An artificial intelligence-based voice recognition method and system for a television stick

Abstract

The application provides a voice recognition method and system for a television stick based on artificial intelligence, and belongs to the field of voice recognition; solves the problem of low voice recognition efficiency of the television stick; and specifically comprises the following steps: S1: obtaining operation voice and calculating the MFCC coefficient sequence of each test voice; S2: calculating the LPCC coefficient sequence of each test voice, combining the MFCC coefficient sequence to construct a voice recognition model, and judging whether the operation voice exists in the model; if yes, the operation voice is converted into a television stick instruction; if not, historical application use data is obtained; and S3: analyzing the historical application use data of a target user and inversely reasoning the television stick instruction of the operation voice; the application improves the voice recognition and instruction generation capacity of the television stick by obtaining, analyzing and processing the language generated when a user operates the television stick.

Description

TECHNICAL FIELD

[0001] The application discloses an artificial intelligence-based voice recognition method and system for a TV stick, and relates to the voice recognition field. BACKGROUND

[0002] The existing voice recognition method or system for a TV stick has the following disadvantages:

[0003] Insufficient semantic understanding capability: most of the existing voice recognition systems can only recognize some simple voice instructions, such as searching for a program or switching a channel; and the recognition capability for some complex semantic expressions or ambiguous instructions needs to be improved.

[0004] Reasoning capability: the existing voice recognition systems lack logical reasoning capability and cannot infer the current intention of a user according to the use habits of the user for different applications; this will result in that the system cannot accurately understand the instructions of the user in some continuous dialogue scenarios.

[0005] Strong technical dependence: the existing voice recognition systems need to depend on stable network and device support to realize normal functions; meanwhile, the artificial intelligence-based voice recognition systems need to invest a large amount of research and development resources and funds to develop and deploy, and the cost is too high for some enterprises or developers. SUMMARY

[0006] In view of the disadvantages of the prior art, the application aims to provide an artificial intelligence-based voice recognition method and system for a TV stick, which aims to solve the problem of low voice recognition efficiency of the TV stick.

[0007] In order to achieve the above-mentioned purpose, the application is implemented by the following technical scheme: an artificial intelligence-based voice recognition method for a TV stick comprises the following steps:

[0008] Step S1: obtaining an operation voice of a target user and a test voice of a tester; performing a first processing of pre-emphasis, windowing and Fourier transform on the test voice, and calculating an MFCC coefficient sequence of each test voice;

[0009] Step S2: performing a second processing of homomorphism and cepstrum on the test voice, calculating an LPCC coefficient sequence of each test voice, and constructing a voice recognition model in combination with the MFCC coefficient sequence of each test voice; taking the operation voice as the output of the voice recognition model, and judging whether the voice recognition model has the output; if yes, converting the operation voice into a TV stick instruction and skipping step S3;

[0010] If not, obtaining the current time and historical application use data of the target user, and entering step S3.

[0011] Step S3: obtaining current time and controllable application quantity of the TV stick; analyzing historical application usage data of the target user to build an application usage model of the target user; obtaining application names that the target user is using or closing; and inversely inferring the TV stick instruction of the operation voice according to the application usage model.

[0012] Further, the specific steps of step S1 are as follows:

[0013] Step S11: counting the quantity vn of test voices; and counting the time length vo of each test voice (1) ~vo (vn) ; extracting the maximum value vo (1) ~vo (vn) and the minimum value vo (max) of vo (min) ; and calculating the average value vo (1) ~vo (vn) of vo (ave) ;

[0014] Step S12: defining the calculation formula A1:

[0015]

[0016] wherein vo (i) represents the time length of the ith test voice, and vo (ve) represents the weighted time length.

[0017] Obtaining the sampling rate fs of the test voice;

[0018] Calculating the frame length fd of a single test voice, fd = vo (ve) / fs;

[0019] Calculating the frame shift length fsd of a single test voice, fsd = fd / 2;

[0020] Step S13: performing enhancement processing on the first test voice to calculate the MFCC coefficient of the first test voice;

[0021] Step S14: repeating the same steps of processing the first test voice to process the second to va th test voices to obtain the MFCC coefficient sequence of the second to vb th test voices, and entering step S2.

[0022] Further, the specific steps of step S13 are as follows:

[0023] Step S131: taking the first test voice as the voice v; removing the low-frequency noise in the voice v to obtain the voice v (1) ; performing frame processing on the voice v (1) with fd and fsd to obtain the voice v (2) ;

[0024] voice v (2) number Fn, Fn=vo (1) / fd;

[0025] Step S132: windowing the voice v (2) using a Hamming window to obtain a windowed voice v (3) ;

[0026] converting the voice v (3) from time domain information to frequency domain information to obtain a voice v (4) ;

[0027] Step S133: obtaining the frequency f (4) ~f(Fn) of the voice v (1) corresponding to the first to Fn timestamps;

[0028] using a triangular filter to receive a non-linear frequency perception of the voice v (4) to obtain a voice v (5) ;

[0029] Step S134: converting the frequency of the voice v (4) corresponding to each timestamp to a Mel frequency, and the conversion formula is as follows:

[0030] wherein f (l) represents the frequency of the voice v (4) of the lth timestamp information; F mel (f (l) ) represents the perceived frequency of f (l) in Mel units; the value range of l is: 1~Fn;

[0031] summarizing F (mel) (f (1) )~F (mel) (f (Fn) ) as the voice v (5) ;

[0032] Step S135: performing a discrete cosine transform on the perceived frequency of the voice v (4) to obtain the MFCC coefficient of the first test voice;

[0033] The formula of the discrete cosine transform is as follows:

[0034]

[0035] wherein MFCC (l) represents the MFCC coefficient corresponding to the voice v (5) of the lth timestamp;

[0036] summarizing MFCC(1) MFCC (Fn) , as the MFCC coefficient sequence of the first test voice.

