A voice recognition method and device and intelligent speaker
Through the noise reduction automatic encoder and deep neural network model, the far-field soundprint feature vector is projected into the near-field soundprint feature space, solving the problem of inaccurate long-distance soundprint recognition in smart speakers and improving the robustness of recognition.
Patent Information
- Application Number
- CN201910265991.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2019-04-03
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2039-04-03
AI Technical Summary
In the voiceprint recognition of smart speakers, users usually register at close range, but changes in distance and spatial position in daily use cause long-distance voiceprint information to be unable to accurately identify, resulting in inaccurate identity recognition.
Using the noise reduction automatic encoder, the voiceprint feature vector obtained through near-field and far-field speech training is used to project the far-field voiceprint feature vector into the near-field voiceprint feature vector space to realize the identity recognition of speech information. .
It improves the robustness of far-field voiceprint recognition, reduces identity identification errors caused by distance changes, and ensures accurate identification of user identity.
Smart Images

Figure CN111785282B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to but is not limited to intelligent voice technology, and in particular to a voice recognition method and device and a smart speaker. Background Art
[0002] In the voiceprint recognition of smart speakers, users usually pre-register the voiceprint information corresponding to the user. In this way, each time the device is woken up, the user will be automatically identified to determine the identity of the speaker based on the pre-registered voiceprint information and the voiceprint information currently input by the user.
[0003] In the above voiceprint recognition scenario, when users register, they usually register at a distance of about 0.5 to 1 meter from the speaker. However, in daily use, users may try to wake up the device from any location in their home. However, changes in distance and spatial position will bring about changes in sound information, resulting in the inability to accurately identify the identity of the speaker from voiceprint information far away from the device based on voiceprint information registered at a close distance. Summary of the invention
[0004] The present application provides a speech recognition method and device and a smart speaker, which can improve the robustness of far-field voiceprint recognition.
[0005] An embodiment of the present invention provides a speech recognition method, comprising:
[0006] A noise reduction autoencoder is used to extract a voiceprint feature vector of the speech information to be recognized; wherein the noise reduction autoencoder is obtained by training based on near-field speech and far-field speech;
[0007] The user identity of the voice information to be recognized is determined based on the extracted voiceprint feature vector.
[0008] In an exemplary embodiment, the method further comprises:
[0009] Extracting a near-field voiceprint feature vector according to the near-field speech, and extracting a far-field voiceprint feature vector according to the far-field speech;
[0010] The near-field voiceprint feature vector and the far-field voiceprint feature vector corresponding to the same user are input into a mathematical model as a group of inputs to train the parameters of the noise reduction automatic encoder corresponding to the mathematical model.
[0011] In an exemplary embodiment, the step of extracting a voiceprint feature vector of speech information to be recognized includes:
[0012] Input the sound energy spectrum information of the voice information to be recognized into a trained speaker recognition neural network model to obtain a voiceprint feature vector of the speaker represented by the voice information to be recognized;
[0013] The speaker's voiceprint feature vector is input into the noise reduction automatic encoder to obtain the speaker's near-field voiceprint feature vector represented by the speech information to be recognized.
[0014] In an exemplary embodiment, the voice information to be recognized is far-field voice information; the voiceprint feature vector of the speaker is a far-field voiceprint feature vector;
[0015] The denoising autoencoder projects the far-field voiceprint feature vector into the near-field voiceprint feature vector.
[0016] In one exemplary instance, the mathematical model comprises a deep neural network model.
[0017] The present application also provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute any of the above-mentioned speech recognition methods.
[0018] The present application further provides a device for implementing speech recognition, comprising a memory and a processor, wherein the memory stores the following instructions executable by the processor: for executing the steps of any of the above-mentioned speech recognition methods.
[0019] The present application further provides a smart speaker, including a memory and a processor, wherein a computer program is stored in the memory; when the computer program is executed by the processor, the following steps are performed:
[0020] A denoising autoencoder is used to extract a voiceprint feature vector of speech information to be recognized; wherein the denoising autoencoder is trained based on near-field speech and far-field speech; and the user identity of the obtained speech information is recognized based on the extracted voiceprint feature vector.
[0021] In an exemplary embodiment, when the computer program is executed by the processor, the following steps are also performed: extracting a near-field voiceprint feature vector based on the near-field speech, and extracting a far-field voiceprint feature vector based on the far-field speech; taking the near-field voiceprint feature vector and the far-field voiceprint feature vector corresponding to the same user as a set of inputs, inputting them into a mathematical model, and training the parameters of the denoising autoencoder corresponding to the mathematical model.
