CN106601229A - Voice awakening method based on soc chip - Google Patents

Abstract

The invention discloses a voice awakening method based on an soc chip. The voice awakening method comprises the steps that: S1, acquiring voice data and sampling the voice data by means of the chip, and converting an analog signal into a digital signal; S2, carrying out MFCC feature extraction on voice data of the digital signal; S3, conducting voice activity detection on MFCC feature values, judging whether a new frame of MFCC data of the current MFCC feature value is a voice frame, if not, jumping to the step S2 and releasing the data, and if so, subjecting the MFCC feature values to processing at the next step; S4, recognizing the MFCC feature values by adopting a voice recognition algorithm based on an HMM model, and awakening control equipment if a recognition result is an effective instruction, otherwise jumping to the step S2. According to the voice awakening method provided by the invention, a real-time system implemented by adopting the algorithm with high robustness has high recognition rate, and achieves the requirements for low power consumption and high performance.

Description

A voice wake-up method based on soc chip

technical field

The invention relates to the technical field of voice recognition, in particular to a voice wake-up method based on a SoC chip.

Background technique

With the development of the times, more and more electronic devices have entered people's daily life. While enjoying the convenience brought by electronic devices, people hope that electronic devices can be more intelligent and realize non-touch interaction.

Voice wake-up, that is, the user speaks the set voice command, so that the device in the dormant state directly enters the state of waiting for the command. Through this technology, anyone can directly speak the preset wake-up word to the device in any environment and at any time to activate the device, thereby achieving low power consumption and touch-free interaction.

However, most of the currently emerging voice wake-up technologies are implemented based on computers and mobile phone terminals, which require powerful processors for support and are not suitable for industrial applications. Although the voice wake-up technology implemented based on the MCU is low in cost, it cannot achieve the desired effect due to the limitation of the processor performance.

Contents of the invention

The technical problem to be solved by the present invention is to provide a voice wake-up method based on a soc chip, and the real-time system realized by using a robust algorithm has a high recognition rate and meets the requirements of low power consumption and high performance.

In order to solve the problems of the technologies described above, the present invention provides the following technical solutions: a voice wake-up method based on a soc chip, comprising the following steps:

S1. The chip collects voice data, samples it, and converts the analog signal into a digital signal;

S2, performing MFCC feature extraction on the voice data of the digital signal;

S3, carry out voice activity detection to MFCC characteristic value, judge whether the new frame MFCC data of current MFCC characteristic value is a voice frame, if otherwise return to step S2 and release data, if then enter next step processing with MFCC characteristic value;

S4. Recognize the MFCC characteristic value through the speech recognition algorithm based on the HMM model. If the recognition result is a valid command, wake up the control device; otherwise, return to step S2.

Further, the MFCC feature extraction in the step S2 is specifically:

1), digital signal preprocessing, including pre-emphasis, framing and windowing;

2), carry out FFT transformation to each frame signal, obtain frequency spectrum, and then obtain amplitude spectrum |X _n (k)|;

3), add Mel filter bank W _l (k) to amplitude spectrum |X _n (k)|, the formula is as follows:

Where k refers to the kth point of FFT; o(l), c(l), h(l) are the lower limit frequency, center frequency and upper limit frequency of the lth triangular filter respectively;

4), perform logarithmic operation on all filter outputs, and then further perform discrete cosine transform DCT to obtain MFCC eigenvalues, the formula is as follows:

Among them, N and L are 26, which refers to the number of filters; i refers to the order of MFCC coefficients, and i takes 12, which means that 12 cepstrum features are obtained; in addition, the logarithmic energy of one frame is added as the 13th feature parameters, defined as follows:

Among them, X _n (k) is the magnitude, from which 13 characteristic parameters can be obtained, including 12 cepstrum features plus 1 logarithmic energy;

5), the cepstrum parameter MFCC of 13 standards that seeks has only reflected the static characteristic of speech parameter, and the dynamic characteristic of speech is described according to the differential spectrum of described static feature; Calculate the first-order difference dtm(i of 13 MFCC features ) and second-order difference dtmm(i):

The 13 standard MFCC features and its 13 first-order differences and 13 second-order differences form 39-dimensional MFCC feature parameters. So far, the MFCC feature extraction is complete.

