A voice wake-up method and device

Abstract

This application provides a voice wake-up method and device. The method includes: receiving an audio sequence; performing audio event detection on the audio sequence; inputting an audio segment containing speech into a first-level wake-up model; when the first-level wake-up model detects a wake word, inputting the audio segment into a second-level wake-up model; when the second-level wake-up model detects a wake word, inputting the audio segment into a voiceprint model; and when the voiceprint model determines that the speaker is the registered user, the system is woken up. The first-level wake-up model is a streaming model modeled on a word-by-word basis; the second-level wake-up model is a non-streaming model, and the convolutional output channels of the second-level wake-up model are assigned attention weights. This application reduces power consumption by constructing a two-level wake-up model. Because the first-level wake-up model uses word-by-word modeling, the second-level wake-up model can obtain accurate wake word boundaries, thereby using an attention mechanism to improve the feature extraction capability of the convolutional channels, ultimately improving wake-up performance.

Description

Technical Field

[0001] This invention relates to speech recognition technology, and more particularly to a voice wake-up method and device. Background Technology

[0002] Voiceprint wake-up technology is an important branch of speech recognition, widely used in voice interaction systems such as mobile terminals, smart homes, in-vehicle navigation, and smart wearables. It allows specific users to wake up devices using voice commands. More specifically, the voiceprint wake-up task involves continuously receiving voice messages in the background and automatically detecting predefined wake words. Once a wake word is detected, it needs to be verified whether it was uttered by a specific speaker. This ensures that only the specific speaker can rouse the device and put it into a specific working state. The detection of wake words is achieved through a wake-up model, while the identification of the speaker is achieved through a voiceprint model.

[0003] Because wake-up models need to run continuously in the background, using a large model for wearable devices, while ensuring performance, significantly impacts battery life. Using a small model, however, makes performance difficult to guarantee, thus negatively affecting user experience. To address this issue, existing technologies propose a two-level model solution: a first-level wake-up model with lower power consumption and lower accuracy, and a second-level model with higher power consumption and higher accuracy. Since the first wake-up model has lower power consumption, it can be used for initial recognition of the speech data stream. When the first wake-up model's recognition result indicates that the speech data stream includes a segment corresponding to the wake word, the speech data stream can be input to the second wake-up model for secondary recognition. Because the second wake-up model has higher accuracy, it can achieve precise recognition of the speech data stream. Furthermore, the first wake-up model can filter out a significant amount of data, thus greatly reducing the amount of data input to the second wake-up model, effectively reducing power consumption.

[0004] However, since existing two-level wake-up models are all based on frame modeling, they cannot obtain accurate wake-up word boundaries, which limits the feature extraction capability of the convolutional channels of the two-level models, thus limiting wake-up performance.

[0005] In addition, the different usage environments of speech recognition devices (especially wearable devices) will cause the voiceprint model to face different room reverberation, background noise with different signal-to-noise ratios, and different channels, all of which will result in a higher false acceptance rate of the voiceprint model. Summary of the Invention

[0006] The purpose of this invention is to provide a voice wake-up method and device with lower power consumption, better wake-up performance, and better robustness.

[0007] The following provides a brief overview of one or more aspects to offer a basic understanding of them. This overview is not an exhaustive summary of all conceived aspects, nor is it intended to identify key or decisive elements of all aspects, nor to define the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form to prepare for the more detailed descriptions that follow.

[0008] According to one aspect of the present invention, a voice wake-up method is provided, comprising:

[0009] Receive audio sequences;

[0010] Audio event detection is performed on the audio sequence, and segments without speech are removed, leaving only audio segments with speech.

[0011] Input an audio clip with speech into the first-level wake-up model;

[0012] When the first-level wake-up model detects a wake word, it inputs the audio segment into the second-level wake-up model;

[0013] When the second-level wake-up model detects a wake word, it inputs an audio clip into the voiceprint model.

[0014] The system is activated when the voiceprint model determines that the speaker is the registered user;

[0015] The first-level wake-up model is a streaming model and is modeled using words as units; the second-level wake-up model is a non-streaming model, and attention weights are assigned to the convolutional output channels of the second-level wake-up model.

[0016] In one embodiment, the first-level wake-up model models the wake-up word syllable by syllable and uses a sliding window to calculate the confidence level h(x) of the wake-up word.

