Voice wake-up method and device, and voice wake-up module training method and device

Abstract

The present invention provides a voice wake-up method and device, and a training method and device for a voice wake-up module. The voice wake-up method comprises: obtaining a voice signal to be recognized, wherein the voice signal to be recognized comprises a wake-up word; performing wake-up word recognition on the voice signal to be recognized using an acoustic model to obtain a wake-up word recognition result, wherein the acoustic model comprises a whole word node, and the whole word node uses the word vector of the wake-up word as a weight; performing voice wake-up according to the wake-up word recognition result, which can improve the wake-up effect.

Description

Technical Field

[0001] The present invention relates to the field of voice wake-up technology, and in particular to a voice wake-up method and device, and a voice wake-up module training method and device. Background Art

[0002] Voice wake-up technology refers to the process in which a smart device detects a user's specific voice (usually a set wake-up word) in sleep mode, causing the device to enter a waiting state and then enter a voice intelligent interaction process. Voice wake-up technology has a wide range of applications, such as robots, mobile phones, wearable devices, smart homes, and cars.

[0003] However, the modeling units of existing voice wake-up models are often syllables or phonemes, but the wake-up words are usually whole words of 3 to 6 syllables. There is a large mismatch between the training target and the test target, and the wake-up effect is poor. Summary of the invention

[0004] In view of this, the embodiments of the present invention provide a voice wake-up method and device, and a voice wake-up module training method and device, which can improve the wake-up effect.

[0005] According to aspects of an embodiment of the present invention, a voice wake-up method is provided, which is applied to a voice wake-up model, including: obtaining a voice signal to be recognized, wherein the voice signal to be recognized includes a wake-up word; using an acoustic model to perform wake-up word recognition on the voice signal to be recognized to obtain a wake-up word recognition result, wherein the acoustic model includes a whole word node, and the whole word node uses the word vector of the wake-up word as a weight; and performing voice wake-up according to the wake-up word recognition result.

[0006] In one embodiment of the present invention, before obtaining the voice signal to be recognized, the above-mentioned voice wake-up method also includes: receiving a custom wake-up word instruction input by the user, wherein the custom wake-up word instruction includes a wake-up word; obtaining a word vector of the wake-up word; adding a whole word node to the acoustic model, and using the word vector of the wake-up word as the weight of the whole word node.

[0007] In one embodiment of the present invention, the speech wake-up model is obtained by the following training method: obtaining a word vector of a wake-up word, and using the word vector of the wake-up word as the weight of the whole word node in the acoustic model; obtaining a training sample, wherein the training sample includes a speech signal sample and annotated data, the speech signal sample includes the wake-up word, and the annotated data includes text information corresponding to the speech signal sample; training the acoustic model based on the training sample to obtain the speech wake-up model.

[0008] In one embodiment of the present invention, the speech wake-up model also includes a word model, wherein the above-mentioned acquisition of the word vector of the wake-up word includes: using the word model to acquire the word vector of the wake-up word; wherein the above-mentioned training of the acoustic model based on the training sample to obtain the speech wake-up model includes: training the word model and the acoustic model based on the training sample to obtain the speech wake-up model.

[0009] In one embodiment of the present invention, the acoustic model also includes multiple syllable nodes, wherein the above training method also includes: using the word model to obtain syllable vectors of multiple syllables; and using the syllable vectors of the multiple syllables as weights of the multiple syllable nodes in the acoustic model.

[0010] In one embodiment of the present invention, the acoustic model includes a neural network-based temporal class classification (CTC) model.

[0011] According to a second aspect of an embodiment of the present invention, a method for training a speech wake-up model is provided, including: obtaining a word vector of a wake-up word, and using the word vector of the wake-up word as the weight of a whole word node in an acoustic model; obtaining a training sample, wherein the training sample includes a speech signal sample and annotated data, the speech signal sample includes the wake-up word, and the annotated data includes text information corresponding to the speech signal sample; training the acoustic model based on the training sample to obtain a speech wake-up model.

[0012] According to a third aspect of an embodiment of the present invention, a voice wake-up device is provided, including: an acquisition module, used to acquire a voice signal to be recognized, wherein the voice signal to be recognized includes a wake-up word; a recognition module, used to perform wake-up word recognition on the voice signal to be recognized using an acoustic model to obtain a wake-up word recognition result, wherein the acoustic model includes a whole word node, and the whole word node uses the word vector of the wake-up word as a weight; a wake-up module, used to perform voice wake-up according to the wake-up word recognition result.

