A voice wake-up decoding method and apparatus

Abstract

The application discloses a voice wake-up decoding method, receives audio data and calculates a posterior probability value of a voice frame; N smoothing windows are used to smooth the posterior probability value, N posterior probability smoothing values are obtained, and different window lengths of the N smoothing windows correspond to different speech speeds; M sub-windows are used to calculate M posterior scores from the N posterior probability smoothing values one by one, and the maximum value in the M posterior scores is a maximum posterior score under a corresponding sub-window; the N posterior probability smoothing values of each voice frame are compared two by two, the maximum value of the N posterior probability smoothing values is obtained, N current wake-up word paths are obtained, the total sum of the number of wake-up words on each path is compared with the number of voice frames of the wake-up word, if the total sum of the number of wake-up words on the path is greater than the number of voice frames, the maximum posterior score corresponding to the posterior probability smoothing value corresponding to the path is a wake-up threshold value; if the wake-up threshold value is greater than or equal to a confidence degree, voice wake-up is performed.

Description

Technical Field

[0001] This invention relates to the field of voice processing technology, and in particular to a voice wake-up decoding method and apparatus. Background Technology

[0002] With the gradual maturation of voice technology, voice control is becoming one of the most attractive human-computer interaction methods for electronic devices. As a result, more and more products are utilizing voice wake-up and command word recognition technology, such as smart headphones, smart speakers, smartwatches, and some IoT products.

[0003] During product use, it was found that the scenarios are complex and varied. The deep learning-based voice wake-up method performs well in normal environments, but the voice wake-up effect will be affected when the speech speed is too fast or too slow. At the same time, when there are a large number of offline command words, it will lead to a surge in computing and storage resources.

[0004] Currently, the commonly used voice wake-up methods mainly train wake-up data through deep learning, obtain a posterior score in the forward process, obtain a decoding score based on the posterior score by smoothing and taking the maximum value, and then compare it with a preset threshold. If it exceeds the threshold, the product is woken up or a certain function of the product is triggered.

[0005] Traditional technical solutions have the following drawbacks:

[0006] 1. Regarding the issue of speaker speed, using a fixed smooth window and a maximum value window to calculate the posterior score will reduce the arousal rate.

[0007] 2. It is difficult to handle a large number of offline command words by using smoothing and taking the maximum value method. After analysis, it was found that when there are hundreds of offline command words, the smoothing and taking the maximum value method has the problems of large cache and large computation. Summary of the Invention

[0008] Based on the above situation, the main objective of this invention is to provide a voice wake-up decoding method and apparatus with high recognition rate, low computational load, low storage resource saving, and high wake-up rate.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A voice wake-up decoding method includes:

[0011] S1: Receive audio data and calculate the posterior probability value of the speech frame through a neural network;

[0012] S2: The posterior probability value of the speech frame is smoothed using N smoothing windows to obtain N smoothed posterior probability values ​​of the speech frame. The N smoothing windows have different window lengths, and different window lengths correspond to different speech rates. N is a positive integer greater than or equal to 3.

[0013] S3: Using M sub-windows, calculate the M posterior scores of the speech frame one by one for the N posterior probability smoothing values. The M sub-windows are obtained by dividing the maximum window equally. The maximum value among the M posterior scores is the maximum posterior score of the speech frame under the corresponding sub-window, and the M maximum posterior scores are obtained. M is a positive integer greater than or equal to 3.

[0014] S4: Compare the pairwise N posterior probability smoothing values ​​of each speech frame to obtain the maximum value of the N posterior probability smoothing values, so as to obtain the path of the N current wake words. Compare the sum of the number of wake words on each path with the number of speech frames of the wake words. If the sum of the number of wake words on the path is greater than the number of speech frames, the posterior score of the maximum value of the posterior probability smoothing value corresponding to the path is the wake-up threshold.

[0015] S5: If the wake-up threshold is greater than or equal to the preset confidence level, then perform voice wake-up.

[0016] Preferably, before S2, the following is also included:

[0017] Based on the pre-collected speech data, the duration of the neural network classification units for the first speech rate, the second speech rate, ... the Nth speech rate are statistically analyzed, and the average duration of each of the N speech rates is calculated.

[0018] Based on the character classification units in the speech data, the duration of the current character is calculated, and the duration of the shortest character is used as the base value.

[0019] Calculate the multiple relationship between the duration of other words in the speech data and the basic value;

[0020] The multiple relationship is weighted and added to the average duration of the N speech rates to obtain the N smooth windows with different window lengths.

