Voice recognition method and system for television stick based on artificial intelligence

Abstract

The invention provides a voice recognition method and system for a television stick based on artificial intelligence, and belongs to the field of voice recognition. The problem of low voice recognition efficiency of the television stick is solved; the method specifically comprises the following steps: S1, obtaining operation voice, and calculating an MFCC coefficient sequence of each test voice; s2, calculating an LPCC coefficient sequence of each test voice, constructing a voice recognition model in combination with an MFCC coefficient sequence, and judging whether the operation voice is output in the model or not; if yes, converting the operation voice into a television stick instruction; if not, acquiring historical application use data; s3, analyzing historical application use data of the target user, and reversely reasoning a television stick instruction of the operation voice; the voice recognition and instruction generation capabilities of the television stick are improved by acquiring, analyzing and processing the language generated when the user operates the television stick.

Description

Technical Field

[0001] The invention discloses a speech recognition method and system for a TV stick based on artificial intelligence, and relates to the field of speech recognition. Background Art

[0002] The existing speech recognition method or system for TV sticks has the following deficiencies: Insufficient semantic understanding ability: Currently, most speech recognition systems can only recognize some simple voice commands, such as searching for programs, switching channels, etc.; and the recognition ability for some complex semantic expressions or ambiguous commands needs to be improved.

[0003] Reasoning ability: Current speech recognition systems lack logical reasoning capabilities and are unable to infer the user's current intentions based on the user's usage habits for different applications; this can result in the system being unable to accurately understand the user's instructions in some continuous conversation scenarios.

[0004] Highly dependent on technology: Existing speech recognition systems rely on stable network and equipment support to achieve normal functions; at the same time, AI-based speech recognition systems require a lot of R&D resources and funds to develop and deploy, which is too costly for some companies or developers. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention aims to provide a voice recognition method and system for a TV stick based on artificial intelligence, aiming to solve the problem of low efficiency of voice recognition in the TV stick.

[0006] In order to achieve the above object, the present invention is implemented by the following technical solution: A voice recognition method for a TV stick based on artificial intelligence includes: Step S1: Acquire the target user's operating voice and the tester's test voice; perform pre-emphasis, windowing and Fourier transform on the test voice, and calculate the MFCC coefficient sequence of each test voice; Step S2: Perform secondary processing of homomorphic and cepstrum processing on the test speech, calculate the LPCC coefficient sequence of each test speech, and build a speech recognition model in combination with the MFCC coefficient sequence of each test speech; take the operation speech as the output of the speech recognition model, and determine whether the speech recognition model has output; if so, convert the operation speech into a TV stick command and skip step S3; If not, the current time and the historical application usage data of the target user are obtained, and the process goes to step S3.

[0007] Step S3: Get the current time and the number of controllable applications on the TV stick; analyze the historical application usage data of the target user and build an application usage model for the target user; get the name of the application that the target user is using or closing, and reversely infer the TV stick instructions of the operation voice based on the application usage model.

[0008] Furthermore, the specific steps of step S1 are as follows: Step S11: Count the number of test voices vn; count the duration of each test voice vo （1） ~vo （vn） Extraction （1） ~vo （vn） The maximum value vo （max） , minimum value vo （min） ; calculatedvo （1） ~vo （vn） Average value （ave） ; Step S12: Define calculation formula A1: ; Among them, （i） Indicates the duration of the i-th test speech, vo （ve） represents weighted duration; Get the sampling rate fs of the test speech; Calculate the frame duration fd of a single test voice, fd = vo （ve） / fs; The frame shift duration fsd of a single test voice is fsd=fd / 2; Step S13: performing enhancement processing on the first test speech and calculating the MFCC coefficients of the first test speech; Step S14: Repeat the same steps of processing the first test speech, process the second to vath test speech, obtain the MFCC coefficient sequence of the second to vbth test speech, and enter step S2.

[0009] Furthermore, the specific steps of step S13 are as follows: Step S131: Take the first test speech as speech v; remove the low-frequency noise in speech v to obtain speech v （1） ; Use fd and fsd to compare the voice v （1） Perform frame processing to obtain the voice v （2） ; Voice （2） The number Fn, Fn = vo （1） / fd; Step S132: Use the Hamming window to （2） Perform windowing processing to obtain the voice v （3） ; The voice v （3） From the time domain information to the frequency domain information, we get the voice v （4） ; Step S133: Obtain the voice v corresponding to the 1st to Fnth timestamps （4） The frequency f （1） ~f （Fn） ; Using a triangular filter, receive the speech v （4） Nonlinear frequency perception, get the speech v （5） ; Step S134: The voice v corresponding to each timestamp （4） The frequency is converted to Mel frequency, and the conversion formula is as follows: ; Among them, f （l） Indicates the lth timestamp information v （4） Frequency; F mel (f （l） ) indicates f in Mel （l） The perceived frequency; the value range of l is: 1~Fn; Summary F （mel） (f （1） )~F （mel） (f （Fn） ) as voice v （5） ; Step S135: （4） Perform discrete cosine transform on the perceived frequency to obtain the MFCC coefficient of the first test speech; The formula for discrete cosine transform is as follows: ; Among them, MFCC （l） Indicates the lth timestamp voice v （5） Corresponding MFCC coefficients; Aggregate MFCC （1） ~MFCC （Fn） , as the MFCC coefficient sequence of the first test speech.