[0013] According to a fourth aspect of an embodiment of the present invention, a training device for a speech wake-up model is provided, including: a word vector module, used to obtain a word vector of a wake-up word, and use the word vector of the wake-up word as the weight of the whole word node in the acoustic model; an acquisition module, used to obtain training samples, wherein the training samples include speech signal samples and labeled data, and the speech signal samples include the wake-up word; a training module, used to train the acoustic model based on the training samples to obtain a speech wake-up model.

[0014] According to a fifth aspect of an embodiment of the present invention, there is provided a computer-readable storage medium on which computer-executable instructions are stored, wherein the executable instructions, when executed by a processor, implement any of the methods described above.

[0015] According to a fifth aspect of an embodiment of the present invention, there is provided an electronic device, characterized in that the electronic device comprises: a processor; a memory for storing instructions executable by the processor; and the processor is used to execute any one of the methods described above.

[0016] According to the technical solution provided by the embodiment of the present invention, a speech signal to be recognized is obtained, wherein the speech signal to be recognized includes a wake-up word; the wake-up word recognition is performed on the speech signal to be recognized using an acoustic model to obtain a wake-up word recognition result, wherein the acoustic model includes a whole word node, and the whole word node uses the word vector of the wake-up word as a weight; and speech wake-up is performed according to the wake-up word recognition result. By adding a whole word node to the acoustic model and using the word vector of the wake-up word as the weight of the whole word node, the acoustic model can obtain a larger posterior on the whole word node corresponding to the wake-up word, thereby improving the wake-up effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 FIG. 1 is a flow chart of a voice wake-up method provided by an embodiment of the present invention.

[0019] Figure 2 FIG. 4 is a flow chart of a voice wake-up method provided by another embodiment of the present invention.

[0020] Figure 3 FIG. 1 is a schematic diagram of a voice wake-up method provided by an embodiment of the present invention.

[0021] Figure 4 FIG. 4 is a schematic diagram of a voice wake-up method provided by another embodiment of the present invention.

[0022] Figure 5 The figure is a flow chart of a method for training a voice wake-up model provided in one embodiment of the present invention.

[0023] Figure 6 FIG. 4 is a block diagram of a voice wake-up device provided by an embodiment of the present invention.

[0024] Figure 7 FIG. 1 is a block diagram of a training device for a voice wake-up model provided by an embodiment of the present invention.

[0025] Figure 8 FIG. 1 is a block diagram of an electronic device provided in accordance with an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] Figure 1 FIG. 1 is a flow chart of a voice wake-up method provided by an embodiment of the present invention. The method can be executed by a robot, a mobile phone, a wearable device, a smart home, a car, or other smart device with a voice wake-up function. Figure 1 As shown, the method includes the following contents.

[0028] S110: Acquire a voice signal to be recognized, wherein the voice signal to be recognized includes a wake-up word.

[0029] The voice signal to be recognized is voice data containing the set wake-up word. The voice signal to be recognized can be collected by a microphone, wherein the microphone can be set on the device to be awakened, or on the controller of the device to be awakened (for example, a remote control). It should be understood that the present invention does not make specific limitations on this.

[0030] S120: Perform wake-up word recognition on the speech signal to be recognized using the acoustic model to obtain a wake-up word recognition result, wherein the acoustic model includes a whole word node, and the whole word node uses the word vector of the wake-up word as a weight.

[0031] Specifically, the speech signal to be recognized can be input into the acoustic model; the acoustic model extracts the acoustic features of the speech signal to be recognized, and outputs the posterior probability of the whole word of the wake-up word and the posterior probability of other syllables according to the acoustic features of the speech signal to be recognized. Taking the speech signal to be recognized as "Xiao Fei Xiao Fei, how is the weather today?" as an example, where "Xiao Fei Xiao Fei" is the wake-up word, the acoustic model in the embodiment of the present invention (the acoustic model has a whole word node with the word vector of the wake-up word "Xiao Fei Xiao Fei" as the weight) can output the posterior probability of the whole word "Xiao Fei Xiao Fei" and the posterior probabilities of multiple syllables such as "today", "day", "of", "day", "air", "how", "what", and "like".