[0037] Further, the specific steps of step S2 are as follows:

[0038] Step S21: Perform homomorphic and cepstrum processing on the test voice to obtain the LPCC coefficient of the test voice information;

[0039] Step S22: Convert the first to the vn test voice information into test voice text, denoted as text wd (1) (vn) ;

[0040] According to the format of test voice text-test voice information MFCC coefficient sequence-test voice information LPCC coefficient sequence, the test voice information is summarized as a sample set;

[0041] Step S23: Take the CNN model as the basic model of the speech recognition model, and define the training steps for supervised training of the basic model:

[0042] Step S231: Take the first to the vn corresponding MFCC coefficient sequence as a set CMF;

[0043] Take the first to the vn corresponding LPCC coefficient sequence as a high-dimensional array, and integrate to obtain a set CLP;

[0044] Step S232: Fuse set Cmf and set Clp to obtain a set of television bar operation instruction voiceprint feature sequences, denoted as set CC;

[0045] Step S233: Fuse text wd (1) (vn) with set CC to obtain a feature set of the test voice, denoted as set a;

[0046] Set the state number in the basic model to 4 and the Gaussian element number to 3;

[0047] Step S234: Take set a into the basic model and iterate until the test voice corresponding to set a is output by the basic model;

[0048] Use cross-validation algorithm to prevent overfitting of the basic model in the process of supervised training;

[0049] Define the convergence condition of the speech recognition model: the probability of incorrect recognition of any test voice is less than 1x10 -3 .

[0050] Further, the subsequent steps of step S23 are as follows: ​​

[0051] Step S24: defining a training step of unsupervised training of the base model:

[0052] Step S241: obtaining a set CC, taking the set CC as a training set 1, taking a test speech text as a training set 2; using a K-means clustering algorithm to iterate the base model until the elements in the training set 1 and the elements in the training set 2 are output at the same time and the output elements correspond;

[0053] Step S242: using a cross-validation algorithm to prevent overfitting of the base model in the process of unsupervised training;

[0054] Defining the determination condition of convergence of the speech recognition model: the probability difference of the same test speech being output by the two consecutive recognitions is less than 1x10 -4 ;

[0055] Step S25: supervised training and unsupervised training of the base model; until the output results of the supervised training and the unsupervised training of the base model are the same, obtaining a speech recognition model;

[0056] Taking the operation speech as the input of the speech recognition model, judging whether the model exists output or not;

[0057] If it exists, converting the operation speech into a TV stick instruction and skipping step S3;

[0058] If it does not exist, obtaining the current time and the historical application use data of the target user, and entering step S3.

[0059] Further, the specific steps of step S21 are as follows:

[0060] Step S211: homomorphic and cepstrum processing of the first test speech information;

[0061] Obtaining the frequency domain information of the first test speech v (4) as information E; obtaining the frequency fx (1) ~ fx (Fn) corresponding to each information E;

[0062] Step S212: Z-transforming the information E corresponding to the first to Fn time stamps to obtain XZ (l) ~ XZ (Fn) ;

[0063] The function expression of Z-transform is:

[0064] Wherein, fx (l) represents the frequency of the information E corresponding to the lth time stamp; ZX (l) represents the Z-transform output of fx (l) ;

[0065] zz (l) representing fx (l) the corresponding complex variable:

[0066] vc represents the speed of sound, and ω represents fx (l) the corresponding phase:

[0067] ω = arctan[(vc x fx (l) ) / fd]; arctan represents the inverse tangent function; and fd represents the frame duration;

[0068] Step S213: Logarithmic conversion is performed on the information E corresponding to the first to Fn time stamps to obtain ln(ZX (1) )~ln(ZX (Fn) );

[0069] The function expression of the logarithmic conversion is:

[0070]

[0071] wherein ln(ZX (l) ) represents the output of the logarithmic conversion of ZX (l) ; fx (q) represents the frequency of the information E corresponding to the qth time stamp; zz (q) represents the complex variable corresponding to fx (q) ;

[0072] Inverse Z-transform processing is performed on ln(ZX (1) )~ln(ZX (Fn) ) to convert the information E corresponding to the first to Fn time stamps from frequency domain information to time domain information, to obtain ts (1) ~ts (Fn) .

[0073] Further, the subsequent steps of the step S213 are as follows:

[0074] Step S214: Cepstrum processing is performed on ts (1) ~ts (Fn) to obtain the LPC coefficients of the information E corresponding to the first to Fn time stamps, denoted as ce (1) ~ce (Fn) .

[0075] The function expression of the cepstrum processing is:

[0076]

[0077] wherein ts (m) represents the time domain information of the information E corresponding to the mth time stamp, fx(l-m+1) represents the frequency of the (l-m+1)th timestamp corresponding information E; m is in the range of 1 to l;

[0078] Using the discrete cosine transform algorithm, the ce (1) ~ ce (Fn) From the time domain information to the frequency domain information, the LPCC coefficient sequence of the first test voice information is obtained.

[0079] Step S215: Repeat the same step of homomorphism and cepstrum processing of the first test voice information, process the second to the va test voice information, and obtain the LPCC coefficient sequence of the second to the va test voice information.

[0080] Further, the specific steps of the step S3 are as follows:

[0081] Step S31: Get the current time ho;

[0082] Get the number of applications un that the TV stick can operate; Get the use time ti of the target user using the first application last month (1，1) ~ ti (td，1) ; Similarly, the use time ti of the unth application

[0083] (td，1) ~ ti (td，un) ; td represents the total number of days last month;

[0084] Step S32: Corresponding calculation formula B-1:

[0085] Wherein, ut (s) represents the time coefficient of the s th application, ti (s，t) represents the time of the target user using the t th application on the s th day;

[0086] Calculate the time coefficient ut (1) ~ ut (un) of the target user using the first to the unth application;

[0087] Step S33: The first to the unth application determination coefficient ha (1，1) ~ ha (td，un) from the first ho time to the td time;

[0088] The first to the unth application determination coefficient hb (1，1) ~ hb (td，un) from the first (ho+1) time to the td time;

[0089] Step S34: According to ha (1，1) ~ ha (td，un)and hb (1，1) ~ hb (td，un) ; analyzing the usage of the target user on the first to the un application, building an application usage model of the target user;

[0090] Step S35: Obtain the application name that the target user is using or closing as a quasi-application; according to the usage habits of the target user on different applications, infer the voice stick instruction of the television bar in reverse.

