TW201216253A - A speech recognition method on sentences in all languages - Google Patents

Abstract

The invention can recognize sentences in all languages. A sentence can be a syllable, word, name or sentence. The most important feature of this invention is that all sentences are represented by E*P=12*12 matrices of linear predict coding cepstra (LPCC), produced by E=12 equal-sized elastic frames without filter and without overlap. The prior speech recognition methods have to compute and compare a series of matrices of features of words, but the invention only computes and compares an E*P matrix of LPCC. 1000 different voices are transformed into 1000 different matrices of LPCC to represent 1000 different databases. The E*P matrices of known sentences after deletion of noise and time intervals between two words and two syllables are put into their closest databases. To classify an unknown sentence, use the distance to find its F closest databases from 1000 different databases and then from known sentences in the F closest databases, find a known sentence to be the unknown one. The invention needs no samples and can find a sentence in one second using Visual Basic. Any person without training can immediately and freely communicate with a computer in any language. It can recognize up to 7200 English words, 500 sentences (any language) and 500 Chinese words and input 4400 Chinese words.

Description

201216253 VI. Description of the Invention: [Technical Field to Which the Invention Is Ascribed] The present invention recognizes sentences in any language. The invention uses 12 elastic frame (window) 'isometric' non-verifiers, does not overlap, converts the sound waves of - a plurality of words into a long sentence: a linear predictive coding cepstrum of ExP=12xl2 (LPCC) ). Will all be touched (four)? _ likeness is first divided into - thousands of different databases. When identifying an unknown sentence, convert it to the ExP linear predictive coding cepstrum matrix, and then use the unknown sentence Εχρ matrix to find the closest database & the closest data from the -1000 database. Known sentences in the library, use distance to find unknown sentences. After the user of the field is pronounced, with Visual Basic, it is possible to identify the desired sentence in less than - seconds. The method is simple, no sample is needed, anyone can use it in time, and the pronunciation is not standard or the wrong sound can be used. To calculate and compare a sentence feature value, the present invention only needs to calculate and compare a sentence secret matrix value. Fast, ^ sure rate. The pronunciation of the official language, Taiwanese, English, Japanese, and German have all been tested with β recognizable large 4 voices. This wealthy book, single mathematics calculation and identification, and quasi-fast. [Prior Art] Generally, an unknown sentence is firstly divided into a plurality of single 201216253 sounds or words, _ is a heterogeneous technique, especially English, and an English single word has multiple syllables, which is difficult to cut. Accurate... The difference between syllables will know the sentence: Admit mistakes. Therefore, when speaking, be careful, be slow, be clear, and have a long word interval. Then, the unknown single sub-words are compared with the known single words of the (4) library, and a single word error will make the unknown sentence recognize the wrong one. Then, the known words found in the single-word database are concatenated into a known sentence according to the order of the unknown sentences, and the most likely known sentence is the unknown sentence from the sentence database. It is often difficult to accurately identify an unknown sentence method. It is time-consuming and cannot talk freely with the computer. - The method of identification must take time to do samples, which must be calculated and identified by statistics. Of course, it is not allowed. Statistics can only be estimated - the pronunciation of a sentence is a silk thread. Sound waves are systems that make non-linear changes over time. - The sentence sound waves contain a dynamic characteristic that also changes nonlinearly with time. When the same sentence is pronounced, there are a series of identical dynamic characteristics that nonlinearly stretch and contract over time, but the same dynamic characteristics are arranged in time-ordered but different in time. When the same sentence is pronounced, it is very difficult to arrange the same dynamic characteristics # at the same time position. A computerized language recognition system first extracts sound waves about language information, that is, dynamic characteristics, transitions, and language-independent noises, such as human voices, tones, speech screams, physiological readings, and linguistic nuances. The same features of the same sentence are then arranged at the same time position. This series of features is represented by an isometric vector, called a sentence _ sign model. At present, the speech recognition system is too complicated to generate a large-scale feature, and it costs 201216253' because the same features of the same sentence are difficult to arrange at the same time position, especially in English, which makes the comparison difficult. A continuous acoustic feature is commonly used in the following ways: energy, zero crossings, 'exfreme c〇unt', peaks, linear predictive coding cepstrum (lpcc) And the Mel Frequency Cepstrum (MFCC), in which the linear predictive coding cepstrum (LpCC) and the Mel frequency cepstrum (MFCC) are the most effective and commonly used. Linear Predictive Coding Spring Frequency (LPCC) is the most reliable, stable, and accurate language feature for a continuous tone.线性 Linear regression mode is used to represent continuous sound waves, and the regression coefficient is calculated by the least squares estimation method, and its estimated value is converted into cepstrum, which becomes linear predictive coding cepstrum (LPCC). The Mel frequency cepstrum (mfcc) converts sound waves into frequencies using the Fourier transform method. The auditory system is estimated based on the ratio of the frequency of the Mel. root

According to scholar S.B, Davisand P. Mermelstein was published in IEEE in 1980.