[0010] Furthermore, the specific steps of step S2 are: Step S21: performing homomorphic and cepstrum processing on the test speech to obtain LPCC coefficients of the test speech information; Step S22: Convert the first to the vnth test voice messages into test voice texts, recorded as text wd （1） ~wd （vn） ; The test speech information is aggregated as a sample set in the format of test speech text—MFCC coefficient sequence of the test speech information—LPCC coefficient sequence of the test speech information; Step S23: Using the CNN model as the basic model of the speech recognition model, define the training steps for supervised training of the basic model: Step S231: taking the MFCC coefficient sequences corresponding to the 1st to the vnth items as the set CMF; The LPCC coefficient sequences corresponding to the 1st to the vnth items are taken as high-dimensional arrays and integrated to obtain the set CLP; Step S232: merging the set Cmf and the set Clp to obtain a set of feature sequences of the TV stick operation command voiceprints, recorded as set CC; Step S233: Write the text wd （1） ~wd （vn） Fuse with the set CC to obtain the feature set of the test speech, denoted as set α; Set the number of states in the base model to 4 and the number of Gaussian elements to 3; Step S234: Bringing the set α into the basic model, and iterating until all the test speech corresponding to the set α is output by the basic model; Use cross-validation algorithms to prevent overfitting of the base model during supervised training; Define the judgment condition for the convergence of the speech recognition model: the probability of any test speech being misrecognized is less than 1×10 -3 .

[0011] Furthermore, the subsequent steps of step S23 are as follows: Step S24: Define the training steps for unsupervised training of the basic model: Step S241: obtaining a set CC, taking the set CC as training set 1, and taking the test speech text as training set 2; iterating the basic model using the K-means clustering algorithm until the elements in the training set 1 and the elements in the training set 2 are output simultaneously and the output elements correspond; Step S242: using a cross-validation algorithm to prevent the basic model from overfitting during unsupervised training; The convergence criteria of the speech recognition model are defined as follows: the probability difference between the two consecutive recognition outputs of the same test speech is less than 1×10 -4 ; Step S25: Perform supervised training and unsupervised training on the basic model until the output results of the supervised training and the unsupervised training of the basic model are the same, thereby obtaining a speech recognition model; Use the operation voice as the input of the speech recognition model to determine whether the model has output; If yes, the operation voice is converted into a TV stick command and step S3 is skipped; If not, the current time and the historical application usage data of the target user are obtained, and the process goes to step S3.

[0012] Furthermore, the specific steps of step S21 are as follows: Step S211: performing homomorphic and cepstrum processing on the first test speech information; Get the frequency domain information of the first test voice v （4） As information E; obtain the frequency fx corresponding to each information E （1） ~fx （Fn） ; Step S212: Perform Z transformation on the information E corresponding to the 1st to Fnth timestamps to obtain XZ （l） ~XZ （Fn） ; The function expression of Z transform is: ; Among them, fx （l） Indicates the frequency of the information E corresponding to the lth timestamp; ZX （l） Indicates fx （l） Z transform output; zz （l） Indicates fx （l） The corresponding complex variables are: ; vc represents the speed of sound, ω represents fx （l） The corresponding phase: ω=arctan[(vc×fx （l） ) / fd]; arctan represents the inverse tangent function; fd represents the frame duration; Step S213: Perform logarithmic transformation on the information E corresponding to the 1st to Fnth timestamps to obtain ln(ZX （1） )~ln(ZX （Fn） ); The function expression of logarithmic transformation is: ; Among them, ln (ZX （l） ) indicates that ZX （l） Logarithmic conversion of the output; fx （q） represents the frequency of information E corresponding to the qth timestamp; zz （q） Indicates fx （q） The corresponding complex variable; For ln (ZX （1） )~ln(ZX （Fn）) performs inverse Z transform processing to convert the information E corresponding to the 1st to Fnth timestamps from frequency domain information to time domain information, and obtains ts （1） ~ts （Fn） .

[0013] Furthermore, the subsequent steps of step S213 are as follows: Step S214: ts （1） ~ts （Fn） Perform cepstrum processing to obtain the LPC coefficients of information E corresponding to the 1st to Fnth timestamps, denoted as ce （1） ~ce （Fn） ; The function expression of cepstrum processing is: ; Among them, ts （m） represents the time domain information of the information E corresponding to the mth timestamp, fx （l-m+1） Indicates the frequency of the information E corresponding to the (l-m+1)th timestamp; the value range of m is: 1~l; Using the discrete cosine transform algorithm, ce （1） ~ce （Fn） Convert the time domain information into the frequency domain information to obtain the LPCC coefficient sequence of the first test speech information; Step S215: Repeat the same steps of performing homomorphic and cepstrum processing on the first test speech information, process the second to vath test speech information, and obtain the LPCC coefficient sequence of the second to vath test speech information.

[0014] Furthermore, the specific steps of step S3 are as follows: Step S31: Get the current time and record it as ho; Get the number of applications that can be operated on the TV stick un; get the usage time ti of the target user for the first application last month （1，1） ~ti （td，1） Similarly, the usage time of the un application is （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） represents the time coefficient of the sth application, ti （s，t） Indicates the time when the target user uses the t-th application on the s-th day; Calculate the time coefficient ut of the target user's first to unth applications （1） ~ut （un） ; Step S33: From time ho on day 1 to time ho on day td, the determination coefficient of the first to un types of applications: ha （1，1） ~ha （td，un） ; From the first day at (ho+1) to the td day at (ho+1), the determination coefficient of the first to un applications is: 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 first to unth applications by the target users and build 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; and reversely infer the TV stick instruction of the operation voice according to the target user's usage habits of different applications.