Further, in the step S3, the voice activity detection is performed on the feature value, and the voice activity detection method based on the GMM model is adopted, which assumes that the voice and background noise conform to the Gaussian mixture distribution in a specific feature space, and respectively construct silence in the feature space model, non-muted model; then calculate the new frame of MFCC data of MFCC features, and calculate the likelihood value P1 of the silent model and the likelihood value P2 of the non-muted model respectively; compare the size of the likelihood value P1 and the likelihood value P2 , if P1 is greater than P2, the current MFCC data frame is a speech frame, otherwise a mute frame.

Further, if the current MFCC data frame is judged as a speech frame, when judging the next frame of MFCC data frame, the likelihood value P1 and the likelihood value P2 are multiplied by the corresponding transition probability respectively, and the two product results are compared, if If the product result of the likelihood value P1 is greater than the product result of the likelihood value P2, the current MFCC data frame is a voice frame, otherwise it is a silent frame;

If after the current MFCC data frame is judged as a silent frame, when judging the next frame of MFCC data frame, the likelihood value P1 and the likelihood value P2 are respectively multiplied by the corresponding transition probability, and the two product results are compared, if the likelihood value If the product result of P1 is greater than the product result of the likelihood value P2, the current MFCC data frame is a speech frame, otherwise it is a silent frame;

The corresponding transition probability is preset model data.

Further, the calculation method of the likelihood value P1 of the silent model and the likelihood value P2 of the non-silent model is specifically:

Among them, the silent model and the non-silent model are composed of 13 39-dimensional Gaussian models; the probability density function of an M-order Gaussian model is obtained by the weighted sum of M Gaussian probability density functions, as shown in the following formula 3.1:

In the formula, M is the number of multidimensional Gaussian models, and M is 13; X is a D-dimensional random vector, which is a 39-dimensional MFCC feature value; b _i (X) is a sub-distribution, and ω _i is a mixing weight; each sub-distribution is The D-dimensional joint Gaussian probability distribution is as follows: 3.2:

Among them, μ _i is the mean value of the i-th dimension; σ _i ² is the variance; x _i is the input MFCC feature value of the i-th dimension; D represents the total dimension, and D is 39;

Since the calculation of formula 3.2 is too complicated, its derivation is simplified:

Take the logarithm on both sides of the formula to get:

It can be seen that the left side of the plus sign is the known parameters in the trained model, which can be trained in advance, so let gconst be used as a parameter of the model:

So formula 3.2 is transformed into the following formula:

Then Equation 3.1 is simplified to:

Putting the MFCC data frame and model parameters into the above formula, the likelihood value of the mute model and the likelihood value of the non-mutation model of the frame data can be obtained.

Further, the MFCC data frame and model parameters will be brought into the above formula to obtain the likelihood value of the silent model of the frame data and the likelihood value of the non-silent model, and the specific steps are:

1) Match and calculate the MFCC eigenvalues of each frame of speech with the silent model and the non-silent model, first perform ( _xi -μ _i ) ² /σ ² calculations, and accumulate the calculation results to obtain the multidimensional model of the two models Exponential parts fa0 and fa1 of a Gaussian distribution:

where mean μ _i and variance Get it directly from the model data;

2), the calculation results of the previous step are calculated as follows, and the likelihood value b of the multidimensional Gaussian distribution can be obtained:

Among them, gconst is the pre-trained data, directly obtained from the model data, so far the calculation of the multidimensional Gaussian distribution likelihood value ln b _i (X) in formula 3.3 is completed;

3) From the above, it can be seen that the silent model and the non-silent model respectively contain 13 multidimensional Gaussian distributions, so the likelihood values ln b _i (X) of 13 multidimensional Gaussian distributions can be obtained after steps 1 and 2 are cycled 13 times, and these The likelihood value and the corresponding weight ω _i are brought into the following formula to obtain the likelihood value P ₁ of the current frame for the silent model and the likelihood value P ₂ for the non-silent model:

Further, the step S4 is based on the speech recognition algorithm of the HMM model, which is specifically:

S41, load the HMM model, construct the recognition network of the HMM chain;

S42. Match the MFCC eigenvalues with the recognition network of the HMM model, and calculate the initial likelihood value;

S43. According to the initial likelihood value, the Token Passing algorithm finds the best path in the HMM chain network, and completes the decoding work;

S45. Determine whether the voice instruction matches the HMM chain, if yes, it is valid voice, otherwise, it is invalid voice.