[0091] Further, the specific steps of step S34 are as follows:

[0092] Step S341: According to ha (1，1) ~ ha (td，un) , build the matrix HAA (1) ~ HAA (td) and the matrix HBB (1) ~ HBB (td) ;

[0093] Step S342: Stack the matrix HAA (1) and the matrix HBB (1) vertically to obtain the first day application change matrix HAB (1) ;

[0094] Similarly, stack the matrix HAA (td) and the matrix HBB (td) vertically to obtain the td-day application change matrix HAB (td) ;

[0095] Step S343: Analyze the usage habits of the first application in the time period from the ho time to the (ho+1) time of the target user, and calculate the usage rate uu (1) and the non-use rate iu (1) of the first application;

[0096] Define a (1x2) matrix {0,1} as matrix ee;

[0097] Define a (unx1) all-one matrix as matrix fo;

[0098] Define the relationship B-2-1: ee x HAB (t) x fo = 0; wherein HAB (t) represents the t-day application change matrix;

[0099] Statistical quantity ill that satisfies the relationship B-2-1;

[0100] Define a (unx1) zero matrix, and fill in 1 in the first row of the matrix to obtain matrix ff;

[0101] Define the relationship B-2-2: ee x HAB(t) Xff = 1

[0102] Counting the number imm satisfying the relationship expression B-2-1

[0103] Calculating the usage rate uu of the first application (1) , uu (1) = imm / (td-ill)

[0104] Non-use rate iu (1) , iu (1) = 1-uu (1) ;

[0105] Step S344: Repeat the same step of calculation, calculate the usage rate uu of the second to the un application (2) ~ uu (un) And the non-use rate iu (2) ~ iu (un) .

[0106] An artificial intelligence-based voice recognition system for a TV stick comprises:

[0107] A voice acquisition module: for acquiring the operation voice of the target user and the test voice of the tester; performing a one-time processing of pre-emphasis, windowing and Fourier transform on the test voice, and calculating the MFCC coefficient sequence of each test voice;

[0108] A voice analysis module: for performing a second processing of homomorphism and cepstrum on the test voice, calculating the LPCC coefficient sequence of each test voice, and constructing a voice recognition model in combination with the MFCC coefficient sequence of each test voice; taking the operation voice as the output of the voice recognition model, and judging whether the voice recognition model exists or not; if it exists, converting the operation voice into a TV stick instruction, and skipping the instruction reasoning module;

[0109] If it does not exist, acquiring the current time and the application usage data of the target user, and entering the instruction reasoning module;

[0110] An instruction reasoning module: for acquiring the current time and the number of controllable applications of the TV stick; analyzing the historical application usage data of the target user, and constructing an application usage model of the target user; acquiring the application name that the target user is using or closing, and inversely reasoning the TV stick instruction of the operation voice according to the application usage model.

[0111] Compared with the prior art, the beneficial effects of the present application are:

[0112] Improve user experience: the present application can make the user without manual operation of the TV stick, just through the voice command can complete the search program, switch channels, adjust the volume and other operations; this interactive way is more natural, in line with the usage habits of most people, especially when the user's hands are occupied or inconvenient, the advantage is particularly obvious.

[0113] Improve interaction efficiency: the voice recognition system can convert voice signals into text or commands within milliseconds, and immediately execute the corresponding operation, greatly shortening the user's waiting time; compared with the operation mode of the traditional remote controller, the present application can accurately recognize the user's oral operation instruction through the voice recognition system, and reduce the occurrence of manual misoperation.

[0114] Personalized service: when encountering user voice that cannot be recognized or voice information error, the present application can analyze the user's usage habits and characteristics, infer the user's operation command, and provide personalized voice service for the user. BRIEF DESCRIPTION OF DRAWINGS

[0115] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments, made with reference to the accompanying drawings:

[0116] Figure 1 The present application is a schematic diagram of the method;

[0117] Figure 2 The present application is a schematic diagram of the system. DETAILED DESCRIPTION

[0118] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0119] Example 1

[0120] Please refer to Figure 1 A voice recognition method for TV stick based on artificial intelligence comprises:

[0121] Step S1: obtaining the operation voice of the target user and the test voice of the tester; pre-emphasizing, windowing and Fourier transforming the test voice once to calculate the MFCC coefficient sequence of each test voice;

[0122] It should be noted that the "target user" in the present application means: the user using the TV stick; the tester means: the volunteer or staff participating in the development of the voice recognition function of the present application;

[0123] The test voice means: the related voice for operating or controlling the TV stick, such as TV stick instruction, TV program name, website or application name, etc.;

[0124] The specific steps of step S1 are as follows:

[0125] Step S11: Count the number of test voice recordings, denoted as vn; count the duration of the first to the vnth test voice recordings, denoted as vo. (1) ~vo (vn) (Unit: seconds);

[0126] Extract vo (1) ~vo (vn) The maximum value in is denoted as vo. (max) The minimum value is denoted as vo. (min) ; calculated vo (1) ~vo (vn) The average value, denoted as vo (ave) ;

[0127] Step S12: Define the calculation formula A1:

[0128]

[0129] Among them, vo (i) Vo represents the duration of the i-th test speech. (ve) It means (vo) (1) ~vo (vn) The corresponding weighted duration; the value range of i is 1 to vn;

[0130] Obtain the sampling rate of the test speech, denoted as fs (unit: kilohertz);

[0131] Calculate the frame duration of a single test speech segment, denoted as fd, where fd = v0. (ve) / fs (fd is in milliseconds);

[0132] The frame shift duration of a single test speech is denoted as fsd, where fsd = fd / 2;

[0133] Step S13: Perform pre-emphasis, windowing, and Fourier transform enhancement processing on the first test speech, and perform Mel filtering and discrete cosine transform to obtain the MFCC coefficients of the first test speech.

[0134] Step S131: Take the first test speech as speech v; use a high-pass filter to remove low-frequency noise from speech v to obtain speech v (1) ;

[0135] Using the frame duration fd and frame shift duration fsd, the speech v (1) Frame segmentation is performed to obtain the speech v (2) ;

[0136] Voice v (2) The number of samples (i.e., the number of samples) is denoted as Fn, Fn = vo (1)Ceiling of fd; Fn is up-rounded;

[0137] Step S132: Hamming window is used to process the voice v (2) to obtain the voice v (3) ; (3) and the voice v (2) ;

[0138] The Fourier transform is used to convert the voice v (3) from time domain information to frequency domain information to obtain the voice v (4) ; the mathematical expression of the Fourier transform is:

[0139]

[0140] wherein k represents the index of the voice v (4) (i.e. frequency domain information), n represents the index of the voice v (4) (i.e. time domain information), and the value range of k and n is: 1~Fn;

[0141] KX (k) represents the voice v (4) (i.e. frequency domain information) of the kth timestamp, x (n) represents the voice v (3) (i.e. time domain information) of the nth timestamp; KX (1) represents the voice v (4) (i.e. frequency domain information) of the 1st timestamp, x (1) represents the voice v (3) (i.e. time domain information) of the 1st timestamp;

[0142] e represents a natural constant, j represents an imaginary unit (i.e. j 2 =-1), and π represents a circular constant;