[0032] It should be understood that the acoustic features may be MFCC (Mel-Frequency Cepstral Coefficients) features, PLP (Linear Predictive Coding) features or Filterbank features, and the present invention does not specifically limit the types and extraction methods of acoustic features.

[0033] The acoustic model may include a neural network-based time series classification CTC model to solve the problem that the input sequence and the output sequence are difficult to correspond one to one. For example, the acoustic model may be a long short-term memory (LSTM) model, a convolutional neural network (CNN) model, etc. It should be understood that the present invention does not limit the specific type of the acoustic model.

[0034] S130: Perform voice wake-up according to the wake-up word recognition result.

[0035] For example, the posterior probability output by the acoustic model may be smoothed with a certain window length. When the smoothed posterior score exceeds a preset threshold, it is considered that the wake-up word is recognized, and then the device to be awakened is awakened.

[0036] According to the technical solution provided by the embodiment of the present invention, by adding a whole word node to the acoustic model and using the word vector of the wake-up word as the weight of the whole word node, the acoustic model can obtain a larger posteriori on the whole word node corresponding to the wake-up word, thereby improving the wake-up effect. At the same time, this solution adopts end-to-end wake-up, and the decoding module logic is simple, which can avoid the problem of increased resource usage caused by too many decoding paths in traditional decoding, and is more suitable for low-power chips.

[0037] Figure 2 FIG. 2 is a flow chart of a voice wake-up method provided by another embodiment of the present invention. The method may be executed by a computer device (eg, a server). Figure 2 As shown, the method includes the following contents.

[0038] S210: Receive a custom wake-up word instruction input by a user, wherein the custom wake-up word instruction includes a wake-up word.

[0039] That is to say, the voice wake-up method provided in the embodiment of the present invention can support the function of customizing the wake-up word. For example, the acoustic model is trained with the wake-up word "Xiao Fei Xiao Fei", and the user wants to customize the wake-up word to "Da Fei Da Fei". At this time, the user can enter the custom wake-up word "Da Fei Da Fei" by entering the custom wake-up word instruction.

[0040] It should be noted that the user can input a custom wake-up word through voice or text, and the present invention does not make any specific limitation to this.

[0041] S220: Obtain the word vector of the wake-up word.

[0042] Specifically, Figure 3 As shown in the figure, the word vector W of the custom wake-up word "Da Fei Da Fei" can be generated through the word model y.

[0043] It should be understood that the word model may be an LSTM model or the like. The present invention does not limit the specific type of the word model, as long as the word vector of the custom wake-up word can be obtained.

[0044] For example, Figure 4 As shown in the figure, for the text sequence "Da Fei Da Fei's weather today", "Da Fei Da Fei" is input into the word model in the form of one-hot encoding to obtain the word vector W of "Da Fei Da Fei". y .

[0045] S230: Add a whole word node to the acoustic model, and use the word vector of the wake-up word as the weight of the whole word node.

[0046] For example, Figure 4 As shown in the figure, the above-mentioned custom wake-up word "Da Fei Da Fei" whole word node can be added to the fully connected layer of the acoustic model, and the word vector W of the wake-up word "Da Fei Da Fei" can be y As the weight of the whole word node of "Da Fei Da Fei", the acoustic model obtains a larger posterior on the whole word node of "Da Fei Da Fei", so as to quickly detect the wake-up word.

[0047] S240: Acquire a voice signal to be recognized, wherein the voice signal to be recognized includes a wake-up word.

[0048] For example, the voice signal X to be recognized is "Da Fei Da Fei, what is the weather like today", where "Da Fei Da Fei" is the wake-up word.

[0049] S250: Using the acoustic model to perform wake-up word recognition on the voice signal to be recognized, and obtain a wake-up word recognition result.

[0050] For example, the speech signal to be recognized X is "Da Fei Da Fei, what is today's weather" and is input into the acoustic model. The feature extraction module in the acoustic model extracts features of the speech signal to be recognized X and obtains the acoustic feature f t am (X), acoustic feature f t am (X) Input the fully connected layer and obtain the text sequence corresponding to the output speech signal to be recognized. Through the softmax activation function, the posterior probability of the whole word "da fei da fei" and the posterior probabilities of multiple syllables such as "jin", "tian", "de", "tian", and "qi" can be obtained, that is, P t (π t |X,W).