After adopting the above technical solution, the present invention has at least the following beneficial effects:

(1) The present invention converts a large number of multiplication operations into addition operations by converting part of the calculations in the original algorithm to the log domain, and successfully reduces the time delay when it is implemented on a microprocessor; the complex calculation of the algorithm is accelerated by special hardware , reduce the delay, and finally achieve the purpose of real-time recognition;

(2) the present invention has higher recognition rate by adopting the real-time system realized by the algorithm with high robustness;

(3) The present invention is easy to upgrade. The algorithm of the present invention is divided into three independent modules: feature extraction, voice activity detection and voice recognition. Subsequent algorithms with better performance can be used to optimize the system by replacing sub-modules .

Description of drawings

Fig. 1 is the overall flowchart of a kind of voice wake-up method based on soc chip of the present invention;

Fig. 2 is the triangular filter schematic diagram of a kind of voice wake-up method based on soc chip of the present invention;

Fig. 3 is a triangular filter bank schematic diagram of a voice wake-up method based on a soc chip of the present invention;

Fig. 4 is the voice activity detection flowchart of a kind of voice wake-up method based on soc chip of the present invention;

Fig. 5 is the parameter composition schematic diagram of the 39-dimensional Gaussian model of a kind of voice wake-up method based on soc chip of the present invention;

Fig. 6 is a flow chart of voice activity detection steps of a voice wake-up method based on a soc chip of the present invention;

Fig. 7 is a schematic diagram of model data pre-trained in voice activity detection of a method for voice wake-up based on a soc chip of the present invention;

Fig. 8 is a flow chart of the overall steps of the voice recognition algorithm of a voice wake-up method based on a soc chip of the present invention;

FIG. 9 is a schematic diagram of an HMM chain in an example of a voice recognition algorithm based on a soc chip in the present invention.

detailed description

It should be noted that, in the case of no conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The present application will be further described in detail below in conjunction with the drawings and specific embodiments.

Fig. 1 is the overall algorithm flow chart of the present invention, and wherein each module calculation process is as follows:

1. Voice front-end processing:

Voice front-end processing is to convert the analog signal into a digital signal by sampling the voice data signal. In this solution, the sampling rate is 16K. Voice digital signal is PCM (Pulse Code Modulation)

Format, that is, pulse code modulation, which samples the sound analog signal to obtain quantized speech data, is the most basic and original speech format. In the present invention, the ADC module is integrated in the soc chip, and a voice detection process is performed every 10 ms. The sampling frequency is 16K data collected per second, and the data bit width is 16 bits.

2. MFCC feature data extraction:

1) Signal preprocessing, including Preemphasis, Frame Blocking, and Windowing; the sampling frequency fs of the voice signal is 16KHz, and since the voice signal is considered stable at 10-30ms, set Each frame is 10ms, so the frame length is 160 points; the frame shift is 1/2 of the frame length, which is 80;

2) Carry out FFT transformation of 256 points for each frame, obtain the frequency spectrum, and then obtain the amplitude spectrum |X _n (k)|;

3) Add Mel filter bank W _l (k) to the amplitude spectrum |X _n (k)|, the formula is as follows:

Wherein k refers to the kth point of FFT; o(l), c(l), h(l) are the lower limit frequency, center frequency and upper limit frequency of the l triangular filter, as shown in Figure 2;

In the present invention, the Mel filter bank is composed of 26 triangular filters, and the parameters are calculated in advance. The triangular filter bank is shown in Figure 3. The abscissa corresponds to the point in the FFT, and the ordinate is W _l (k). Since it is symmetrical, only half of the points in front of the FFT are taken to calculate the spectrum, and then added to the triangular filter;

4) Perform logarithmic operation (Logarlithm) on all filter outputs, and then further perform discrete cosine transform DCT to obtain MFCC, the formula is as follows.

