Voice wake-up method, device, storage medium and electronic device

Abstract

The present invention discloses a method, device, storage medium, and electronic device for voice wake-up. The method comprises: obtaining a voice wake-up model, setting a wake-up sensitivity and a wake-up threshold; obtaining a voice to be determined, and outputting a predicted probability through the obtained voice wake-up model; specifying a normal distribution probability density function, and calculating a cumulative probability distribution through the normal distribution probability density function; converting the output predicted probability into a distribution point, and obtaining a smoothed probability by taking a corresponding value from the distribution point in the obtained cumulative probability distribution; and determining whether to wake up based on the smoothed probability, the wake-up sensitivity, and the wake-up threshold. The present invention avoids the problem of false activation caused by rapid jumps in predicted values ​​when the voice wake-up model has a very low confidence level for a certain predicted value, and at the same time provides a relatively linear threshold adjustment for voice wake-up, thereby responding to voice wake-up more accurately.

Description

Technical Field

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

[0002] Voice wake-up technology, which allows users to activate smart devices from sleep mode by speaking a wake-up word, has been widely adopted. Traditional voice wake-up methods primarily compare the output of the voice wake-up model with a set threshold to determine whether to wake the device.

[0003] Forecasting confidence is the probability of a predicted value falling within an acceptable error range, obtained using interval estimation techniques in mathematical statistics. However, when the prediction confidence is very low, rapid jumps in the predicted value can cause false activations. Furthermore, traditional methods cannot linearly adjust the threshold for voice wake-up. Summary of the Invention

[0004] The object of the present invention is to provide a method and apparatus for voice wake-up to solve the problems of false activation caused by rapid jumps in model prediction values ​​and lack of linear threshold adjustment in voice wake-up proposed in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A voice wake-up method, comprising:

[0007] S1. Obtain the voice wake-up model and set the wake-up sensitivity and wake-up threshold;

[0008] S2. Obtain the speech to be judged and output the predicted probability through the speech wake-up model obtained in S1;

[0009] S3. specify a normal distribution probability density function and calculate the cumulative probability distribution using the normal distribution probability density function;

[0010] S4. Convert the predicted probability output by S2 into a distribution point, and take the corresponding value of the distribution point in the cumulative probability distribution obtained by S3 to obtain the smoothed probability;

[0011] S5. Determine whether to wake up based on the smoothed probability obtained in S4 and the wake-up sensitivity and wake-up threshold set in S1.

[0012] A further technical solution is: the voice wake-up model is a two-classification model.

[0013] A further technical solution is: S2 specifically comprises: obtaining voice data frame by frame from the data stream of the speech to be determined according to the set data block size, storing and updating the data in the cache; and updating the cache according to the first-in-first-out principle, inputting the speech frames in the cache into the speech wake-up model to obtain the predicted probability.

[0014] A further technical solution is: S3 is specifically:

[0015] The normal distribution probability density function is obtained by specifying the mean and variance.

[0016] The distribution point sequence is obtained by specifying the distribution range and resolution.

[0017] Then the cumulative probability distribution is obtained through the normal distribution probability density function and the distribution point sequence.

[0018] A further technical solution is: the method for converting the predicted probability into distribution points is:

[0019] The predicted probability is obtained as a scalar value through the inverse Sigmoid function, and then the scalar value is normalized by the specified distribution range. The distribution point is calculated based on the normalized scalar value according to the specified resolution.

[0020] A further technical solution is: the S5 is specifically:

[0021] The activation number is initialized to 0, and the smoothed probability obtained by S4 is compared with the wake-up sensitivity set by S1. If the smoothed probability is greater than the wake-up sensitivity, the activation number is increased by 1; otherwise, the activation number is reduced by 1. When the activation number is greater than the wake-up threshold, it is judged to be in the awakening state.

[0022] At the same time, the present invention also provides the following technical solutions:

[0023] A voice wake-up device, comprising:

[0024] A voice data acquisition module is used to receive voice data and store and update it in a buffer area;

[0025] A model processing module for predicting the probability of the acquired speech output;

[0026] Normal distribution smoothing module, used to convert predicted probabilities into smoothed probabilities;

[0027] The activation decision module is used to set the wake-up sensitivity, wake-up threshold, and determine whether to wake up.

[0028] In addition, the present invention also provides the following technical solutions:

[0029] A computer-readable storage medium stores a computer program, wherein the computer program is used to implement the voice wake-up method described above.

