TW201106340A - A speech recognition method for all languages without using samples - Google Patents

Abstract

The invention can recognize any several languages at the same time without using samples. The important skill is that features of known words in any language are extracted from unknown words or continuous voices. These unknown words represented by matrices are spread in the 144-dimensional space. The feature of a known word of any language represented by a matrix is simulated by the surrounding unknown words. The invention includes 12 elastic frames of equal length without filter and without overlap to normalize the signal waveform of variable length for a word, which has one to several syllables, into a 12x12 matrix as a feature of the word. The invention can improve the feature such that the speech recognition of an unknown sentence is correct. The invention can correctly recognize any languages without samples, such as English, Chinese, German, French, Japanese, Korean, Russian, Cantonese, Taiwanese, etc.

Description

201106340, invention description: [Technical field to which the invention pertains] - A continuous tone contains - one syllable (mono tone). The present invention can recognize all languages without a sample of continuous sound. The invention uses 12 elastic frames (windows), equal length, no chopper, no overlap, and the sound waves of different lengths and one continuous sound are converted into a matrix of 12x12 linear predictive coding cepstrum (LPCC), Unknown continuous sounds are represented by a matrix of 12χ12 linear predictive coding cepstrum. A 12χΐ2 matrix is considered to be a vector of (4) degrees of space. Many vectors of unknown continuous sounds are scattered in 144 degrees. When the speaker sends a known continuous sound, the characteristics of the known continuous sound are simulated and calculated by the characteristics of the surrounding continuous sound (LPCC). The present invention comprises a sigmoidal box normalization-sound sound wave, and the Bayesian=pair method finds a known continuous sound for the speaker's unknown continuous sound, which will be - a speaker's unknown The sentence is divided into D unknown continuous sounds, and - a visual ff帛·'s screening - a known sentence is the successor of the successor [Prior Art] 4 gentleman hair - a continuous sound, its pronunciation is expressed by sound waves. The sound wave is a kind of nonlinear change (4) with (4), and a continuous sound helps to have a kind of dynamic, and also makes a nonlinear continuous change with time. When the same continuous sound is pronounced, there are two consecutive dynamic characteristics, which are nonlinearly stretched and contracted with time, but the phase is arranged in time-like order but the time is different. When the same continuous sound is pronounced: It is very difficult to arrange the same dynamic characteristics at the same time position. It is even more difficult to identify because of the similarity of 201106340. - a brain-based 浯 ^ identification system, the first to extract sound waves related language resources