[0143] Step S133: the frequency of the voice v (4) corresponding to the 1st to Fnth timestamps is obtained, denoted as f (1) ~f (Fn) ;

[0144] The Mel scale triangular filter is used to receive the voice v (4) (i.e. frequency domain information) to simulate the nonlinear frequency perception of human hearing to obtain the voice v (5) ;

[0145] The number of triangular filters is Fn, and the frequency scanning range is: 0 Hz~(2 / Fn) Hz;

[0146] The reason for setting the number of triangular filters to F here is to prevent two voices v (4)Interference occurs when Mel frequency conversion is performed; the number of triangular filters can be adjusted by the user or relevant technician as needed;

[0147] Step S134: the voice v (4) corresponding to each timestamp is converted into Mel frequency, and the conversion formula is as follows:

[0148] Wherein, f (l) represents the frequency of the lth timestamp information v (4) ; F mel (f (l) ) represents the perceived frequency (i.e. Mel frequency) of f (l) in Mel units; the value range of l is: 1~Fn;

[0149] F (mel) (f (1) )~F (mel) (f (Fn) ) are summarized as the voice v (5) ;

[0150] Step S135: the perceived frequency of the voice v (4) is subjected to discrete cosine transform to obtain the MFCC coefficient of the 1st test voice;

[0151] The formula of the discrete cosine transform is as follows:

[0152]

[0153] Wherein, MFCC (l) represents the MFCC coefficient corresponding to the lth timestamp voice v (5) ;

[0154] MFCC (1) ~MFCC (Fn) are summarized as the MFCC coefficient sequence of the 1st test voice;

[0155] Step S14: the same steps (i.e. steps S131~S135) of the 1st test voice are repeated to process the 2nd to va-th test voices to obtain the MFCC coefficient sequence of the 2nd to vb-th test voices;

[0156] Step S2: the test voice is subjected to secondary processing of homomorphism and cepstrum processing, the LPCC coefficient sequence of each test voice is calculated, and the voice recognition model is constructed in combination with the MFCC coefficient sequence of each test voice; the operation voice is taken as the output of the voice recognition model, and it is judged whether the voice recognition model exists or not; if it exists, the (target user's) operation voice is converted into a TV stick instruction, (the TV stick is operated), and step S3 is skipped;

[0157] If not, obtain the current time and the historical application usage data of the target user (in the past month), and go to step S3;

[0158] The specific steps of step S2 are as follows:

[0159] Step S21: Perform homomorphic and cepstrum processing on the test voice to obtain the LPCC coefficient of the test voice information;

[0160] Step S211: Perform homomorphic and cepstrum processing on the first test voice information;

[0161] Obtain the frequency domain information of the first test voice v (4) As information E; obtain the frequency corresponding to each information E (i.e., the voice v (4) corresponding to the first to Fn timestamps), denoted as fx (1) ~ fx (Fn) ;

[0162] Step S212: Perform Z transform on the information E corresponding to the first to Fn timestamps to obtain XZ (l) ~ XZ (Fn) ;

[0163] The function expression of Z transform is:

[0164] wherein, fx (l) represents the frequency of the information E corresponding to the lth timestamp; ZX (l) represents the Z transform output of fx (l) ;

[0165] zz (l) represents the complex variable corresponding to fx (l) ;

[0166] wherein, vc represents the speed of sound, e represents the natural constant, j represents the imaginary unit (j2=-1), and ω represents the phase corresponding to fx (l) ;

[0167] ω=arctan[(vc×fx (l) ) / fd]; wherein, arctan represents the inverse tangent function, and fd represents the frame length;

[0168] Step S213: Perform logarithmic conversion on the information E corresponding to the first to Fn timestamps to obtain ln(ZX (1) )~ln(ZX (Fn) );

[0169] The function expression of logarithmic conversion is:

[0170]

[0171] wherein, ln(ZX (l) ) represents the output of the logarithmic conversion of ZX (l) ; fx (q) represents the frequency of the information E corresponding to the qth timestamp; zz (q) represents the complex variable corresponding to fx (q) ; the value range of q is: 1~Fn;

[0172] The iztrans function is used to perform inverse Z transform processing on ln(ZX (1) )~ln(ZX (Fn) ), so as to convert the information E corresponding to the 1st to the Fnth timestamp from frequency domain information into time domain information, and obtain ts (1) ~ts (Fn) ;

[0173] Step S214: cepstrum processing is performed on ts (1) ~ts (Fn) , so as to obtain the LPC coefficients of the information E corresponding to the 1st to the Fnth timestamp, denoted as ce (1) ~ce (Fn) ;

[0174] The function expression of the cepstrum processing is:

[0175]

[0176] wherein, ts (m) represents the time domain information of the information E corresponding to the mth timestamp, and fx (l-m+1) represents the frequency of the information E corresponding to the (l-m+1)th timestamp; the value range of m is: 1~l;

[0177] The discrete cosine transform algorithm is used to convert ce (1) ~ce (Fn) from time domain information into frequency domain information, so as to obtain the LPCC coefficient sequence of the 1st test voice information;

[0178] Step S215: the same steps (i.e., steps S211~S214) of homomorphism and cepstrum processing are repeated to process the 2nd to the va-th test voice information, so as to obtain the LPCC coefficient sequence of the 2nd to the va-th test voice information;

[0179] Step S22: the recognize_google function and the recognize_sphinx function in the SpeechRecognition library are used to convert the 1st to the vn-th test voice information into test voice text, denoted as text wd (1) ~wd(vn) ;

[0180] The test voice information is summarized as a sample set in the format of "test voice text - MFCC coefficient sequence of test voice information - LPCC coefficient sequence of test voice information";

[0181] Step S23: Defining a training step for supervised training of the base model, taking the CNN model as the base model of the speech recognition model;

[0182] Step S231: Taking the MFCC coefficient sequence corresponding to the first to the vn th as a high-dimensional array, denoted as Cmf (1) ~ Cmf (vn) ; define a set CMF, CMF = {Cmf (1) ~ Cmf (vn)};

[0183] Taking the LPCC coefficient sequence corresponding to the first to the vn th as a high-dimensional array, denoted as Clp (1) ~ Clp (vn) ; define a set CLP, CLPF = {Clp (1) ~ Clp (vn)};

[0184] Step S232: Fusing the set Cmf and the set Clp to obtain a set of voiceprint feature sequences of remote control operation instructions, denoted as set CC, CC = {(Cmf (1) , Clp (1) ) ~ (Cmf (vn) , Clp ( vn )};