[0051] S260: Perform voice wake-up according to the wake-up word recognition result.

[0052] The posterior probability output by the acoustic model is then smoothed with a certain window length. When the smoothed posterior score exceeds a preset threshold, it is considered that the wake-up word "Da Fei Da Fei" is recognized, and the device to be awakened can be awakened.

[0053] It should be noted that when the user wants to continue to add custom wake-up words (such as "xiaobai xiaobai"), the acoustic model can continue to add whole word nodes (such as "xiaobai xiaobai"), and use the word vector of the newly added wake-up word (such as "xiaobai xiaobai") as the weight of the newly added (such as "xiaobai xiaobai") whole word node, so that the user can use the newly added wake-up word (such as "xiaobai xiaobai") to wake up the device to be woken up.

[0054] According to the technical solution provided by the implementation of the present invention, by receiving a custom wake-up word instruction input by the user; obtaining the word vector of the wake-up word; adding a whole word node in the acoustic model, and using the word vector of the wake-up word as the weight of the whole word node; obtaining a voice signal to be recognized; using the acoustic model to perform wake-up word recognition on the voice signal to be recognized, and obtaining a wake-up word recognition result; and performing voice wake-up according to the wake-up word recognition result, it is possible to support a custom wake-up function, and facilitate users to use custom wake-up words to wake up the device to be awakened. In addition, the embodiment of the present invention adopts end-to-end voice wake-up, and the decoding module logic is simple, which can avoid the problem of increased resource usage caused by too many decoding paths in traditional decoding. Even if the number of custom wake-up words increases, the resource usage will not increase significantly.

[0055] Figure 5 FIG. 1 is a flow chart of a method for training a voice wake-up model provided by an embodiment of the present invention. The method may be executed by a computer device (eg, a server). Figure 5 As shown, the method includes the following contents.

[0056] S510: Obtain a word vector of the wake-up word, and use the word vector of the wake-up word as the weight of the whole word node in the acoustic model.

[0057] That is to say, compared with the traditional acoustic model with only syllable nodes, the acoustic model in the embodiment of the present application adds whole word nodes and converts the word vector W of the wake-up word into y As the weight of the whole word node. For example, if the wake-up word is "Xiao Fei Xiao Fei", then the whole word node "Xiao Fei Xiao Fei" is added to the acoustic model, and the word vector of "Xiao Fei Xiao Fei" is used as the weight of the whole word node "Xiao Fei Xiao Fei".

[0058] It should be noted that when the wake-up word "Da Fei Da Fei" is continued to be added, the whole word node of "Da Fei Da Fei" can also be added to the acoustic model, and the word vector of "Da Fei Da Fei" can be used as the weight of the whole word node of "Da Fei Da Fei".

[0059] S520: Acquire a training sample, wherein the training sample includes a speech signal sample and annotation data, the speech signal sample includes a wake-up word, and the annotation data includes text information corresponding to the speech signal sample.

[0060] Specifically, audio and annotation (X, Y) can be input to the acoustic model, where X = {x1, ..., x s} is the acoustic feature of the speech signal sample, Y = {y1,…y N} can be a markup with whole words and syllables as modeling units, Taking the voice signal sample "Xiao Fei Xiao Fei, how is the weather today?" as an example, the labeled data can be: "Xiao Fei Xiao Fei", "today", "day", "of", "weather", "air", "how", "what", "how". It should be noted that the entire wake-up word can be labeled as a whole.

[0061] S530: Train the acoustic model based on the training samples to obtain a voice wake-up model.

[0062] The acoustic model may be a CTC (Connectionist Temporal Classification) model, which maps X to Y.

[0063] Specifically, the input T frame acoustic features X enter the acoustic model and the output sequence is obtained If the continuous outputs in π are removed and can be the same as the labeled sequence Y, then π is considered to be a valid output. In the model training phase, the acoustic model training can be completed with the goal of maximizing the likelihood of all valid sequences π∈B, where B is all valid output sequences.