[0017]

[0018] in Let T be the input sequence within the sliding window. conf Let t be the length of the sliding window, 1≤t1≤t2...≤T conf This indicates that the path of the entire decoding network can only be monotonic. Represents time t i The model outputs the wake word as ω i The probability is given by the fact that the wake word has a total of M characters, i.e.:

[0019] ω = {ω1, ω2, ..., ω} M},

[0020] h(x) represents the maximum posterior probability that can be achieved within the sliding window under the premise of satisfying the decoding network; when the confidence h(x) is greater than the first threshold, it is considered that a wake word has been detected within the sliding window.

[0021] In one embodiment, the second-level wake-up model adopts a multi-layer FSMN stacked architecture, whose convolutional receptive field is equivalent to the receptive field of the wake word. The output nodes of each layer of the model represent the probability value of the receptive field containing the wake word. When the probability value is greater than a second threshold, it is determined that the wake word has been detected.

[0022] In one embodiment, the convolutional output channels of the second-level wake-up model are assigned attention weights using the following method:

[0023] Assume the channel output of the convolution is U∈R o×L Calculate the attention weights Atten∈R for each channel of U. o×1 :

[0024] Atten=sigmoid(linear(avgpool(U)))

[0025] Where avgpool represents average pooling on the second dimension of U, linear represents a fully connected layer, and sigmoid represents the sigmoid activation function;

[0026] Use Attention (Attenuation) to enhance the output channels of convolutions:

[0027]

[0028] in This represents the Hadamard product of a matrix.

[0029] In one embodiment, the voiceprint model is a three-layer LSTM model with ASP pooling as the output. The voiceprint model extracts a voiceprint vector for each wake word and compares the voiceprint vector of the wake word with the pre-stored voiceprint vector. If the similarity between the voiceprint vector of the wake word and the pre-stored voiceprint vector is greater than a third threshold, then the speaker is determined to be the registrant.

[0030] In one embodiment, before comparing the voiceprint vector of the wake-up word with the pre-stored voiceprint vector, the method further includes: projecting the voiceprint vector of the wake-up word onto the pre-stored voiceprint vector using a projection matrix, wherein the training process of the projection matrix is ​​as follows:

[0031] Input the voiceprint of the training audio;

[0032] Calculate the global mean of the voiceprint vec of all training audios.

[0033]

[0034] Subtract the mean from the voiceprint vec of all training audios;

[0035] The EM algorithm is used to estimate the covariance Φ of the within-class components. w Covariance Φ of inter-class components b ;

[0036] The calculation for step E is as follows:

[0037]

[0038]

[0039] The training samples contain K speakers (K classes). This represents the inter-class variance of the training samples for the k-th speaker. Let represent the intraclass variance of the k-th speaker. and All of these originate from the previous EM iteration. The M-step makes Φ w，k Φ b，k Based on the previous iteration, move closer to the global optimum:

[0040]

[0041]

[0042] Where n represents the number of training samples for the k-th speaker, and m represents the mean of the training samples for the k-th speaker. These represent Φ after M steps of optimization. w，k With Φ b，k ;

[0043] After multiple rounds of EM iterations, the covariance Φ of the intra-class components that maximizes the likelihood of the voiceprint is obtained. w Covariance Φ between class components b ;

[0044] For Φ w Perform Cholesky decomposition:

[0045]

[0046] For A -1 Φ b (A -1 ) T Perform singular value decomposition:

[0047]

[0048] Where A matrix is ​​Φ w The Cholesky decomposition, where the U matrix is ​​A -1Φ b (A -1 ) T The affine matrix of the singular value decomposition, where I is the identity matrix and Ψ is the diagonal matrix;

[0049] The projection matrix T = U is obtained. T A -1 .

[0050] In one embodiment, projecting the voiceprint vector of the wake-up word onto a pre-stored voiceprint vector using a projection matrix includes:

[0051] Voiceprint vectors are extracted from n wake-up phrases of the registrant using a voiceprint model;

[0052] Subtracting the global mean μ from the voiceprint vector yields: u = {u1, u2, ..., u} n};

[0053] Projecting u using the projection matrix T, we get: in

[0054] For u g The average value is obtained as follows: It refers to the voiceprint registered by the target speaker;

[0055] Subtract the global mean μ from the voiceprint of the wake word fragment to obtain the voiceprint. The voiceprint of the wake word is obtained by projecting using the projection matrix T.

[0056] In one embodiment, the similarity is represented by the log-likelihood ratio. The calculation method is as follows:

[0057]

[0058] In one embodiment, the method further includes: whenever a similarity value close to but not exceeding a third threshold is detected by the voiceprint model, the third threshold is reduced by a preset amount; when the system is woken up, the third threshold is restored to its initial value.

