn the voice signal into words through recognition and understanding, is an important branch of modern artificial intelligence development.
  • ASR Automatic Speech Recognition
  • the realization of speech recognition technology is the premise of natural language processing, and can effectively promote the development of voice-activated interaction related fields and greatly facilitate people's lives, such as smart home, voice input, making some people who are not suitable for using hands and eyes, such as middle-aged People, and the environment, such as driving, on the road and other scenes to make command operations become a reality.
  • the accuracy of speech recognition directly determines the effectiveness of technical applications. In recent years, with the development of deep learning technology, the accuracy of speech recognition has been greatly improved.
  • LSTM-RNN Long Short-term Memory Recurrent Neural Network
  • LSTM-RNN Long Short-term Memory Recurrent Neural Network
  • BLSTM Bidirectional Long Short-term Memory
  • the embodiment of the present invention provides a voice recognition method, device, device, and computer readable storage medium, which can greatly improve the effect of voice recognition, effectively shorten the delay of voice decoding, and enable real-time voice based on two-way long-term recurrent neural network.
  • the identified application is possible.
  • an embodiment of the present application provides a voice recognition method, where the method includes:
  • each The information block includes a master information block, each master information block is spliced with a slave information block, each master information block and its corresponding slave information block include a plurality of frame voice feature information, and the last frame voice feature of the master information block
  • the information is adjacent to the stitched first frame speech feature information of the information block; all the information blocks are input into the preset bidirectional long-term recurrent neural network model in parallel to calculate the main information block in each information block separately
  • the forward recognition result under the corresponding neuron node of the hidden layer of the preset bidirectional long-term recurrent neural network model, and the hidden state of each information block in the preset bidirectional long-term recurrent neural network model The result of the backward recognition under the corresponding neuron node of the layer; obtaining
  • the embodiment of the present application further provides a voice recognition device, where the device includes:
  • a first processing unit configured to perform framing and acoustic feature extraction on the speech information to be measured according to a preset rule to obtain a frame-level speech feature sequence
  • a second processing unit configured to sequentially divide the frame-level speech feature sequence into n An information block, where n is an integer greater than or equal to 2, each information block includes a primary information block, each primary information block is spliced with a secondary information block, and each primary information block and its corresponding secondary information block include a plurality of frame voices.
  • the identification unit is configured to input all the information blocks into the preset bidirectional length and time in parallel
  • the forward recognition result of the main information block in each information block under the corresponding neuron node of the hidden layer of the preset bidirectional long-term recurrent neural network model is respectively calculated, and a backward recognition result of each information block under the corresponding neuron node of the hidden layer of the recursive neural network model in the preset two-way length; an output unit for obtaining The forward recognition result corresponding to the information block and the output result of the backward recognition result under the corresponding neuron node of the output layer of the preset two-way long recursive neural network model to obtain the voice information to be tested The speech recognition result;
  • the preset bidirectional long-term recurrent neural network model comprises an input layer composed of neuron nodes corresponding to all information blocks, a hidden layer
  • the embodiment of the present application further provides a voice recognition device, including: a memory, configured to store a program for implementing voice recognition; and a processor, configured to run a program for implementing voice recognition stored in the memory, to The method as described in the first aspect above is performed.
  • a voice recognition device including: a memory, configured to store a program for implementing voice recognition; and a processor, configured to run a program for implementing voice recognition stored in the memory, to The method as described in the first aspect above is performed.
  • an embodiment of the present application further provides a computer readable storage medium storing one or more programs, the one or more programs being executable by one or more processors, To achieve the method as described in the first aspect above.
  • the embodiment of the present application introduces a delay control method to enable the "past" and "future” voice information in the two-way long-term recurrent neural network to be effectively utilized at the same time, which greatly improves the recognition effect, and because of the preset two-way length
  • the recursive neural network model can start processing when the whole sentence is just read in, which effectively shortens the decoding delay, making it possible to apply real-time speech recognition based on two-way long-term recursive neural network.
  • FIG. 1 is a schematic flowchart of a voice recognition method according to an embodiment of the present application
  • 1a is a block diagram of voice information to be tested in the embodiment of the present application.
  • FIG. 1b is a schematic diagram of a recognition direction of voice information to be tested in the embodiment of the present application.
  • FIG. 2 is another schematic flowchart of a voice recognition method according to an embodiment of the present application.
  • FIG. 3 is another schematic flowchart of a voice recognition method according to an embodiment of the present application.
  • FIG. 5 is a schematic flowchart of a voice recognition method according to another embodiment of the present application.
  • FIG. 6 is a schematic block diagram of a voice recognition apparatus according to an embodiment of the present application.
  • FIG. 7 is another schematic block diagram of a voice recognition apparatus according to an embodiment of the present application.
  • FIG. 8 is another schematic block diagram of a voice recognition apparatus according to an embodiment of the present application.
  • FIG. 9 is another schematic block diagram of a voice recognition apparatus according to an embodiment of the present application.
  • FIG. 10 is a schematic block diagram of a voice recognition apparatus according to another embodiment of the present application.
  • FIG. 11 is a schematic structural diagram of a voice recognition device according to an embodiment of the present application.
  • FIG. 1 is a schematic flowchart of a voice recognition method according to an embodiment of the present application.
  • the method can be run on terminals such as smart phones (such as Android phones, IOS phones, etc.), tablets, laptops, and smart devices.
  • terminals such as smart phones (such as Android phones, IOS phones, etc.), tablets, laptops, and smart devices.
  • the steps of the method include S101 ⁇ S104.
  • S101 Perform framing and acoustic feature extraction on the voice information to be tested according to a preset rule to obtain a frame-level voice feature sequence.
  • the voice information to be tested needs to be processed according to a preset rule to obtain an identifiable voice feature, that is, a frame-level voice feature sequence is obtained.
  • acoustic feature extraction is also required to obtain a frame-level speech feature sequence.
  • the step S101 includes S201 ⁇ S203.
  • the components having discriminative characteristics in each speech frame after the frame processing are extracted, and the feature sequence of the Mel frequency cepstral coefficients corresponding to each speech frame is obtained.
  • a fast Fourier transform can be performed on all input speech frames, and the logarithmic energy of each filter bank output can be calculated.
  • the Mel frequency cepstrum coefficient can be obtained by discrete cosine transform (mel Frequency cepstrum Coefficient, MFCC), which is the MFCC acoustic feature, simultaneously outputs the Mel frequency cepstral coefficient feature sequence.
  • MFCC discrete cosine transform
  • the characteristic sequence of the Mel frequency cepstrum coefficient is a feature sequence of "Dimensions of MFCC ⁇ Number of Framings".
  • the Mel frequency cepstral coefficient characteristic sequence may be input into a preset phoneme model, thereby obtaining a corresponding frame-level speech feature sequence.
  • the phoneme model may be a hidden Markov and a mixed Gaussian model, and the hidden Markov and the mixed Gaussian model may be a decoded network model trained by existing speech sample data, the Mel frequency cepstrum coefficient feature sequence.
  • the monophone model, the context triphone model, and the triphone model based on linear discriminant analysis and maximum likelihood linear feature conversion are different preset hidden Markovs and mixed Gaussians. model.
  • the step S203 may include S301 ⁇ S303.
  • S301 Input the Mel frequency cepstral coefficient feature sequence into a preset monophone model to obtain aligned first feature information.
  • the Mel frequency cepstral coefficient feature sequence is input into the preset single phoneme model, and a first feature information is obtained, and the first feature information is more accurate.
  • the second feature information is input into a preset three-phone model based on linear discriminant analysis and maximum likelihood linear feature conversion to obtain an aligned frame-level speech feature sequence.
  • a final frame-level speech feature sequence can be obtained, and the frame-level speech feature sequence is obtained.
  • the frame-level speech feature sequence may be sequentially divided into n information blocks according to chronological order, and each information block includes a main information block Ai and a corresponding sub-information.
  • the value of the block Bi, i ranges from 1 ⁇ i ⁇ n, wherein the main information block A1 to the main information block An are divided according to time, and each main information block Ai and the corresponding sub-information block Bi are both Includes several frames of speech feature information.
  • the first information block includes a main information block A1 and a slave information block B1
  • the second information block includes a main information block A2 and a slave information block B2
  • the nth information block includes a main information block An and a slave information.
  • S103 Enter all the information blocks into the preset two-way long-time recurrent neural network model in parallel, to respectively calculate the hidden layer of the main information block in each information block in the preset bidirectional long-term recurrent neural network model.
  • all information blocks can be input into the preset two-way long-term recurrent neural network model in parallel.
  • the speech information to be tested is related to time, and each frame of the speech feature information in the frame-level speech feature sequence also corresponds to a corresponding time. t, so when the forward recognition is performed, the voice feature information corresponding to the time T can be forwardly recognized by the voice feature information corresponding to the time T+1, and the voice feature information corresponding to the T+1 is immediately after the recognition is completed.
  • the speech feature information corresponding to T is back-recognized, thereby improving the accuracy of speech information recognition and greatly shortening the delay time of speech recognition.
  • Parallel input of all information blocks into a preset two-way long-term recurrent neural network model means that all information blocks are input into the neuron nodes of the input layer of the preset two-way recursive neural network model in time sequence.
  • Each of the information blocks corresponds to a neuron node of the input layer, and the main information block of each information block passes through the neuron node of the input layer and then recursively neurons of the hidden layer of the neural network model in the preset two-way length
  • the forward recognition is performed in the node. After the forward recognition is completed, each information block can pass through the neuron node of the input layer and then recursively the neuron node of the hidden layer of the neural network model in the preset two-way length.
  • the first frame of speech feature information for backward recognition and backward recognition is the last frame of speech feature information in the information block.
  • the main information block A1 in the first information block enters the recursive neural network model after the preset two-way length
  • the specific two-dimensional long-term recurrent recurrent neural network model is hidden in the neuron node.
  • the voice feature information in the entire first information block needs to be reversely input into the preset two-way length from the last frame of the voice feature information in the information block B1.
  • the backward recognition is performed and the backward recognition result is obtained.
  • other information blocks of the two-way long-term recursive neural network model of the parallel input preset also perform corresponding forward recognition and backward recognition, so as to realize real-time recognition of speech.
  • the preset bidirectional long-term recursive neural network model is constructed including steps S401-S404.
  • S401 Perform framed and acoustic feature extraction on the selected voice sample data to obtain a Mel frequency cepstral coefficient feature sequence.
  • a fast Fourier transform is performed on each speech frame, and the time domain speech signal is converted into a frequency domain energy spectrum for analysis.
  • the energy spectrum is then passed through a set of Mel-scale triangular filter banks to highlight the formant characteristics of the speech.
  • the logarithmic energy of each filter bank output is then calculated. After the logarithmic energy is calculated, the characteristic sequence of the Mel frequency cepstral coefficient can be obtained by discrete cosine transform.
  • the triphone model after training through a monophone model, usually a complete hidden Markov and mixed Gaussian model of the speech recognition process will also perform several rounds of triphone model training.
  • the triphone model training considers the current phoneme before and after. Related phoneme effects can result in more accurate alignment and better recognition results.
  • the triphone model generally uses context-dependent (delta+delta-delta, ⁇ + ⁇ ) features, linear discriminant analysis and maximum likelihood linear feature conversion (linear Discriminant analysis + maximum likelihood linear transform, LDA+MLLT) and speaker adaptive and feature space maximum likelihood probability linear regression (speaker adaptive training + feature space The maximum likelihood linear regression, SAT+fMLLR) triathlon feature model was trained.
  • the triphone model based on the delta+delta-delta feature is a model training that adds delta features to the original Mel frequency cepstral coefficient feature. For example, if the input original Mel frequency cepstral coefficient characteristic is 13-dimensional, the feature input after adding the delta+delta-delta feature becomes 39-dimensional.
  • the triphone model based on linear discriminant analysis and maximum likelihood linear feature transformation is to perform frame superposition on the improved delta+delta-delta feature data, increase the feature dimension, and then analyze the feature change by dimensionality analysis. Perform model training for triphones.
  • the triphone feature model based on SAT+fMLLR features requires that the original training data samples can not be affected by the speaker's phonetic intonation and the environment, so the same speaker in the case of different voices when the voice input is helpful, the practical application Also less.
  • the primary phoneme training, the triphone training based on the delta+delta-delta feature, the triphone training based on the delta+delta-delta feature, and the linear discriminant analysis+maximum likelihood linear feature conversion are used. Hidden Markov and mixed Gaussian model training and alignment in automatic speech recognition.
  • the general framework for training two-way long-term recurrent neural network models is to train hidden Markov and mixed Gaussian models to perform forced alignment, one or more, and then combine the results of forced alignment with the input features of the original speech samples as a two-way length.
  • the training samples of the recurrent neural network are trained.
  • the research shows that the two-way long-term recurrent neural network model obtained by training can reduce the error rate of speech recognition based on hidden Markov and mixed Gaussian models. For example, the error rate can be reduced by 20%-30%.
  • the aligned speech sample feature sequences are sequentially divided into n sample information blocks, where n is an integer greater than or equal to 2, each sample information block includes a main sample information block, and the last one of each main sample information block
  • the frame speech feature information is spliced with a corresponding slave sample information block, and each main sample information block and its corresponding slave sample information block include several frame speech feature information, and the last frame speech of each main sample information block
  • the feature information is adjacent to the corresponding first frame speech feature information of the sample information block.
  • the frame-level speech feature sequence is read by a chunk processing method in the forward and backward calculation of the neural network model.
  • the frame-level speech feature sequences are equally divided into a plurality of main information blocks (edge-cut), and in the reverse calculation, the fixed-length speech feature information is spliced to
  • the speech feature information of the fixed frame length is composed here to form a corresponding slave information block, and each main information block and the corresponding slave information block constitute an information block.
  • the control makes these later spliced speech feature information correspondingly not output, so no additional error is generated for the calculation of the original main information block, which makes the training speed relatively and needs to be read.