[0030] In addition, the present invention also provides the following technical solutions:

[0031] An electronic device, comprising:

[0032] a storage device for storing a computer program;

[0033] A processing device is used to execute the computer program in the storage device to implement the steps of the voice wake-up method as described above.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] By estimating the normal distribution of the voice wake-up model output and using the mean and standard deviation to calculate the cumulative probability distribution, we avoid the problem of false activation caused by rapid jumps in predicted values ​​when the voice wake-up model has low confidence in a certain predicted value. At the same time, we provide a relatively linear threshold adjustment for voice wake-up, thereby responding to voice wake-up more accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Schematic diagram of the flow of the voice wake-up method in an embodiment of the present invention;

[0037] Figure 2 A schematic diagram of obtaining speech to be determined in an embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram of the wake-up decision process in an embodiment of the present invention;

[0039] Figure 4 This is a structural block diagram of a voice wake-up device in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0041] Example 1

[0042] The embodiment of the present invention provides a voice wake-up method, such as Figure 1 As shown, the following steps are included:

[0043] S1. Obtain the voice wake-up model and set the wake-up sensitivity and wake-up threshold;

[0044] Optionally, the voice wake-up model includes but is not limited to a binary classification model; the wake-up sensitivity is used to control the sensitivity of the wake-up; the wake-up threshold is used to determine whether to activate;

[0045] S2. Obtain the speech to be judged and output the predicted probability through the speech wake-up model obtained in S1;

[0046] like Figure 2 As shown, the speech data stream to be determined obtains speech data frame by frame according to the manually set data block size, and stores and updates it to the buffer area at the same time. The buffer area is updated according to the first-in-first-out principle, and the speech frames in the buffer area are input into the voice wake-up model to obtain the predicted probability;

[0047] S3. specify a normal distribution probability density function and calculate the cumulative probability distribution for the normal distribution probability density function;

[0048] Optionally, the normal distribution probability density function is obtained by a specified mean and variance, a distribution point sequence is obtained by a specified distribution range and resolution, and a cumulative probability distribution is calculated by the normal distribution density function and the distribution point sequence;

[0049] For example, specifying a mean of 4 and a variance of 6 can yield a probability density function. Specifying a distribution range of -10 to 22 and a resolution of 200 yields a distribution point sequence of [-10.00, -9.99, -9.98, ..., 21.994, 22.00], totaling 32 × 200 points, and then using the distribution point sequence and probability density function to calculate the cumulative probability distribution.

[0050] S4. Convert the predicted probability of S2 into a distribution point, take the corresponding value in the cumulative probability distribution obtained in S3, and obtain the smoothed probability;

[0051] Optionally, the predicted probability is converted into a distribution point by: applying a Sigmoid inverse function to the predicted probability to obtain a scalar value, normalizing the scalar value by a specified distribution range, calculating the normalized scalar value according to a specified resolution to obtain a distribution point, and subjecting the distribution point to a cumulative probability distribution to obtain a smoothed probability;

[0052] S5. Determine whether to wake up based on the smoothed probability obtained in S4 and the wake-up sensitivity and wake-up threshold set in S1;

[0053] like Figure 3 As shown, the activation number is initialized to 0, that is, the smoothed probability obtained by S4 is compared with the awakening sensitivity. If it is greater than the awakening sensitivity, the activation number is increased by 1, otherwise the activation number is reduced by 1. When the activation number is greater than the awakening threshold, it is judged to be in the awakening state;

[0054] The voice wake-up method provided in Example 1 avoids the problem of false activation caused by rapid jumps in prediction values ​​when the confidence of the activation word model for a certain prediction value is very low, and at the same time provides relatively linear threshold adjustment for voice wake-up.

[0055] Example 2

[0056] A voice wake-up device according to an embodiment of the present invention, such as Figure 4 Shown, including:

[0057] A voice data acquisition module is used to receive voice data and store and update it in a buffer area;

[0058] like Figure 3 As shown, the speech to be determined obtains speech data by frame according to the artificially set data block size, and stores and updates it to the buffer area at the same time. The update of the buffer area follows the first-in-first-out principle;

[0059] A model processing module for predicting the probability of the acquired speech output;

[0060] Optionally, the speech frames in the buffer are input into a speech wake-up model to obtain a predicted probability;

[0061] Normal distribution smoothing module, used to convert predicted probabilities into smoothed probabilities;

[0062] Optionally, a normal distribution density function is derived from the specified mean and variance, and a distribution point sequence is derived from the specified distribution range and resolution. The cumulative probability distribution is calculated from the normal distribution density function and the distribution point sequence. The predicted probability is then applied to a scalar value using the inverse Sigmoid function. The scalar value is then normalized using the specified distribution range. The normalized scalar value is then calculated using the specified resolution to obtain the distribution points. The distribution points are then substituted into the cumulative probability distribution to obtain the smoothed probability.