[0021] Preferably, in step S4, the pairwise comparison of the N posterior probability smoothing values ​​for each speech frame to obtain the maximum value of the N posterior probability smoothing values, in order to obtain the paths of the N current wake words, includes:

[0022] S41: The N posterior probability smoothing values ​​p0, p1...p(N-1) of the speech frame are compared pairwise to obtain the maximum value of the N posterior probability smoothing values;

[0023] S42: Obtain the index number of the maximum value;

[0024] S43: Repeat steps S41 and S42;

[0025] S44: Obtain the paths of the N current wake words, where each path corresponds to one of the N smoothing windows.

[0026] This invention also discloses a voice wake-up decoding device, including a voice acquisition module, a neural network module, a posterior probability smoothing module, a sub-window-based maximum value calculation module, a wake-up threshold calculation module, and a wake-up judgment module.

[0027] The voice acquisition module is used to receive audio data;

[0028] The neural network module is used to calculate the posterior probability value of the speech frame in the audio data;

[0029] The posterior probability smoothing module is used to smooth the posterior probability value of the speech frame using N smoothing windows to obtain N smoothed posterior probability values ​​of the speech frame. The N smoothing windows have different window lengths, and different window lengths correspond to different speech rates. N is a positive integer greater than or equal to 3.

[0030] The sub-window-based maximum value calculation module is used to calculate the M posterior scores of the speech frame by applying the M sub-windows to the N posterior probability smoothing values ​​one by one. The M sub-windows are obtained by dividing the maximum window equally. The maximum value among the M posterior scores is the maximum posterior score of the speech frame under the corresponding sub-window, and the M maximum posterior scores are obtained. M is a positive integer greater than or equal to 3.

[0031] The wake-up threshold calculation module is used to compare the pairwise N posterior probability smoothing values ​​of each speech frame to obtain the maximum value of the N posterior probability smoothing values, so as to obtain the path of the N current wake-up words, and compare the sum of the number of wake-up words on each path with the number of speech frames of the wake-up words. If the sum of the number of wake-up words on the path is greater than the number of speech frames, the posterior score of the maximum value of the posterior probability smoothing value corresponding to the path is the wake-up threshold.

[0032] The wake-up determination module is used to perform voice wake-up when the wake-up threshold is greater than or equal to a preset confidence level.

[0033] Preferably, it also includes a smoothing window calculation module, comprising a mean duration calculation unit, a base value calculation unit, and a smoothing window length calculation unit.

[0034] The mean duration calculation unit is used to calculate the duration of the neural network classification units for the first speech rate, the second speech rate, ... the Nth speech rate based on the pre-collected speech data, and to calculate the mean duration of the N speech rates respectively.

[0035] The basic value calculation unit is used to calculate the duration of the current character based on the character classification unit in the speech data, take the duration of the shortest character as the basic value, and calculate the multiple relationship between the duration of other characters in the speech data and the basic value.

[0036] The smooth window length calculation unit is used to weight the multiple relationship into the average duration of the N speech rates to obtain the N smooth windows with different window lengths.

[0037] Preferably, the wake-up threshold calculation module includes a maximum value comparison unit and a wake-up word path calculation unit.

[0038] The maximum value comparison unit is used to compare the N posterior probability smoothing values ​​p0, p1...p(N-1) of each speech frame in pairs to obtain the maximum value of the N posterior probability smoothing values.

[0039] The wake word path calculation unit is used to obtain the paths of the N current wake words according to the index number of the maximum value, and the paths correspond to the N smooth windows.

[0040] The present invention also discloses a voice processing chip capable of executing the voice wake-up decoding method described herein.

[0041] The present invention also discloses a voice wake-up device, including the voice wake-up decoding device of the present invention or the voice processing chip of the present invention.

[0042] The present invention also discloses a storage medium storing a program, wherein the program is used to be executed to implement the voice wake-up decoding method of the present invention.

[0043] The voice wake-up decoding method of this invention sets N different smoothing window length parameters corresponding to different speech rates, preventing the decline in recognition rate caused by excessively fast or slow speech rates or unreasonable window lengths. Simultaneously, it divides the maximum window in existing technologies into M sub-windows and reduces computational load and storage resources by employing a method based on the maximum smoothing value of the sub-windows. By obtaining different wake-up thresholds at different speech rates, it can effectively compare with preset confidence levels. Compared with traditional methods, this method reduces computational load and solves the problem of inconsistent classification unit durations, thereby improving the product's wake-up rate.