[0059] According to another aspect of the present invention, a voice wake-up device is provided, including a memory and a processor; the memory is used to store a software program; the processor is used to implement the voice wake-up method as described in any of the preceding claims when the software program is executed.

[0060] The beneficial effects of this invention are as follows: by constructing a two-level wake-up model, the first-level wake-up model uses a streaming model to coarsely detect wake words, and the second-level wake-up model uses a non-streaming model with an attention mechanism. Furthermore, the first-level wake-up model uses character modeling, which enables the second-level wake-up model to obtain accurate wake word boundaries. Thus, the second-level wake-up model can use the attention mechanism to improve the feature extraction capability of the convolutional channels, thereby ultimately improving wake-up performance. Attached Figure Description

[0061] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0062] The above-described features and advantages of the present invention will be better understood after reading the following detailed description of embodiments of the present disclosure in conjunction with the accompanying drawings. In the drawings, components are not necessarily drawn to scale, and components having similar related properties or features may have the same or similar reference numerals.

[0063] Figure 1 This is a flowchart of an embodiment of the method of this application. Detailed Implementation

[0064] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should be noted that the aspects described below with reference to the accompanying drawings and specific embodiments are merely exemplary and should not be construed as limiting the scope of protection of the present invention in any way.

[0065] like Figure 1 As shown, this application provides a voice wake-up method, including:

[0066] S1, Receive audio sequence;

[0067] S2. Perform audio event detection on the audio sequence, namely VAD (Voice Activity Detection), remove segments of the audio sequence that do not contain speech, and retain only audio segments that contain speech, thereby ensuring that the input of the voiceprint model contains human voice, thus reducing system power consumption.

[0068] S3. Input the audio clip with speech into the first-level wake-up model;

[0069] The first-level wake-up model uses a streaming model with a convolutional and fully connected structure to coarsely filter wake-up words, eliminating most non-wake-up words. This reduces the number of calls to the second-level wake-up model and the voiceprint model, significantly lowering system power consumption. To ensure the overall system's false alarm rate isn't too high and thus doesn't negatively impact user experience, the wake-up threshold of the first-level wake-up model is typically set very low. However, the subsequent second-level wake-up model can perform secondary verification of false wake-ups from the first-level model, thereby filtering out false wake-ups.

[0070] In this embodiment, the first-level wake-up model models the wake-up word character by character, meaning the number of nodes in the output layer corresponds to the number of characters in the wake-up word, and each node represents the probability of each character at each time step; and a sliding window is used to calculate the confidence level h(x) of the wake-up word.

[0071]

[0072] in Let T be the input sequence within the sliding window. conf Let t be the length of the sliding window, 1≤t1≤t2...≤T conf This indicates that the path of the entire decoding network can only be monotonic. Represents time t i The model outputs the wake word as ω i The probability is given by the fact that the wake word has a total of M characters, i.e.:

[0073] ω = {ω1, ω2, ..., ω M},

[0074] h(x) represents the maximum posterior probability that can be achieved within the sliding window, provided that the decoding network is satisfied. When the confidence level h(x) is greater than the first threshold, it is considered that a wake-up word has been detected within the sliding window. In addition, the first-level wake-up model can also obtain the temporal boundary of the wake-up word by backtracking the decoding network, enabling the second-level wake-up model to adopt a non-streaming structure.

[0075] S4. When the first-level wake-up model detects a wake word, the audio segment is input into the second-level wake-up model. The second-level wake-up model adopts a multi-layer FSMN (Feedforward Sequential Memory Networks) stacked architecture. Its convolutional receptive field is comparable to that of the wake word. The output nodes of each layer of the model represent the probability value that the receptive field contains the wake word. When the probability value is greater than the second threshold, it is determined that the wake word has been detected.

[0076] Since the first-level wake-up model can obtain the temporal boundary of the wake word, thus enabling the acquisition of complete wake word fragments, and the second-level wake-up model can use a non-streaming structure, this method also uses SE (squeeze and excitation) to add attention to different channels of the convolution on the output of each layer of FSMN, thereby improving the performance of the wake-up model.