Transactions on Acoustics, Speech Signal Processing, φ V〇丨.28, No. 4 Published by Parametric representations for monosyllabic word recognition in continuous speech Sentences using dynamic time warping (DTW), Mel frequency cepstrum (MFCC) The feature is higher than the linear predictive coding cepstrum (LPCC) feature recognition rate. However, after many speech recognition experiments (including my previous invention), the Bayesian classification method, the linear prediction coding cepstrum (LPCC) feature recognition rate is higher than the Mel frequency cepstrum (MFCC) feature, and saves time. [s] 7 201216253 There are many methods for language rhyme _ 1 dynamic time-warping, vector quantization (vect〇r quantizati〇n) and hidden Markov mode (HMM). If the same pronunciation changes in time, one side compares the same feature to the same time position. The recognition rate will be good' but it is difficult and too long to pull the same feature to the same position and cannot be applied. Vector quantization, such as identifying a large number of tones, is not only inaccurate, but also time consuming. Recently, the hidden Markov model method (HMM) identification method is good, but the method is complicated, too many unknown parameters need to be estimated, and the estimated value and the time limit for identification are calculated. T. F. Li (Li Zifen) published in 2003 in Pattern Recognition, vol. 36 Speech recognition of mandarin monosyllables’ Li, TzeFen (Li Zifen) in 1997 in the US Patent Certificate, Apparatus and Method for Normalizing and Categorizing

Linear Prediction Code Vectors using Bayesian Categorization

Technique,USA· Patent No. 5,704,004, Dec. 30, 1997, Li Zifen • In 2008, in the Republic of China Patent Certificate I 297487 (2008, 6, 1) name speech recognition method and Li Zifen in 2009 in the Republic of China patent certificate I 310543 (2009, 6, 1) Name A continuous quadratic Bayesian classification method for identifying similar Mandarin singles. Using the Bayesian classification method, the same database is used to compress a series of different LPCC vectors in various ways. Character model of the same size, identification results than YK·Chen, CYLiu, GH Chiang, MT Lin published in 1990 in proceedings 〇f Telecommunication [si 8 201216253