News, that is, dynamic characteristics, filtering and language-independent noise, such as human voices, tones, psychological, physiological and emotional and speech recognition when speaking, do not know the details ^ Align the same features of the same g at the same time position . This series of features is represented by a series of equal-length eigenvectors called a continuous eigenmodel. At present, the speech recognition secret size is too complicated and time consuming, because the same features of the same continuous sound are difficult to arrange in the same time position, especially in English, which makes the identification difficult. - General sentence or name recognition method - even red main work: unknown sentence or name cut into D unknown continuous sounds, extracted features, feature normalization (feature model size - and the same feature arrangement of the same continuous sound At the same time position), unknown continuous sounds, and find a suitable sentence or name in the sentence or name of the bribe. - A continuous sound wave feature is commonly used to shun kind of energy · bribe gy, zero crossing points (zero. Secretary), the extreme number (mail · c shout), peak (f〇rmants) 'linear prediction The coded cepstrum (10) and the Mel frequency cepstrum (MFCC) are most effective in linearly difficult coding _哉 (Lpcc) and Mel's frequency (10) C). The mis-predictive coding cepstrum (lpct) is the most stable and accurate linguistic feature of a continuous tone. Its secretive regression mode represents continuous sound waves, and the small square estimation method calculates Jingsong. _ (4), the gorge is the linear pre-201106340 estimated coded cepstrum (LPCC), while the Mel frequency cepstrum (MFCC) converts the sound wave into a frequency using the Fourier transform method, and then estimates the auditory system according to the ratio of the Mel frequency. SB Davis and P. Mermelstein published in 1980 in IEEE Transactions on Acoustics, Speech Signal Processing, Vol. 28, No. 4, "Comparison of parametric representations for monosyllabic word recognition in the continuous sentences" using dynamic time warping (DTW) The 'Mel frequency cepstrum (MFCC) feature is higher than the linear predictive coding cepstrum (LPCC) feature recognition rate. However, after repeated speech recognition experiments (including my invention), using Bayesian classification, linear pre- The estimated coded cepstrum (LPCC) feature recognition rate is higher than the Mel frequency cepstrum (MFCC) feature and saves time. Identify the 'Wei Xiang method lion. There are secret time-warping (vector), vector quantization (vect〇rquantizati〇n) and hidden Markov model (method). If the same pronunciation is in time There is a difference in the change, one side compares the same feature to the same time position. The recognition rate will be very good, but the same feature is pulled to the same position and the time is difficult to report and the distortion time is too long to be applied. Vector quantization method, such as identifying a large number of Continuous sound is not only inaccurate, but also time consuming. Recently, the hidden Markov model method is good, but the method is complicated. 'Taixiang unknown parameters are needed, and the calculation seems to be time-consuming. Recently TF Li (Li Zifen) was published in Pattern ReCognition in 2 years. , 201106340 vo 1· 36 Published in the paper Speech recognition of mandarin monosyllables using Bayesian classification, with the same database, a series of LPCC vectors of various lengths and compressions into the same size classification model, the recognition result ratio γ. κ.

Chen, C.Y. Liu, G.H. Chiang, Μ·Τ. Lin published in 1990

Proceedings of Telec〇mmUnicati〇n Symp〇sium, the paper of the remuneration of the paper The rec〇gnition of mandarin m〇n〇syllables based on the hidden discrete hidden Markov model with the hidden Markov model HMM method is better. However, the compression process is complicated and time consuming, and it is difficult for the same continuous tone to press the same feature and the ___ position, which is difficult to recognize for similar continuous sounds. The speech-cutting financial method of the present invention touches the frequency point, and from the learning plane, according to the sound characteristics of the sound wave, the nonlinear change with time, naturally derives a set of methods for extracting the voice feature. The gradual normalization of the syllable wave is converted into a size-transformation model sufficient to represent the continuous sound, and the phase of the same time position in the symmetry model has a phase, and in our special selection day (4) (10), the human domain experiment is not required to adjust the sound. The known continuous sound standard model in the feature database is better than the clear speech recognition party. Therefore, the present invention [Summary] ‘speed element into feature extraction, feature normalization and recognition. (1) The purpose of this & Ming (4) is to calculate and calculate any of the words in the township. Any feature can be used to model the invention. Any feature can be created without the sample, so any one of the -s- consecutive 201106340 sounds The feature that the present invention can correctly recognize various languages without using samples. In detail, the present invention simulates and calculates the known continuous tone for any known continuous tone of any language using a Bayesian distance of 144 degrees of spatially unknown continuous tone matrix to achieve no known continuous tone. The sample is still able to establish the characteristics of any known continuous sound. Therefore, any language can be identified. (2) The present invention provides a language recognition green. It can delete non-verbal sound waves. (3) The present invention provides a method for normalizing and extracting continuous sound waves. it makes

•...a~ is the same as the wind in the main 邗冏吋 邗冏吋 position. Please recognize the day and reach the computer's instant recognition effect.