[0063] The activation decision module is used to set the wake-up sensitivity and wake-up threshold and determine whether to wake up.

[0064] It can be understood that the activation number is initialized to 0, that is, the obtained smoothed probability is compared with the awakening sensitivity. If it is greater than the awakening sensitivity, the activation number is increased by 1, otherwise the activation number is reduced by 1. When the activation number is greater than the awakening threshold, it is judged to be in the awakening state;

[0065] The voice wake-up device provided in this embodiment 2 avoids the problem of false activation caused by rapid jumps in the prediction value when the activation word model has a very low confidence in a certain prediction value, and provides a relatively linear threshold adjustment for voice wake-up.

[0066] Example 3

[0067] Based on the same inventive concept, an embodiment of the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program is used to implement the steps of a voice wake-up method as described in Example 1.

[0068] Based on the same inventive concept, an embodiment of the present invention further provides an electronic device, including:

[0069] a storage device for storing a computer program;

[0070] A processing device is used to execute the computer program in the storage device to implement the steps of a voice wake-up method as described in Example 1.

[0071] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A voice wake-up method, characterized in that: include: S1. Obtain a voice wake-up model and set the wake-up sensitivity and threshold. S2. Obtain the speech to be determined and output the predicted probability through the voice wake-up model obtained in S1; Specifically, S2 includes: obtaining speech data frame by frame from the data stream of the speech to be determined according to a set data block size, storing and updating the data in a buffer; and updating the buffer according to a first-in-first-out principle, inputting the speech frames in the buffer into the speech wake-up model to obtain a predicted probability; S3. Specify a normal distribution probability density function and calculate the cumulative probability distribution using the normal distribution probability density function; The S3 is specifically: The normal distribution probability density function is obtained by specifying the mean and variance. The distribution point sequence is obtained by specifying the distribution range and resolution. Then, the cumulative probability distribution is obtained through the normal distribution probability density function and the distribution point sequence; S4. Convert the predicted probability output from S2 into a distribution point. Then, use the corresponding value of the distribution point in the cumulative probability distribution obtained in S3 to obtain the smoothed probability. S5. Determine whether to wake up based on the smoothed probability obtained in S4 and the wake-up sensitivity and wake-up threshold set in S1.

2. A voice wake-up method according to claim 1, characterized in that: The voice wake-up model is a two-class model.

3. The voice wake-up method according to claim 1, characterized in that: The method for converting the predicted probability into distribution points is: The predicted probability is obtained as a scalar value through the inverse Sigmoid function, and then the scalar value is normalized by the specified distribution range. The distribution point is calculated based on the normalized scalar value according to the specified resolution.

4. The voice wake-up method according to claim 1, characterized in that: Specifically, step S5 is as follows: the activation number is initialized to 0, and the smoothing probability obtained in step S4 is compared with the wake-up sensitivity set in step S1. If the smoothing probability is greater than the wake-up sensitivity, the activation number is increased by 1; otherwise, the activation number is decreased by 1. When the activation number is greater than the wake-up threshold, it is determined to be in the wake-up state.

5. A voice wake-up device, used to implement a voice wake-up method according to any one of claims 1 to 4, characterized in that: include: A voice data acquisition module is used to receive voice data and store and update it in a buffer area; A model processing module for predicting the probability of the acquired speech output; Normal distribution smoothing module, used to convert predicted probabilities into smoothed probabilities; The activation decision module is used to set the wake-up sensitivity, wake-up threshold, and determine whether to wake up.

6. A computer-readable storage medium storing a computer program, characterized in that: The computer program is used to implement the voice wake-up method according to any one of claims 1 to 4.

7. An electronic device, characterized in that: include: a storage device for storing a computer program; A processing device, configured to execute the computer program in the storage device to implement the steps of the voice wake-up method according to any one of claims 1 to 4.

Images