  • the traditional two-way long-term recurrent neural network training into the whole voice has been improved by more than 10 times in the test environment.
  • the initial state of the memory unit forwarded at time T can be directly calculated by the main information block at time T-1, and the initial state of the backward memory unit at time T is the fixed frame spliced by the main information block.
  • the long slave block is generated, so the delay of the decoding application is determined only by the voice feature information in the block, so that the total time taken for voice recognition is reduced. Therefore, it is not necessary to wait until the entire voice information to be tested can be read to start the backward calculation, and only wait for the time taken from the information block, so that the voice recognition of the recursive neural network application in real-time two-way length is possible;
  • the mini-batch method is used for parallel training, which reduces the computational memory consumption. In addition, the accuracy of speech recognition can be effectively improved, which is practical and expandable.
  • S404 input all the sample information blocks into a bidirectional long-term recurrent neural network in parallel, and perform forward training on the bidirectional long-term recurrent neural network by using the main sample information block in each sample information block, and each sample information.
  • the block performs backward training on the bidirectional long-term recurrent neural network to obtain a preset bidirectional long-term recurrent neural network model.
  • the forward training process of the two-way long-term recurrent neural network is the same as the processing process of the one-way long-term recurrent neural network, and covers all the "history" information related to the input frame.
  • the backward training process in order to reduce the delay factor caused by the reverse reading of the entire sentence information, it is necessary to adopt the manner of splicing the speech frame with the "future" information in the above steps.
  • a 3-5 layer long and short recursive neural network model can be trained first, so that the output of the model is forcedly aligned, and the result is forced to be sent to the bidirectional long-term recurrent neural network, and finally the decoded output is obtained.
  • a two-way long-term recurrent neural network model is obtained.
  • the bidirectional long-term recurrent neural network is forwardly trained by the main sample information block in each sample information block.
  • Each sample information block performs backward training on the two-way long-term recurrent neural network to obtain a preset bidirectional long-term recurrent neural network model, and each information block is trained by parallel input bidirectional long-term recurrent neural network, and reversed.
  • the delay caused by the training is only the time spent in training each block, so the time spent on the entire training process is greatly reduced.
  • the structure of the two-way long-term recursive neural network model based on delay control can be as follows: there are 1024 neuron nodes in each hidden layer of the model, of which 512 are forward-recognizing and 512 are backward-recognizing.
  • the hidden layer is followed by a projection layer with 300 neuron nodes, and the final output layer is a Softmax layer. All neuronal nodes are used
  • the uniform random weight is initialized in the interval [-0.05 ⁇ 0.05], and the offset is initially 0.
  • Neural network training uses cross-entropy (corss-entropy) criteria and truncated back-interval (truncated) Back-propagation-through-time) optimization training method.
  • Each information block identifiable by the model may contain 20 frames of voice information, and each minibatch contains 40 pronunciation sentences.
  • the first iteration (epoch) uses 0, and the subsequent iteration uses 0.9.
  • the learning rate update if the iteration does not increase, the learning rate of the last iteration is halved as the learning rate of the next iteration training.
  • the corresponding neuron nodes in the output layer of the recursive neural network model of the preset two-way length are performed.
  • Output that is, an output result, which is the result of speech recognition.
  • the embodiment of the present application effectively shortens the decoding delay, and makes it possible to apply real-time speech recognition based on the reciprocal neural network in the two-way length.
  • FIG. 5 is a schematic flowchart of a voice recognition method according to another embodiment of the present application. As shown in FIG. 5, the steps of the method include S501a ⁇ S504.
  • the voice information to be tested is first pre-emphasized, and the high-frequency part in the voice information to be tested is improved, so that the spectrum is smoother. Then, the framed windowing is performed to convert the non-stationary speech signal into a short-time stationary signal analysis process. Then through the endpoint detection, the speech and noise are distinguished, and the effective speech part is extracted, thereby improving the accuracy of the identification of the speech information to be tested.
  • S501 Perform framed and acoustic feature extraction on the pre-processed speech information according to a preset rule to obtain a frame-level speech feature sequence.
  • the frame-level speech feature sequence is sequentially divided into n information blocks, where n is an integer greater than or equal to 2, each information block includes a main information block, and each main information block is spliced with a sub-information block, each The master information block and its corresponding slave information block each include a plurality of frame voice feature information, and the last frame voice feature information of the master information block is adjacent to the stitched first frame voice feature information of the slave information block.
  • S503 input all the information blocks into the preset bidirectional long-term recurrent neural network model in parallel, to respectively calculate the hidden layer of the main information block in each information block in the preset bidirectional long-term recurrent neural network model The forward recognition result under the corresponding neuron node, and the backward recognition result of each information block under the corresponding neuron node of the hidden layer of the preset bidirectional long-term recurrent neural network model.
  • the preset bidirectional long-term recurrent neural network model includes an input layer composed of neuron nodes corresponding to all information blocks, a hidden layer composed of nonlinear mapping units, and a neuron node corresponding to the output result of the hidden layer.
  • the output layer consisting of points.
  • the embodiment of the present application further provides a voice recognition device, where the device 100 includes: a first processing unit 101, a second processing unit 102, an identification unit 103, and an output unit 104.
  • the first processing unit 101 is configured to perform framing and acoustic feature extraction on the voice information to be tested according to a preset rule to obtain a frame-level voice feature sequence.
  • the first processing unit 101 includes a framing unit 201, a feature extraction unit 202, and an adjustment unit 203.
  • the framing unit 201 is configured to perform framing on the voice information to be tested to obtain a plurality of voice frames.
  • the feature extraction unit 202 is configured to perform feature extraction on each speech frame to obtain a Mel frequency cepstral coefficient feature sequence corresponding to each speech frame.
  • the adjusting unit 203 is configured to input the Mel frequency cepstral coefficient feature sequence into a preset phoneme model to obtain an aligned frame-level speech feature sequence.
  • the adjustment unit 203 includes a first input unit 301 , a second input unit 302 , and a third input unit 303 .
  • the first input unit 301 is configured to input the Mel frequency cepstral coefficient feature sequence into a preset monophone model to obtain aligned first feature information.
  • the second input unit 302 is configured to input the first feature information into a preset context triphone model to obtain aligned second feature information.
  • the third input unit 303 is configured to input the second feature information into a preset three-phone model based on linear discriminant analysis and maximum likelihood linear feature conversion to obtain an aligned frame-level speech feature sequence.
  • the monophone model, the context triphone model, and the triphone model based on linear discriminant analysis and maximum likelihood linear feature transformation are different preset hidden Markov and mixed Gaussian models.
  • the second processing unit 102 is configured to sequentially divide the frame-level speech feature sequence into n information blocks, where n is an integer greater than or equal to 2, each information block includes a main information block, and each main information block is spliced From the information block, each of the main information blocks and the corresponding sub-blocks thereof comprise a plurality of frame speech feature information, a last frame speech feature information of the main information block and a first frame speech feature of the spliced sub-block Information is adjacent.
  • the identifying unit 103 is configured to input all the information blocks into the preset two-way long-term recursive neural network model in parallel, to respectively calculate the main information block in each information block in the preset bidirectional long-term recurrent neural network model.
  • the forward recognition result under the corresponding neuron node of the hidden layer, and the backward direction of each information block under the corresponding neuron node of the hidden layer of the preset bidirectional long-term recurrent neural network model Identify the results.
  • the output unit 104 is configured to obtain a forward recognition result corresponding to each information block and an output result of the backward recognition result under the corresponding neuron node of the output layer of the preset bidirectionally long recursive neural network model, Obtaining a voice recognition result of the voice information to be tested.
  • the preset bidirectional long-term recurrent neural network model includes an input layer composed of neuron nodes corresponding to all information blocks, a hidden layer composed of nonlinear mapping units, and a neuron node corresponding to the output result of the hidden layer.
  • the output layer consisting of points.
  • the embodiment of the present application effectively shortens the decoding delay, making it possible to apply real-time speech recognition based on the recursive neural network in the two-way length.
  • the devices in the embodiments of the present application are all corresponding to the methods in the foregoing embodiments, so the specific analysis and principles are not described herein again.
  • the apparatus 100 further includes a third processing unit 401, a first training unit 402, a fourth processing unit 403, and a second training unit 404, for implementing a preset bidirectional length and time. Construction of a recurrent neural network model.
  • the third processing unit 401 is configured to perform framing and acoustic feature extraction on the selected voice sample data to obtain a Mel frequency cepstral coefficient feature sequence.
  • the first training unit 402 is configured to train a hidden Markov and a mixed Gaussian model by using a Mel frequency cepstral coefficient feature sequence to achieve alignment of a frame-level speech sample feature sequence with text content, wherein the hidden Markov The model is in one-to-one correspondence with the speech sample data, and the mixed Gaussian model is used to describe an output probability distribution of the hidden Markov model state.
  • the fourth processing unit 403 is configured to sequentially divide the aligned speech sample feature sequences into n sample information blocks, where n is an integer greater than or equal to 2, and each sample information block includes a main sample information block, each The last frame of the main sample information block is spliced with a corresponding sub-sample information block, and each main sample information block and its corresponding sub-sample information block include several frames of speech feature information, and each main sample The last frame of speech information of the information block is adjacent to the corresponding first frame speech feature information of the sample information block.
  • the second training unit 404 is configured to input all the sample information blocks into a bidirectional long-term recurrent neural network in parallel, and respectively perform the bidirectional long-term recurrent neural network by using the main sample information block in each sample information block. The training is performed, and each of the sample information blocks is back-trained to the two-way long-term recurrent neural network to obtain a preset bidirectional long-term recurrent neural network model.
  • another embodiment of the present application further provides a voice recognition apparatus, where the apparatus 500 includes: a pre-processing unit 501a, a first processing unit 501, a second processing unit 502, and identification. Unit 503 and output unit 504.
  • the pre-processing unit 501a performs pre-processing on the voice information to be tested.
  • the first processing unit 501 is configured to perform framing and acoustic feature extraction on the pre-processed speech information according to a preset rule to obtain a frame-level speech feature sequence.
  • the second processing unit 502 is configured to sequentially divide the frame-level speech feature sequence into n information blocks, where n is an integer greater than or equal to 2, each information block includes a main information block, and each main information block is spliced From the information block, each of the main information blocks and the corresponding sub-blocks thereof comprise a plurality of frame speech feature information, a last frame speech feature information of the main information block and a first frame speech feature of the spliced sub-block Information is adjacent.
  • the identifying unit 503 is configured to input all the information blocks into the preset bidirectional long-term recursive neural network model in parallel, to respectively calculate the main information block in each information block in the preset bidirectional long-term recurrent neural network model.
  • the forward recognition result under the corresponding neuron node of the hidden layer, and the backward direction of each information block under the corresponding neuron node of the hidden layer of the preset bidirectional long-term recurrent neural network model Identify the results.
  • the output unit 504 is configured to obtain a forward recognition result corresponding to each information block and an output result of the backward recognition result under the corresponding neuron node of the output layer of the preset bidirectionally long recursive neural network model, Obtaining a voice recognition result of the voice information to be tested.
  • the preset bidirectional long-term recurrent neural network model includes an input layer composed of neuron nodes corresponding to all information blocks, a hidden layer composed of nonlinear mapping units, and a neuron node corresponding to the output result of the hidden layer.
  • the output layer consisting of points.
  • FIG. 11 is a schematic structural diagram of a voice recognition device according to the present application.
  • the device may be a terminal or a server, wherein the terminal may be a communication-enabled electronic device such as a smart phone, a tablet computer, a notebook computer, a desktop computer, a personal digital assistant, and a wearable device.
  • the server can be a standalone server or a server cluster consisting of multiple servers.
  • the computer device 600 includes a processor 602, a non-volatile storage medium 603, an internal memory 604, and a network interface 605 connected by a system bus 601.
  • the non-volatile storage medium 603 of the computer device 600 can store an operating system 6031 and a computer program 6032.
  • the processor 602 can be caused to perform a voice recognition method.
  • the processor 602 of the computer device 600 is used to provide computing and control capabilities to support the operation of the entire computer device 600.
  • the internal memory 604 provides an environment for the operation of a computer program in a non-volatile storage medium that, when executed by the processor, causes the processor 602 to perform the speech recognition method of the above-described embodiments.
  • the network interface 605 of the computer device 600 is used to perform network communications, such as sending assigned tasks and the like. It will be understood by those skilled in the art that the structure shown in FIG. 11 is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation of the computer device to which the solution of the present application is applied.
  • the specific computer device may It includes more or fewer components than those shown in the figures, or some components are combined, or have different component arrangements.
  • the application further provides a computer readable storage medium storing one or more programs, the one or more programs being executable by one or more processors to implement the above-described embodiments Speech recognition method.
  • the foregoing storage medium of the present application includes: a magnetic disk, an optical disk, and a read-only storage memory (Read-Only) Memory, ROM, and other media that can store program code.
  • the units in all embodiments of the present application may be through a general-purpose integrated circuit, such as a CPU (Central Processing) Unit, central processing unit, or through ASIC (Application Specific Integrated Circuit, ASIC) to achieve.
  • the steps in the voice recognition method of the embodiment of the present application may be sequentially adjusted, merged, and deleted according to actual needs.
  • the units in the voice recognition terminal of the embodiment of the present application may be combined, divided, and deleted according to actual needs.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Multimedia (AREA)
  • Acoustics & Sound (AREA)
  • Human Computer Interaction (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Molecular Biology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Mathematical Optimization (AREA)
  • Mathematical Analysis (AREA)
  • Medical Informatics (AREA)
  • Pure & Applied Mathematics (AREA)
  • Computational Mathematics (AREA)
  • Algebra (AREA)
  • Probability & Statistics with Applications (AREA)
  • Telephonic Communication Services (AREA)
  • Machine Translation (AREA)