[0015] Furthermore, the specific steps of step S34 are as follows: Step S341: According to ha （1，1） ~ha （td，un） , construct the matrix HAA （1） ~HAA （td） and matrix HBB （1） ~HBB （td） ; Step S342: Matrix HAA （1） and matrix HBB （1） Vertical stacking to get the application change matrix HAB on day 1 （1） ; Similarly, the matrix HAA （td） and matrix HBB （td） Stack vertically to get the application change matrix HAB on the tdth day （td） ; Step S343: Analyze the usage habits of the first application of the target user 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}, denoted as matrix ee; Define a (un×1) all-one matrix, denoted as matrix fo; Define the relationship B-2-1: Among them, HAB （t） represents the application change matrix on day t; Count the number of items that satisfy the relationship B-2-1; Define a zero matrix (un×1) and fill the first row of the matrix with 1 to obtain the matrix ff; Define the relationship B-2-2: ; Count the number imm that satisfies the relation B-2-1; Calculate the usage rate of the first application uu （1） ,uu （1） =imm / (td-ill); Non-use rate （1） , iu （1） =1-uu （1） ; Step S344: Repeat the same calculation steps to calculate the usage rate uu of the second to unth applications （2） ~uu （un） and non-use rate iu （2） ~iu （un） .

[0016] A speech recognition system for a TV stick based on artificial intelligence includes: Voice acquisition module: used to acquire the target user's operating voice and the tester's test voice; perform pre-emphasis, windowing and Fourier transform on the test voice, and calculate the MFCC coefficient sequence of each test voice; Speech analysis module: used to perform secondary processing of homomorphic and cepstrum processing on the test speech, calculate the LPCC coefficient sequence of each test speech, and build a speech recognition model in combination with the MFCC coefficient sequence of each test speech; use the operation speech as the output of the speech recognition model to determine whether the speech recognition model has output; if so, convert the operation speech into a TV stick command and skip the command reasoning module; If it does not exist, the current time and the target user's application usage data are obtained and the instruction reasoning module is entered; Command reasoning module: used to obtain the current time and the number of controllable applications of the TV stick; analyze the historical application usage data of the target user and build an application usage model for the target user; obtain the name of the application that the target user is currently using or closing, and reversely infer the TV stick command of the operation voice based on the application usage model.

[0017] Compared with the prior art, the present invention has the following beneficial effects: Improve user experience: The present invention allows users to search for programs, switch channels, adjust volume, etc. through voice commands without having to manually operate the TV stick; this interactive method is more natural and conforms to the usage habits of most people, especially when the user's hands are occupied or the user has difficulty in moving.

[0018] Improve interaction efficiency: The voice recognition system can convert voice signals into text or commands within milliseconds and immediately execute corresponding operations, greatly shortening the user's waiting time; compared with the operation method of traditional remote controls, the present invention can accurately recognize the user's verbal operation instructions through the voice recognition system, reducing the occurrence of manual misoperations.

[0019] Personalized service: When the user's voice cannot be recognized or the voice information is wrong, the present invention can analyze the user's usage habits and characteristics, infer the user's operation commands, and provide the user with personalized voice services. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings: Figure 1 It is a schematic diagram of the method of the present invention; Figure 2 Schematic diagram of the system of the present invention. DETAILED DESCRIPTION