[0022] 10. A speech recognition device, comprising: a processing module and a recognition module; wherein:
[0023] A processing module, used to extract a voiceprint feature vector of the speech information to be recognized by using a noise reduction autoencoder; wherein the noise reduction autoencoder is trained based on near-field speech and far-field speech;
[0024] The recognition module is used to determine the user identity of the voice information to be recognized based on the extracted voiceprint feature vector.
[0025] 11. The speech recognition device according to claim 10, further comprising a training module for:
[0026] A near-field voiceprint feature vector is extracted according to the near-field speech, and a far-field voiceprint feature vector is extracted according to the far-field speech; the near-field voiceprint feature vector and the far-field voiceprint feature vector corresponding to the same user are input into a mathematical model as a group of inputs to train the parameters of the denoising autoencoder corresponding to the mathematical model.
[0027] This application uses a noise reduction autoencoder trained based on near-field speech and far-field speech to project the voiceprint feature vectors of the speech information to be recognized, especially the far-field voiceprint feature vectors, into a space that is as similar as possible to the near-field voiceprint features, and then uses the voiceprint feature vectors output by the noise reduction autoencoder to realize subsequent recognition of the voice user's identity, thereby reducing the problem of not being able to accurately identify the identity of a speaker far away from the device due to the distance between the speaker and the device, and improving the robustness of far-field voiceprint recognition.
[0028] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] The accompanying drawings are used to provide further understanding of the technical solution of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the technical solution of the present application and do not constitute a limitation on the technical solution of the present application.
[0030] Figure 1 A flowchart of an embodiment of the speech recognition method of the present application;
[0031] Figure 2 Schematic diagram of the structure of a speech recognition device according to an embodiment of the present invention. DETAILED DESCRIPTION
[0032] In order to make the purpose, technical solution and advantages of the present application more clear, the embodiments of the present application will be described in detail below with reference to the accompanying drawings. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other arbitrarily without conflict.
[0033] In a typical configuration of the present application, a computing device includes one or more processors (CPU), an input / output interface, a network interface, and a memory.
[0034] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.
[0035] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include non-transitory media such as modulated data signals and carrier waves.
[0036] The steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions. Also, although a logical sequence is shown in the flowchart, in some cases, the steps shown or described can be performed in a sequence different from that shown here.
[0037] Figure 1 This is a flow chart of an embodiment of the speech recognition method of the present application, such as Figure 1 As shown, including:
[0038] Step 100: extracting the voiceprint feature vector of the speech information to be recognized by using a noise reduction autoencoder; wherein the noise reduction autoencoder is trained based on near-field speech and far-field speech.
[0039] In an exemplary embodiment, the speech recognition method further includes:
[0040] Train a denoising autoencoder on near-field and far-field speech.
[0041] In an exemplary embodiment, training a noise reduction autoencoder based on near-field speech and far-field speech may include:
[0042] Extracting near-field voiceprint feature vectors based on near-field speech, and extracting far-field voiceprint feature vectors based on far-field speech;
[0043] The near-field voiceprint feature vector and the far-field voiceprint feature vector corresponding to the same user are input into a mathematical model as a group of inputs, and the parameters of the denoising autoencoder corresponding to the mathematical model are trained.
[0044] In an exemplary embodiment, the mathematical model can be a deep neural network model, in which the number of neurons in the input layer is equal to the number of neurons in the output layer, so that the output signal is as similar as possible to the input signal. This mathematical model can include but is not limited to: time-delay neural network (TDNN), convolutional neural network (CNN), recurrent neural network (RNN), long short-term memory neural network (LSTM), etc.
[0045] The neural network model is retrained using pairs of near-field voiceprint feature vectors and far-field voiceprint feature vectors of multiple groups of different speakers. The neural network model is a denoising autoencoder for near-field and far-field speech.
[0046] In an exemplary embodiment, extracting the voiceprint feature vector of the voice information to be recognized in this step is to perform a forward propagation algorithm on the voiceprint feature vector of the voice information to be recognized, which may include:
[0047] First, the sound energy spectrum information of the speech information to be recognized is input into the trained speaker recognition neural network model; the sound energy spectrum information passes through the speaker recognition neural network, and then the voiceprint feature vector of the speaker represented by the audio can be output. This part can be achieved through relevant technologies, that is, if the speech input into the speaker recognition neural network is far-field speech, the far-field voiceprint feature vector of the speaker is obtained; if the speech input into the speaker recognition neural network is near-field speech, the near-field voiceprint feature vector of the speaker is obtained.