Abstract

一种语音识别方法、装置、设备以及计算机可读存储介质，其中语音识别方法包括根据预设规则对待测语音信息进行分帧和声学特征提取以得到帧级语音特征序列（S101）；将所述帧级语音特征序列依次分为n个信息块（S102）；将所有信息块并行输入预设的双向长短时递归神经网络模型中；获取每个信息块对应的前向识别结果以及后向识别结果在所述预设的双向长短时递归神经网络模型的输出层的相应的神经元结点下的输出结果，以得到所述待测语音信息的语音识别结果。该方法可以提高语音识别的效果，缩短语音解码的时延。

Description

语音识别方法、装置、设备以及计算机可读存储介质
本申请要求于2017年10月23日提交中国专利局、申请号为CN 2017109942685 、申请名称为“语音识别方法、装置、设备以及计算机可读存储介质”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
本申请涉及计算机技术领域，尤其涉及一种语音识别方法、装置、设备以及计算机可读存储介质。
背景技术
语音识别技术，也被称为自动语音识别(Automatic Speech Recognition, ASR)，其目标是让机器通过识别和理解，把语音信号变成文字，是现代人工智能发展的重要分支。语音识别技术的实现是自然语言处理的前提，并能有效推动声控交互相关领域的发展并极大方便人们的生活，如智能家居、语音输入，使得一些不适宜使用手和眼睛的人群如中老年人，和环境如开车、路上等场景进行命令操作成为现实。语音识别的准确度直接决定了技术应用的有效性。近年来，随着深度学习技术的发展，语音识别准确度得到了极大的提升。
长短时递归神经网络(Long Short-term Memory Recurrent Neural Network，LSTM-RNN)被认为是一种高效并且应用广泛的语音识别领域的深度学习框架。LSTM通过在传统递归神经网络RNN基础上引入长短时记忆模块，有效解决了传统简单RNN训练梯度消失等问题，此外RNN的网络结构使得语音时序序列中时间维度的信息得到利用，因而较之传统深度学习框架如，实用性和准确度上有了很大提升。而双向长短时递归神经网络(Bidirectional Long Short-term Memory，BLSTM)在对当前语音帧进行判断时，不仅可以有效利用“历史”的语音信息，还可以利用“未来”的语音信息，这使得原始输入的语音特征能够得到更有效的挖掘，从而进行更加准确的决策。尽管在研究领域已经取得极大的准确率提升，BLSTM在实用性上却存在先天不足，因为不像单向LSTM，BLSTM的解码器需要等到整条语句都读入后才能进行双向的处理计算，亦即神经网络中的前向传播在反向处理时存在极大的时延，想象输入的语音越长时延越大，这让实时语音识别变得不再可能。
发明内容
本申请实施例提供一种语音识别方法、装置、设备以及计算机可读存储介质，能够极大地提高语音识别的效果，有效缩短了语音解码的时延，使得基于双向长短时递归神经网络的语音实时识别的应用成为可能。
第一方面，本申请实施例提供了一种语音识别方法，该方法包括：
根据预设规则对待测语音信息进行分帧和声学特征提取以得到帧级语音特征序列；将所述帧级语音特征序列依次分为n个信息块，其中n为大于等于2的整数，每个信息块包括一主信息块，每一主信息块拼接有一从信息块，每个主信息块以及与其相应的从信息块均包括若干帧语音特征信息，所述主信息块的最后一帧语音特征信息与拼接的所述从信息块的第一帧语音特征信息相邻；将所有信息块并行输入预设的双向长短时递归神经网络模型中，以分别计算每个信息块中的主信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的前向识别结果，以及每个信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的后向识别结果；获取每个信息块对应的前向识别结果以及后向识别结果在所述预设的双向长短时递归神经网络模型的输出层的相应的神经元结点下的输出结果，以得到所述待测语音信息的语音识别结果；其中，所述预设的双向长短时递归神经网络模型包括由所有信息块对应的神经元结点组成的输入层、由非线性映射单元组成的隐藏层以及与隐藏层输出结果相应的神经元结点组成的输出层。
第二方面，本申请实施例还提供了一种语音识别装置，该装置包括：
第一处理单元，用于根据预设规则对待测语音信息进行分帧和声学特征提取以得到帧级语音特征序列；第二处理单元，用于将所述帧级语音特征序列依次分为n个信息块，其中n为大于等于2的整数，每个信息块包括一主信息块，每一主信息块拼接有一从信息块，每个主信息块以及与其相应的从信息块均包括若干帧语音特征信息，所述主信息块的最后一帧语音特征信息与拼接的所述从信息块的第一帧语音特征信息相邻；识别单元，用于将所有信息块并行输入预设的双向长短时递归神经网络模型中，以分别计算每个信息块中的主信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的前向识别结果，以及每个信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的后向识别结果；输出单元，用于获取每个信息块对应的前向识别结果以及后向识别结果在所述预设的双向长短时递归神经网络模型的输出层的相应的神经元结点下的输出结果，以得到所述待测语音信息的语音识别结果；其中，所述预设的双向长短时递归神经网络模型包括由所有信息块对应的神经元结点组成的输入层、由非线性映射单元组成的隐藏层以及与隐藏层输出结果相应的神经元结点组成的输出层。
第三方面，本申请实施例还提供了一种语音识别设备，包括：存储器，用于存储实现语音识别的程序；以及处理器，用于运行所述存储器中存储的实现语音识别的程序，以执行如上第一方面所述的方法。
再一方面，本申请实施例还提供了一种计算机可读存储介质，计算机可读存储介质存储有一个或者一个以上程序，所述一个或者一个以上程序可被一个或者一个以上的处理器执行，以实现如上第一方面所述的方法。
本申请实施例通过引入了时延控制方法使得双向长短时递归神经网络中的“过去”和“未来”的语音信息能够同时得到有效利用，极大提高了识别效果，同时因为预设的双向长短时递归神经网络模型可以在整条语句刚读入时就开始进行处理，有效缩短了解码时延，使得基于双向长短时递归神经网络实时语音识别的应用成为可能。
附图说明
为了更清楚地说明本申请实施例技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1是本申请实施例提供的一种语音识别方法的示意流程图；
图1a是本申请实施例中待测语音信息的分块示意图；
图1b是本申请实施例中待测语音信息的识别方向示意图；
图2是本申请实施例提供的一种语音识别方法的另一示意流程图；
图3是本申请实施例提供的一种语音识别方法的另一示意流程图；
图4是本申请实施例提供的一种语音识别方法的另一示意流程图；
图5是本申请另一实施例提供的一种语音识别方法的示意流程图；
图6是本申请实施例提供的一种语音识别装置的示意性框图；
图7是本申请实施例提供的一种语音识别装置的另一示意性框图；
图8是本申请实施例提供的一种语音识别装置的另一示意性框图；
图9是本申请实施例提供的一种语音识别装置的另一示意性框图；
图10是本申请另一实施例提供的一种语音识别装置的示意性框图；
图11是本申请实施例提供的一种语音识别设备结构组成示意图。
具体实施方式
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述。
请参阅图1，图1是本申请实施例提供的一种语音识别方法的示意流程图。该方法可以运行在智能手机（如Android手机、IOS手机等）、平板电脑、笔记本电脑以及智能设备等终端中。如图1所示，该方法的步骤包括S101~S104。
S101，根据预设规则对待测语音信息进行分帧和声学特征提取以得到帧级语音特征序列。在本申请实施例中，需要根据预设的规则对待测语音信息进行相应的处理从而得到可识别的语音特征，也即得到帧级语音特征序列。对待测语音信息进行分帧后，还需要进行声学特征提取，从而得到一个帧级语音特征序列。
进一步地，如图2所示，所述步骤S101包括S201~S203。
S201，对待测语音信息进行分帧以得到多个语音帧。在本申请实施例中，一段待测语音信息是根据时间顺序排列构成的，对所述待测语音信息进行分帧后可以得到若干语音帧，其中每一语音帧的时长可以为10ms~30ms。在一些实施例中，每一语音帧的时长可以为25ms，帧移为10ms。
S202，对每个语音帧进行特征提取，以得到每个语音帧对应的梅尔频率倒谱系数特征序列。在本申请实施例中，对分帧处理后的每个语音帧中具有辨识性的成分提取出来，得到每个语音帧对应的梅尔频率倒谱系数特征序列。具体的，可以对输入的所有语音帧进行快速傅里叶变换，并计算每个滤波器组输出的对数能量。对对数能量进行计算后，经离散余弦变换就可得到梅尔频率倒谱系数(mel frequency cepstrum coefficient，MFCC)，亦即MFCC声学特征，同时输出梅尔频率倒谱系数特征序列。梅尔频率倒谱系数特征序列为“MFCC的维度×分帧数”的特征序列。
S203，将所述梅尔频率倒谱系数特征序列输入预设的音素模型中以得到对齐的帧级语音特征序列。在本申请实施例中，为了提高语音识别的准确度，可以将所述梅尔频率倒谱系数特征序列输入预设的音素模型中，从而得到一个相应的帧级语音特征序列，此时，这里的音素模型可以为隐马尔可夫及混合高斯模型，该隐马尔可夫及混合高斯模型可以为通过现有的语音样本数据进行训练后的解码网络模型，所述梅尔频率倒谱系数特征序列通过上述解码网络模型后得到的解码结果即为帧级语音特征序列。
进一步地，如图3所示，其中所述单音素模型、上下文三音素模型以及基于线性判别分析及最大似然线性特征转换的三音素模型均为不同的预设的隐马尔可夫及混合高斯模型。所述步骤S203可以包括S301~S303。
S301，将所述梅尔频率倒谱系数特征序列输入预设的单音素模型中以得到对齐的第一特征信息。在本申请实施例中，将梅尔频率倒谱系数特征序列输入预设的单音素模型，可以得到一个第一特征信息，该第一特征信息更为精准。
S302，将所述第一特征信息输入预设的上下文三音素模型中以得到对齐的第二特征信息。在本申请实施例中，将第一特征信息输入预设的上下文三音素模型中后，可以得到一个第二特征信息，该第二特征信息相对而言也更为精准。
S303，将所述第二特征信息输入预设的基于线性判别分析及最大似然线性特征转换的三音素模型中以得到对齐的帧级语音特征序列。在本申请实施例中，将第二特征信息输入预设的基于线性判别分析及最大似然线性特征转换的三音素模型中后，可以得到最终的帧级语音特征序列，该帧级语音特征序列为更为精准的待测语音信息中包含的声学特征，更方便后续的识别及识别的准确度。