[0185] Step S233: Fusing the text wd (1) ~ wd (vn) and the set CC to obtain a feature set of the test voice, denoted as set α; α = {(wd (1) , Cmf (1) , Clp (1) ) ~ (wd (vn) , Cmf (vn) , Clp (vn)};

[0186] Setting the number of states in the base model to 4 and the number of Gaussians to 3;

[0187] Step S234: Bringing the set α into the base model and iterating until the test voice corresponding to the set α is output by the base model;

[0188] Using cross-validation algorithm to prevent overfitting of the base model in the process of supervised training;

[0189] Define the condition for determining the convergence of the speech recognition model: the probability of any one test speech being misrecognized is less than 1 x 10 -3 ;

[0190] Step S24: define the training step of unsupervised training of the base model:

[0191] Step S241: obtain the set CC (i.e., the set of voiceprint feature sequences of the television stick operation instruction), take the set CC as the training set 1, and take the test speech text as the training set 2; use the K-means clustering algorithm to iterate the base model until the elements in the training set 1 and the elements in the training set 2 are output at the same time and the output elements correspond (i.e., the speech text output by the training set 1 corresponds to the voiceprint feature sequence output by the training set 2);

[0192] Step S242: use the cross-validation algorithm to prevent overfitting of the base model during unsupervised training;

[0193] Define the condition for determining the convergence of the speech recognition model: the probability difference of the same test speech being output by the two consecutive recognitions is less than 1 x 10 -4 ;

[0194] Step S25: perform supervised training and unsupervised training on the base model; until the output results of the supervised training and the unsupervised training of the base model are the same (the model training is completed), the speech recognition model is obtained;

[0195] Take the operation speech as the input of the speech recognition model, and determine whether there is an output of the speech recognition model;

[0196] If there is, convert the operation speech of the target user into a television stick instruction, operate the television stick, and skip step S3;

[0197] If there is not, obtain the current time and the historical application usage data of the target user, and enter step S3.

[0198] Step S3: obtain the current time and the number of controllable applications of the television stick; analyze the historical application usage data of the target user (in the past month) to construct an application usage (habit) model of the target user; obtain the application name that the target user is using or closing, and infer the television stick instruction of the operation speech according to the application usage (habit) model;

[0199] The specific steps of step S3 are as follows:

[0200] Step S31: obtain the current time as (mo month da day) ho;

[0201] The historical application usage data includes: the usage time and usage of the target user using the 1st to the unth application in the time period from the ho hour to the (ho+1) hour last month;

[0202] The number of applications operable by the TV stick is obtained, denoted as un; the usage time ti of the target user using the 1st application (through the TV stick) in the time period from the ho hour to the (ho+1) hour last month is obtained (1，1) ~ti (td，1) ; the usage time ti of the target user using the 2nd application is obtained (2，1) ~ti (td，2) ;

[0203] By analogy, the usage time ti of the target user using the unth application is obtained (td，1) ~ti (td，un) (unit: minute)

[0204] ; wherein, td represents the total number of days last month;

[0205] It should be noted that if the ho hour is 23 o'clock of a day, then the (ho+1) hour is 0 o'clock of the next day; for example, if the ho hour is 23 o'clock on March 1st, then the (ho+1) hour is 0 o'clock on March 2nd;

[0206] Step S32: corresponding calculation formula B-1 is:

[0207] Wherein, ut (s) represents the time coefficient of the s-th application (used by the target user), ti (s，t) represents the time of the target user using the t-th application on the s-th day (last month); the value range of s is 1-td, and the value range of t is 1-un;

[0208] The time coefficients ut (1) ~ut (un) of the target user using the 1st to the unth application (last month) are calculated according to the calculation formula B-1;

[0209] Step S33: define the binary tuple ha (s，t) represents the determination coefficient of the target user using the t-th application at the ho hour on the s-th day (last month); define the binary tuple hb (s，t) represents the determination coefficient of the target user using the t-th application at the (ho+1) hour on the s-th day (last month); ha (s，t) and hb (s，t) ∈{0,1}; (ha (s，t) is 0, indicating that the target user does not use the t-th application at the ho hour on the s-th day (last month); ha (s，t)A value of 1 indicates that the target user used the t-th application on the s-th day at the ho-th hour (in the previous month).

[0210] hb (s，t) A value of 0 indicates that the target user (in the previous month) did not use the t-th application on the s-th day at the (ho+1)-th time; hb (s，t) A value of 1 indicates that the target user used the t-th application on the s-th day (ho+1) of the previous month.

[0211] (using binary ha) (s，t) Define the decision coefficients for the target user (at time ho on day 1 of the previous month) to time ho on day td, for the first to the unth application: ha (1，1) ~ha (td，un) ;

[0212] (using the binary hb) (s，t) Define the decision coefficients for the first to the last application for the target user (from day 1, hour (ho+1) to day td, hour (ho+1) of the previous month: hb (1，1) ~hb(td,un);

[0213] Step S34: According to ha (1，1) ~ha (td，un) and hb (1，1) ~hb (td，un) ; Analyze the usage of the target users (from time ho to time (ho+1)) on the first to the unth applications, and construct an application usage (habit) model for the target users;

[0214] Step S341: Define an empty matrix of (1×un) and fill it with ha sequentially. (1，1) ~ha (1，un) and hb (1，1) ~hb (1，un) The matrix HAA is obtained. (1) Sum matrix HBB (1) Among them, ha (1，un) and hb (1，un) , representing the determination coefficients of the target user at the ho hour on day 1 and at the (ho+1) hour on day td, respectively;

[0215] Fill in ha (2，1) ~ha (2，un) and hb (2，1) ~hb (2，un) The matrix HAA is obtained. (2) Sum matrix HBB (2) Among them, ha (2，1) ~ha (2，un) and hb (2，1) ~hb(2，un) , respectively, represent the determination coefficients of the first to the un-th applications from the ho-th hour to the (ho+1)-th hour and from the ho-th hour to the (ho+1)-th hour on the second day of the target user (in the last month);

[0216] Similarly, fill in ha (td，1) ~ ha (td，un) and hb (td，1) ~ hb (td，un) to obtain the matrix HAA ( td ) and the matrix HBB (td) ; wherein ha (td，1) ~ ha (td，un) and hb (td，1) ~ hb (td，un) , respectively, represent the determination coefficients of the first to the un-th applications from the ho-th hour to the (ho+1)-th hour and from the ho-th hour to the (ho+1)-th hour on the second day of the target user (in the last month);

[0217] Step S342: vertically stack the matrix HAA (1) and the matrix HBB (1) to obtain the first day application change matrix HAB (1) ;