[0048] Then, the voiceprint feature vector of the speaker represented by the audio obtained after passing through the speaker recognition neural network is input into the noise reduction autoencoder to obtain the near-field voiceprint feature vector of the speaker represented by the audio, that is, the far-field voiceprint feature vector is converted into the corresponding near-field voiceprint feature vector.
[0049] Among them, the speaker recognition neural network model is trained based on near-field speech and far-field speech, using the speaker identity (ID) as a label. That is to say, after inputting the near-field sound capability spectrum information and the far-field sound capability spectrum information into the speaker recognition neural network, the speaker's feature vector in the far field and the feature vector in the near field can be obtained respectively. The trained speaker recognition neural network model has the ability to distinguish speaker characteristics.
[0050] The present application uses a noise reduction autoencoder to project the voiceprint feature vector of the voice information to be recognized into a space that is as similar as possible to the near-field voiceprint feature. In particular, for scenarios where the voice information to be recognized is far-field voice information, the far-field voiceprint feature vector to be recognized is converted into a corresponding near-field voiceprint feature vector, thereby reducing the problem of not being able to accurately identify the identity of a speaker far away from the device due to the distance between the speaker and the device.
[0051] Step 101: Determine the user identity of the voice information to be recognized based on the extracted voiceprint feature vector.
[0052] In an exemplary embodiment, identifying the user identity of the acquired voice information includes:
[0053] The obtained voiceprint feature vector is matched with the voiceprint information corresponding to the user identity during registration to determine the user identity corresponding to the obtained voiceprint feature vector.
[0054] This application uses a noise reduction autoencoder trained based on near-field speech and far-field speech to project the voiceprint feature vectors of the speech information to be recognized, especially the far-field voiceprint feature vectors, into a space that is as similar as possible to the near-field voiceprint features, and then uses the voiceprint feature vectors output by the noise reduction autoencoder to realize subsequent recognition of the voice user's identity, thereby reducing the problem of not being able to accurately identify the identity of a speaker far away from the device due to the distance between the speaker and the device, and improving the robustness of far-field voiceprint recognition.
[0055] The present application also provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute any of the above-mentioned speech recognition methods.
[0056] The present application further provides a speech recognition device, comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor; when the computer program is executed by the processor, the steps of any one of the above-mentioned speech recognition methods are implemented.
[0057] The present application also provides a smart speaker, including a memory and a processor, wherein a computer program is stored in the memory; when the computer program is executed by the processor, the following steps are performed: using a noise reduction autoencoder to extract a voiceprint feature vector of voice information to be recognized; wherein the noise reduction autoencoder is obtained based on near-field speech and far-field speech training; and identifying the user identity of the obtained voice information based on the extracted voiceprint feature vector.
[0058] In an exemplary embodiment, when the computer program is executed by the processor, the following steps are also performed: extracting a near-field voiceprint feature vector based on the near-field speech, and extracting a far-field voiceprint feature vector based on the far-field speech; taking the near-field voiceprint feature vector and the far-field voiceprint feature vector corresponding to the same user as a set of inputs, inputting them into a mathematical model, and training the parameters of the denoising autoencoder corresponding to the mathematical model.
[0059] Figure 2 Schematic diagram of the structure of the embodiment of the speech recognition device of the present application. Figure 2 As shown, it at least includes: a processing module and an identification module; wherein,
[0060] A processing module, used to extract a voiceprint feature vector of the speech information to be recognized by using a noise reduction autoencoder; wherein the noise reduction autoencoder is trained based on near-field speech and far-field speech;
[0061] The recognition module is used to determine the user identity of the voice information to be recognized based on the extracted voiceprint feature vector.
[0062] In an exemplary embodiment, the speech recognition device of the present application also includes: a training module for extracting a near-field voiceprint feature vector based on near-field speech, and extracting a far-field voiceprint feature vector based on far-field speech; inputting the near-field voiceprint feature vector and the far-field voiceprint feature vector corresponding to the same user as a set of inputs into a mathematical model, and training the parameters of the denoising autoencoder corresponding to the mathematical model.
[0063] Although the embodiments disclosed in this application are as above, the contents described are only embodiments adopted to facilitate understanding of this application and are not intended to limit this application. Any technician in the field to which this application belongs can make any modifications and changes in the form and details of implementation without departing from the spirit and scope disclosed in this application, but the scope of patent protection of this application shall still be based on the scope defined in the attached claims.