Symposium, Taiwan published paper The recognition of mandarin 〇 osyllables based on the discrete hidden Markov model HMM method with hidden Markov mode is better. However, the compression process is complicated and time consuming and the same tone is difficult to compress the same feature to the same time position, which is difficult to recognize for similar tones. The speech recognition method of the present invention is directed to the above shortcomings. From the academic point of view, according to the sound characteristics of the sound wave, the nonlinear change with time, naturally derives a set of extracted feature green, which is equal to any language. Εχ ρ = ΐΜ 2 matrix ” [Abstract] (1) The fourth objective of the present invention is to quickly and accurately identify any sentence in any language, previously calculated and compared - a single sub-word eigenvalue 'the present invention as long as Calculate and compare the value of a sentence Εχρ matrix to achieve the goal of free conversation with the computer. (2) In order to achieve the purpose of (1), the present invention applies a sentence sound wave normalization and extraction feature method. A small number of Ε ==12 _ rotational box , isometric, non-overlapping, no filter, can be according to a sentence with a short wavelength, freely adjust all the sentences containing the cover, return to all the dragons - to the length of multiple words - (four) all turn to honest scales Εχ ΐ = 2 2 linear prediction Encoding the cepstrum matrix. The intra-sentence-series changes nonlinearly with time. The 201216253 state property is converted into an equal-sized ExP linear prediction. Cepstrum matrix, and the feature model of the same sentence has the same feature at the same time position. It can be compared in time to achieve the real-time recognition effect of the computer. (3) The application of the invention - thousands of different databases, can identify a large number of sentences, and catch fast The accuracy rate is also greatly improved. The main known sentences are scattered in the database of the nearest sounds of thousands of databases. When the ship is unknown, first find the closest database with the unknown sentence sounds, and then from ρ The known sentences in the closest database find the unknown sentences to be identified. There are not many known sentences in the F closest database, and it is easy to identify and accurate. Previously calculated and compared For all single-character eigenvalue matrices in a sentence, the present invention only needs to calculate and compare an ExP matrix value of a sentence. (4) The present invention does not use samples for statistical calculation and recognition, and uses mathematical calculations and misinterpretation of sentences to encode cepstrums. The distance between the matrices is recognized. (5) The text of the confession can recognize the sentence in which the speech is too slow. When the speech is too fast, the sentence sound is very short, the present invention The length of the isometric elastic frame can be reduced to 'still use the same number of elastic frames with the same length to cover the short sound waves. Generate a linear predictive coding side spectrum (Lpcc) vector. Speak too slowly - a sentence is longer (4) The length of the elastic frame of the same length will be elongated. The same number of E linear predictive coding cepstrums 10 (6) 201216253 (LPCC) vector can also effectively represent the long sentence. The present invention provides a correction technique. Clearly pronounce the sentence that is being erroneous. [Embodiment] Use the first and second diagrams to invent the execution program. The first diagram is a table • A database is established, each database has similar known sentences. The figure shows that the user recognizes the unknown sentence execution program. There is a different sound, and a sound sound wave is converted into a digital sound point 10 to remove the noise or mute 2〇. The secret should be normalization of the sound sound wave = the extraction feature, the - scale sound sound wave all the money age into a handsome time slaves - each time frame - frame. A sound has a total of E equal length frames, no check n, * overlap, according to the length of the sound cast money, E phase • The length of the frame is free to adjust all the signal points covered. Therefore, the frame is called the same length elastic frame 'length freely stretchable, but E, the length of the elastic frame is the same." Unlike the Hamming window, there are filters, semi-overlaps, fixed lengths, and cannot be freely adjusted with wavelength. Because a sound wave changes nonlinearly with time, the sound wave contains a dynamic feature of the voice, which also changes nonlinearly with time. Because it does not overlap, the present invention uses less (E=12) equal-length elastic frames, covering the whole P sound sound wave 'Because the signal point can be estimated from the previous signal point, use the regression mode which changes linearly with time to closely estimate the nonlinear change sound wave, use 201216253 to take,], square silk phase county know silk. Using the top ten P-12 linear prediction codes to encode the cepstrum 4〇, one sound is represented by the ExP linear prediction coding cepstrum matrix, and the secret linear prediction coding of the sound is represented by a database. Thousands of funds;, 'Library 5Q clearly pronounces the known sentences to be recognized - times, except for mute and miscellaneous 9 before and after the sentence 'between words and two syllables, delete all mute and To convert the sentence to be recognized into an exp-predictive coded cepstrum (Lpcc) matrix 60 using E equal elastic boxes. Use the distance to classify the Εχρ linear predictive coding spectrum ((4)) matrix of the known sentence. To the nearest database (4). All must be identified _ known sentences are divided into Μ-1000 different databases. There are Μ = 1 database each database contains similar known sentences 8 &. > second The figure shows the process of identifying the unknown sentence method. First, the unknown #子β楚 pronunciation 2 is not. The (4) sub-sonic is digitized into signal points to remove the silence and murmur 2〇, before the unknown sentence and after the unknown sentence, Between the two words and the 1^ syllables, the mute and murmurs are deleted. The equal-length elastic frame normalizes the sound wave 'extraction feature'. The unknown sentences that are to be recognized are all divided into Ε _ segments, which are formed every time period— Elastic frame 3〇° - There are a total of rotating frames, no filters, no overlap, freely telescopic s cover Wang Zai points. In each pivot, because the signal points can be estimated by the front signal 12 201216253, using the minimum payment method Regression unknown system _ estimated value. Each box The resulting M2 minimum estimate is called linear test coding (10). To the summer, the linear predictive coding (LPC) vector is transformed into a more stable linear estimate, which is a vector, and an unknown (four)-a Εχρ linear estimate. The coded cepstrum matrix represents 4 本 发 _ 要 辨 辨 要 要 要 要 要 要 要 要 要 要 要 要 要 要 要 要 要 要 要 要 要 要 要 要 要 要 要 _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ The library, that is, the shot (4) library from the linear predictive coding cepstrum moment to identify the unknown sentence _ F minimum set 84. The distance or weighted distance in the closest sentence to the known sentence in the database, find the unknown sentence The present invention is described in detail later: (1) After a sound (sentence) is pronounced clearly, the sound (sentence) sound wave is converted into a series of digitized sound wave signal points (signai s (four) points) 10 . Again, there is no voice tone, and before the sound (sentence) and after the sound (sentence), between the two words and the two syllables, all the mute and noise are deleted. After the voice signal point is deleted, the remaining signal points represent all the signal points of a sound (sentence). The sound wave is normalized and then the feature is extracted, and all the signal points are divided into (four) time periods, and a frame is formed every time period. A sound (sentence) has a total of e-length elastic frames, and no waver 'does not overlap, freely stretch' covers all signal points 30. In each equal-length elastic frame, the signal 13 [S3 201216253 point changes nonlinearly with time and is difficult to represent with a mathematical model. Because J. Markhoui—published in proceedings of IEEE in 1975, Vol. 63' Νο·4 published papers Linear Prediction: A tutorial review and LU Tze Fen (Li Zifen) in 1997 in the US Patent Certificate, Apparatus and Method for Normalizing and Categorizing Linear Prediction Code Vectors using Bayesian Categorization Technique, USA Patent No. 5,704,004, Dec?. 30, 1997 illustrates that the signal point has a linear relationship with the previous signal point, and the nonlinearly varying signal can be estimated using a regression model that varies linearly with time. point. Signal point 5(8) can be estimated from the previous signal point 'The estimated value of 'S, (8) is represented by the following regression pattern:

P s,(.n) = ^i,S(n-k), n>〇 (i)

Ar=| In the formula (1), '~, Α = 1,...,Ρ, is the estimated value of the regression unknown coefficient, and P is the number of previous signal points. In 1993, L. Rabiner and BH Juang published Fundamentals of Speech Recognition, Prentice Hall PTR, Englewood Cliffs, New Jersey and Li, TzeFen (Li Zifen) in 1997 in the US Patent Certificate, Apparatus and Method for Normalizing and Categorizing Linear Prediction. Code Vectors using Bayesian Categorization Technique, USA Patent No. 5,704,004, Dec. 30, 1997, Durbin's formula for finding the least squares estimate [s] 201216253, this set of estimates is called the linear predictive coding (LPC) vector. The Linear Predictive Coding (LPC) vector method for signal points in Quanshu is detailed below. To represent the sum of the squared differences between signal point S(8) and its estimated value s, (8). ^ι=Σ[5(Λ)-Σα*5 (»-^)]2 rM) PL ka\ (2) Find the regression coefficient to minimize the sum of squares. For each unknown regression system

The number ~, =1, ···, pleading for the partial differential of (2), and dividing the partial differential into 〇, the normal equation of P group is obtained: 2a^S(n~mn-i) = ^SWin-i), Imp (3) rt After expanding (2), substituting (3), the minimum total squared difference is £/>=Z52(w)-Ea*Z5(«)^(«'^) (4) n η (3) and (4) are converted to ρ Σα* and ('·- moxibustion)= and (0, (5)

(five households = chat ~ Σ ~ away) (6) In (5) and (6), use the wire to indicate the number of signal points in the frame, N, i ^ (0 = X 5(ri)S(« + /), i > 〇(?) l|s〇 Use the Durbin's loop to quickly calculate the linear predictive coding (Lpc) vector as follows: (8) E〇= R(0) [Si 15 (9)201216253 kt ^(,) =k,.

(10) 01) End = (1 food) (12)

(8~12) Formula bribe calculation, get the regression coefficient least squares estimation value of ''1', '···', (linear predictive coding (LPC) vector) as follows: Μτπ# will be more stable prediction of LPC vector conversion Code scrambling 3 Lu (LPCC) vector household a'=a, +|(7)a^ (14) ... 'in (f) W P<i (15) elastic boxes produce a linear coded cepstrum ( Lpc orientation

The amount ('' ’ P) 40 ° according to the speech recognition method of the present invention, P = 12, because the final linear coding code (grandfather) is almost G. A cepstrum (10) c) vector with E linear predictive coding represents a sonic (sentence) feature, that is, a matrix containing Εχρ=1Μ2 linear predictive coding cepstrums (LPCC) represents a sound of 5〇. (2) After the known sentence of the bribe is pronounced, the sentence is known, the two words and the two θ 4 are 'deleted all the mute and murmur, and the (10)-(5) formula converts the sentence into a linear predictive coding scrambling error. (8) Moment [si- 201216253 array 60. Linearly predict the coded cepstrum with a known sentence (the distance between the lpcqexP matrix and the linear predictive coding cepstrum Εχρ matrix of all M=1000 different sounds or the weighted distance to find the closest database, the linearity of the known sentence The estimated coded cepstrum (LPCC) matrix is assigned to the nearest database. 70. There are .1000 databases, each containing 8 known similar sentences.