[0021] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Embodiment 1 See also Figure 1 , a speech recognition method for a TV stick based on artificial intelligence includes: Step S1: Acquire the target user's operating voice and the tester's test voice; perform pre-emphasis, windowing and Fourier transform on the test voice, and calculate the MFCC coefficient sequence of each test voice; It should be noted that the "target users" in the present invention refer to users who use TV sticks; the testers refer to volunteers or staff who participate in the development of the speech recognition function of the present invention; Test voice representation: voice related to operating or controlling the TV stick, such as TV stick instructions, TV program names, website or application names, etc. The specific steps of step S1 are as follows: Step S11: Count the number of test voices, recorded as vn; count the duration of the 1st to vnth test voices, recorded as vo （1） ~vo （vn） (Unit: seconds); Extraction （1） ~vo （vn） The maximum value in is denoted as vo （max） , the minimum value is denoted as vo （min） ; calculatedvo （1） ~vo（vn） The average value, denoted as vo （ave） ; Step S12: Define calculation formula A1: ; Among them, （i） Indicates the duration of the i-th test speech, vo （ve） indicates (vo （1） ~vo （vn） The corresponding weighted duration; the value range of i is: 1~vn; Get the sampling rate of the test speech, recorded as fs (unit: kilohertz); Calculate the frame duration of a single test speech, recorded as fd, fd = vo （ve） / fs (fd unit: milliseconds); The frame shift duration of a single test speech is recorded as fsd, fsd = fd / 2; Step S13: performing pre-emphasis, windowing and Fourier transform enhancement processing on the first test speech, and performing Mel filtering and discrete cosine transform to obtain MFCC coefficients of the first test speech; Step S131: Take the first test speech as speech v; use a high-pass filter to remove low-frequency noise in speech v to obtain speech v （1） ; Based on the frame length fd and frame shift length fsd, the speech v （1） Perform frame processing to obtain the voice v （2） ; Voice （2） The number of samples (i.e., the number of samples) is recorded as Fn, Fn = vo （1） / fd; Fn rounds up; Step S132: Use the Hamming window to （2） Perform windowing processing to obtain the voice v （3） ; Voice （3） With voice v （2） The same number of Using Fourier transform, the speech v （3） From the time domain information to the frequency domain information, we get the voice v （4） ; The mathematical expression of Fourier transform is: ; Where k represents the voice v （4） (i.e. frequency domain information), n represents the index of speech v （4） (i.e., time domain information) index, the value range of k and n is: 1 to Fn; KX （k） Represents the speech v at the kth timestamp （4）(i.e. frequency domain information), x （n） Represents the voice v at the nth timestamp （3） (i.e. real-time information); KX （1） Indicates the first timestamp of the voice v （4） (i.e. frequency domain information), x （1） Indicates the first time stamp of the voice v （3） (i.e., time domain information); e represents a natural constant, j represents an imaginary unit (i.e. j 2 =-1), π represents pi; Step S133: Obtain the voice v corresponding to the 1st to Fnth timestamps （4） The frequency, denoted as f （1） ~f （Fn） ; Using a Mel-scale triangular filter, the received speech v （4） (i.e. frequency domain information) simulates the nonlinear frequency perception of human hearing and obtains the speech v （5） ; The number of triangular filters is Fn, and the frequency scanning range is: 0 Hz ~ (2 / Fn) Hz; The reason for setting the number of triangular filters to F here is to prevent two voices v at adjacent timestamps from （4） When performing Mel frequency conversion, interference occurs; users or relevant technicians can adjust the number of triangular filters according to actual needs; Step S134: The voice v corresponding to each timestamp （4） The frequency is converted to Mel frequency, and the conversion formula is as follows: ; Among them, f （l） Indicates the lth timestamp information v （4） Frequency; F mel (f （l） ) indicates f in Mel （l） The perceived frequency (i.e., Mel frequency); the value range of l is: 1 to Fn; Summary F （mel） (f （1） )~F （mel） (f （Fn） ) as voice v （5） ; Step S135: （4） Perform discrete cosine transform on the perceived frequency to obtain the MFCC coefficient of the first test speech; The formula for discrete cosine transform is as follows: ; Among them, MFCC （l）Indicates the lth timestamp voice v （5） Corresponding MFCC coefficients; Aggregate MFCC （1） ~MFCC （Fn） , as the MFCC coefficient sequence of the first test speech; Step S14: Repeat the same steps of processing the first test speech (i.e., steps S131 to S135), process the second to vath test speech, and obtain the MFCC coefficient sequences of the second to vbth test speech; Step S2: Perform secondary processing of homomorphic and cepstrum processing on the test speech, calculate the LPCC coefficient sequence of each test speech, and build a speech recognition model in combination with the MFCC coefficient sequence of each test speech; take the operation speech as the output of the speech recognition model, and determine whether the speech recognition model has output; if it does, convert the operation speech (of the target user) into a TV stick command (operate the TV stick), and skip step S3; If it does not exist, then obtain the historical application usage data of the current time and the target user (in the past month) and proceed to step S3; The specific steps of step S2 are: Step S21: performing homomorphic and cepstrum processing on the test speech to obtain LPCC coefficients of the test speech information; Step S211: performing homomorphic and cepstrum processing on the first test speech information; Get the frequency domain information