[0044] Other beneficial effects of the present invention will be explained in detail through the introduction of specific technical features and technical solutions in specific embodiments. Those skilled in the art should be able to understand the beneficial technical effects brought about by these technical features and technical solutions through the introduction of these technical features and technical solutions. Attached Figure Description

[0045] The preferred embodiments of the voice wake-up decoding method and apparatus according to the present invention will now be described with reference to the accompanying drawings. In the drawings:

[0046] Figure 1 This is a voice wake-up decoding method according to a preferred embodiment of the present invention;

[0047] Figure 2 This is a block diagram of a voice wake-up decoding device according to a preferred embodiment of the present invention;

[0048] Figure 3 This is a block diagram of a voice wake-up decoding device according to another preferred embodiment of the present invention;

[0049] Figure 4 This is a block diagram of a wake-up threshold calculation module according to a preferred embodiment of the present invention. Detailed Implementation

[0050] Figure 1 A voice wake-up decoding method according to a preferred embodiment of the present invention includes:

[0051] Step S1: Receive audio data and calculate the posterior probability value of each speech frame using a neural network. For example, audio data is collected using a microphone or other audio acquisition module. The collected audio data is then fed into the neural network to calculate the posterior probability value. In a specific implementation, the neural network uses deep learning to train the model on the wake-up data, and can obtain the posterior probability value of each speech frame during forward computation.

[0052] Step S2: Smooth the posterior probability value of the speech frame using N smoothing windows to obtain N smoothed posterior probability values ​​of the speech frame. The N smoothing windows have different window lengths, and different window lengths correspond to different speech rates. N is a positive integer greater than or equal to 3.

[0053] In one embodiment, the window lengths of the N smooth windows can be determined according to the following method:

[0054] Based on pre-collected speech data, the duration of neural network classification units for the first, second, ..., Nth speech rates are statistically analyzed, and the average duration of each of the N speech rates is calculated. Based on the character classification units in the speech data, the duration of the current character is statistically analyzed, and the duration of the shortest character is taken as the baseline value. The multiple relationship between the duration of other characters in the speech data and the baseline value is calculated. The multiple relationship is weighted into the average duration of the N speech rates to obtain the N smooth windows with different window lengths.

[0055] For example, based on a large amount of pre-collected training data, the duration of neural network classification units at fast, normal, and slow speech rates is statistically analyzed, and the average duration of each of the three speech rates is calculated. Then, based on the character classification unit, the duration of the current character is statistically analyzed. Using the duration of the shortest character as the baseline value, the multiple relationship between the duration of other characters and the baseline value is calculated. Finally, this multiple relationship is weighted and applied to the average duration of the three speech rates, thus obtaining the three smoothing window lengths s. f ,s n ,s s .

[0056] By statistically analyzing a large amount of data with different speech rates and classification units of different durations, the corresponding durations based on speech rate and classification units are obtained. Based on the duration of the statistically derived classification units, different smoothing parameters are set to correspond to different speech rates, thus preventing the decline in recognition rate caused by excessively fast or slow speech rates and unreasonable window lengths.

[0057] In a specific implementation, after obtaining smoothed windows of different window lengths based on the above method, the corresponding posterior probability smoothing value can be calculated using the following formula:

[0058]

[0059] Where i takes values ​​of 0, 1, ..., n-1, n represents the number of classifiers in the model output, j is the number of speech frames, and k represents a smooth window of different lengths.

[0060] h smooth =max{1,jw smooth +1}, let w smooth These are equal to different smoothing window lengths, for example, s. f ,s n ,s s This yields three posterior smoothing probability values.

[0061] Step S3: Calculate M posterior scores for each of the N posterior probability smoothing values ​​using M sub-windows. The M sub-windows are obtained by equally dividing the maximum window. The maximum value among the M posterior scores is the maximum posterior score of the speech frame within its corresponding sub-window, resulting in M ​​maximum posterior scores. M is a positive integer greater than or equal to 3. By dividing the maximum window into M sub-windows and employing a single judgment process, the computational load and storage resources are reduced while improving the accuracy of the wake-up word.

[0062] The maximum posterior score is calculated based on the smoothed posterior probability value obtained in step S2. In a specific implementation, the following formula can be used for calculation:

[0063]

[0064] Among them, h max =max{1,jw max +1},w max This is the largest window.