[0077] Specifically, attention weights can be assigned using the following methods:

[0078] Assume the channel output of the convolution is U∈R o×L Where o represents the number of output channels and L represents the time length of the output, the attention weights Atten∈R for each channel of U are calculated. o×1 :

[0079] Atten=sigmoid(linear(avgpool(U)))

[0080] Where avgpool represents average pooling on the second dimension of U, linear represents a fully connected layer, and sigmoid represents the sigmoid activation function;

[0081] Use Attention (Attenuation) to enhance the output channels of convolutions:

[0082]

[0083] in This represents the Hadamard product of a matrix.

[0084] S5. When the second-level wake-up model detects a wake-up word, it is necessary to use the voiceprint model to further confirm the speaker in the wake-up word fragment. Therefore, the audio fragment is input into the voiceprint model. When the voiceprint model determines that the speaker is the registered user, the system is woken up.

[0085] In practical applications, we pre-calculate and store the registrant's voiceprint. When the second-level wake-up model sends in a wake-up word fragment, it only needs to extract the voiceprint of that fragment and compare it with the pre-stored voiceprint. Considering that voiceprint wake-up uses fixed speech content and that the terminal device has limited computing power, the voiceprint model used in this embodiment is a three-layer LSTM with ASP pooling as the output.

[0086] The voiceprint model extracts a voiceprint vector for each wake word and compares the voiceprint vector of the wake word with the pre-stored voiceprint vector. If the similarity between the voiceprint vector of the wake word and the pre-stored voiceprint vector is greater than the third threshold, the speaker is determined to be the registrant.

[0087] Furthermore, considering the complex and varied usage environment of voiceprint models in wearable devices such as smartwatches, the differences in channel transfer functions on different deployment devices, and the fact that the training data of voiceprint models cannot cover all usage scenarios and channel distortion, the mismatch between training and testing domains is unavoidable. This will cause the performance of voiceprint models to drop sharply in some scenarios.

[0088] To address the aforementioned issues, this method employs a dual covariance assumption (between-class covariance and within-class covariance) for voiceprints. Before comparing the voiceprint vector of the wake-up word with the pre-stored voiceprint vector, a projection matrix is ​​used to project the voiceprint vector of the wake-up word onto the pre-stored voiceprint vector. The projection matrix improves the intra-class aggregation of voiceprints from the same speaker and the inter-class discriminative power of voiceprints from different speakers, thus exhibiting good robustness even in different usage scenarios.

[0089] Specifically, assume that the voiceprint z of the training data (after subtracting the global mean) consists of between-class components x and within-class components y, i.e., z = x + y. The covariance of the within-class components is Φ. w The covariance of the inter-class components is Φ b We hope to find θ = {Φ} w , Φ b} maximizes p(z, x|θ).

[0090] Therefore, the specific steps for training the projection matrix SO are as follows:

[0091] SOO, Input training data;

[0092] S01. Calculate the global mean of the voiceprint vec of all training data.

[0093] S02, voiceprint vec minus this mean.

[0094] S03. Estimate the covariance Φ of the within-class components using the EM (Expectation-Maximum) algorithm. w Covariance Φ of inter-class components b The EM algorithm consists of an expectation step (E-step) and a maximization step (M-step). The E-step is calculated as follows:

[0095]

[0096]

[0097] The training sample contains K speakers, meaning there are K classes of voiceprints. This represents the inter-class variance of the training samples for the k-th speaker. Let the variance of the k-th speaker be the within-class variance. and All are from the previous EM iteration;

[0098] M steps make Φ w，k Φ b，k Based on the previous iteration, move closer to the global optimum:

[0099]

[0100]

[0101] Where n represents the number of training samples for the k-th speaker, and m represents the mean of the training samples for the k-th speaker. These represent Φ after M steps of optimization. w，k With Φ b，k ;

[0102] After multiple rounds of EM iterations, the covariance Φ of the intra-class components that maximizes the likelihood of the voiceprint is obtained. w Covariance Φ between class components b .

[0103] S04. Calculate the projection matrix. For ease of storage and computation, we need to obtain a projection matrix T such that the in-class covariance Φ of T*z is... w Transform into identity matrix I, inter-class covariance Φ b It becomes a diagonal matrix Ψ.

[0104] For Φ w Perform Cholesky decomposition:

[0105]

[0106] For A -1 Φ b (A -1 ) T Perform singular value decomposition:

[0107]

[0108] Where I is the identity matrix and Ψ is the diagonal matrix;

[0109] The projection matrix T = U is obtained. T A -1 .

[0110] Once the projection matrix is ​​obtained, it can be used to improve intra-class cohesion and inter-class differentiation during the voiceprint registration phase of the target speaker.