S102，将所述帧级语音特征序列依次分为n个信息块，其中n为大于等于2的整数，每个信息块包括一主信息块，每一主信息块拼接有一从信息块，每个主信息块以及与其相应的从信息块均包括若干帧语音特征信息，所述主信息块的最后一帧语音特征信息与拼接的所述从信息块的第一帧语音特征信息相邻。
在本申请实施例中，如图1a所示，可以根据时间顺序将所述帧级语音特征序列依次分为n个信息块，每个信息块均包括一主信息块Ai和一对应的从信息块Bi，i的取值范围为1≤i≤n，其中，从主信息块A1到主信息块An是根据时间的先后进行划分的，每个主信息块Ai和对应的从信息块Bi均包括若干帧语音特征信息。例如，第一个信息块包括主信息块A1以及从信息块B1，第二个信息块包括主信息块A2以及从信息块B2，……，第n个信息块包括主信息块An以及从信息块Bn。
S103，将所有信息块并行输入预设的双向长短时递归神经网络模型中，以分别计算每个信息块中的主信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的前向识别结果，以及每个信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的后向识别结果。
在本申请实施例中，可以将所有信息块并行输入预设的双向长短时递归神经网络模型中。如图1b所示，对于预设的双向长短时递归神经网络模型而言，因待测语音信息跟时间有关，且帧级语音特征序列中的每一帧语音特征信息也均对应一个相应的时间t，故当进行前向识别的时候，可以从时间T对应的语音特征信息到时间T+1对应的语音特征信息进行前向识别，识别完成后立马从T+1对应的语音特征信息到时间T对应的语音特征信息进行后向识别，从而提高了语音信息识别的准确度，同时也极大地缩短了语音识别的时延时间。
将所有信息块并行输入预设的双向长短时递归神经网络模型中是指，将所有的信息块按时间的顺序输入预设的双向长短时递归神经网络模型的输入层的神经元结点中，其中每个信息块对应一个输入层的神经元结点，每个信息块的主信息块经过输入层的神经元结点后再在预设的双向长短时递归神经网络模型的隐藏层的神经元结点中做前向识别，完成前向识别后，即可以让每个信息块经过输入层的神经元结点后再在预设的双向长短时递归神经网络模型的隐藏层的神经元结点中做后向识别，且做后向识别的第一帧语音特征信息即为从信息块中的最后一帧语音特征信息。
例如，当第一个信息块中的主信息块A1进入经预设的双向长短时递归神经网络模型后，具体经预设的双向长短时递归神经网络模型的隐藏层的神经元结点中做前向识别并得到前向识别结果后，此时，需要从从信息块B1中的最后一帧语音特征信息开始将整个第一个信息块中的语音特征信息反向输入预设的双向长短时递归神经网络模型的隐藏层的神经元结点中做后向识别并得到后向识别结果。同理，可知并行输入预设的双向长短时递归神经网络模型的其他信息块也进行相应的前向识别和后向识别，以便于实现语音的实时识别。
进一步地，如图4所示，所述预设的双向长短时递归神经网络模型的构建包括步骤S401~S404。
S401，对选取的语音样本数据进行分帧和声学特征提取以得到梅尔频率倒谱系数特征序列。其中，对选取的语音样本数据进行分帧后，对每个语音帧进行快速傅里叶变换，将时域的语音信号转换为频域的能量谱进行分析。然后将能量谱通过一组梅尔尺度的三角滤波器组，突出语音的共振峰特征。之后计算每个滤波器组输出的对数能量。对对数能量计算后，经离散余弦变换就可得到梅尔频率倒谱系数特征序列。
S402，利用梅尔频率倒谱系数特征序列训练一隐马尔可夫及混合高斯模型，以实现帧级语音样本特征序列与文本内容的对齐，其中所述隐马尔可夫模型与所述语音样本数据是一一对应的，所述混合高斯模型用来描述所述隐马尔可夫模型状态的输出概率分布。具体的，文本内容即与选取的语音样本数据像对应的文本文件中含有的内容。实现帧级语音样本特征序列与文本内容的对齐则指实现状态和音素级别的对齐。
首先可以对输入的梅尔频率倒谱系数特征序列进行归一化，默认对方差进行归一化；然后进行快速启动，即利用少量特征数据快速得到一个初始化的隐马尔可夫及混合高斯模型和决策树；之后再构造训练用的网络，对每个句子构造一个音素级别的FST网络用于解码，在训练的过程中通过不断对齐特征序列，反复训练得到中间统计量，其中隐马尔可夫模型的统计量是两个音素在FST网络中互联的边中出现的次数，高斯模型的统计量是每个概率密度函数pdf-id对应的特征累计值和特征平方累计值，统计量跟高斯模型的方差、均值的更新相关，通过不断的更新模型完成解码网络的训练，即实现一个单音素模型的训练；最后再做一次强制对齐，用于解码输出结果或者进行下一阶段的模型训练。
在一些实施例中，在通过单音素模型训练后，通常一个完整的隐马尔可夫及混合高斯模型的语音识别过程还会进行几轮三音素模型训练，三音素模型训练考虑了当前音素的前后相关音素影响，能够得到更加准确的对齐效果，也就能产生更好的识别结果。根据特征和作用的不同，三音素模型一般采用基于上下文相关(delta+delta-delta，Δ+ΔΔ)特征、线性判别分析及最大似然线性特征转换(linear discriminant analysis + maximum likelihood linear transform, LDA+MLLT)以及说话人自适应及特征空间最大似然概率的线性回归(speaker adaptive training + feature space maximum likelihood linear regression, SAT+fMLLR)三音素特征模型进行训练。
基于delta+delta-delta特征的三音素模型是对原始的梅尔频率倒谱系数特征添加了delta特征进行的模型训练。如，若输入的原始的梅尔频率倒谱系数特征是13维，添加delta+delta-delta特征后输入的特征即变成39维。基于线性判别分析及最大似然线性特征转换的三音素模型是对改进后的delta+delta-delta特征数据先进行帧叠加，增加特征维度，然后通过线性判别式分析降维进行的特征变化，之后进行三音素的模型训练。基于SAT+fMLLR特征的三音素特征模型需要原始的训练数据样本不能受说话人的语音语调以及环境等因素的影响，故对语音输入时相同说话人在不同环境的情况下帮助甚微，实际应用也较少。在一些实施例中，主要依次采用单音素训练、基于delta+delta-delta特征的三音素训练、基于delta+delta-delta特征的三音素训练和基于线性判别分析+最大似然线性特征转换，来进行自动语音识别中的隐马尔可夫及混合高斯模型训练与对齐。
训练双向长短时递归神经网络模型的一般框架是先训练隐马尔可夫及混合高斯模型来做强制对齐，可以一个也可以多个，然后将强制对齐的结果结合原始语音样本的输入特征作为双向长短时递归神经网络的训练样本进行训练。研究表明，训练得到的双向长短时递归神经网络模型能够在隐马尔可夫及混合高斯模型基础上，使得系语音识别的错误率下降，例如，错误率可以下降20%-30%。
S403，将对齐后的语音样本特征序列依次分为n个样本信息块，其中n为大于等于2的整数，每个样本信息块包括一主样本信息块，每个主样本信息块的最末一帧语音特征信息后均拼接有一相应的从样本信息块，每个主样本信息块以及与其相应的从样本信息块均包括若干帧语音特征信息，且每个主样本信息块的最末一帧语音特征信息跟相应的从样本信息块的第一帧语音特征信息相邻。
在本申请实施例中，如图1a所示，在神经网络模型的前后向计算中均采用块(chunk)处理的方式读取帧级语音特征序列的。首先根据按帧长固定将读入帧级语音特征序列无重叠的等间隔的切分为多个主信息块（边读边切），在反向计算中则拼接固定帧长的语音特征信息到每个主信息块的后面，此处拼接固定帧长的语音特征信息即组成相应的从信息块，每个主信息块以及相应的从信息块构成一信息块。在训练双向长短时递归神经网络的过程中控制使得这些后来拼接的语音特征信息对应地并不产生输出，因而针对本来的主信息块的计算不会产生额外误差，这使得训练速度相对与需读入整条语音的传统双向长短时递归神经网络训练在测试环境下提高了10倍以上。
同时，在解码过程中，T时刻前向的记忆单元初始状态可以直接由T-1时刻的主信息块计算来得到，T时刻后向的记忆单元初始状态则由主信息块后拼接的固定帧长的从信息块产生，因而解码应用的时延只由从信息块中的语音特征信息决定，从而使得语音识别的总耗费时间却得到了缩减。故无须等到读入整条待测语音信息才能开始后向计算，而只需等待从信息块占用的时长，使得实时双向长短时递归神经网络应用的语音识别成为可能；同时也可以采用很多条语音组成小批（mini-batch）的方式进行并行训练，减少了计算内存消耗；另外，也可以有效提高语音识别的准确率，有实用意义并且可拓展性好。
S404，将所有样本信息块并行输入一双向长短时递归神经网络中，分别通过每个样本信息块中的主样本信息块对所述双向长短时递归神经网络进行前向训练，以及每个样本信息块对所述双向长短时递归神经网络进行后向训练以得到预设的双向长短时递归神经网络模型。在本申请实施例中，双向长短时递归神经网络的前向训练过程与单向长短时递归神经网络的处理过程一样，都覆盖了所有输入帧相关的“历史”信息。而在后向训练过程中，为了减少反向读取整条语句信息时产生的时延影响因素，需要采用上述步骤中拼接带有“将来”信息的语音帧的方式。
例如，可以首先训练了一个3-5层长短时递归神经网络模型，从而实现对该模型的输出做强制对齐，将强制对其的结果送入双向长短时递归神经网络中，最后得到解码输出，同时得到双向长短时递归神经网络模型。具体的，分别通过每个样本信息块中的主样本信息块对所述双向长短时递归神经网络进行前向训练。每个样本信息块对所述双向长短时递归神经网络进行后向训练以得到预设的双向长短时递归神经网络模型，各个信息块是并行输入双向长短时递归神经网络进行训练的，同时反向训练时造成的时延只是每个从信息块训练时所耗的时间，所以整个训练过程所耗的时间大大地减少了。
例如，基于时延控制的双向长短时递归神经网络模型的结构可以如下：该模型的每个隐含层有1024个神经元结点，其中前向识别的为512个，后向识别的为512个，隐含层后为带有含有300个神经元结点的投影层，最后的输出层采用Softmax层。所有神经元结点均采用 [-0.05~0.05]区间的均匀随机权重初始化，偏置初始为0。神经网络的训练采用交叉熵（corss-entropy）评判准则，以及沿时间截断反向传播（truncated back-propagation-through-time）的优化训练方法。其中该模型可识别的每个信息块可以包含20帧语音信息，每个minibatch包含40条发音语句。此外，在动量因子的选择上，第一个迭代（epoch）采用0，之后的迭代采用0.9。在学习率更新上，如果本次迭代没有增加，则将上个迭代的学习率减半作为下个迭代训练的学习率。