Among them, N and L are 26, referring to the number of filters; i refers to the order of MFCC coefficients, and the present invention takes 12, that is, 12 cepstrum features are obtained; in addition, the logarithmic energy of one frame is added as the 13th characteristic parameter , defined as follows:

From this, 13 characteristic parameters (12 cepstrum features plus 1 logarithmic energy) can be obtained;

5), the cepstrum parameter MFCC of these 13 standards only reflects the static characteristic of speech parameter, the dynamic characteristic of speech can be described by the differential spectrum of these static features; Calculate the first-order difference dtm(i) of 13 MFCC features and Second order difference dtmm(i):

The 13 standard MFCC features and its 13 first-order differences and 13 second-order differences form 39-dimensional MFCC feature parameters. So far, the MFCC feature extraction is complete.

3. Voice Activity Detection (VAD):

The present invention adopts the sound activity detection method based on GMM model, and this method assumes voice and background noise meet Gaussian mixture distribution in specific feature space, builds their GMM model respectively in feature space, then uses the method for model matching in being Effective speech segments are detected in the test signal; the algorithm flow is as shown in Figure 4:

The model is trained in advance through the HTK toolbox. A 39-dimensional Gaussian model consists of 1 weight (MIXTURE), 39 means (MEAN), 39 variances (VARIANCE) and 1 gconst, as shown in Figure 5:

The silent model and the non-silent model are respectively composed of 13 multidimensional Gaussian models as shown in Figure 5; when a new frame of speech data is collected into the system, the new frame of 39-dimensional MFCC eigenvalues are respectively compared with the silent and non-silent models Carry out the likelihood value calculation, compare the size of the two likelihood values, the model with the larger likelihood value is the matching model of the current frame, so as to judge whether the current frame is a speech frame, and the detailed processing flow of VAD is shown in Figure 6:

Among them, the transfer coefficients a ₁₁ , a ₁₂ , a ₂₁ , and a ₂₂ are pre-trained model data, as shown in Figure 7, a ₁₁ is the transition probability that the previous frame is a silent frame and the current frame is also a silent frame; a ₁₂ is The previous frame is a silent frame, but the current frame is the transition probability of a speech frame; a ₂₁ is the transition probability that the previous frame is a speech frame, but the current frame is a silent frame; a ₂₂ is a speech frame for the previous frame, and the current frame is also The transition probability of the speech frame;

The most complicated calculation in the whole process is the calculation of the likelihood value. The calculation of the likelihood value is introduced below:

The probability density function of the 13th-order multidimensional Gaussian mixture model is obtained by the weighted sum of 13 multidimensional Gaussian probability density functions, as shown in the following formula 3.1:

In the formula, M is the number of multidimensional Gaussian models, which is 13 in the present invention; X is a D-dimensional random vector (that is, the 39-dimensional MFCC eigenvalue mentioned above), b _i (X) is a sub-distribution, and ω _i is a mixed Weights. Each subdistribution is a D-dimensional joint Gaussian probability distribution, as follows:

For 1-dimensional, μ is the expectation, and σ ² is the variance; for multi-dimensional, D represents the dimension of X, and represents the covariance matrix of D*D, which is defined as ∑=E[(x-μ)(x-μ ) ^T ], |∑| is the value of the determinant of the covariance;

Therefore, the specific calculation steps of the VAD algorithm are as follows:

1) Match and calculate the 39-dimensional MFCC eigenvalues of each frame of speech with the silent and non-silent models, first perform (X-μ _i ) ² /σ ² calculations, and accumulate the 39 results to obtain two models The exponential parts fa0 and fa1 of the multidimensional Gaussian distribution of (this calculation is done by the hardware acceleration IP):

where mean μ _i and variance Get it directly from the model data;

2) Perform the following calculation on the results of the previous step to obtain the likelihood value of the multidimensional Gaussian distribution:

b=exp(fa0)