[0065] The value of i is 0, 1...m-1, where m represents the number of small windows after the maximum window is divided equally, j is the number of speech frames, and k represents smooth windows of different window lengths.

[0066] For w max Divide the data into m equal parts, for example, if m equals 3, then there's no need to store w all at once. max The frame posterior probability smoothing value only needs to be stored as w. max / m frame data, then to w max The maximum posterior score is calculated from the / m frame data, and the result value λ is saved. nf1 ,λ nf2 ,λ nf3 Then, the three values ​​are compared to obtain the maximum posterior score λ. fmax .

[0067] Next, the smoothed value of the second posterior probability is calculated to obtain the result value λ. nn1 ,λ nn2 ,λ nn3 The three values ​​are compared to obtain the maximum posterior score λ. nmax For each type, a third posterior probability smoothing value is calculated, resulting in the value λ. ns1 ,λ ns2 ,λ ns3 The three values ​​are compared to obtain the maximum posterior score λ. smax .

[0068] Step S4: Compare the pairwise N posterior probability smoothing values ​​of each speech frame to obtain the maximum value of the N posterior probability smoothing values, so as to obtain the paths of the N current wake words. Compare the sum of the number of wake words on each path with the number of speech frames of the wake words. If the sum of the number of wake words on the path is greater than the number of speech frames, then the posterior score of the maximum value of the posterior probability smoothing value corresponding to the path is the wake-up threshold.

[0069] In one implementation, the paths of the N current wake words can be calculated using the following method:

[0070] Step S41: Compare the N posterior probability smoothing values ​​p0, p1...p(N-1) of the speech frame pairwise to obtain the maximum value of the N posterior probability smoothing values;

[0071] Step S42: Obtain the index number of the maximum value;

[0072] Step S43: Repeat steps S41 and S42;

[0073] Step S44: Obtain the paths of the N current wake words, where each path corresponds to one of the N smoothing windows.

[0074] For example, in step S41, when calculating the maximum value, the index number corresponding to the maximum value is obtained. For instance, if the current wake word is classified into 5 categories (0, 1, 2, 3, 4), and the posterior probability smoothing value of the second character is the largest, then the output is 1. By comparing the maximum value for each speech frame, the path of the current wake word can be obtained, for example, 111111222223333344444. For the posterior probability smoothing values ​​of the above N smoothing windows, N path values ​​can be obtained. By comparing the number of wake words on each path with the number of speech frames for the wake word, if the number of wake words on the path is greater than the number of speech frames, the maximum posterior score corresponding to the posterior probability smoothing value of the path is taken as the wake-up threshold.

[0075] Step S5: If the wake-up threshold is greater than or equal to the preset confidence level, then perform voice wake-up.

[0076] In a specific implementation, the confidence level is a preset value that can be obtained by running the test set or waking up the model in advance.

[0077] The voice wake-up decoding method of this invention sets N different smoothing window length parameters corresponding to different speech rates, preventing the decline in recognition rate caused by excessively fast or slow speech rates or unreasonable window lengths. Simultaneously, it divides the maximum window in existing technologies into M sub-windows and reduces computational load and storage resources by employing a method based on the maximum smoothing value of the sub-windows. By obtaining different wake-up thresholds at different speech rates, it can effectively compare with preset confidence levels. Compared with traditional methods, this method reduces computational load and solves the problem of inconsistent classification unit durations, thereby improving the product's wake-up rate.

[0078] Figure 2The block diagram of a voice wake-up decoding device according to a preferred embodiment of the present invention includes a voice acquisition module 100, a neural network module 200, a posterior probability smoothing module 300, a sub-window-based maximum value calculation module 400, a wake-up threshold calculation module 500, and a wake-up judgment module 600. The voice acquisition module 100 is used to receive audio data; the neural network module 200 is used to calculate the posterior probability value of a voice frame in the audio data; the posterior probability smoothing module 300 is used to smooth the posterior probability value of the voice frame using N smoothing windows to obtain N smoothed posterior probability values ​​of the voice frame, wherein the N smoothing windows have different window lengths, and different window lengths correspond to different speech rates, and N is a positive integer greater than or equal to 3; the sub-window-based maximum value calculation module 400 is used to apply the maximum value calculation to each of the N smoothed posterior probability values ​​using M sub-windows. The window calculates M posterior scores for the speech frame. The M sub-windows are obtained by equally dividing the maximum window. The maximum value among the M posterior scores is the maximum posterior score of the speech frame under the corresponding sub-window, resulting in M ​​maximum posterior scores, where M is a positive integer greater than or equal to 3. The wake-up threshold calculation module 500 compares the N posterior probability smoothing values ​​of each speech frame pairwise to obtain the maximum value of the N posterior probability smoothing values, thereby obtaining the paths of the N current wake-up words. The module compares the sum of the number of wake-up words on each path with the duration of the wake-up words. If the sum of the number of wake-up words on the path is greater than the duration, the maximum posterior score corresponding to the posterior probability smoothing value of the path is the wake-up threshold. The wake-up judgment module 600 outputs a wake-up signal and performs voice wake-up when the wake-up threshold is greater than or equal to a preset confidence level.