S104，获取每个信息块对应的前向识别结果以及后向识别结果在所述预设的双向长短时递归神经网络模型的输出层的相应的神经元结点下的输出结果，以得到所述待测语音信息的语音识别结果。其中，所述预设的双向长短时递归神经网络模型包括由所有信息块对应的神经元结点组成的输入层、由非线性映射单元组成的隐藏层以及与隐藏层输出结果相应的神经元结点组成的输出层。
在本申请实施例中，根据每个信息块对应的前向识别结果以及后向识别结果，在所述预设的双向长短时递归神经网络模型的输出层的相应的神经元结点下的进行输出，即得到一个输出结果，该输出结果即为语音识别的结果。本申请实施例有效缩短了解码时延，使得基于双向长短时递归神经网络实时语音识别的应用成为可能。
请参阅图5，图5是本申请另一实施例提供的一种语音识别方法的示意流程图。如图5所示，该方法的步骤包括S501a~S504。
S501a，对待测语音信息进行预处理。
在本申请实施例中，首先对待测语音信息做预加重，提升待测语音信息中的高频部分，使得频谱更平滑。然后进行分帧加窗，将非平稳的语音信号转变为短时平稳信号分析处理。接着通过端点检测，区分语音与噪声，并提取出有效的语音部分，从而提高对待测语音信息识别的精确度。
S501，根据预设规则对进行预处理后的待测语音信息进行分帧和声学特征提取以得到帧级语音特征序列。
S502，将所述帧级语音特征序列依次分为n个信息块，其中n为大于等于2的整数，每个信息块包括一主信息块，每一主信息块拼接有一从信息块，每个主信息块以及与其相应的从信息块均包括若干帧语音特征信息，所述主信息块的最后一帧语音特征信息与拼接的所述从信息块的第一帧语音特征信息相邻。
S503，将所有信息块并行输入预设的双向长短时递归神经网络模型中，以分别计算每个信息块中的主信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的前向识别结果，以及每个信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的后向识别结果。
S504，获取每个信息块对应的前向识别结果以及后向识别结果在所述预设的双向长短时递归神经网络模型的输出层的相应的神经元结点下的输出结果，以得到所述待测语音信息的语音识别结果。
其中，所述预设的双向长短时递归神经网络模型包括由所有信息块对应的神经元结点组成的输入层、由非线性映射单元组成的隐藏层以及与隐藏层输出结果相应的神经元结点组成的输出层。
请参阅图6，对应上述一种语音识别方法，本申请实施例还提出一种语音识别装置，该装置100包括：第一处理单元101、第二处理单元102、识别单元103、输出单元104。
第一处理单元101，用于根据预设规则对待测语音信息进行分帧和声学特征提取以得到帧级语音特征序列。
进一步地，如图7所示，所述第一处理单元101包括：分帧单元201，特征提取单元202，调整单元203。其中，所述分帧单元201用于对待测语音信息进行分帧以得到多个语音帧。特征提取单元202用于对每个语音帧进行特征提取，以得到每个语音帧对应的梅尔频率倒谱系数特征序列。调整单元203用于将所述梅尔频率倒谱系数特征序列输入预设的音素模型中以得到对齐的帧级语音特征序列。
进一步地，如图8所示，所述调整单元203包括第一输入单元301、第二输入单元302以及第三输入单元303。第一输入单元301用于将所述梅尔频率倒谱系数特征序列输入预设的单音素模型中以得到对齐的第一特征信息。第二输入单元302，用于将所述第一特征信息输入预设的上下文三音素模型中以得到对齐的第二特征信息。第三输入单元303用于将所述第二特征信息输入预设的基于线性判别分析及最大似然线性特征转换的三音素模型中以得到对齐的帧级语音特征序列。其中所述单音素模型、上下文三音素模型以及基于线性判别分析及最大似然线性特征转换的三音素模型均为不同的预设的隐马尔可夫及混合高斯模型。
第二处理单元102，用于将所述帧级语音特征序列依次分为n个信息块，其中n为大于等于2的整数，每个信息块包括一主信息块，每一主信息块拼接有一从信息块，每个主信息块以及与其相应的从信息块均包括若干帧语音特征信息，所述主信息块的最后一帧语音特征信息与拼接的所述从信息块的第一帧语音特征信息相邻。
识别单元103，用于将所有信息块并行输入预设的双向长短时递归神经网络模型中，以分别计算每个信息块中的主信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的前向识别结果，以及每个信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的后向识别结果。
输出单元104，用于获取每个信息块对应的前向识别结果以及后向识别结果在所述预设的双向长短时递归神经网络模型的输出层的相应的神经元结点下的输出结果，以得到所述待测语音信息的语音识别结果。
其中，所述预设的双向长短时递归神经网络模型包括由所有信息块对应的神经元结点组成的输入层、由非线性映射单元组成的隐藏层以及与隐藏层输出结果相应的神经元结点组成的输出层。
综上，本申请实施例有效缩短了解码时延，使得基于双向长短时递归神经网络实时语音识别的应用成为可能。本申请实施例中的装置，均与上述实施例中的方法相对应，故具体的分析和原理在此不再赘述。
进一步地，如图9所示，所述装置100还包括第三处理单元401、第一训练单元402、第四处理单元403以及第二训练单元404，以用于实现对预设的双向长短时递归神经网络模型的构建。
其中，所述第三处理单元401，用于对选取的语音样本数据进行分帧和声学特征提取以得到梅尔频率倒谱系数特征序列。所述第一训练单元402，用于利用梅尔频率倒谱系数特征序列训练一隐马尔可夫及混合高斯模型，以实现帧级语音样本特征序列与文本内容的对齐，其中所述隐马尔可夫模型与所述语音样本数据是一一对应的，所述混合高斯模型用来描述所述隐马尔可夫模型状态的输出概率分布。所述第四处理单元403，用于将对齐后的语音样本特征序列依次分为n个样本信息块，其中n为大于等于2的整数，每个样本信息块包括一主样本信息块，每个主样本信息块的最末一帧语音特征信息后均拼接有一相应的从样本信息块，每个主样本信息块以及与其相应的从样本信息块均包括若干帧语音特征信息，且每个主样本信息块的最末一帧语音特征信息跟相应的从样本信息块的第一帧语音特征信息相邻。所述第二训练单元404，用于将所有样本信息块并行输入一双向长短时递归神经网络中，分别通过每个样本信息块中的主样本信息块对所述双向长短时递归神经网络进行前向训练，以及每个样本信息块对所述双向长短时递归神经网络进行后向训练以得到预设的双向长短时递归神经网络模型。
请参阅图10，对应上述一种语音识别方法，本申请另一实施例还提出一种语音识别装置，该装置500包括：预处理单元501a、第一处理单元501、第二处理单元502、识别单元503以及输出单元504。
预处理单元501a，对待测语音信息进行预处理。第一处理单元501，用于根据预设规则对进行预处理后的待测语音信息进行分帧和声学特征提取以得到帧级语音特征序列。第二处理单元502，用于将所述帧级语音特征序列依次分为n个信息块，其中n为大于等于2的整数，每个信息块包括一主信息块，每一主信息块拼接有一从信息块，每个主信息块以及与其相应的从信息块均包括若干帧语音特征信息，所述主信息块的最后一帧语音特征信息与拼接的所述从信息块的第一帧语音特征信息相邻。识别单元503，用于将所有信息块并行输入预设的双向长短时递归神经网络模型中，以分别计算每个信息块中的主信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的前向识别结果，以及每个信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的后向识别结果。输出单元504，用于获取每个信息块对应的前向识别结果以及后向识别结果在所述预设的双向长短时递归神经网络模型的输出层的相应的神经元结点下的输出结果，以得到所述待测语音信息的语音识别结果。其中，所述预设的双向长短时递归神经网络模型包括由所有信息块对应的神经元结点组成的输入层、由非线性映射单元组成的隐藏层以及与隐藏层输出结果相应的神经元结点组成的输出层。
上述语音识别装置可以实现为一种计算机程序的形式，计算机程序可以在如图11所示的计算机设备上运行。图11为本申请一种语音识别设备的结构组成示意图。该设备可以是终端，也可以是服务器，其中，终端可以是智能手机、平板电脑、笔记本电脑、台式电脑、个人数字助理和穿戴式设备等具有通信功能的电子设备。服务器可以是独立的服务器，也可以是多个服务器组成的服务器集群。
参照图11，该计算机设备600包括通过系统总线601连接的处理器602、非易失性存储介质603、内存储器604和网络接口605。其中，该计算机设备600的非易失性存储介质603可存储操作系统6031和计算机程序6032，该计算机程序6032被执行时，可使得处理器602执行一种语音识别方法。该计算机设备600的处理器602用于提供计算和控制能力，支撑整个计算机设备600的运行。该内存储器604为非易失性存储介质中的计算机程序的运行提供环境，该计算机程序被处理器执行时，可使得处理器602执行上述实施例的语音识别方法。计算机设备600的网络接口605用于进行网络通信，如发送分配的任务等。本领域技术人员可以理解，图11中示出的结构，仅仅是与本申请方案相关的部分结构的框图，并不构成对本申请方案所应用于其上的计算机设备的限定，具体的计算机设备可以包括比图中所示更多或更少的部件，或者组合某些部件，或者具有不同的部件布置。
本申请还提供了一种计算机可读存储介质，该计算机可读存储介质存储有一个或者一个以上程序，所述一个或者一个以上程序可被一个或者一个以上的处理器执行，以实现上述实施例的语音识别方法。
本申请前述的存储介质包括：磁碟、光盘、只读存储记忆体（Read-Only Memory，ROM）等各种可以存储程序代码的介质。本申请所有实施例中的单元可以通过通用集成电路，例如CPU（Central Processing Unit，中央处理器），或通过ASIC (Application Specific Integrated Circuit，专用集成电路）来实现。本申请实施例语音识别方法中的步骤可以根据实际需要进行顺序调整、合并和删减。本申请实施例语音识别终端中的单元可以根据实际需要进行合并、划分和删减。
以上所述，仅为本申请的具体实施方式，但本申请的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，可轻易想到各种等效的修改或替换，这些修改或替换都应涵盖在本申请的保护范围之内。因此，本申请的保护范围应以权利要求的保护范围为准。