of the first test voice v （4） As information E; obtain the frequency corresponding to each information E (that is, the voice v corresponding to the 1st to Fnth timestamps （4） frequency), denoted as fx （1） ~fx （Fn） ; Step S212: Perform Z transformation on the information E corresponding to the 1st to Fnth timestamps to obtain XZ （l） ~XZ （Fn） ; The function expression of Z transform is: ; Among them, fx （l） Indicates the frequency of the information E corresponding to the lth timestamp; ZX （l） Indicates fx （l） Z transform output; zz （l） Indicates fx （l） The corresponding complex variables are: ; where vc represents the speed of sound, e represents a natural constant; j represents an imaginary unit (j2 = -1); ω represents fx （l） The corresponding phase: ω=arctan[(vc×fx （l） ) / fd]; where arctan represents the inverse tangent function; fd represents the frame duration; Step S213: Perform logarithmic transformation on the information E corresponding to the 1st to Fnth timestamps to obtain ln(ZX （1） )~ln(ZX （Fn） ); The function expression of logarithmic transformation is: ; Among them, ln (ZX （l） ) indicates that ZX （l） Logarithmic conversion of the output; fx （q） represents the frequency of information E corresponding to the qth timestamp; zz （q） Indicates fx （q） The corresponding complex variable; the value range of q is: 1~Fn; Use the iztrans function to transform ln(ZX （1） )~ln(ZX （Fn） ) performs inverse Z transform processing to convert the information E corresponding to the 1st to Fnth timestamps from frequency domain information to time domain information, and obtains ts （1） ~ts （Fn） ; Step S214: ts （1） ~ts （Fn） Perform cepstrum processing to obtain the LPC coefficients of information E corresponding to the 1st to Fnth timestamps, denoted as ce （1） ~ce （Fn） ; The function expression of cepstrum processing is: ; Among them, ts （m） represents the time domain information of the information E corresponding to the mth timestamp, fx （l-m+1） Indicates the frequency of the information E corresponding to the (l-m+1)th timestamp; the value range of m is: 1~l; Using the discrete cosine transform algorithm, ce （1） ~ce （Fn） Convert the time domain information into the frequency domain information to obtain the LPCC coefficient sequence of the first test speech information; Step S215: repeating the same steps of performing homomorphic and cepstrum processing on the first test speech information (i.e., steps S211 to S214), processing the second to vath test speech information, and obtaining the LPCC coefficient sequences of the second to vath test speech information; Step S22: Use the recognize_google function and recognize_sphinx function in the SpeechRecognition library to convert the test speech information from the 1st to the vnth pieces into test speech text, recorded as text wd （1） ~wd （vn） ; The test speech information is summarized as a sample set in 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 basic model of the speech recognition model, define the training steps for supervised training of the basic model: Step S231: The MFCC coefficient sequences corresponding to the 1st to the vnth items are taken as a high-dimensional array, denoted as Cmf （1） ~Cmf （vn） ; Define the set CMF, CMF = {Cmf （1） ~Cmf （vn）}; The LPCC coefficient sequence corresponding to the 1st to vnth items is taken as a high-dimensional array, denoted as Clp （1） ~Clp （vn） ; Define the set CLP, CLPF = {Clp （1） ~Clp （vn）}; Step S232: Fusing the set Cmf and the set Clp, we get the set of feature sequences of the TV stick operation command voiceprints, which is recorded as set CC, where CC = {(Cmf （1） , Clp （1） )~(Cmf （vn） , Clp （vn） )}; Step S233: Write the text wd （1） ~wd （vn） Fusion with the set CC yields the feature set of the test speech, denoted as set α; α = {(wd （1） , Cmf （1） , Clp （1） )~(wd （vn） , Cmf （vn） , Clp （vn） )}; Set the number of states in the base model to 4 and the number of Gaussian elements to 3; Step S234: Bringing the set α into the basic model, and iterating until all the test speech corresponding to the set α is output by the basic model; Use cross-validation algorithms to prevent overfitting of the base model during supervised training; Define the judgment condition for the convergence of the speech recognition model: the probability of any test speech being misrecognized is less than 1×10 -3 ; Step S24: Define the training steps for unsupervised training of the basic model: Step S241: Get set CC (i.e., the voiceprint feature sequence set of the TV stick operation command), use set CC as training set 1, and use the test voice text as training set 2; iterate the basic model using the K-means clustering algorithm until the elements in training set 1 and the elements in training set 2 are output simultaneously and the output elements correspond (i.e., the voice text output from training set 1 corresponds to the voiceprint feature sequence output from training set 2); Step S242: using a cross-validation algorithm to prevent the basic model from overfitting during unsupervised training; The convergence criteria of the speech recognition model are defined as follows: the probability difference between the two consecutive recognition outputs of the same test speech is less than 1×10 -4 ; Step S25: Perform supervised training and unsupervised training on the basic model until the output results of the supervised training and the unsupervised training of the basic model are the same (model training is completed), thereby obtaining a speech recognition model; Use the operation voice as the input of the speech recognition model to determine whether the (speech recognition model) has output; If yes, then convert the target user's operation voice into a TV stick command (operate the TV stick), and skip step S3; If not, the current time and the historical application usage data of the target user are obtained, and the process goes to step S3.