[0064]

[0065] According to the technical solution provided by the embodiment of the present invention, the word vector of the wake-up word is obtained, and the word vector of the wake-up word is used as the weight of the whole word node in the acoustic model; a training sample is obtained; and the acoustic model is trained based on the training sample to obtain a speech wake-up model. By adding a whole word node to the acoustic model and using the word vector of the wake-up word as the weight of the whole word node in the acoustic model, the syllable and the whole word can be forced to be mapped to the same space, which can make the training target and the test target more matched and improve the wake-up effect.

[0066] In one embodiment of the present invention, the output sequence is For example, y represents the whole word and y′ represents the syllable. The likelihood probability of the whole word y at time t can be expressed as follows:

[0067]

[0068] The likelihood probability of syllable node y′ at time t can be expressed as follows:

[0069]

[0070] The likelihood probability of the labeled sequence Y can be expressed as follows:

[0071]

[0072] Among them, W y is the weight of the whole word node, W y′ is the weight of the syllable node, f t am (X) is the acoustic feature of the speech signal sample.

[0073] It should be understood that the above description is merely an exemplary description, and the present invention does not impose any specific limitation on the specific formula representation.

[0074] In one embodiment of the present invention, the speech wake-up model also includes a word model, wherein the above-mentioned acquisition of the word vector of the wake-up word includes: using the word model to acquire the word vector of the wake-up word; wherein the above-mentioned training of the acoustic model based on the training sample to obtain the speech wake-up model includes: training the word model and the acoustic model based on the training sample to obtain the speech wake-up model.

[0075] Specifically, the text information of the wake-up word can be input into the word model to obtain the word vector W of the wake-up word y ; The word vector W of the wake-up word y Input the acoustic model and replace the whole word node in the acoustic model with the word vector W of the wake-up word y The syllable nodes in the acoustic model are represented by W y′ By jointly training the acoustic model and the word model and forcing syllables and whole words to be mapped to the same space, the training and testing objectives can be more closely matched.

[0076] In one embodiment of the present invention, the acoustic model also includes multiple syllable nodes, wherein the above training method also includes: using the word model to obtain syllable vectors of multiple syllables; and using the syllable vectors of the multiple syllables as weights of the multiple syllable nodes in the acoustic model.

[0077] That is, the weights W of multiple syllable nodes in the acoustic model y′ It can also be determined by a word model. Specifically, the text information corresponding to each syllable is input into the word model, and the word model outputs the syllable vector W corresponding to the syllable y′ ; The syllable vector W output by the word model y′ Input the acoustic model and convert the syllable vector W y′ As the weight of the syllable node in the acoustic model.

[0078] In the embodiment of the present invention, the whole word node of the acoustic model is represented by W y is the weight, and the syllable node is W y′ By jointly training the acoustic model and the word model and forcing the syllables and the whole word to be mapped to the same space, the training target and the test target can be more closely matched.

[0079] All the above optional technical solutions can be arbitrarily combined to form optional embodiments of the present invention, which will not be described in detail here.

[0080] The following are embodiments of the device of the present invention, which can be used to implement the method embodiments of the present invention. For details not disclosed in the device embodiments of the present invention, please refer to the method embodiments of the present invention.

[0081] Figure 6 FIG. 1 is a block diagram of a voice wake-up device provided by an embodiment of the present invention. Figure 6 As shown, the voice wake-up device 600 includes:

[0082] An acquisition module 610 is used to acquire a voice signal to be recognized, wherein the voice signal to be recognized includes a wake-up word;

[0083] The recognition module 620 is used to perform wake-up word recognition on the speech signal to be recognized by using the acoustic model to obtain the wake-up word recognition result, wherein the acoustic model includes a whole word node, and the whole word node uses the word vector of the wake-up word as the weight;

[0084] The wake-up module 630 is used to perform voice wake-up according to the wake-up word recognition result.

[0085] According to the technical solution provided by the embodiment of the present invention, a speech signal to be recognized is obtained, wherein the speech signal to be recognized includes a wake-up word; wake-up word recognition is performed on the speech signal to be recognized using an acoustic model to obtain a wake-up word recognition result, wherein the acoustic model includes a whole word node, and the whole word node uses the word vector of the wake-up word as a weight; speech wake-up is performed according to the wake-up word recognition result, and by adding a whole word node to the acoustic model and using the word vector of the wake-up word as the weight of the whole word node, the acoustic model can obtain a larger posterior on the whole word node corresponding to the wake-up word, thereby improving the wake-up effect.