Claims (20)

  1. 一种语音识别方法，其特征在于，所述方法包括：
    根据预设规则对待测语音信息进行分帧和声学特征提取以得到帧级语音特征序列；
    将所述帧级语音特征序列依次分为n个信息块，其中n为大于等于2的整数，每个信息块包括一主信息块，每一主信息块拼接有一从信息块，每个主信息块以及与其相应的从信息块均包括若干帧语音特征信息，所述主信息块的最后一帧语音特征信息与拼接的所述从信息块的第一帧语音特征信息相邻；
    将所有信息块并行输入预设的双向长短时递归神经网络模型中，以分别计算每个信息块中的主信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的前向识别结果，以及每个信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的后向识别结果；
    获取每个信息块对应的前向识别结果以及后向识别结果在所述预设的双向长短时递归神经网络模型的输出层的相应的神经元结点下的输出结果，以得到所述待测语音信息的语音识别结果；
    其中，所述预设的双向长短时递归神经网络模型包括由所有信息块对应的神经元结点组成的输入层、由非线性映射单元组成的隐藏层以及与隐藏层输出结果相应的神经元结点组成的输出层。
  2. 如权利要求1所述的方法，其特征在于，所述根据预设规则对待测语音信息进行分帧和声学特征提取以得到帧级语音特征序列包括：
    对待测语音信息进行分帧以得到多个语音帧；
    对每个语音帧进行特征提取，以得到每个语音帧对应的梅尔频率倒谱系数特征序列；
    将所述梅尔频率倒谱系数特征序列输入预设的音素模型中以得到对齐的帧级语音特征序列。
  3. 如权利要求2所述的方法，其特征在于，所述将所述梅尔频率倒谱系数特征序列输入预设的音素模型中以得到对齐的帧级语音特征序列包括：
    将所述梅尔频率倒谱系数特征序列输入预设的单音素模型中以得到对齐的第一特征信息；
    将所述第一特征信息输入预设的上下文三音素模型中以得到对齐的第二特征信息；
    将所述第二特征信息输入预设的基于线性判别分析及最大似然线性特征转换的三音素模型中以得到对齐的帧级语音特征序列；
    其中所述单音素模型、上下文三音素模型以及基于线性判别分析及最大似然线性特征转换的三音素模型均为不同的预设的隐马尔可夫及混合高斯模型。
  4. 如权利要求2所述的方法，其特征在于，所述预设的双向长短时递归神经网络模型的构建包括：
    对选取的语音样本数据进行分帧和声学特征提取以得到梅尔频率倒谱系数特征序列；
    利用梅尔频率倒谱系数特征序列训练一隐马尔可夫及混合高斯模型，以实现帧级语音样本特征序列与文本内容的对齐，其中所述隐马尔可夫模型与所述语音样本数据是一一对应的，所述混合高斯模型用来描述所述隐马尔可夫模型状态的输出概率分布；
    将对齐后的语音样本特征序列依次分为n个样本信息块，其中n为大于等于2的整数，每个样本信息块包括一主样本信息块，每个主样本信息块的最末一帧语音特征信息后均拼接有一相应的从样本信息块，每个主样本信息块以及与其相应的从样本信息块均包括若干帧语音特征信息，且每个主样本信息块的最末一帧语音特征信息跟相应的从样本信息块的第一帧语音特征信息相邻；
    将所有样本信息块并行输入一双向长短时递归神经网络中，分别通过每个样本信息块中的主样本信息块对所述双向长短时递归神经网络进行前向训练，以及每个样本信息块对所述双向长短时递归神经网络进行后向训练以得到预设的双向长短时递归神经网络模型。
  5. 如权利要求1所述的方法，其特征在于，所述方法还包括：
    对待测语音信息进行预处理；
    其中，根据预设规则对进行预处理后的待测语音信息进行分帧和声学特征提取以得到帧级语音特征序列。
  6. 一种语音识别装置，其特征在于，所述装置包括：
    第一处理单元，用于根据预设规则对待测语音信息进行分帧和声学特征提取以得到帧级语音特征序列；
    第二处理单元，用于将所述帧级语音特征序列依次分为n个信息块，其中n为大于等于2的整数，每个信息块包括一主信息块，每一主信息块拼接有一从信息块，每个主信息块以及与其相应的从信息块均包括若干帧语音特征信息，所述主信息块的最后一帧语音特征信息与拼接的所述从信息块的第一帧语音特征信息相邻；
    识别单元，用于将所有信息块并行输入预设的双向长短时递归神经网络模型中，以分别计算每个信息块中的主信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的前向识别结果，以及每个信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的后向识别结果；
    输出单元，用于获取每个信息块对应的前向识别结果以及后向识别结果在所述预设的双向长短时递归神经网络模型的输出层的相应的神经元结点下的输出结果，以得到所述待测语音信息的语音识别结果；
    其中，所述预设的双向长短时递归神经网络模型包括由所有信息块对应的神经元结点组成的输入层、由非线性映射单元组成的隐藏层以及与隐藏层输出结果相应的神经元结点组成的输出层。
  7. 如权利要求6所述的装置，其特征在于，所述第一处理单元，包括：
    分帧单元，用于对待测语音信息进行分帧以得到多个语音帧；
    特征提取单元，用于对每个语音帧进行特征提取，以得到每个语音帧对应的梅尔频率倒谱系数特征序列；
    调整单元，用于将所述梅尔频率倒谱系数特征序列输入预设的音素模型中以得到对齐的帧级语音特征序列。
  8. 如权利要求7所述的装置，其特征在于，所述调整单元，包括：
    第一输入单元，用于将所述梅尔频率倒谱系数特征序列输入预设的单音素模型中以得到对齐的第一特征信息；
    第二输入单元，用于将所述第一特征信息输入预设的上下文三音素模型中以得到对齐的第二特征信息；
    第三输入单元，用于将所述第二特征信息输入预设的基于线性判别分析及最大似然线性特征转换的三音素模型中以得到对齐的帧级语音特征序列；
    其中所述单音素模型、上下文三音素模型以及基于线性判别分析及最大似然线性特征转换的三音素模型均为不同的预设的隐马尔可夫及混合高斯模型。
  9. 如权利要求7所述的装置，其特征在于，所述装置还包括：
    第三处理单元，用于对选取的语音样本数据进行分帧和声学特征提取以得到梅尔频率倒谱系数特征序列；
    第一训练单元，用于利用梅尔频率倒谱系数特征序列训练一隐马尔可夫及混合高斯模型，以实现帧级语音样本特征序列与文本内容的对齐，其中所述隐马尔可夫模型与所述语音样本数据是一一对应的，所述混合高斯模型用来描述所述隐马尔可夫模型状态的输出概率分布；
    第四处理单元，用于将对齐后的语音样本特征序列依次分为n个样本信息块，其中n为大于等于2的整数，每个样本信息块包括一主样本信息块，每个主样本信息块的最末一帧语音特征信息后均拼接有一相应的从样本信息块，每个主样本信息块以及与其相应的从样本信息块均包括若干帧语音特征信息，且每个主样本信息块的最末一帧语音特征信息跟相应的从样本信息块的第一帧语音特征信息相邻；
    第二训练单元，用于将所有样本信息块并行输入一双向长短时递归神经网络中，分别通过每个样本信息块中的主样本信息块对所述双向长短时递归神经网络进行前向训练，以及每个样本信息块对所述双向长短时递归神经网络进行后向训练以得到预设的双向长短时递归神经网络模型。
  10. 如权利要求6所述的装置，其特征在于，所述装置还包括：
    预处理单元，用于对待测语音信息进行预处理；
    其中，第一处理单元用于根据预设规则对进行预处理后的待测语音信息进行分帧和声学特征提取以得到帧级语音特征序列。
  11. 一种语音识别设备，其特征在于，包括：
    存储器，用于存储实现语音识别的程序；以及
    处理器，用于运行所述存储器中存储的实现语音识别的程序，以执行以下操作：
    根据预设规则对待测语音信息进行分帧和声学特征提取以得到帧级语音特征序列；
    将所述帧级语音特征序列依次分为n个信息块，其中n为大于等于2的整数，每个信息块包括一主信息块，每一主信息块拼接有一从信息块，每个主信息块以及与其相应的从信息块均包括若干帧语音特征信息，所述主信息块的最后一帧语音特征信息与拼接的所述从信息块的第一帧语音特征信息相邻；
    将所有信息块并行输入预设的双向长短时递归神经网络模型中，以分别计算每个信息块中的主信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的前向识别结果，以及每个信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的后向识别结果；
    获取每个信息块对应的前向识别结果以及后向识别结果在所述预设的双向长短时递归神经网络模型的输出层的相应的神经元结点下的输出结果，以得到所述待测语音信息的语音识别结果；
    其中，所述预设的双向长短时递归神经网络模型包括由所有信息块对应的神经元结点组成的输入层、由非线性映射单元组成的隐藏层以及与隐藏层输出结果相应的神经元结点组成的输出层。
  12. 如权利要求11所述的设备，其特征在于，所述根据预设规则对待测语音信息进行分帧和声学特征提取以得到帧级语音特征序列包括：
    对待测语音信息进行分帧以得到多个语音帧；
    对每个语音帧进行特征提取，以得到每个语音帧对应的梅尔频率倒谱系数特征序列；
    将所述梅尔频率倒谱系数特征序列输入预设的音素模型中以得到对齐的帧级语音特征序列。
  13. 如权利要求12所述的设备，其特征在于，所述将所述梅尔频率倒谱系数特征序列输入预设的音素模型中以得到对齐的帧级语音特征序列包括：
    将所述梅尔频率倒谱系数特征序列输入预设的单音素模型中以得到对齐的第一特征信息；
    将所述第一特征信息输入预设的上下文三音素模型中以得到对齐的第二特征信息；
    将所述第二特征信息输入预设的基于线性判别分析及最大似然线性特征转换的三音素模型中以得到对齐的帧级语音特征序列；
    其中所述单音素模型、上下文三音素模型以及基于线性判别分析及最大似然线性特征转换的三音素模型均为不同的预设的隐马尔可夫及混合高斯模型。
  14. 如权利要求12所述的设备，其特征在于，所述预设的双向长短时递归神经网络模型的构建包括：
    对选取的语音样本数据进行分帧和声学特征提取以得到梅尔频率倒谱系数特征序列；
    利用梅尔频率倒谱系数特征序列训练一隐马尔可夫及混合高斯模型，以实现帧级语音样本特征序列与文本内容的对齐，其中所述隐马尔可夫模型与所述语音样本数据是一一对应的，所述混合高斯模型用来描述所述隐马尔可夫模型状态的输出概率分布；
    将对齐后的语音样本特征序列依次分为n个样本信息块，其中n为大于等于2的整数，每个样本信息块包括一主样本信息块，每个主样本信息块的最末一帧语音特征信息后均拼接有一相应的从样本信息块，每个主样本信息块以及与其相应的从样本信息块均包括若干帧语音特征信息，且每个主样本信息块的最末一帧语音特征信息跟相应的从样本信息块的第一帧语音特征信息相邻；
    将所有样本信息块并行输入一双向长短时递归神经网络中，分别通过每个样本信息块中的主样本信息块对所述双向长短时递归神经网络进行前向训练，以及每个样本信息块对所述双向长短时递归神经网络进行后向训练以得到预设的双向长短时递归神经网络模型。
  15. 如权利要求11所述的设备，其特征在于，所述处理器还执行以下操作：
    对待测语音信息进行预处理；
    其中，根据预设规则对进行预处理后的待测语音信息进行分帧和声学特征提取以得到帧级语音特征序列。
  16. 一种计算机可读存储介质，其特征在于，计算机可读存储介质存储有一个或者一个以上程序，所述一个或者一个以上程序可被一个或者一个以上的处理器执行，以实现以下步骤：
    根据预设规则对待测语音信息进行分帧和声学特征提取以得到帧级语音特征序列；
    将所述帧级语音特征序列依次分为n个信息块，其中n为大于等于2的整数，每个信息块包括一主信息块，每一主信息块拼接有一从信息块，每个主信息块以及与其相应的从信息块均包括若干帧语音特征信息，所述主信息块的最后一帧语音特征信息与拼接的所述从信息块的第一帧语音特征信息相邻；
    将所有信息块并行输入预设的双向长短时递归神经网络模型中，以分别计算每个信息块中的主信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的前向识别结果，以及每个信息块在所述预设的双向长短时递归神经网络模型的隐含层的相应的神经元结点下的后向识别结果；
    获取每个信息块对应的前向识别结果以及后向识别结果在所述预设的双向长短时递归神经网络模型的输出层的相应的神经元结点下的输出结果，以得到所述待测语音信息的语音识别结果；
    其中，所述预设的双向长短时递归神经网络模型包括由所有信息块对应的神经元结点组成的输入层、由非线性映射单元组成的隐藏层以及与隐藏层输出结果相应的神经元结点组成的输出层。
  17. 如权利要求16所述的计算机可读存储介质，其特征在于，所述根据预设规则对待测语音信息进行分帧和声学特征提取以得到帧级语音特征序列包括：
    对待测语音信息进行分帧以得到多个语音帧；
    对每个语音帧进行特征提取，以得到每个语音帧对应的梅尔频率倒谱系数特征序列；
    将所述梅尔频率倒谱系数特征序列输入预设的音素模型中以得到对齐的帧级语音特征序列。
  18. 如权利要求17所述的计算机可读存储介质，其特征在于，所述将所述梅尔频率倒谱系数特征序列输入预设的音素模型中以得到对齐的帧级语音特征序列包括：
    将所述梅尔频率倒谱系数特征序列输入预设的单音素模型中以得到对齐的第一特征信息；
    将所述第一特征信息输入预设的上下文三音素模型中以得到对齐的第二特征信息；
    将所述第二特征信息输入预设的基于线性判别分析及最大似然线性特征转换的三音素模型中以得到对齐的帧级语音特征序列；
    其中所述单音素模型、上下文三音素模型以及基于线性判别分析及最大似然线性特征转换的三音素模型均为不同的预设的隐马尔可夫及混合高斯模型。
  19. 如权利要求17所述的计算机可读存储介质，其特征在于，所述预设的双向长短时递归神经网络模型的构建包括：
    对选取的语音样本数据进行分帧和声学特征提取以得到梅尔频率倒谱系数特征序列；
    利用梅尔频率倒谱系数特征序列训练一隐马尔可夫及混合高斯模型，以实现帧级语音样本特征序列与文本内容的对齐，其中所述隐马尔可夫模型与所述语音样本数据是一一对应的，所述混合高斯模型用来描述所述隐马尔可夫模型状态的输出概率分布；
    将对齐后的语音样本特征序列依次分为n个样本信息块，其中n为大于等于2的整数，每个样本信息块包括一主样本信息块，每个主样本信息块的最末一帧语音特征信息后均拼接有一相应的从样本信息块，每个主样本信息块以及与其相应的从样本信息块均包括若干帧语音特征信息，且每个主样本信息块的最末一帧语音特征信息跟相应的从样本信息块的第一帧语音特征信息相邻；
    将所有样本信息块并行输入一双向长短时递归神经网络中，分别通过每个样本信息块中的主样本信息块对所述双向长短时递归神经网络进行前向训练，以及每个样本信息块对所述双向长短时递归神经网络进行后向训练以得到预设的双向长短时递归神经网络模型。
  20. 如权利要求16所述的计算机可读存储介质，其特征在于，还包括实现以下步骤：
    对待测语音信息进行预处理；
    其中，根据预设规则对进行预处理后的待测语音信息进行分帧和声学特征提取以得到帧级语音特征序列。
PCT/CN2017/113230 2017-10-23 2017-11-28 语音识别方法、装置、设备以及计算机可读存储介质 Ceased WO2019080248A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
SG11202001627XA SG11202001627XA (en) 2017-10-23 2017-11-28 Speech recognition method, apparatus, and computer readable storage medium
US16/642,371 US11081103B2 (en) 2017-10-23 2017-11-28 Speech recognition method, apparatus, and computer readable storage medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201710994268.5A CN107680597B (zh) 2017-10-23 2017-10-23 语音识别方法、装置、设备以及计算机可读存储介质
CN201710994268.5 2017-10-23