[0111] Specifically, the registration step S7 is as follows:

[0112] S71. Input the n wake-up word fragments of the target speaker into the voiceprint model;

[0113] S72. Extracting voiceprint vectors using a voiceprint model;

[0114] S73. Subtract the global mean μ from the voiceprint vector to obtain: u = {u1, u2, ..., u} n};

[0115] S74. Projecting u using the projection matrix T, we get: in

[0116] S75, the average value of ug is obtained as follows: It refers to the voiceprint registered by the target speaker.

[0117] Similarly, the voiceprint recognition stage also includes:

[0118] S6. Subtract the global mean μ from the voiceprint of the wake-up word fragment to obtain the voiceprint.

[0119] S8. Project using projection matrix T to obtain the voiceprint of the wake word.

[0120] S9. After obtaining the projected wake-up word voiceprint and the pre-stored target speaker's voiceprint, calculate the similarity between the two. In this embodiment, the similarity is calculated using the log-likelihood ratio. The calculation method is as follows:

[0121]

[0122] If the similarity exceeds the third threshold, the system is considered to be the same person as the wake-up word and the registered speaker, and the system is activated.

[0123] Preferably, the method further includes S10, dynamic adjustment of the threshold.

[0124] When a user attempts to use a voiceprint wake-up system, they will typically try multiple times if the system fails to wake them up. For command words that fail to wake up, both the voiceprint system's score and the wake-up system's output probability will be close to the third threshold. Therefore, whenever the system detects a value close to the third threshold in the above model, it appropriately lowers the third threshold. This makes it easier for the user to activate the voiceprint wake-up system during subsequent attempts.

[0125] Therefore, whenever a similarity value close to but not exceeding the third threshold is detected by the voiceprint model, the third threshold is reduced by a preset amount; after the user activates the voiceprint wake-up system and transmits the command, the third threshold is restored to its initial value to avoid false triggering.

[0126] Corresponding to the above method, this application embodiment also provides a voice wake-up device, including a memory and a processor; the memory is used to store a software program; the processor is used to implement the voice wake-up method as described above when the software program is executed.

[0127] Furthermore, embodiments of this application also provide a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform any of the above-described implementations of the voice wake-up method.

[0128] In summary, the embodiments of this application have the following beneficial effects:

[0129] 1) To address the low power consumption of wearable devices, a two-level wake-up model is designed. A first-level small model intercepts most non-wake-up words, while a second-level large model performs secondary confirmation of the wake-up of the small model. This greatly reduces system power consumption while ensuring wake-up performance.

[0130] 2) The first and second level wake-up models use different modeling units: the first level wake-up model uses word modeling to obtain the time boundary of the wake word, and the second level wake-up model predicts the complete wake word based on the time boundary. Therefore, the second level wake-up model can incorporate an attention mechanism to greatly improve the performance of the second level wake-up model.

[0131] 3) In view of the complex and varied usage scenarios and channel transfer functions of wearable devices, a projection matrix is ​​used to increase the inter-class distance between different speakers and reduce the intra-class distance between the same speaker, thereby greatly improving the wake-up performance of the voiceprint model.

[0132] 4) Based on the characteristic that users will try multiple times after failing to wake up via voiceprint, the threshold is dynamically adjusted to ensure user experience.

[0133] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0134] The prior description of this disclosure is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not intended to be limited to the examples and designs described herein, but should be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0135] The above description is merely a preferred example of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application shall be included within the scope of protection of this application.

Claims

1. A voice wake-up method, characterized in that, include: Receive audio sequences; Audio event detection is performed on the audio sequence, and segments without speech are removed, leaving only audio segments with speech. Input an audio clip with speech into the first-level wake-up model; When the first-level wake-up model detects a wake word, it inputs an audio clip into the second-level wake-up model; when the second-level wake-up model detects a wake word, it inputs an audio clip into the voiceprint model; when the voiceprint model determines that the speaker is the registered user, the system is woken up. The first-level wake-up model is a streaming model that models words as units; the second-level wake-up model is a non-streaming model, and attention weights are assigned to the convolutional output channels of the second-level wake-up model; the first-level wake-up model models the wake-up word word by word and uses a sliding window to calculate the confidence h(x) of the wake-up word. in Let T be the input sequence within the sliding window. conf Let t be the length of the sliding window, 1≤t1≤t2...≤T conf This indicates that the path of the entire decoding network can only be monotonic. Represents time t i The model outputs the wake word as ω i The probability is given by the fact that the wake word has a total of M characters, i.e.: ω={ω1,ω2,....,ω M }, h(x) represents the maximum posterior probability that can be achieved within the sliding window under the premise of satisfying the decoding network; when the confidence h(x) is greater than the first threshold, it is considered that a wake word has been detected within the sliding window; wherein, the first-level wake-up model uses character modeling to obtain the time boundary of the wake word, and the second-level wake-up model is also used to predict the complete wake word based on the time boundary.