[0023] Step S3: Obtain the current time and the number of controllable applications on the TV stick; analyze the historical application usage data of the target user (in the past month) and build an application usage (habit) model for the target user; obtain the names of applications that the target user is using or closing, and reversely infer the TV stick instructions of the operation voice based on the application usage (habit) model; The specific steps of step S3 are as follows: Step S31: Get the current time and record it as (mo month da day) ho hour; The historical application usage data includes: the target user's usage time and usage of the 1st to unth applications during the time period from the hoth hour to the (ho+1th hour) last month; Get the number of applications that can be operated on the TV stick, recorded as un; get the usage time ti of the target user (through the TV stick) using the first application from time ho to time (ho+1) last month （1，1） ~ti （td，1） ; Usage time of the second application（2，1） ~ti （td，2） ; And so on, the usage time of the un application is （td，1） ~ti （td，un） (Unit: minutes); where td represents the total number of days in the previous month; It should be noted that if ho is 23:00 on a certain day, then (ho+1) is 0:00 on the next day; for example, if ho is 23:00 on March 1, then (ho+1) is 0:00 on March 2. Step S32: Corresponding calculation formula B-1: ; Among them, ut （s） represents the time coefficient of the sth application (used by the target user), ti （s，t） Indicates the time the target user used the t-th application on the s-th day (in the last month); the value range of s is: 1~td, and the value range of t is: 1~un; (According to calculation formula B-1) calculate the time coefficient ut of the target user (in the last month) for the first to unth applications （1） ~ut （un） ; Step S33: define the tuple ha （s，t） The determination coefficient of the target user (in the last month) using the t-th application at the ho-th time on the s-th day; define the bigram hb （s，t） represents the determination coefficient of the target user (in the last month) using the t-th application at the (ho+1)th time on the s-th day; ha （s，t） and hb （s，t） ∈{0,1}; (ha （s，t） is 0, indicating that the target user did not use the t-th application at the ho-th time on the s-th day (in the last month); （s，t） 1 means the target user used the t-th application at the ho-th time on the s-th day (last month); hb （s，t） is 0, indicating that the target user (in the last month) did not use the tth application at the (ho+1)th time on the sth day; hb （s，t） 1 means the target user used the t-th application at (ho+1) on the s-th day (last month); (Use the two-tuple ha （s，t） ) Define the determination coefficient of the target user (in the last month) from the first day at time ho to the td day at time ho, the 1st to the unth application: ha （1，1） ~ha （td，un） ; (Use the binary hb （s，t）) Define the determination coefficient of the target user (in the last month) from the first day (ho+1) to the td day (ho+1), the first to the unth application: 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 (in the time period from time ho to time (ho+1)) for the 1st to unth applications, and build the application usage (habit) model of the target users; Step S341: define an empty matrix of (1×un) and fill in ha （1，1） ~ha （1，un） and hb （1，1） ~hb （1，un） , we get the matrix HAA （1） and matrix HBB （1） ; Among them, ha （1，un） and hb （1，un） , respectively represent the determination coefficient of the target user (in the last month) at time ho on the first day and at time (ho+1) on the td day; Fill in ha （2，1） ~ha （2，un） and hb （2，1） ~hb （2，un） , we get the matrix HAA （2） and 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 target user (in the last month) for the 1st to unth applications from time ho to time (ho+1) on the second day and from time ho to time (ho+1) on the second day; And so on, fill in ha （td，1） ~ha （td，un） and hb （td，1） ~hb （td，un） , we get the matrix HAA （td） and matrix HBB （td） ; Among them, ha （td，1） ~ha （td，un） and hb （td，1） ~hb （td，un） , represent the determination coefficients of the target user (in the last month) from the time ho to the time (ho+1) on the td day and from the time ho to the time (ho+1) on the td day for the 1st to unth applications respectively; Step S342: Matrix HAA （1）and matrix HBB （1） Vertical stacking to get the application change matrix HAB on day 1 （1） ; The matrix HAA （2） and matrix HBB （2） Vertical stacking to obtain the application change matrix HAB on the second day （2） ; By analogy, the matrix HAA （td） and matrix HBB （td） Stack vertically to get the application change matrix HAB on the tdth day （td） ; Step S343: Analyze the usage habits of the first application of the target user from time ho to time (ho+1) (last month), and calculate the usage rate uu of the first application （1） and non-use rate iu （1） ; Define a (1×2) matrix {0,1}, denoted as matrix ee; Define a (un×1) all-one matrix, denoted as matrix fo; Define the relationship B-2-1: Among them, HAB （t） represents the application change matrix on day t; HAB （1） ~HAB （td） Substitute into equation B-2-1, count the number of values ​​that satisfy equation B-2-1, and record it as ill; Define a zero matrix of (un×1) and fill the first row of the matrix with 1 to obtain the matrix ff; Define the relationship B-2-2: ; HAB （1） ~HAB （td） Substitute into relation B-2-2, count the number of values ​​that satisfy relation B-2-1, and record it as imm; Calculate the usage rate of the first application uu （1） ,uu （1） =imm / (td-ill); Non-use rate （1） , iu （1） =1-uu （1） ; Step S344: Repeat the same calculation steps to calculate the usage rate uu of the second to unth applications （2） ~uu （un） and non-use rate iu （2） ~iu （un） ; Step S345: Analyze the influence of different applications on the first application, and calculate the collaborative utilization rate lx of different applications on the first application. （1-1） ~lx （un-1） ; Step S3451: Calculate the collaborative usage rate lx of the first application to the second application （1-1） ; Define a (1×2) matrix {1,0}, denoted as matrix rr; define a (un×1) zero matrix, denoted as matrix zf; Fill the first row of the matrix zf with 1 to get the matrix zf （1） ; Define relation B-3-1: ; HAB （1） ~HAB （td） Substitute into equation B-3-1 and count the number of numbers that satisfy equation B-3-1, denoted as oo （1） ; Calculate the collaborative usage rate of the first application to the first application lx （1-1） , lx （1-1） =oo （1） / (td×un); Step S3452: Calculate the collaborative usage ratio lx of the second application to the first application （2-1） ; Fill the second row of the matrix zf with 1 to get the matrix zf （2） ; Define relation B-3-2: ; HAB （1） ~HAB （td） Substitute into equation B-3-2 and count the number of numbers that satisfy equation B-3-2, denoted as oo （2） ; Calculate the collaborative usage rate of the second application to the first application lx （2-1） , lx （2-1） =oo （2） / (td×un); Step S3453: Similarly, calculate the collaborative usage rate lx of the unth application to the first application （un-1） ; Fill the unth row in the matrix zf with 1 to get the matrix zf （un） ; Define the relationship B-3-3: ; HAB （1） ~HAB （td） Substitute into equation B-3-3 and count the number of numbers that satisfy equation B-3-3, denoted as oo（un） ; Calculate the collaborative usage rate of the unth application to the first application lx （un-1） , lx （un-1） =oo （un） / (td×un); Step S346: Repeat the calculation of lx （1-1） ~lx （un-1） The same steps as in , calculate the collaborative usage rate lx of the 1st to unth applications to the 2nd to unth applications （1-2） ~lx （un-un） ; Summarize the usage rate, non-usage rate, and collaborative usage rate of the first to unth applications as the target user's usage (habit) model of the application; Step S35: obtaining the name of the application that the target user is using or closing (at the time of mo month da day ho) as the quasi-application; and reversely inferring the TV stick instruction of the operation voice according to the target user's usage habits of different applications; Step S351: Obtain the usage rate of the quasi-application, denoted as quu, and the time coefficient base as qu; Among them, quu∈{uu （2） ~uu （un）}, qu∈{ut （1） ~ut （un）}; Calculate the probability that the target user continues to use the quasi-application, denoted as ccg, ccg = qu × (quu) 2 ; Step S352: Obtain the usage rate of the non-quasi-application, denoted as iqu （1） ~iqu （an） (Among them, {iqu （1） ~iqu （an）}⊆{uu （2） ~uu （un）}), the non-use rate is recorded as nqu （1） ~nqu （an） (Among them, {nqu （1） ~nqu （an）}⊆{iu （2） ~iu （un）}), the time coefficient is recorded as iut （1） ~iut （an） ({iut （1） ~iut （an）}⊆{ut （1） ~ut （un）}); Where an represents the number of non-quasi-applications, an<un; Get the collaborative usage rate of the quasi-application to the first to anth applications, denoted as qlx （1） ~qlx （an） ; Among them, {qlx （1） ~qlx （an）}⊆{lx （1-1） ~lx （un-un）}; Define calculation formula B-4: ; Among them, （y） represents the probability that the target user selects the yth non-quasi application in (ho+1), qlx （y） ,iqu （y） and iut （y） , respectively represent the collaborative utilization rate, utilization rate and time coefficient of the quasi-application to the yth non-quasi-application; qx （p） It represents the non-usage rate of the quasi-application to the pth non-quasi-application, and y and p cannot take the same value at the same time; Step S353: (according to calculation formula B-4) calculate the probability that the target user selects the first to anth non-quasi-applications in (ho+1), qqg （1） ~qqg （an） ; Extract ccg and qqg （1） ~qqg （an） The maximum value among is used as maa; and a TV stick instruction for controlling the application corresponding to maa is generated.