Publications (1)

Publication Number Publication Date
WO2019080248A1 true WO2019080248A1 (zh) 2019-05-02

Family

ID=61141446

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2017/113230 Ceased WO2019080248A1 (zh) 2017-10-23 2017-11-28 语音识别方法、装置、设备以及计算机可读存储介质

Country Status (4)

Country Link
US (1) US11081103B2 (zh)
CN (1) CN107680597B (zh)
SG (1) SG11202001627XA (zh)
WO (1) WO2019080248A1 (zh)

Families Citing this family (90)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9318108B2 (en) 2010-01-18 2016-04-19 Apple Inc. Intelligent automated assistant
US8977255B2 (en) 2007-04-03 2015-03-10 Apple Inc. Method and system for operating a multi-function portable electronic device using voice-activation
US8676904B2 (en) 2008-10-02 2014-03-18 Apple Inc. Electronic devices with voice command and contextual data processing capabilities
US10057736B2 (en) 2011-06-03 2018-08-21 Apple Inc. Active transport based notifications
US10417037B2 (en) 2012-05-15 2019-09-17 Apple Inc. Systems and methods for integrating third party services with a digital assistant
DE112014000709B4 (de) 2013-02-07 2021-12-30 Apple Inc. Verfahren und vorrichtung zum betrieb eines sprachtriggers für einen digitalen assistenten
US10170123B2 (en) 2014-05-30 2019-01-01 Apple Inc. Intelligent assistant for home automation
US9715875B2 (en) 2014-05-30 2017-07-25 Apple Inc. Reducing the need for manual start/end-pointing and trigger phrases
US9338493B2 (en) 2014-06-30 2016-05-10 Apple Inc. Intelligent automated assistant for TV user interactions
US9886953B2 (en) 2015-03-08 2018-02-06 Apple Inc. Virtual assistant activation
US10460227B2 (en) 2015-05-15 2019-10-29 Apple Inc. Virtual assistant in a communication session
US10331312B2 (en) 2015-09-08 2019-06-25 Apple Inc. Intelligent automated assistant in a media environment
US10671428B2 (en) 2015-09-08 2020-06-02 Apple Inc. Distributed personal assistant
US10747498B2 (en) 2015-09-08 2020-08-18 Apple Inc. Zero latency digital assistant
US11587559B2 (en) 2015-09-30 2023-02-21 Apple Inc. Intelligent device identification
US10691473B2 (en) 2015-11-06 2020-06-23 Apple Inc. Intelligent automated assistant in a messaging environment
US10586535B2 (en) 2016-06-10 2020-03-10 Apple Inc. Intelligent digital assistant in a multi-tasking environment
DK201670540A1 (en) 2016-06-11 2018-01-08 Apple Inc Application integration with a digital assistant
US12197817B2 (en) 2016-06-11 2025-01-14 Apple Inc. Intelligent device arbitration and control
US11204787B2 (en) 2017-01-09 2021-12-21 Apple Inc. Application integration with a digital assistant
DK180048B1 (en) 2017-05-11 2020-02-04 Apple Inc. MAINTAINING THE DATA PROTECTION OF PERSONAL INFORMATION
DK179496B1 (en) 2017-05-12 2019-01-15 Apple Inc. USER-SPECIFIC Acoustic Models
DK201770428A1 (en) 2017-05-12 2019-02-18 Apple Inc. LOW-LATENCY INTELLIGENT AUTOMATED ASSISTANT
DK201770411A1 (en) 2017-05-15 2018-12-20 Apple Inc. Multi-modal interfaces
DK179560B1 (en) 2017-05-16 2019-02-18 Apple Inc. FAR-FIELD EXTENSION FOR DIGITAL ASSISTANT SERVICES
US10303715B2 (en) 2017-05-16 2019-05-28 Apple Inc. Intelligent automated assistant for media exploration
US10818288B2 (en) 2018-03-26 2020-10-27 Apple Inc. Natural assistant interaction
CN108520741B (zh) 2018-04-12 2021-05-04 科大讯飞股份有限公司 一种耳语音恢复方法、装置、设备及可读存储介质
CN108664460A (zh) * 2018-04-16 2018-10-16 北京天使软件技术有限公司 语音填表装置、方法、系统及存储介质
US10928918B2 (en) 2018-05-07 2021-02-23 Apple Inc. Raise to speak
DK180639B1 (en) 2018-06-01 2021-11-04 Apple Inc DISABILITY OF ATTENTION-ATTENTIVE VIRTUAL ASSISTANT
DK201870355A1 (en) 2018-06-01 2019-12-16 Apple Inc. VIRTUAL ASSISTANT OPERATION IN MULTI-DEVICE ENVIRONMENTS
CN108877775B (zh) * 2018-06-04 2023-03-31 平安科技（深圳）有限公司 语音数据处理方法、装置、计算机设备及存储介质
CN108922513B (zh) * 2018-06-04 2023-03-17 平安科技（深圳）有限公司 语音区分方法、装置、计算机设备及存储介质
CN108776795A (zh) * 2018-06-20 2018-11-09 邯郸学院 用户身份识别方法、装置及终端设备
CN110619871B (zh) * 2018-06-20 2023-06-30 阿里巴巴集团控股有限公司 语音唤醒检测方法、装置、设备以及存储介质
CN108962223A (zh) * 2018-06-25 2018-12-07 厦门快商通信息技术有限公司 一种基于深度学习的语音性别识别方法、设备及介质
CN108935188A (zh) * 2018-07-05 2018-12-07 平安科技（深圳）有限公司 猪只疾病识别方法、装置及电子设备
CN108922521B (zh) * 2018-08-15 2021-07-06 合肥讯飞数码科技有限公司 一种语音关键词检索方法、装置、设备及存储介质
US11462215B2 (en) 2018-09-28 2022-10-04 Apple Inc. Multi-modal inputs for voice commands
CN109559735B (zh) * 2018-10-11 2023-10-27 平安科技（深圳）有限公司 一种基于神经网络的语音识别方法、终端设备及介质
CN109346103B (zh) * 2018-10-30 2023-03-28 交通运输部公路科学研究所 一种用于公路隧道交通事件的音频检测方法
CN110517679B (zh) * 2018-11-15 2022-03-08 腾讯科技（深圳）有限公司 一种人工智能的音频数据处理方法及装置、存储介质
CN110166826B (zh) * 2018-11-21 2021-10-08 腾讯科技（深圳）有限公司 视频的场景识别方法、装置、存储介质及计算机设备
US11114103B2 (en) 2018-12-28 2021-09-07 Alibaba Group Holding Limited Systems, methods, and computer-readable storage media for audio signal processing
CN109658921B (zh) * 2019-01-04 2024-05-28 平安科技（深圳）有限公司 一种语音信号处理方法、设备及计算机可读存储介质
CN109872713A (zh) * 2019-03-05 2019-06-11 深圳市友杰智新科技有限公司 一种语音唤醒方法及装置
US11348573B2 (en) 2019-03-18 2022-05-31 Apple Inc. Multimodality in digital assistant systems
US11468879B2 (en) * 2019-04-29 2022-10-11 Tencent America LLC Duration informed attention network for text-to-speech analysis
US11307752B2 (en) 2019-05-06 2022-04-19 Apple Inc. User configurable task triggers
DK201970509A1 (en) 2019-05-06 2021-01-15 Apple Inc Spoken notifications
EP3970141B1 (en) * 2019-05-14 2024-02-28 Dolby Laboratories Licensing Corporation Method and apparatus for speech source separation based on a convolutional neural network
CN110277088B (zh) * 2019-05-29 2024-04-09 平安科技（深圳）有限公司 智能语音识别方法、装置及计算机可读存储介质
US11289073B2 (en) * 2019-05-31 2022-03-29 Apple Inc. Device text to speech
US11227599B2 (en) 2019-06-01 2022-01-18 Apple Inc. Methods and user interfaces for voice-based control of electronic devices
US20220230630A1 (en) * 2019-06-10 2022-07-21 Nippon Telegraph And Telephone Corporation Model learning apparatus, method and program
CN110288999B (zh) * 2019-07-02 2020-12-11 腾讯科技（深圳）有限公司 语音识别方法、装置、计算机设备及存储介质
KR102885234B1 (ko) * 2019-07-19 2025-11-12 삼성전자주식회사 음성 인식 방법, 음성 인식을 위한 학습 방법 및 그 장치들
CN110534098A (zh) * 2019-10-09 2019-12-03 国家电网有限公司客户服务中心 一种年龄增强的语音识别增强方法和装置
CN111128235A (zh) * 2019-12-05 2020-05-08 厦门快商通科技股份有限公司 一种基于语音的年龄预测方法和装置以及设备
CN111145765B (zh) * 2019-12-31 2022-04-15 思必驰科技股份有限公司 一种音频处理方法、装置、电子设备及存储介质
CN112750425B (zh) * 2020-01-22 2023-11-03 腾讯科技（深圳）有限公司 语音识别方法、装置、计算机设备及计算机可读存储介质