[0079] In one implementation, such as Figure 3 As shown, the voice wake-up decoding device may further include a smooth window calculation module 700. The smooth window calculation module 700 includes an average duration calculation unit 701, a base value calculation unit 702, and a smooth window length calculation unit 703. The average duration calculation unit 701 is used to calculate the duration of the neural network classification units for the first speech rate, the second speech rate, ..., the Nth speech rate based on the pre-collected speech data, and calculate the average duration of the N speech rates respectively. The base value calculation unit 702 is used to calculate the duration of the current character based on the character classification units in the speech data, take the duration of the shortest character as the base value, and calculate the multiple relationship between the duration of other characters in the speech data and the base value. The smooth window length calculation unit 703 is used to weight the multiple relationship into the average duration of the N speech rates to obtain the N smooth windows with different window lengths.

[0080] In one implementation, such as Figure 4As shown, the wake-up threshold calculation module 500 may include a maximum value comparison unit 501 and a wake-up word path calculation unit 502. The maximum value comparison unit 501 is used to compare the N posterior probability smoothing values ​​p0, p1...p(N-1) of each speech frame pairwise to obtain the maximum value of the N posterior probability smoothing values. The wake-up word path calculation module 502 is used to obtain the path of the N current wake-up words according to the index number of the maximum value, and the path corresponds to the N smoothing windows.

[0081] The present invention also provides a voice processing chip capable of executing the voice wake-up decoding method of the present invention.

[0082] The present invention also provides a voice wake-up device, which includes the voice wake-up decoding device of the present invention or the voice processing chip of the present invention.

[0083] The present invention also provides a storage medium storing a program, wherein the program is used to be executed to implement the voice wake-up decoding method of the present invention.

[0084] It should be noted that the use of step numbers (letters or numbers) to refer to certain specific method steps in this invention is merely for the purpose of convenience and brevity in description, and is by no means intended to restrict the order of these method steps. Those skilled in the art will understand that the order of the relevant method steps should be determined by the technology itself and should not be unduly restricted by the existence of step numbers.

[0085] Those skilled in the art will understand that, without conflict, the above-mentioned preferred solutions can be freely combined and superimposed.

[0086] It should be understood that the above embodiments are merely exemplary and not restrictive. Various obvious or equivalent modifications or substitutions that can be made by those skilled in the art regarding the above details without departing from the basic principles of the present invention will be included within the scope of the claims of the present invention.

Claims

1. A voice wake-up decoding method, characterized in that, include: S1: Receive audio data and calculate the posterior probability value of the speech frame through a neural network; S2: The posterior probability value of the speech frame is smoothed using N smoothing windows to obtain N smoothed posterior probability values ​​of the speech frame. The N smoothing windows have different window lengths, and different window lengths correspond to different speech rates. N is a positive integer greater than or equal to 3. S3: Using M sub-windows, calculate the M posterior scores of the speech frame one by one for the N posterior probability smoothing values. The M sub-windows are obtained by dividing the maximum window equally. The maximum value among the M posterior scores is the maximum posterior score of the speech frame under the corresponding sub-window, and the M maximum posterior scores are obtained. M is a positive integer greater than or equal to 3. S4: Compare the pairwise N posterior probability smoothing values ​​of each speech frame to obtain the maximum value of the N posterior probability smoothing values, so as to obtain the path of the N current wake words. Compare the sum of the number of wake words on each path with the number of speech frames of the wake words. If the sum of the number of wake words on the path is greater than the number of speech frames, the posterior score of the maximum value of the posterior probability smoothing value corresponding to the path is the wake-up threshold. S5: If the wake-up threshold is greater than or equal to the preset confidence level, then perform voice wake-up.