[0218] Vertically stack the matrix HAA (2) and the matrix HBB (2) to obtain the second day application change matrix HAB(2);

[0219] Similarly, vertically stack the matrix HAA (td) and the matrix HBB (td) to obtain the td-th day application change matrix HAB (td) ;

[0220] Step S343: analyze the usage habit of the first application in the time period from the ho-th hour to the (ho+1)-th hour of the target user (in the last month), and calculate the usage rate uu (1) and the non-usage rate iu (1) of the first application;

[0221] Define a (1x2) matrix {0,1} as the matrix ee;

[0222] Define a (unx1) all-one matrix as the matrix fo;

[0223] Define the relationship B-2-1: ee x HAB (t) x fo = 0; wherein HAB (t) represents the t-th day application change matrix;

[0224] HAB(1) ~HAB (td) Substitute the relationship B-2-1, and count the number of relationship B-2-1, recorded as ill;

[0225] Define the zero matrix of (unx1), and fill in 1 in the first row of the matrix, get the matrix ff;

[0226] Define the relationship B-2-2: ee x HAB (t) X ff = 1;

[0227] HAB (1) ~HAB (td) Substitute the relationship B-2-2, and count the number of relationship B-2-1, recorded as imm;

[0228] Calculate the usage rate of the first application uu (1) , uu (1) = imm / (td-ill);

[0229] The non-use rate iu (1) , iu (1) = 1-uu (1) ;

[0230] Step S344: repeat the calculation of the same step, calculate the usage rate of the second to the un application uu (2) ~uu (un) And the non-use rate iu (2) ~iu (un) ;

[0231] Step S345: analyze the influence relationship of different applications on the first application, calculate the collaborative use rate of different applications on the first application lx (1-1) ~lx (un-1) ;

[0232] Step S3451: calculate the collaborative use rate of the first application on the first application lx (1-1) ;

[0233] Define the matrix {1,0} of (1x2), recorded as matrix rr; define the zero matrix of (unx1), recorded as matrix zf;

[0234] Fill in 1 in the first row of the matrix zf (1) ;

[0235] Define the relationship B-3-1: (rr x HAB (t) X zf (1) )+(ee x HAB (t) X zf (1) )=2;

[0236] HAB (1) ~ HAB (td) Substitute HAB (1) into relation B-3-1, and count the number of times that relation B-3-1 is satisfied, and denote it as oo (1-1) ;

[0237] Calculate the collaborative use rate lx of the first application to the first application (1-1) ; (1) = oo (2-1) / (td x un);

[0238] Step S3452: Calculate the collaborative use rate lx of the second application to the first application (2) ;

[0239] Fill in 1 in the second row of matrix zf (t) ;

[0240] Define relation B-3-2: (rr x HAB (2) x zf (t) ) + (ee x HAB (1) x zf (1) ) = 2;

[0241] Substitute HAB (td) ~ HAB (2) into relation B-3-2, and count the number of times that relation B-3-2 is satisfied, and denote it as oo (2-1) ;

[0242] Calculate the collaborative use rate lx of the second application to the first application (2-1) ; (2) = oo (un) / (td x un);

[0243] Step S3453: Similarly, calculate the collaborative use rate lx(un-1) of the un application to the first application

[0244] Fill in 1 in the un row of matrix zf (t) ;

[0245] Define relation B-3-3: (rr x HAB (un) x zf (t) ) + (ee x HAB (1) x zf (1) ) = 2;

[0246] Substitute HAB (td) ~ HAB (un) into relation B-3-3, and count the number of times that relation B-3-3 is satisfied, and denote it as oo ;

[0247] lx (un-1) , lx (un-1) = oo (un) / (td x un);

[0248] Step S346: repeat the calculation of lx (1-1) ~ lx (un-1) , lx (1-2) ~ lx (un-un) ;

[0249] Summarize the usage rate, non-usage rate and the synergistic usage rate of the first to the nth application as the usage (habit) model of the target user to the application;

[0250] Step S35: obtain the application name that the target user (at mo month da day ho hour) is using or closing as the quasi-application; according to the usage habit of the target user to different applications, inversely infer the voice stick command of the television stick;

[0251] Step S351: obtain the usage rate of the quasi-application as quu, and the time coefficient base qu;

[0252] wherein quu∈{uu (2) ~ uu (un)}, qu∈{ut (1) ~ ut (un)};

[0253] Calculate the probability that the target user continues to use the quasi-application as ccg, ccg=qu×(quu) 2 ;

[0254] Step S352: obtain the usage rate of the non-quasi-application as iqu (1) ~ iqu (an) (wherein, ), the non-usage rate as nqu (1) ~ nqu (an) (wherein, ), and the time coefficient as iut (1) ~ iut (an) (wherein, );

[0255] wherein an represents the number of non-quasi-applications, an

[0256] Obtain the synergistic usage rate of the quasi-application to the first to the an application as qlx (1) ~ qlx

[0257] (an) ; wherein,

[0258] The calculation formula B-4 is defined as:

[0259]

[0260] wherein, qqg (y) represents the probability of the target user selecting the yth non-qualified application at (ho+1), qlx (y) , iqu (y) and iut (y) respectively represent the cooperative use rate, use rate and time coefficient of the qualified application to the yth non-qualified application;

[0261] qlx (p) represents the non-use rate of the qualified application to the pth non-qualified application, and y and p cannot take the same value at the same time;

[0262] Step S353: (According to the calculation formula B-4) calculate the probability of the target user selecting the 1st to anth non-qualified application at (ho+1), qqg (1) ~ qqg (an) ;

[0263] Extract the maximum value of ccg and qqg (1) ~ qqg (an) as maa; generate the TV stick instruction for controlling the application corresponding to maa.

[0264] Embodiment two

[0265] Please refer to Figure 2 , an artificial intelligence-based voice recognition system for a TV stick includes:

[0266] A voice acquisition module: for acquiring the operation voice of the target user and the test voice of the tester; performing a one-time processing of pre-emphasis, windowing and Fourier transform on the test voice, and calculating the MFCC coefficient sequence of each test voice;

[0267] A voice analysis module: for performing a second processing of homomorphism and cepstrum on the test voice, calculating the LPCC coefficient sequence of each test voice, and constructing a voice recognition model in combination with the MFCC coefficient sequence of each test voice; taking the operation voice as the output of the voice recognition model, judging whether the voice recognition model exists or not; if it exists, converting the operation voice of the target user into a TV stick instruction, operating the TV stick, and skipping the instruction reasoning module;

[0268] If it does not exist, acquire the current time and the application use data of the target user, and enter the instruction reasoning module;

[0269] Instruction inference module: for acquiring current time and the number of controllable applications of the TV stick; analyzing the historical application use data of the target user (in the past month) to build an application use (habit) model of the target user; acquiring the application name that the target user is using or closing, and inversely inferring the TV stick instruction of the operation voice according to the application use (habit) model.