CN113470662B (zh) 2020-03-31 2024-08-27 微软技术许可有限责任公司 生成和使用用于关键词检出系统的文本到语音数据和语音识别系统中的说话者适配
CN113593539B (zh) * 2020-04-30 2024-08-02 阿里巴巴集团控股有限公司 流式端到端语音识别方法、装置及电子设备
US11061543B1 (en) 2020-05-11 2021-07-13 Apple Inc. Providing relevant data items based on context
US12301635B2 (en) 2020-05-11 2025-05-13 Apple Inc. Digital assistant hardware abstraction
CN111667817A (zh) * 2020-06-22 2020-09-15 平安资产管理有限责任公司 一种语音识别方法、装置、计算机系统及可读存储介质
CN111696526B (zh) * 2020-06-22 2021-09-10 北京达佳互联信息技术有限公司 语音识别模型的生成方法、语音识别方法、装置
US11490204B2 (en) 2020-07-20 2022-11-01 Apple Inc. Multi-device audio adjustment coordination
US11438683B2 (en) 2020-07-21 2022-09-06 Apple Inc. User identification using headphones
CN112002305B (zh) * 2020-07-29 2024-06-18 北京大米科技有限公司 语音合成方法、装置、存储介质及电子设备
CN112216270B (zh) * 2020-10-09 2024-02-06 携程计算机技术（上海）有限公司 语音音素的识别方法及系统、电子设备及存储介质
US11942078B2 (en) * 2021-02-26 2024-03-26 International Business Machines Corporation Chunking and overlap decoding strategy for streaming RNN transducers for speech recognition
CN112820279B (zh) * 2021-03-12 2024-02-09 深圳市臻络科技有限公司 基于语音上下文动态特征的帕金森检测模型构建方法
CN114038465B (zh) * 2021-04-28 2022-08-23 北京有竹居网络技术有限公司 语音处理方法、装置和电子设备
US11948550B2 (en) * 2021-05-06 2024-04-02 Sanas.ai Inc. Real-time accent conversion model
CN113724718B (zh) * 2021-09-01 2022-07-29 宿迁硅基智能科技有限公司 目标音频的输出方法及装置、系统
CN113724690B (zh) * 2021-09-01 2023-01-03 宿迁硅基智能科技有限公司 Ppg特征的输出方法、目标音频的输出方法及装置
CN113611285B (zh) * 2021-09-03 2023-11-24 哈尔滨理工大学 基于层叠双向时序池化的语种识别方法
CN113851111B (zh) * 2021-09-13 2025-04-22 联想(北京)有限公司 一种语音识别方法和语音识别装置
CN113918394A (zh) * 2021-09-22 2022-01-11 百信信息技术有限公司 一种机箱前置音频接口检测方法及系统
CN116415166B (zh) * 2021-12-28 2025-11-21 深圳大学 多键盘混合按键声音的识别方法、装置、设备及存储介质
CN114566154B (zh) * 2022-03-01 2025-12-23 清华大学苏州汽车研究院（相城） 一种音频识别方法、装置、设备及存储介质
CN115035887B (zh) * 2022-05-20 2025-05-16 京东方科技集团股份有限公司 语音信号的处理方法、装置、设备及介质
US12424215B2 (en) * 2022-05-27 2025-09-23 SoundHound AI IP, LLC. Pre-wakeword speech processing with reverse automatic speech recognition
CN116312458B (zh) * 2023-02-17 2025-11-21 海信视像科技股份有限公司 声学模型训练方法、语音合成方法、装置及计算机设备
US12556651B2 (en) 2023-04-10 2026-02-17 Microsoft Technology Licensing, Llc Systems and methods for real-time meeting summarization
WO2025047998A1 (ko) * 2023-08-29 2025-03-06 주식회사 엔씨소프트 지정된 텍스트에 대응하는 음성 신호를 식별하기 위한 전자 장치, 방법, 및 컴퓨터 판독 가능 저장 매체
CN118380011B (zh) * 2024-04-16 2024-10-25 泰德网聚(北京)科技股份有限公司 基于多模型的语音数据分析方法及装置
CN119993125B (zh) * 2025-04-10 2025-07-04 青岛丹香投资管理有限公司 一种基于人工智能的语音数据处理方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120166194A1 (en) * 2010-12-23 2012-06-28 Electronics And Telecommunications Research Institute Method and apparatus for recognizing speech
CN104952448A (zh) * 2015-05-04 2015-09-30 张爱英 一种双向长短时记忆递归神经网络的特征增强方法及系统
CN105679316A (zh) * 2015-12-29 2016-06-15 深圳微服机器人科技有限公司 一种基于深度神经网络的语音关键词识别方法及装置
CN105869624A (zh) * 2016-03-29 2016-08-17 腾讯科技（深圳）有限公司 数字语音识别中语音解码网络的构建方法及装置
CN106803422A (zh) * 2015-11-26 2017-06-06 中国科学院声学研究所 一种基于长短时记忆网络的语言模型重估方法

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10332509B2 (en) * 2015-11-25 2019-06-25 Baidu USA, LLC End-to-end speech recognition
US10949736B2 (en) * 2016-11-03 2021-03-16 Intel Corporation Flexible neural network accelerator and methods therefor
US20180330718A1 (en) * 2017-05-11 2018-11-15 Mitsubishi Electric Research Laboratories, Inc. System and Method for End-to-End speech recognition

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120166194A1 (en) * 2010-12-23 2012-06-28 Electronics And Telecommunications Research Institute Method and apparatus for recognizing speech
CN104952448A (zh) * 2015-05-04 2015-09-30 张爱英 一种双向长短时记忆递归神经网络的特征增强方法及系统
CN106803422A (zh) * 2015-11-26 2017-06-06 中国科学院声学研究所 一种基于长短时记忆网络的语言模型重估方法
CN105679316A (zh) * 2015-12-29 2016-06-15 深圳微服机器人科技有限公司 一种基于深度神经网络的语音关键词识别方法及装置
CN105869624A (zh) * 2016-03-29 2016-08-17 腾讯科技（深圳）有限公司 数字语音识别中语音解码网络的构建方法及装置

Also Published As

Publication number Publication date
US20210074264A1 (en) 2021-03-11
CN107680597A (zh) 2018-02-09
US11081103B2 (en) 2021-08-03
CN107680597B (zh) 2019-07-09
SG11202001627XA (en) 2020-03-30

Similar Documents

Publication Publication Date Title
WO2019080248A1 (zh) 语音识别方法、装置、设备以及计算机可读存储介质
WO2020101263A1 (en) Electronic apparatus and method for controlling thereof
EP3818518A1 (en) Electronic apparatus and method for controlling thereof
WO2020027619A1 (ko) 순차적 운율 특징을 기초로 기계학습을 이용한 텍스트-음성 합성 방법, 장치 및 컴퓨터 판독가능한 저장매체
EP3977446A1 (en) Device for recognizing speech input from user and operating method thereof
WO2022260432A1 (ko) 자연어로 표현된 스타일 태그를 이용한 합성 음성 생성 방법 및 시스템
EP3877975A1 (en) Electronic device and method of controlling speech recognition by electronic device
WO2019139431A1 (ko) 다중 언어 텍스트-음성 합성 모델을 이용한 음성 번역 방법 및 시스템
WO2021010617A1 (ko) 강인한 음성인식을 위한 음향 및 언어모델링 정보를 이용한 음성 끝점 검출 방법 및 장치
WO2020145472A1 (ko) 화자 적응형 모델을 구현하고 합성 음성 신호를 생성하는 뉴럴 보코더 및 뉴럴 보코더의 훈련 방법
WO2022035183A1 (ko) 사용자의 음성 입력을 인식하는 디바이스 및 그 동작 방법
Mantena et al. Use of articulatory bottle-neck features for query-by-example spoken term detection in low resource scenarios
EP3980991A1 (en) System and method for recognizing user's speech
CN110648655A (zh) 一种语音识别方法、装置、系统及存储介质
US5758022A (en) Method and apparatus for improved speech recognition from stress-induced pronunciation variations with a neural network utilizing non-linear imaging characteristics
Zhao et al. Phone-aware multi-task learning and length expanding for short-duration language recognition
CN112216270A (zh) 语音音素的识别方法及系统、电子设备及存储介质
Zeng et al. Adamax online training for speech recognition
CN117238297A (zh) 声音信号处理的方法、装置、设备、介质和程序产品
Singh et al. A critical review on automatic speaker recognition
JP2012053218A (ja) 音響処理装置および音響処理プログラム
Kaur et al. Power-Normalized Cepstral Coefficients (PNCC) for Punjabi automatic speech recognition using phone based modelling in HTK
Darling et al. Feature extraction in speech recognition using linear predictive coding: An overview
WO2022031061A1 (ko) 심화신경망 기반의 가상 채널 확장을 이용한 wpe 기반 잔향 제거 장치
WO2025023415A1 (en) Method and apparatus for registering and updating audio information associated with a user

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: 17929702

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 24.09.2020)

122 Ep: pct application non-entry in european phase

Ref document number: 17929702

Country of ref document: EP

Kind code of ref document: A1