(3) When an unknown sentence is to be recognized, the user first pronounces a known sentence that is to be recognized. The unknown sentence is digitized into a signal point of 1〇, except for mute and murmur 20. After the county knows the sentence is unknown, the mute and murmur are deleted between the two words and the two syllables.

E equal length frame normalizes the sound wave, extracts the average, and the unknown sentences to be identified all have the time of the time into the £_ segment, forming a transition box every time period... the financial E turns the long box, there is no filter, no overlap , freely telescopic cover all signal points 3 〇. In each frame, the signal point can be estimated from the previous signal, and the estimated value of the 2 knowledge coefficient is obtained by the least square method. Each box uses (4)) formula to convert to line 1 to estimate the coded inverse (_ΕχΡ matrix, - an unknown sentence with f ΕχΡ linear prediction coded spectral matrix represents illusion. The invention ^ M the unknown sentence linear prediction coding cepstrum ΕχΡ The distance between the matrix and the ="____ copper-sense spectrum secret matrix 0 or the weighted distance is found in the closest database of 201216253. From the closest database of F, the linear prediction of the unknown sentence is encoded. The distance or weighted distance between the spectrum ExP matrix and the linear 雠 code 隹4 spectrum Εχρ matrix of the known sentences in the closest F databases is used to find the unknown sentence 90 desired by the user. (4) To confirm the speed of the invention And accurately identify any sentence in any language, the purpose of the car and the fine free conversation, the inventor uses 1 English single-word voice to represent 1 different databases, the inventor pronounces _ 928 sentences (8G English sentences, 284 Chinese sentences, 3 sentences, 2 sentences, 160 English words, 398 Chinese words, and German characters. After the test, the sentences and English words are all ranked first, and the previous calculations and comparisons - sentences All single word feature values, the present invention only needs to calculate and compare an ExP matrix value, the Chinese word is also in the top two, because the homophone is too much, the recognition time is less than 1 second; the inventor pronounces the English word, after the test 'before The five recognition time is less than 2 seconds; • The inventor pronounces 4400 Chinese words, and after the test, it is also in the top 20, and the recognition time is less than 2 seconds. The 4400 Chinese word is used as the voice input Chinese software. The present specification is input by the software. (5) Figure 6 and Figure 7 are used to input the fragment of the present invention. Figure 3 to Figure 5 identify the Chinese and English sentences with the invention. [Simplified illustration] The first picture and the second picture 5 The invention is executed. The first figure shows the establishment of M=1〇〇0 [S] 201216253 different databases, each containing similar known sentences. The second figure shows the process of identifying unknown sentences. The first is to use the yis (10) software to input the segment of the present invention and to identify Chinese and English sentences. [Main component symbol description] (1) First M=1000 different sounds Lu (1〇) sound wave digitization (20) Remove noise and mute Time Segment (30) E equal length elastic frames normalized all sound waves (40) every fresh long _, Saki small flat wire glaring linear prediction coding cepstrum (50) - a linear prediction of the sound encoding cepstrum Εχ ρ matrix Represents a database with a total of one thousand databases. • (60) Clearly pronounces known sentences - times, removes mutes and noises, converts them into linear predictive coding cepstrum (Lpcc) ΕχΡ matrix (70) with distance The known sentence linear predictive coding cepstrum (LpCc) matrix is divided into the closest database (80) with Μ=1000 databases, each containing similar known sentences (2) Forbearance of unknown sentences is clearly pronounced (41) Within each equal-length elastic frame, ρ linear predictive coding cepstrums are calculated by the least square method, and an unknown sentence is coded by linear predictive cepstrum ε 201216253 χΡ matrix representation (84) distance Find the unknown sentences in the F closest database with the closest known database (90) in the database of 1000=1000 databases, and find the unknown sentences to be identified by distance.