[0270] The above formulas are all dimensionless values calculated, the formulas are obtained by collecting a large amount of data to simulate a formula of the most recent real situation, and the preset parameters in the formula are set by a person skilled in the art according to the actual situation, such as the weight coefficient and the proportion coefficient, the size of which is to obtain a specific value of each parameter after quantization, for subsequent comparison, and as long as the proportion relationship between the parameters and the quantized value is not affected, the size of the weight coefficient and the proportion coefficient can be.

[0271] Finally, it should be noted that: the above-described embodiments are only specific implementations of the present application, which are used to illustrate the technical solutions of the present application, but not to limit it, the protection scope of the present application is not limited to this, although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand: any person skilled in the art within the technical range disclosed by the present application, can still modify or easily think of changes to the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part of the technical features; and these modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and all should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A voice recognition method for a TV stick based on artificial intelligence, characterized in that, The method includes: Step S1: Obtain the target user's operation voice and the tester's test voice; perform pre-emphasis, windowing and Fourier transform processing on the test voice, and calculate the MFCC coefficient sequence for each test voice; Step S2: Perform secondary processing of homomorphic and cepstral processing on the test speech, calculate the LPCC coefficient sequence of each test speech, and construct a speech recognition model by combining the MFCC coefficient sequence of each test speech; use the operation speech as the output of the speech recognition model, and determine whether the speech recognition model has an output; if it does, convert the operation speech into a TV stick command and skip step S3. If it does not exist, obtain the current time and the target user's historical application usage data, and proceed to step S3; Step S3: Obtain the current time and the number of applications controllable by the TV stick; analyze the target user's historical application usage data and build an application usage model for the target user; obtain the names of applications that the target user is currently using or closing, and infer the TV stick commands for voice operation based on the application usage model; The specific steps of step S2 are as follows: Step S21: Perform homomorphic and cepstral processing on the test speech to obtain the LPCC coefficients of the test speech information; Step S22: Convert the 1st to the vnth test speech information into test speech text, denoted as text wd(1)~wd(vn); The test speech information is compiled into a sample set according to the format of test speech text - MFCC coefficient sequence of test speech information - LPCC coefficient sequence of test speech information; Step S23: Using the CNN model as the base model for speech recognition, define the training steps for supervised training of the base model: Step S231: Take the MFCC coefficient sequence corresponding to the first to the nth item as the set CMF; The LPCC coefficient sequences corresponding to the first to the vnth items are used as high-dimensional arrays and integrated to obtain the set CLP; Step S232: Merge set Cmf and set Clp to obtain the set of voiceprint feature sequences of TV stick operation commands, denoted as set CC; Step S233: Merge the texts wd(1)~wd(vn) with the set CC to obtain the feature set of the test speech, denoted as set α; Set the number of states in the basic model to 4 and the number of Gaussian elements to 3; Step S234: Substitute set α into the basic model and iterate until all test voices corresponding to set α are output by the basic model; The convergence criterion for a speech recognition model is defined as follows: the probability that any test speech is misrecognized is less than 1 × 10⁻⁶. -3 ; The subsequent steps of step S23 are as follows: Step S24: Define the training steps for unsupervised training of the base model: Step S241: Obtain set CC, use set CC as training set 1, and use test speech text as training set 2; use K-means clustering algorithm to iterate the basic model until the elements in training set 1 and the elements in training set 2 are output simultaneously and the output elements correspond. Step S242: Define the convergence criterion for the speech recognition model: the probability difference between two consecutive recognition outputs of the same test speech is less than 1×10. -4 ; Step S25: Perform supervised and unsupervised training on the base model until the outputs of supervised and unsupervised training of the base model are the same, thus obtaining the speech recognition model; The operation speech is used as the input to the speech recognition model to determine whether the model has an output; If it exists, the voice command will be converted into a TV stick command, and step S3 will be skipped. If it does not exist, obtain the current time and the target user's historical application usage data, and proceed to step S3; The specific steps of step S21 are as follows: Step S211: Perform homomorphic and cepstral processing on the first test speech information; Obtain the frequency domain information of the first test speech, speech v(4), as information E; obtain the frequencies fx(1)~fx(Fn) corresponding to each information E; Step S212: Perform Z-transformation on the information E corresponding to the 1st to Fnth timestamps to obtain XZ(l)~XZ(Fn); The functional expression for the Z-transform is: Where fx(l) represents the frequency of information E corresponding to the l-th timestamp; ZX(l) represents the Z-transform output of fx(l); zz(l) represents the complex variable corresponding to fx(l): vc represents the speed of sound, and ω represents the phase corresponding to fx(l): ω = arctan[(vc × fx(l)) / fd]; fd represents the frame duration; Step S213: Perform a logarithmic transformation on the information E corresponding to the 1st to Fnth timestamps to obtain ln(ZX(1))~ln(ZX(Fn)); The function expression for the logarithmic transformation is: Where ln(ZX(l)) represents the output of the logarithmic transformation of ZX(l); fx(q) represents the frequency of the information E corresponding to the q-th timestamp; zz(q) represents the complex variable corresponding to fx(q); Perform inverse Z-transform on ln(ZX(1))~ln(ZX(Fn)) to convert the information E corresponding to the 1st to Fnth timestamps from frequency domain information to time domain information, and obtain ts(1)~ts(Fn); The subsequent steps of step S213 are as follows: Step S214: Perform cepstral processing on ts(1)~ts(Fn) to obtain the LPC coefficients of information E corresponding to the 1st to Fnth timestamps, denoted as ce(1)~ce(Fn); The functional expression for cepstral processing is: Where ts(m) represents the time-domain information of information E corresponding to the m-th timestamp, and fx(l-m+1) represents the frequency of information E corresponding to the (l-m+1)-th timestamp; Using the discrete cosine transform algorithm, ce(1)~ce(Fn) is transformed from time domain information to frequency domain information to obtain the LPCC coefficient sequence of the first test speech information; Step S215: Process the 2nd to vath test speech information to obtain the LPCC coefficient sequence of the 2nd to vath test speech information.