Claims
1. A speech recognition method, comprising: A voiceprint feature vector of speech information to be recognized is extracted by using a noise reduction autoencoder; wherein the noise reduction autoencoder is obtained by training based on near-field speech and far-field speech, the voiceprint feature vector is obtained by a speaker recognition neural network, the speaker recognition neural network is obtained by training based on the near-field speech, the far-field speech and the speaker identity, the noise reduction autoencoder is used to project the far-field voiceprint feature vector into a near-field voiceprint feature vector, and the noise reduction autoencoder is obtained by training a neural network model using pairs of near-field voiceprint feature vectors and far-field voiceprint feature vectors of multiple groups of different users; The user identity of the voice information to be recognized is determined based on the extracted voiceprint feature vector.
2. The speech recognition method according to claim 1, further comprising: Extracting a near-field voiceprint feature vector according to the near-field speech, and extracting a far-field voiceprint feature vector according to the far-field speech; The near-field voiceprint feature vector and the far-field voiceprint feature vector corresponding to the same user are taken as a group of inputs and input into a mathematical model to train the parameters of the noise reduction automatic encoder corresponding to the mathematical model.
3. The speech recognition method according to claim 1 or 2, wherein: The step of extracting the voiceprint feature vector of the voice information to be recognized includes: Input the sound energy spectrum information of the voice information to be recognized into a trained speaker recognition neural network model to obtain a voiceprint feature vector of the speaker represented by the voice information to be recognized; The speaker's voiceprint feature vector is input into the noise reduction automatic encoder to obtain the speaker's near-field voiceprint feature vector represented by the speech information to be recognized.
4. The speech recognition method according to claim 3, wherein: The voice information to be recognized is far-field voice information; the voiceprint feature vector of the speaker is a far-field voiceprint feature vector;.
5. The speech recognition method according to claim 2, wherein: The mathematical model includes a deep neural network model.
6. A computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute the speech recognition method according to any one of claims 1 to 5.
7. A device for implementing speech recognition, comprising a memory and a processor, wherein: The memory stores the following instructions executable by the processor: used to execute the steps of the speech recognition method described in any one of claims 1 to 5.
8. A smart speaker, comprising a memory and a processor, wherein: A computer program is stored in the memory; when the computer program is executed by the processor, the following steps are performed: A noise reduction autoencoder is used to extract a voiceprint feature vector of speech information to be recognized; wherein the noise reduction autoencoder is obtained by training based on near-field speech and far-field speech, the voiceprint feature vector is obtained by a speaker recognition neural network, the speaker recognition neural network is trained by the near-field speech, the far-field speech and the speaker identity identifier, the noise reduction autoencoder is used to project the far-field voiceprint feature vector into a near-field voiceprint feature vector, the noise reduction autoencoder is obtained by training a neural network model using pairs of near-field voiceprint feature vectors and far-field voiceprint feature vectors of multiple groups of different users; the user identity of the speech information obtained is recognized based on the extracted voiceprint feature vector.
9. According to the smart speaker of claim 8, when the computer program is executed by the processor, the following steps are also performed: extracting a near-field voiceprint feature vector based on the near-field speech, and extracting a far-field voiceprint feature vector based on the far-field speech; taking the near-field voiceprint feature vector and the far-field voiceprint feature vector corresponding to the same user as a set of inputs, inputting them into a mathematical model, and training the parameters of the denoising autoencoder corresponding to the mathematical model.
10. A speech recognition device, comprising: Processing module, identification module; wherein, A processing module, configured to extract a voiceprint feature vector of speech information to be recognized by using a noise reduction autoencoder; wherein the noise reduction autoencoder is obtained by training based on near-field speech and far-field speech, the voiceprint feature vector is obtained by a speaker recognition neural network, the speaker recognition neural network is obtained by training based on the near-field speech, the far-field speech and a speaker identity identifier, the noise reduction autoencoder is configured to project the far-field voiceprint feature vector into a near-field voiceprint feature vector, and the noise reduction autoencoder is obtained by training a neural network model using pairs of near-field voiceprint feature vectors and far-field voiceprint feature vectors of multiple groups of different users; The recognition module is used to determine the user identity of the voice information to be recognized based on the extracted voiceprint feature vector.
11. The speech recognition device according to claim 10, further comprising a training module for: A near-field voiceprint feature vector is extracted according to the near-field speech, and a far-field voiceprint feature vector is extracted according to the far-field speech; the near-field voiceprint feature vector and the far-field voiceprint feature vector corresponding to the same user are taken as a group of inputs, input into a mathematical model, and the parameters of the denoising autoencoder corresponding to the mathematical model are trained.