Among them, gconst is the pre-trained data, which is directly obtained from the model data. So far, the calculation of the likelihood value of the multidimensional Gaussian distribution in formula 3.2 is completed;

3) From the above, it can be seen that the silent model and the non-silent model contain 13 multidimensional Gaussian distributions respectively, so the likelihood values of 13 multidimensional Gaussian distributions can be obtained after steps 1 and 2 are cycled 13 times, and these likelihood values are multiplied by the model weight And adding them together is Equation 3.1 to get the likelihood values of the silent model and the non-silent model; therefore, after 13 cycles of steps 1 and 2, the likelihood values ln b _i (X) of 13 multidimensional Gaussian distributions can be obtained, and these likelihoods The likelihood value and the corresponding weight ω _i are brought into the following formula, and the likelihood value P ₁ of the current frame to the silent model and the likelihood value P ₂ to the non-silent model can be obtained:

4) Finally multiply the transition probability a:

If the previous frame data is a speech frame, then the probability that the current frame is a speech frame= _a22 *P2 _;

Probability that the current frame is a silent frame=a ₂₁ *P ₁ ;

If the previous frame data is a silent frame, then the probability that the current frame is a voice frame=a ₁₂ *P ₂ ;

Probability that the current frame is a silent frame=a ₁₁ *P ₁ ;

Compare the probability of a speech frame with the probability of a silent frame. If the probability of a speech frame is high, the current frame is considered a speech frame, otherwise it is a silent frame. So far, the VAD algorithm is completed.

4. Speech recognition algorithm:

The flow of this module is shown in Figure 8. The loading of the model and the construction of the HMM chain are completed at the initial initialization of the program, and there is no need to repeat it later; when the upper-level VAD module detects a valid voice, it enters this module for calculation. Each state of the HMM model called by this module is composed of 24 GMMs. The process is as follows:

(1) Load the HMM model and construct the recognition network of the HMM chain;

(2), match the MFCC eigenvalue with the recognition network of the HMM model, and calculate the initial likelihood value;

(3) According to the initial likelihood value, the Token Passing algorithm finds the best path in the HMM chain network and completes the decoding work;

(4), judging whether the voice command matches the HMM chain, if so, it is a valid voice, otherwise it is an invalid voice.

The whole process is described below: taking "shutdown" as an example, the following is the HMM chain corresponding to "shutdown" (the actual HMM chain is longer, and each syllable is composed of multiple states, which are simplified here for the convenience of explanation). "Shutdown" can be divided into syllables "g", "uan", "j" and "i", and the HMM model is used to describe the 4 syllables into 4 states, and connect them to obtain the HMM chain, as shown in Figure 9;

A. Initialize the token value P _g =0 at the starting point of this network (namely "g");

B. When the first frame of MFCC data arrives, token-passing starts. The first frame only has the token value P _g , and the token value P _g will be passed to the state "g" and "uan", specifically as follows:

P _g ＝P _g +a ₁₁ +log(GMM _g )

P _uan =P _g +a ₁₂ +log(GMM _uan )

log(GMM _g ) is the likelihood value of MFCC data to state “g”, log(GMM _uan ) is the likelihood value of MFCC data to state “uan”, the calculation method of likelihood value is consistent with vad, see formula 3.3 and 3.4.

C. When the second frame of data arrives, the states "g" and "uan" both have token values, so the tokens are passed to the state connected to these two states.

Token value update for state "g":

P _g ＝P _g +a ₁₁ +log(GMM _g )

Token value update for state "uan":

P _g→uan =P _g +a ₁₂

P _uan→uan ＝P _uan +a ₂₂

After update: P _uan = max(P _g→uan ，P _uan→uan )+log(GMM _uan )

Since the left side of the state "uan" is connected to "g", and at the same time, it is connected to itself, so two token values will be obtained. At this time, the two token values must be compared, and the larger one should be selected and kept.