2. The voice wake-up method according to claim 1, characterized in that: The second-level wake-up model adopts a multi-layer FSM N-stacked architecture, whose convolutional receptive field is comparable to that of the wake word. The output nodes of each layer of the model represent the probability value of the wake word being contained in the receptive field. When the probability value is greater than the second threshold, it is determined that the wake word has been detected.

3. The voice wake-up method according to claim 2, characterized in that, The convolutional output channels of the second-level wake-up model are weighted with attention using the following method: Assume the channel output of the convolution is U∈R o×L Where o represents the number of output channels and L represents the output time sequence length, the attention weights Atten∈R for each channel of U are calculated. o×1 : Atten=sigmoid(linear(avgpool(U))) Where avgpool represents average pooling on the second dimension of U, linear represents a fully connected layer, and sigmoid represents the sigmoid activation function; Use Attention (Attenuation) to enhance the output channels of convolutions: in This represents the Hadamard product operation of a matrix.

4. The voice wake-up method according to claim 1, characterized in that, The voiceprint model is a three-layer LSTM model with ASP pooling as the output. The voiceprint model extracts a voiceprint vector for each wake word and compares the voiceprint vector of the wake word with the pre-stored voiceprint vector. If the similarity between the voiceprint vector of the wake word and the pre-stored voiceprint vector is greater than the third threshold, then the speaker is determined to be the registrant.

5. The voice wake-up method according to claim 4, characterized in that, Before comparing the wake-up word's voiceprint vector with the pre-stored voiceprint vector, the process further includes: projecting the wake-up word's voiceprint vector onto the pre-stored voiceprint vector using a projection matrix. The training process for the projection matrix is ​​as follows: Input the voiceprint of the training audio; Calculate the global mean of the voiceprint vec of all training audios. Where N represents the number of voiceprints in the training audio, vec i This represents the voiceprint of the i-th training audio; the voiceprint vec of all training audios minus this mean; The EM algorithm is used to estimate the covariance Φ of the within-class components. w Covariance Φ of inter-class components b The calculation for step E is as follows: The training samples contain K speakers, meaning there are K types of voiceprints. This represents the inter-class variance of the training samples for the k-th speaker. This represents the within-class variance of the training sample for the k-th speaker. and All are from the previous EM iteration; M steps make Φ w，k Φ b，k Based on the previous iteration, move closer to the global optimum: Where n represents the number of training samples for the k-th speaker, and m represents the mean of the training samples for the k-th speaker. These represent Φ after M steps of optimization. w，k With Φ b，k After multiple rounds of EM iterations, the covariance Φ of the intra-class components that maximizes the likelihood of the voiceprint is obtained. w Covariance Φ between class components b ; For Φ w Perform Cholesky decomposition: For A -1 Φ b (A -1 ) T Perform singular value decomposition: The projection matrix T = U is obtained. T A -1 .

6. The voice wake-up method according to claim 5, characterized in that, The projection of the wake word's voiceprint vector onto the pre-stored voiceprint vector using a projection matrix includes: Voiceprint vectors are extracted from n wake-up phrases of the registrant using a voiceprint model; Subtracting the global mean μ from the voiceprint vector yields: u = {u1, u2, ..., u} n }; Projecting u using the projection matrix T, we get: in For u g The average value is obtained as follows: It refers to the voiceprint registered by the target speaker; Subtract the global mean μ from the voiceprint of the wake word fragment to obtain the voiceprint. The voiceprint of the wake word is obtained by projecting using the projection matrix T.

7. The voice wake-up method according to claim 6, characterized in that, The similarity is expressed as the log-likelihood ratio. The calculation method is as follows:

8. The voice wake-up method according to claim 7, characterized in that, Also includes: Whenever a similarity value close to but not exceeding the third threshold is detected from the voiceprint model, the third threshold is reduced by a preset amount. When the system is woken up, the third threshold is restored to its initial value.

9. A voice wake-up device, characterized in that, Includes a memory and a processor; the memory is used to store software programs; The processor is configured to implement the voice wake-up method as described in any one of claims 1 to 8 when executing the software program.

Images