[0086] In another embodiment of the present invention, the above-mentioned voice wake-up device also includes a receiving module 640, which is used to receive a custom wake-up word instruction input by the user before obtaining the voice signal to be recognized, wherein the custom wake-up word instruction includes a wake-up word; obtain the word vector of the wake-up word; add a whole word node to the acoustic model, and use the word vector of the wake-up word as the weight of the whole word node.

[0087] Figure 7FIG. 1 is a block diagram of a training device for a voice wake-up model provided by an embodiment of the present invention. Figure 7 As shown, the training device 700 of the voice wake-up model includes:

[0088] A word vector module 710 is used to obtain a word vector of the wake-up word and use the word vector of the wake-up word as the weight of the whole word node in the acoustic model;

[0089] An acquisition module 720 is used to acquire a training sample, wherein the training sample includes a speech signal sample and annotated data, the speech signal sample includes a wake-up word, and the annotated data includes text information corresponding to the speech signal sample;

[0090] The training module 730 is used to train the acoustic model based on the training samples to obtain a voice wake-up model.

[0091] According to the technical solution provided by the embodiment of the present invention, the word vector of the wake-up word is obtained, and the word vector of the wake-up word is used as the weight of the whole word node in the acoustic model; a training sample is obtained; and the acoustic model is trained based on the training sample to obtain a speech wake-up model. By adding a whole word node to the acoustic model and using the word vector of the wake-up word as the weight of the whole word node in the acoustic model, the syllable and the whole word can be forced to be mapped to the same space, which can make the training target and the test target more matched and improve the wake-up effect.

[0092] In another embodiment of the present invention, the speech wake-up model also includes a word model, wherein the word vector module 710 is used to obtain the word vector of the wake-up word using the word model; wherein the training module 730 is used to train the word model and the acoustic model based on the training samples to obtain the speech wake-up model.

[0093] In another embodiment of the present invention, the acoustic model also includes multiple syllable nodes, wherein the above-mentioned training device also includes a syllable vector module 740, which is used to obtain syllable vectors of multiple syllables using the word model; and use the syllable vectors of the multiple syllables as weights of multiple syllable nodes in the acoustic model.

[0094] In another embodiment of the present invention, the acoustic model includes a neural network-based temporal class classification (CTC) model.

[0095] The implementation process of the functions and effects of each module in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, which will not be repeated here.

[0096] Figure 8 FIG. 8 is a block diagram of an electronic device 800 provided according to an embodiment of the present invention.

[0097] Reference Figure 8The electronic device 800 includes a processing component 810, which further includes one or more processors, and a memory resource represented by a memory 820 for storing instructions executable by the processing component 810, such as an application. The application stored in the memory 820 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 810 is configured to execute instructions to perform the above-mentioned voice wake-up method or the training method of the voice wake-up model.

[0098] The electronic device 800 may also include a power supply component configured to perform power management of the electronic device 800, a wired or wireless network interface configured to connect the electronic device 800 to a network, and an input / output (I / O) interface. The electronic device 800 may operate based on an operating system stored in the memory 820, such as Windows Server 2000. TM , Mac OSX TM , Unix TM , Linux TM , FreeBSD TM or similar.

[0099] A non-temporary computer-readable storage medium, when the instructions in the storage medium are executed by the processor of the above-mentioned electronic device 700, enables the above-mentioned electronic device 800 to execute a voice wake-up method or a training method for a voice wake-up model.

[0100] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0101] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0102] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0103] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0104] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0105] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program check codes.

[0106] In addition, it should be noted that the combination of the various technical features in this case is not limited to the combination described in the claims of this case or the combination described in the specific embodiments. All technical features recorded in this case can be freely combined or combined in any way unless there is a contradiction between them.

[0107] It should be noted that the above examples are only specific embodiments of the present invention, and the present invention is obviously not limited to the above examples, and there are many similar variations. All variations directly derived or associated from the contents disclosed by the technicians in this field should fall within the protection scope of the present invention.