Claims (1)

201216253 VII. Patent application scope: 1. A method for recognizing all language sentences, the steps of which include: (1) A sentence may be a single tone, a single word, a name or a sentence in any language, preceded by ι=ιοοο a different voice; ) - a pre-processor deletes the sentence <sound) and after the sentence (sound), between the two words and between the two syllables, all φ does not have the voiced signal points (sampled points) of the mute and (3) - Sound or sentence sound normalization and extraction feature method: use e equal elastic frames, no filter, no overlap, normalize a sound or sentence sound wave, and convert to equal size Linear predictive coding cepstrum (LPCC) ExP matrix; (4) M=1000 linear prediction coding cepstrum for different sounds (Lpcc) Εχρ matrix stands for Μ=1000 different databases; (5) user to known sentences Clearly pronounce once, delete the mute and murmur of the sampled points before and after the sentence, between the two words and the two syllables. The equal-elastic frame normalizes the sound waves of a known sentence with speech and converts them into a linear predictive coding cepstrum (LPCC) ExP matrix of equal size; (6) Linearly predicts the encoded cepstrum αρ with a known sentence (χ) Εχρ matrix and all M=1000 different sound linear prediction coding cepstrum Εχρ matrix distance or weighted distance to find the closest database to the linear pre-preparation of the known sentence [S3 201216253 Estimated Coded Cepstrum (LPCC) The ExP matrix is divided into the closest database. Then, using the distance or weighted distance, the linear predictive coded cepstrum ExP matrix of any known language is recognized and the linear predictive code representing the sound of the database is divided. The censored ExP matrix is located in the nearest database, and similarly known sentences are placed in the same database; (When the unknown sentence is to be recognized, after the user pronounces the unknown sentence, the present invention also uses the known sentence to encode the sentence. The linear equation of the inverse Εχρ matrix and all Μ-1000 different sounds is used to find the distance or weighted distance between the cepstrum gyp matrices to find the F closest to the library, and then use the unknown sentence. The linear predictive coding cepstrum ExP matrix and the F closest database, similar to the distance or weighted distance between the linear predictive coding cepstrum ΕχΡ matrix of the known sentence, find the unknown sentence desired by the user; (8) If If the recognition is unsuccessful, the user will pronounce the sentence once again, and convert the sentence into a linear predictive coding cepstrum by using the special elastic box (and Εχ p matrix' uses the distance to encode the linear predictive coding of the sentence (Lpcc) The Exp matrix is assigned to the closest database and the sentence will be recognized successfully. 2. According to the application of the suspects around the first item - carefully identify all language sentence methods 'where step (3) contains E equal elastic boxes, equal length, no filter, no overlap, will be - To normalize the sound or sentence sound waves and extract the feature matrix of the same size, the steps are as follows: (a) Before the sentence (sound) and after the sentence (sound), between the two words and between m 22 2〇1216253 two syllables, All mutes and murmurs that do not have a voice signal point (Q) ed fine (6), use a method of equally dividing a sentence (a sound) with all sound wave signal points. 'In order to closely estimate the nonlinearly varying sound waves with a linearly varying regression pattern, The full length has a sound wave point divided into £=12 equal time periods. Each phase of the material segments forms a bullet. It is difficult. One sentence (one sound ") has a total of E equal length elastic frames. There is no wave (4), no overlap, and it can be freely stretched and covered. Full-length sound wave, not a fixed-length Hamming window; (b) In each #long box, use a regression model that varies linearly with time to estimate the sound wave that changes nonlinearly with time. (C) a cyclic manner with Durbin's ^ 0) = J ^ S (n) S (n + /), / &gt; = and fifty square ⑼ k^W)-J:arR^j)]/E y Si-1 aj = aT\ κβΡ, called linear prediction P=12 regression coefficient least squares estimate e coding (LPC) vector, reuse 23 201216253 ^ =^+Σ(ζΚ-&gt;αν 1 ρ y=i 1 a', = Σ (τΚ-&gt;°'7- p&lt;ij=i-^ 1 Convert Linear Predictive Coding (LPC) Vector to Stable The linear predictive coding cepstrum (LPCC) vector &lt;, 匕; (d) expresses a sentence using E linear predictive coding cepstrum (LPCC) vectors (a linear predictive coding cepstrum (LPCC) ExP matrix) (or a voice). [S3 24