Make an update to the token in state "j"

P _j ＝P _uan +a ₂₃ +log(GMM _j )

D. When the third frame arrives, update the token value of state "g":

P _g ＝P _g +a ₁₁ +log(GMM _g )

Token value update for state "uan":

P _g→uan =P _g +a ₁₂

P _uan→uan ＝P _uan +a ₂₂

After update: P _uan = max(P _g→uan ，P _uan→uan )+log(GMM _uan )

Make an update to the token in state "j"

P _uan→j ＝P _uan +a ₂₃

P _j→j ＝P _j +a ₃₃

After update: P _j = max(P _uan→j ，P _j→j )+log(GMM _j )

To update the token for state "i":

P _i =P _j +a ₃₄ +log(GMM _i )

E. When the fourth frame arrives, update the token value of state "g":

P _g ＝P _g +a ₁₁ +log(GMM _g )

Token value update for state "uan":

P _g→uan =P _g +a ₁₂

P _uan→uan ＝P _uan +a ₂₂

After update: P _uan = max(P _g→uan ，P _uan→uan )+log(GMM _uan )

Make an update to the token in state "j"

P _uan→j ＝P _uan +a ₂₃

P _j→j ＝P _j +a ₃₃

After update: P _j = max(P _uan→j ，P _j→j )+log(GMM _j )

To update the token for state "i":

P _j→i = P _j +a ₃₄

P _i→i =P _i +a ₄₄

After update: P _i ＝max(P _j→i ，P _i→i )+log(GMM _i )

So far, all voice instruction frames have been input, start token comparison, and sort the token values of the four states. If the token value of the last state of the HMM chain (that is, "i") is the largest, it means that the input voice The instruction matches the HMM chain of "shutdown", and the decoding result is "shutdown". Otherwise, it is considered that the input is an invalid voice.

The whole decoding process can be seen that as the number of frames increases, tokens spread from the left end to the right end. In this process, each state has a token, and the token will be passed to the adjacent state and calculated. When the specified The number of frames (the number of frames is determined by the length of the preset voice command, such as "shutdown" is shorter, and the number of frames of "Open Sesame" is longer due to the longer voice), sort the tokens of all states, If the token value on the end state of the HMM chain is the largest, it means that the input voice matches this HMM chain. In practical applications, the number of recognizable voice commands can be increased. At this time, there will be multiple HMM chains. In this way, in the last frame, all states of all HMM chains will be sorted to determine which command it is.

While embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various equivalents can be made to these embodiments without departing from the principles and spirit of the invention. Changes, modifications, substitutions and variations, the scope of the present invention is defined by the appended claims and their equivalents.

Claims (7)