2. The voice wake-up decoding method according to claim 1, characterized in that, Before S2, it also includes: Based on the pre-collected speech data, the duration of the neural network classification units for the first speech rate, the second speech rate, ... the Nth speech rate are statistically analyzed, and the average duration of each of the N speech rates is calculated. Based on the character classification units in the speech data, the duration of the current character is calculated, and the duration of the shortest character is used as the base value. Calculate the multiple relationship between the duration of other words in the speech data and the basic value; The multiple relationship is weighted and added to the average duration of the N speech rates to obtain the N smooth windows with different window lengths.

3. The voice wake-up decoding method according to claim 1, characterized in that, In step S4, the pairwise comparison of the N posterior probability smoothing values ​​for each speech frame yields the maximum value of the N posterior probability smoothing values, thus obtaining the paths for the N current wake-up words, including: S41: The N posterior probability smoothing values ​​p0, p1...p(N-1) of the speech frame are compared pairwise to obtain the maximum value of the N posterior probability smoothing values; S42: Obtain the index number of the maximum value; S43: Repeat steps S41 and S42; S44: Obtain the paths of the N current wake words, where each path corresponds to one of the N smoothing windows.

4. A voice wake-up decoding device, characterized in that, It includes a voice acquisition module, a neural network module, a posterior probability smoothing module, a sub-window-based maximum value calculation module, a wake-up threshold calculation module, and a wake-up judgment module. The voice acquisition module is used to receive audio data; The neural network module is used to calculate the posterior probability value of the speech frame in the audio data; The posterior probability smoothing module is used to smooth the posterior probability value of the speech frame using N smoothing windows to obtain N smoothed posterior probability values ​​of the speech frame. The N smoothing windows have different window lengths, and different window lengths correspond to different speech rates. N is a positive integer greater than or equal to 3. The sub-window-based maximum value calculation module is used to calculate the M posterior scores of the speech frame by applying the M sub-windows to the N posterior probability smoothing values ​​one by one. The M sub-windows are obtained by dividing the maximum window equally. The maximum value among the M posterior scores is the maximum posterior score of the speech frame under the corresponding sub-window, and the M maximum posterior scores are obtained. M is a positive integer greater than or equal to 3. The wake-up threshold calculation module is used to compare the pairwise N posterior probability smoothing values ​​of each speech frame to obtain the maximum value of the N posterior probability smoothing values, so as to obtain the path of the N current wake-up words, and compare the sum of the number of wake-up words on each path with the number of speech frames of the wake-up words. If the sum of the number of wake-up words on the path is greater than the number of speech frames, the posterior score of the maximum value of the posterior probability smoothing value corresponding to the path is the wake-up threshold. The wake-up determination module is used to perform voice wake-up when the wake-up threshold is greater than or equal to a preset confidence level.

5. The voice wake-up decoding device according to claim 4, characterized in that, It also includes a smoothing window calculation module, comprising a mean duration calculation unit, a base value calculation unit, and a smoothing window length calculation unit. The mean duration calculation unit is used to calculate the duration of the neural network classification units for the first speech rate, the second speech rate, ... the Nth speech rate based on the pre-collected speech data, and to calculate the mean duration of the N speech rates respectively. The basic value calculation unit is used to calculate the duration of the current character based on the character classification unit in the speech data, take the duration of the shortest character as the basic value, and calculate the multiple relationship between the duration of other characters in the speech data and the basic value. The smooth window length calculation unit is used to weight the multiple relationship into the average duration of the N speech rates to obtain the N smooth windows with different window lengths.

6. The voice wake-up decoding device according to claim 4, characterized in that, The wake-up threshold calculation module includes a maximum value comparison unit and a wake-up word path calculation unit. The maximum value comparison unit is used to compare the N posterior probability smoothing values ​​p0, p1...p(N-1) of each speech frame in pairs to obtain the maximum value of the N posterior probability smoothing values. The wake word path calculation unit is used to obtain the paths of the N current wake words according to the index number of the maximum value, and the paths correspond to the N smooth windows.

7. A voice processing chip, characterized in that, Perform the voice wake-up decoding method as described in any one of claims 1-3.

8. A voice wake-up device, characterized in that, Includes the voice wake-up decoding device as described in any one of claims 4-6, or the voice processing chip as described in claim 7.

9. A storage medium, characterized in that, The storage medium stores a program, wherein the program is executed to implement the voice wake-up decoding method as described in any one of claims 1-3.

Images