[0024] Embodiment 2 See also Figure 2 , a speech recognition system for a TV stick based on artificial intelligence includes: Voice acquisition module: used to acquire the target user's operating voice and the tester's test voice; perform pre-emphasis, windowing and Fourier transform on the test voice, and calculate the MFCC coefficient sequence of each test voice; Speech analysis module: used to perform secondary processing of homomorphic and cepstrum processing on the test speech, calculate the LPCC coefficient sequence of each test speech, and build a speech recognition model in combination with the MFCC coefficient sequence of each test speech; take the operation speech as the output of the speech recognition model, and determine whether the speech recognition model has output; if it does, convert the operation speech (of the target user) into a TV stick command (operate the TV stick), and skip the command reasoning module; If it does not exist, the current time and the target user's application usage data are obtained and the instruction reasoning module is entered; Command reasoning module: used to obtain the current time and the number of controllable applications on the TV stick; analyze the historical application usage data of the target user (in the past month) and build an application usage (habit) model for the target user; obtain the name of the application that the target user is using or closing, and reversely infer the TV stick command of the operation voice based on the application usage (habit) model.

[0025] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technical personnel in this field according to actual conditions. For example, if there are weight coefficients and proportional coefficients, their set sizes are to quantify each parameter to obtain a specific value, which is convenient for subsequent comparison. Regarding the size of the weight coefficient and the proportional coefficient, it is sufficient as long as it does not affect the proportional relationship between the parameter and the quantized value.

[0026] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A speech recognition method for a TV stick based on artificial intelligence, characterized in that: The method comprises: Step S1: Acquire the target user's operating voice and the tester's test voice; perform pre-emphasis, windowing and Fourier transform on the test voice, and calculate the MFCC coefficient sequence of each test voice; Step S2: Perform secondary processing of homomorphic and cepstrum processing on the test speech, calculate the LPCC coefficient sequence of each test speech, and build a speech recognition model in combination with the MFCC coefficient sequence of each test speech; take the operation speech as the output of the speech recognition model, and determine whether the speech recognition model has output; if so, convert the operation speech into a TV stick command and skip step S3; If not, then obtain the current time and the historical application usage data of the target user and proceed to step S3; Step S3: Get the current time and the number of controllable applications on the TV stick; analyze the historical application usage data of the target user and build an application usage model for the target user; get the name of the application that the target user is using or closing, and reversely infer the TV stick instructions of the operation voice based on the application usage model.

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

3. The method for speech recognition based on artificial intelligence for a TV stick 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 the low-frequency noise in speech v to obtain speech v （1） ; Use fd and fsd to compare the voice v （1） Perform frame processing to obtain the voice v （2） ; Voice （2） The number Fn, Fn = vo （1） / fd; Step S132: Use the Hamming window to （2） Perform windowing processing to obtain the voice v （3） ; The voice v （3） From the time domain information to the frequency domain information, we get the voice v （4） ; Step S133: Obtain the voice v corresponding to the 1st to Fnth timestamps （4） The frequency f （1） ~f （Fn） ; Using a triangular filter, receive the speech v （4） Nonlinear frequency perception, get the speech v （5） ; Step S134: The voice v corresponding to each timestamp （4） The frequency is converted to Mel frequency, and the conversion formula is as follows: ; Among them, f （l） Indicates the lth timestamp information v （4） Frequency; F mel (f （l） ) indicates f in Mel （l） The perceived frequency; Summary F （mel） (f （1） )~F （mel） (f （Fn） ) as voice v （5） ; Step S135: （4） Perform discrete cosine transform on the perceived frequency to obtain the MFCC coefficient of the first test speech; The formula for discrete cosine transform is as follows: ； Among them, MFCC （l） Indicates the lth timestamp voice v （5） Corresponding MFCC coefficients; Aggregate MFCC （1） ~MFCC （Fn） , as the MFCC coefficient sequence of the first test speech.

4. The method for speech recognition based on artificial intelligence for a TV stick according to claim 2, characterized in that: The specific steps of step S2 are: Step S21: performing homomorphic and cepstrum processing on the test speech to obtain LPCC coefficients of the test speech information; Step S22: Convert the first to the vnth test voice messages into test voice texts, recorded as text wd （1） ~wd （vn） ; The test speech information is aggregated as a sample set in the format of test speech text—MFCC coefficient sequence of the test speech information—LPCC coefficient sequence of the test speech information; Step S23: Using the CNN model as the basic model of the speech recognition model, define the training steps for supervised training of the basic model: Step S231: taking the MFCC coefficient sequences corresponding to the 1st to the vnth items as the set CMF; The LPCC coefficient sequences corresponding to the 1st to the vnth items are taken as high-dimensional arrays and integrated to obtain the set CLP; Step S232: merging the set Cmf and the set Clp to obtain a set of feature sequences of the TV stick operation command voiceprints, recorded as set CC; Step S233: Write the text wd （1） ~wd （vn） Fuse with the set CC to obtain the feature set of the test speech, denoted as set α; Set the number of states in the base model to 4 and the number of Gaussian elements to 3; Step S234: Bringing the set α into the basic model, and iterating until all the test speech corresponding to the set α is output by the basic model; Define the judgment condition for the convergence of the speech recognition model: the probability of any test speech being misrecognized is less than 1×10 -3 .

5. The method for speech recognition based on artificial intelligence for a TV stick according to claim 4, characterized in that: The subsequent steps of step S23 are as follows: Step S24: Define the training steps for unsupervised training of the basic model: Step S241: obtaining a set CC, taking the set CC as training set 1, and taking the test speech text as training set 2; iterating the basic model using the K-means clustering algorithm until the elements in the training set 1 and the elements in the training set 2 are output simultaneously and the output elements correspond; Step S242: define the judgment condition for the convergence of the speech recognition model: the probability difference between the two consecutive recognition outputs of the same test speech is less than 1×10 -4 ; Step S25: Perform supervised training and unsupervised training on the basic model until the output results of the supervised training and the unsupervised training of the basic model are the same, thereby obtaining a speech recognition model; Use the operation voice as the input of the speech recognition model to determine whether the model has output; If yes, the operation voice is converted into a TV stick command and step S3 is skipped; If not, the current time and the historical application usage data of the target user are obtained, and the process goes to step S3.