1. a voice wake-up method based on soc chip, is characterized in that, comprises the following steps: S1. The chip collects voice data, samples it, and converts the analog signal into a digital signal; S2, performing MFCC feature extraction on the voice data of the digital signal; S3, carry out voice activity detection to MFCC characteristic value, judge whether the new frame MFCC data of current MFCC characteristic value is a voice frame, if otherwise return to step S2 and release data, if then enter next step processing with MFCC characteristic value; S4. Recognize the MFCC characteristic value through the speech recognition algorithm based on the HMM model. If the recognition result is a valid command, wake up the control device; otherwise, return to step S2. 2. a kind of voice wake-up method based on soc chip as claimed in claim 1, is characterized in that, in the described step S2, MFCC feature extraction, it is specially: 1), digital signal preprocessing, including pre-emphasis, framing and windowing; 2), carry out FFT transformation to each frame signal, obtain frequency spectrum, and then obtain amplitude spectrum |X _n (k)|; 3), add Mel filter bank W _l (k) to amplitude spectrum |X _n (k)|, the formula is as follows: $\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; o</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; )</span>}{\overset{\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; l</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; ...</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 26</span>}$ ${\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; l</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{c}\frac{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; o</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; o</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; o</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; )</span>\\ \frac{\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; )</span>\end{array}\right.$ Where k refers to the kth point of FFT; o(l), c(l), h(l) are the lower limit frequency, center frequency and upper limit frequency of the lth triangular filter respectively; 4), perform logarithmic operation on all filter outputs, and then further perform discrete cosine transform DCT to obtain MFCC eigenvalues, the formula is as follows: $\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\sqrt{\frac{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; L</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; log</span>}\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; \{</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>\frac{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; \&pi;</span>}}{\mathrm{<span\; class="notranslate">\; L</span>}}<span\; class="notranslate">\; \}</span>$ Among them, N and L are 26, which refers to the number of filters; i refers to the order of MFCC coefficients, and i takes 12, that is, 12 cepstrum features are obtained; in addition, the logarithmic energy of one frame is added as the 13th feature parameters, defined as follows: $\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 13</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 10</span>}\mathrm{<span\; class="notranslate">\; lg</span>}\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 256</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$ Among them, X _n (k) is the magnitude, from which 13 characteristic parameters can be obtained, including 12 cepstrum features plus 1 logarithmic energy; 5), the cepstrum parameter MFCC of 13 standards that seeks has only reflected the static characteristic of speech parameter, and the dynamic characteristic of speech is described according to the differential spectrum of described static feature; Calculate the first-order difference dtm(i of 13 MFCC features ) and second-order difference dtmm(i): $\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; 3</span>}}$ $\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; 3</span>}}$ The 13 standard MFCC features and its 13 first-order differences and 13 second-order differences form 39-dimensional MFCC feature parameters. So far, the MFCC feature extraction is complete. 3. a kind of voice wake-up method based on soc chip as claimed in claim 1, it is characterized in that, in described step S3, feature value is carried out voice activity detection, adopts the voice activity detection method based on GMM model, it assumes voice and The background noise conforms to the Gaussian mixture distribution in a specific feature space, and the silent model and the non-silent model are respectively constructed in the feature space; then the new frame of MFCC data of the MFCC feature is calculated, and the likelihood value P1 of the silent model and the non-silent model are calculated respectively. The likelihood value P2 of the silence model; compare the likelihood value P1 and the likelihood value P2, if P1 is greater than P2, the current MFCC data frame is a speech frame, otherwise a mute frame. 4. a kind of voice wake-up method based on soc chip as claimed in claim 3, is characterized in that, if after described current MFCC data frame is judged as speech frame, when judging next frame MFCC data frame, likelihood value P1 Multiply the corresponding transition probability with the likelihood value P2 respectively, compare two product results, if the product result of the likelihood value P1 is greater than the product result of the likelihood value P2, then the current MFCC data frame is a speech frame, otherwise it is a silent frame; If after the current MFCC data frame is judged as a silent frame, when judging the next frame of MFCC data frame, the likelihood value P1 and the likelihood value P2 are respectively multiplied by the corresponding transition probability, and the two product results are compared, if the likelihood value If the product result of P1 is greater than the product result of the likelihood value P2, the current MFCC data frame is a speech frame, otherwise it is a silent frame; The corresponding transition probability is preset model data. 5. a kind of voice wake-up method based on soc chip as claimed in claim 3, is characterized in that, the calculation method of the likelihood value P1 of described mute model, the likelihood value P2 of non-quiet model, specifically: Among them, the silent model and the non-silent model are composed of 13 39-dimensional Gaussian models; the probability density function of an M-order Gaussian model is obtained by the weighted sum of M Gaussian probability density functions, as shown in the following formula 3.1: $\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; \&lambda;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 3.1</span>}$ In the formula, M is the number of multidimensional Gaussian models, and M is 13; X is a D-dimensional random vector, which is a 39-dimensional MFCC feature value; b _i (X) is a sub-distribution, and ω _i is a mixing weight; each sub-distribution is The D-dimensional joint Gaussian probability distribution is as follows: 3.2: ${\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&pi;</span>}<span\; class="notranslate">\; )</span>}^{\frac{\mathrm{<span\; class="notranslate">\; D.