[0108] It should be understood that the first, second, etc. qualifiers mentioned in the embodiments of the present invention are only used to more clearly describe the technical solutions of the embodiments of the present invention and cannot be used to limit the protection scope of the present invention.

[0109] The above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A voice wake-up method, characterized in that: Applied to voice wake-up models, including: Acquire a voice signal to be recognized, wherein the voice signal to be recognized includes a wake-up word; Using an acoustic model to perform wake-up word recognition on the speech signal to be recognized, and obtaining a wake-up word recognition result, wherein the acoustic model includes a whole word node, the whole word node uses a word vector of the wake-up word as a weight, and the word vector is obtained by embedding the wake-up word in the form of one-hot encoding using the word model; Voice wake-up is performed according to the wake-up word recognition result.

2. The voice wake-up method according to claim 1, characterized in that: Before acquiring the speech signal to be recognized, the method further includes: Receiving a custom wake-up word instruction input by a user, wherein the custom wake-up word instruction includes the wake-up word; Obtaining a word vector for the wake-up word; The whole word node is added to the acoustic model, and the word vector of the wake-up word is used as the weight of the whole word node.

3. The voice wake-up method according to claim 1, characterized in that: The voice wake-up model is obtained by the following training method: Obtaining a word vector of the wake-up word, and using the word vector of the wake-up word as the weight of the whole word node in the acoustic model; Acquire a training sample, wherein the training sample includes a speech signal sample and annotation data, the speech signal sample includes the wake-up word, and the annotation data includes text information corresponding to the speech signal sample; The acoustic model is trained based on the training samples to obtain the voice wake-up model.

4. The voice wake-up method according to claim 3, characterized in that: The voice wake-up model also includes a word model, wherein obtaining the word vector of the wake-up word includes: Using the word model to obtain a word vector of the wake-up word; The step of training the acoustic model based on the training sample to obtain the voice wake-up model includes: The word model and the acoustic model are trained based on the training samples to obtain the voice wake-up model.

5. The voice wake-up method according to claim 4, characterized in that: The acoustic model also includes a plurality of syllable nodes, wherein the training method further includes: Using the word model to obtain syllable vectors of multiple syllables; The syllable vectors of the multiple syllables are respectively used as weights of the multiple syllable nodes in the acoustic model.

6. The voice wake-up method according to any one of claims 1 to 5, characterized in that: The acoustic model includes a neural network-based time series classification (CTC) model.

7. A method for training a voice wake-up model, characterized in that: include: Obtain a word vector of the wake-up word, and use the word vector of the wake-up word as the weight of the whole word node in the acoustic model. The word vector is obtained by embedding the wake-up word in the form of one-hot encoding using the word model; Acquire a training sample, wherein the training sample includes a speech signal sample and annotation data, the speech signal sample includes the wake-up word, and the annotation data includes text information corresponding to the speech signal sample; The acoustic model is trained based on the training samples to obtain a voice wake-up model.

8. A voice wake-up device, characterized in that: include: An acquisition module, used to acquire a voice signal to be recognized, wherein the voice signal to be recognized includes a wake-up word; A recognition module, configured to use an acoustic model to perform wake-up word recognition on the speech signal to be recognized, and obtain a wake-up word recognition result, wherein the acoustic model includes a whole word node, the whole word node uses a word vector of the wake-up word as a weight, and the word vector is obtained by embedding the wake-up word in the form of one-hot encoding using the word model; The wake-up module is used to perform voice wake-up according to the wake-up word recognition result.

9. A training device for a voice wake-up model, characterized in that: include: A word vector module is used to obtain a word vector of a wake-up word and use the word vector of the wake-up word as the weight of the whole word node in the acoustic model. The word vector is obtained by embedding the wake-up word in the form of one-hot encoding using the word model; An acquisition module, configured to acquire a training sample, wherein the training sample includes a speech signal sample and annotated data, the speech signal sample includes a wake-up word, and the annotated data includes text information corresponding to the speech signal sample; A training module is used to train the acoustic model based on the training samples to obtain a voice wake-up model.

10. A computer-readable storage medium having computer-executable instructions stored thereon, characterized in that: When the executable instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

11. An electronic device, characterized in that: The electronic device comprises: processor; a memory for storing instructions executable by the processor; The processor is used to execute the method according to any one of claims 1 to 7.

Images