6. The method for speech recognition based on artificial intelligence for a TV stick according to claim 4, characterized in that: The specific steps of step S21 are as follows: Step S211: performing homomorphic and cepstrum processing on the first test speech information; Get the frequency domain information of the first test voice v （4） As information E; obtain the frequency fx corresponding to each information E （1） ~fx （Fn） ; Step S212: Perform Z transformation on the information E corresponding to the 1st to Fnth timestamps to obtain XZ （l） ~XZ （Fn） ; The function expression of Z transform is: ; Among them, fx （l） Indicates the frequency of the information E corresponding to the lth timestamp; ZX （l） Indicates fx （l） Z transform output; zz （l） Indicates fx （l） The corresponding complex variables are: ; vc represents the speed of sound, ω represents fx （l） The corresponding phase: ω=arctan[(vc×fx （l） ) / fd]; fd indicates the frame duration; Step S213: Perform logarithmic transformation on the information E corresponding to the 1st to Fnth timestamps to obtain ln(ZX （1） )~ln(ZX （Fn） ); The function expression of logarithmic transformation is: ； Among them, ln (ZX （l） ) indicates that ZX （l） Logarithmic conversion of the output; fx （q） represents the frequency of information E corresponding to the qth timestamp; zz （q） Indicates fx （q） The corresponding complex variable; For ln (ZX （1） )~ln(ZX （Fn） ) performs inverse Z transform processing to convert the information E corresponding to the 1st to Fnth timestamps from frequency domain information to time domain information, and obtains ts （1） ~ts （Fn） .

7. The artificial intelligence-based voice recognition method for a TV stick according to claim 6, characterized in that: The subsequent steps of step S213 are as follows: Step S214: ts （1） ~ts （Fn） Perform cepstrum processing to obtain the LPC coefficients of information E corresponding to the 1st to Fnth timestamps, denoted as ce （1） ~ce （Fn） ; The function expression of cepstrum processing is: ； Among them, ts （m） represents the time domain information of the information E corresponding to the mth timestamp, fx （l-m+1） Indicates the frequency of the information E corresponding to the (l-m+1)th timestamp; Using the discrete cosine transform algorithm, ce （1） ~ce （Fn） Convert the time domain information into the 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.

8. The method for speech recognition based on artificial intelligence for a TV stick according to claim 5, characterized in that: The specific steps of step S3 are as follows: Step S31: Obtain the number of applications that can be operated on the TV stick un; obtain the usage time ti of the target user using the first application last month （1，1） ~ti （td，1） Similarly, the usage time of the un application is （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） represents the time coefficient of the sth application, ti （s，t） Indicates the time when the target user uses the t-th application on the s-th day; Calculate the time coefficient ut of the target user's first to unth applications （1） ~ut （un） ; Step S33: From time ho on day 1 to time ho on day td, the determination coefficient of the first to un types of applications: ha （1，1） ~ha （td，un） ; From the first day at (ho+1) to the td day at (ho+1), the determination coefficient of the first to un applications is: 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 first to unth applications by the target users and build 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; and reversely infer the TV stick instruction of the operation voice according to the target user's usage habits of different applications.

9. The method for speech recognition based on artificial intelligence for a TV stick according to claim 8, characterized in that: The specific steps of step S34 are as follows: Step S341: According to ha （1，1） ~ha （td，un） , construct the matrix HAA （1） ~HAA （td） and matrix HBB （1） ~HBB （td） ; Step S342: Matrix HAA （1） and matrix HBB （1） Vertical stacking to get the application change matrix HAB on day 1 （1） ; Similarly, the matrix HAA （td） and matrix HBB （td） Stack vertically to get the application change matrix HAB on the tdth day （td） ; Step S343: Analyze the usage habits of the first application of the target user 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}, denoted as matrix ee; Define a (un×1) all-one matrix, denoted as matrix fo; Define the relationship B-2-1: Among them, HAB （t） represents the application change matrix on day t; Count the number of items that satisfy the relationship B-2-1; Define a zero matrix of (un×1) and fill the first row of the matrix with 1 to obtain the matrix ff; Define the relationship B-2-2: ; Count the number imm that satisfies the relation B-2-1; Calculate the usage rate of the first application uu （1） ,uu （1） =imm / (td-ill); Non-use rate （1） , iu （1） =1-uu （1） ; Step S344: Calculate the usage rate and non-usage rate of the 2nd to unth applications.

10. A speech recognition system for a TV stick based on artificial intelligence, applicable to a speech recognition method for a TV stick based on artificial intelligence as claimed in any one of claims 1 to 9, characterized in that: The system comprises: Voice acquisition module: used to acquire the target user's operating voice and the tester's test voice; perform pre-emphasis, windowing and Fourier transform on the test voice, and calculate the MFCC coefficient sequence of each test voice; Speech analysis module: used to perform secondary processing of homomorphic and cepstrum processing on the test speech, calculate the LPCC coefficient sequence of each test speech, and build a speech recognition model in combination with the MFCC coefficient sequence of each test speech; use the operation speech as the output of the speech recognition model to determine whether the speech recognition model has output; if so, convert the operation speech into a TV stick command and skip the command reasoning module; If it does not exist, the current time and the target user's application usage data are obtained and the instruction reasoning module is entered; Command reasoning module: used to obtain the current time and the number of controllable applications of the TV stick; analyze the historical application usage data of the target user and build an application usage model for the target user; obtain the name of the application that the target user is currently using or closing, and reversely infer the TV stick command of the operation voice based on the application usage model.

Images