</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; \{</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; T</span>}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \}</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 3.2</span>}$ Among them, μ _i is the mean value of the i-th dimension; σ _i ² is the variance; x _i is the input MFCC feature value of the i-th dimension; D represents the total dimension, and D is 39; Since the calculation of formula 3.2 is too complicated, its derivation is simplified: ${\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&pi;</span>}<span\; class="notranslate">\; )</span>}^{\frac{\mathrm{<span\; class="notranslate">\; D.</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}{<span\; class="notranslate">\; \&Pi;</span>}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; 39</span>}}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; 39</span>}}\frac{{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; 2</span>}{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}}$ Take the logarithm on both sides of the formula to get: $\mathrm{<span\; class="notranslate">\; ln</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; \{</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; ln</span>}<span\; class="notranslate">\; \&lsqb;</span>{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&pi;</span>}<span\; class="notranslate">\; )</span>}^{\frac{\mathrm{<span\; class="notranslate">\; D.</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; |</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 39</span>}}{<span\; class="notranslate">\; \&Pi;</span>}}{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}{<span\; class="notranslate">\; |</span>}^{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; \&rsqb;</span><span\; class="notranslate">\; +</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 39</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}\frac{{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; \}</span>$ It can be seen that the left side of the plus sign is the known parameters in the trained model, which can be trained in advance, so let gconst be used as a parameter of the model: $\mathrm{<span\; class="notranslate">\; g</span>}\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; the\; s</span>}\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; ln</span>}<span\; class="notranslate">\; \&lsqb;</span>{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&pi;</span>}<span\; class="notranslate">\; )</span>}^{\frac{\mathrm{<span\; class="notranslate">\; D.</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; |</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 39</span>}}{<span\; class="notranslate">\; \&Pi;</span>}}{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}{<span\; class="notranslate">\; |</span>}^{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; \&rsqb;</span>$ So formula 3.2 is transformed into the following formula: $\mathrm{<span\; class="notranslate">\; ln</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; \&lsqb;</span>\mathrm{<span\; class="notranslate">\; g</span>}\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; the\; s</span>}\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; +</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 39</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}\frac{{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; \&rsqb;</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 3.3</span>}$ Then Equation 3.1 is simplified to: Putting the MFCC data frame and model parameters into the above formula, the likelihood value of the mute model and the likelihood value of the non-mutation model of the frame data can be obtained. 6. a kind of voice wake-up method based on soc chip as claimed in claim 5, it is characterized in that, described will bring MFCC data frame and model parameter in the above formula, can obtain the similarity of the mute model of this frame data The likelihood value and the likelihood value of the non-silent model, the specific steps are: 1) Match and calculate the MFCC eigenvalues of each frame of speech with the silent model and the non-silent model, first perform ( _xi -μ _i ) ² /σ ² calculations, and accumulate the calculation results to obtain the multidimensional model of the two models Exponential parts fa0 and fa1 of a Gaussian distribution: $\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 39</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}\frac{{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$ where mean μ _i and variance Get it directly from the model data; 2), the calculation results of the previous step are calculated as follows, and the likelihood value b of the multidimensional Gaussian distribution can be obtained: $\mathrm{<span\; class="notranslate">\; b</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; g</span>}\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; the\; s</span>}\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; )</span>$ Among them, gconst is the pre-trained data, directly obtained from the model data, so far the calculation of the multidimensional Gaussian distribution likelihood value ln b _i (X) in formula 3.3 is completed; 3) From the above, it can be seen that the silent model and the non-silent model respectively contain 13 multidimensional Gaussian distributions, so the likelihood values ln b _i (X) of 13 multidimensional Gaussian distributions can be obtained after steps 1 and 2 are cycled 13 times, and these The likelihood value and the corresponding weight ω _i are brought into the following formula to obtain the likelihood value P ₁ of the current frame for the silent model and the likelihood value P ₂ for the non-silent model: 7. a kind of voice wake-up method based on soc chip as claimed in claim 1, is characterized in that, described step S4 is based on the speech recognition algorithm of HMM model, and it is specially: S41, load the HMM model, construct the recognition network of the HMM chain; S42. Match the MFCC eigenvalues with the recognition network of the HMM model, and calculate the initial likelihood value; S43. According to the initial likelihood value, the Token Passing algorithm finds the best path in the HMM chain network, and completes the decoding work; S45. Determine whether the voice command matches the HMM chain, if so, it is a valid voice, otherwise, it is an invalid voice.