2. The voice recognition method for a TV stick based on artificial intelligence according to claim 1, characterized in that, The specific steps of step S1 are as follows: Step S11: Count the number of test voice recordings (vn); Count the duration (vo) of each test voice recording. (1) ~vo (vn) Extract vo (1) ~vo (vn) The maximum value of vo (max) Minimum value vo (min) ; calculated vo (1) ~vo (vn) Average value vo (ave) ; Step S12: Define calculation formula A1: Among them, vo (i) Vo represents the duration of the i-th test speech. (ve) Indicates the weighted duration; Obtain the sampling rate fs of the test speech; Calculate the frame duration fd of a single test speech, fd = vo (ve) / fs; The frame shift duration fsd for a single test speech, fsd = fd / 2; Step S13: Perform enhancement processing on the first test speech and calculate the MFCC coefficients of the first test speech; Step S14: Process test speech from the 2nd to the vath test speech to obtain the MFCC coefficient sequence of test speech from the 2nd to the vbth test speech, and proceed to step S2.

3. The voice recognition method for a TV stick based on artificial intelligence according to claim 2, characterized in that, The specific steps of step S13 are as follows: Step S131: Take the first test speech as speech v; remove low-frequency noise from speech v to obtain speech v (1) ; using fd and fsd to represent speech v (1) Frame segmentation is performed to obtain the speech v (2) ; Voice v (2) The quantity Fn, Fn = vo (1) / fd; Step S132: Use Hamming window to process speech v (2) Windowing is applied to obtain the voice voice. (3) ; Voice v (3) The time-domain information is converted into the frequency-domain information to obtain the speech v. (4) ; Step S133: Obtain the voice v corresponding to the 1st to Fnth timestamps. (4) frequency f (1) ~f(Fn); Using a triangular filter, the voice signal v is received. (4) Nonlinear frequency sensing to obtain speech v (5) ; Step S134: Record the voice v corresponding to each timestamp (4) The conversion formula for frequency to Mel frequency is as follows: Among them, f (l) Represents the l-th timestamp information v (4) The frequency; F mel (f (l) f represents units in Mel. (l) The perceived frequency; Summary F (mel) (f (1) )~F (mel) (f (Fn) (as a voice) (5) ; Step S135: For voice v (4) The MFCC coefficients of the first test speech are obtained by performing discrete cosine transform on the perceived frequency. The formula for the discrete cosine transform is as follows: Among them, MFCC (l) Represents the l-th timestamp voice v (5) The corresponding MFCC coefficients; Summary of MFCC (1) ~MFCC (Fn) , which serves as the MFCC coefficient sequence for the first test speech.

4. The voice recognition method for a TV stick based on artificial intelligence according to claim 1, characterized in that, The specific steps of step S3 are as follows: Step S31: Obtain the number of applications that the TV stick can operate (un); obtain the usage time (ti) of the target user using the first type of application last month. (1，1) ~ti (td，1) Similarly, the usage time ti of the unth application... (td，1) ~ti (td，un) ;td represents the total number of days in the previous month; Step S32: Corresponding calculation formula B-1: Among them, ut (s) Let ti represent the time coefficient for the s-th application. (s，t) This represents the time the target user uses the t-th application on day s; Calculate the time coefficient ut for the target user's first to unth applications. (1) ~ut (un) ; Step S33: From day 1, hour ho to day td, hour ho, the determination coefficients for applications 1 to un: ha (1，1) ~ha (td，un) ; From hour (ho+1) on day 1 to hour (ho+1) on day td, the determination coefficients for applications 1 to un are: hb (1，1) ~hb (td，un) ; Step S34: According to ha (1，1) ~ha (td，un) and hb (1，1) ~hb (td，un) ; Analyze the usage of the target users for applications from type 1 to type u, and construct an application usage model for the target users; Step S35: Obtain the name of the application that the target user is using or closing, as the quasi-application; based on the target user's usage habits of different applications, reverse-engineer the TV stick commands for voice operation.

5. The voice recognition method for a TV stick based on artificial intelligence according to claim 4, characterized in that, The specific steps of step S34 are as follows: Step S341: According to ha (1，1) ~ha (td，un) Construct matrix HAA (1) ~HAA (td) Sum matrix HBB (1) ~HBB (td) ; Step S342: Convert matrix HAA (1) Sum matrix HBB (1) Vertical stacking yields the transformation matrix HAB applied on day 1. (1) ; Similarly, let matrix HAA (td) Sum matrix HBB (td) Vertical stacking yields the applied transformation matrix HAB for day td. (td) ; Step S343: Analyze the target user's usage habits of the first application during the time period from time ho to time (ho+1), and calculate the usage rate uu of the first application. (1) and non-use rate iu (1) ; Define a (1×2) matrix {0,1} as matrix ee; Define an all-one matrix of size (un×1) as matrix fo; Define relation B-2-1: ee × HAB (t) ×fo=0; where HAB (t) This indicates the change matrix applied on day t; Count the number of items that satisfy relation B-2-1. Define a zero matrix of size (un×1) and fill the first row of the matrix with 1s to obtain matrix ff; Define relation B-2-2: ee × HAB (t) ×ff = 1; Count the number of expressions that satisfy relation B-2-1, i.e., imm. Calculate the usage rate of the first type of application. (1) ,uu (1) =imm / (td-ill); Usage rate (1) iu (1) =1-uu (1) ; Step S344: Calculate the usage rate and non-use rate of applications 2 to u.

6. An AI-based voice recognition system for a TV stick, applicable to the AI-based voice recognition method for a TV stick as described in any one of claims 1-5, characterized in that, The system includes: Voice acquisition module: used to acquire the target user's operation voice and the tester's test voice; to perform pre-emphasis, windowing and Fourier transform on the test voice, and to calculate the MFCC coefficient sequence of each test voice; Speech Analysis Module: This module performs secondary processing on the test speech using homomorphic and cepstral processing, calculates the LPCC coefficient sequence for each test speech, and constructs a speech recognition model by combining the MFCC coefficient sequence for each test speech. It uses the operation speech as the output of the speech recognition model and determines whether the speech recognition model has an output. If it does, it converts the operation speech into a TV stick command and skips the command inference module. If it does not exist, obtain the current time and the target user's application usage data, and enter the instruction reasoning module; Command reasoning module: used to obtain the current time and the number of TV stick controllable applications; analyze the target user's historical application usage data to build the target user's application usage model; obtain the names of applications that the target user is using or closing, and infer the TV stick commands for voice operation based on the application usage model.

Images