(6) This (4) makes it possible for Saki to handle sound waves (sounds). Use less

"No. Point." Cover all letters 201106340 with less overlap. Not because a continuous sound wave is too short, delete the continuous sound, and not because it is too long to cut or compress part of the signal point. As long as human The auditory can discriminate the continuous sound, and the present invention can extract the continuous sound. Therefore: the inventive speech recognition method applies each of the signal points with speech, and can capture the speech features as much as possible. Since #=12 elastic frames do not overlap, The number of frames is small, which greatly reduces the feature extraction and calculation of the misGc) time. (7) The invention of the invention can recognize the continuous sound of the fast or the age of the lecture. When the speech is too fast, the continuous sound is short. The length of the elastic frame of the present invention can be reduced, and the short sound waves are still covered by the same number of elastic frames of the same length. A square linear predictive coding cepstrum (Lpcc) vector is generated. As long as the short sound can be discerned, the miscellaneous The linear predictive coding cepstrum ap(6) vector can effectively represent the characteristic model of the short sound. The continuous sound waves emitted by the too slow are longer. The elastic frame will be elongated. The generated linear predictive coding scrambling 瑨 (LP) The CC) vector can also effectively represent the long sound. (8) The present invention provides a method for stabilizing and adjusting the characteristics of all known continuous sounds in the database 'make all the continuous sound features have their own health in a 144 degree space, In order to touch iL. (9) When identifying the name or name, the New Zealand (four) child or name is cut into an unknown continuous sound. The present invention uses the Bayesian method in the continuous sound feature database to select the most similar F. One known continuous sound. One sentence is represented by _ known continuous sounds, which may be cut into more or less than the number of consecutive sounds of 201106340 due to the difficulty of cutting. The present invention is similar to three consecutive columns of F for each unknown continuous sound. It is known that continuous tones compare a known continuous sound in a sentence or name, that is, in a sentence and name database, 'screen a known continuous sound for each sentence or name with a known similar continuous sound of 3xF. Find the most likely sentence or name from the sentence and name database. The method is simple and the success rate is the same (identify 70 English sentences and names and 4〇7 country sentences and names).

(10) The present invention provides two kinds of unique continuous sounds of unknown and continuous speech sounds and unknown sentence or name recognition success. (11) ^Invented - a Mandarin single as a - only - syllable continuous 曰 Chinese and foreign features are represented by the same size matrix. Therefore, the present invention can recognize various languages at the same time. [Embodiment] The condition of the invention is invented. The first picture is a permanent database of scales, known as continuous tone feature,1, with a continuous sound wave of 10, '3' to convert continuous sound waves into a sequence to receive 20. There are two ways to delete the digital gas: 45 (1) Signal points of interest. Previously processed i-like noise variations. For example, the former: the variation of the back domain point should be deleted. (2) Counting time, no time in speech, and the distance between two consecutive signal points in the kiss segment and 1 201106340, after the H I heart 45 is finished, the obtained sequence has the known continuous 曰k point. Normalize the sound wave and then extract the feature, which will be known

The signal points are divided into turn-time periods, and each time frame is composed of a frame. All have _ long box 50, no filter, no overlap, according to the length of:: = the length of the point, (four) the length of the frame freely adjust all the signal points with the cover.咖二匡^ is a flexible frame, the length is free to stretch, but the length of the _ sex box. Unlike the Hamming (H milli lng) window, there are fortune, semi-overlapping, fixed length Lu can follow, long free adjustment. Because a continuous sound wave follows the _ non-two change 匕'. The sound wave contains a dynamic feature of the voice, and also changes nonlinearly with time. Because they do not overlap, the present invention uses less (five) 2) miscellaneous frames, culminating all continuous sound waves ' _ ship can be _ money points, the time of the linear change of the regression mode to closely estimate the nonlinear changes of the sound waves, with the most The Xiaoping method estimates the regression unknown coefficients. The minimum squared estimate of the scale-to-score coefficient is called the linear predictive coding (10) vector. The linear predictive coding (LPC) vector is then converted into a more stable linear prediction and the cepstrum is degraded. The speech transition sign containing the _ sequence with non-linear variation over time is converted into a nonlinear transition in the present invention. The equal-precision linear predictive calculus ^ item day (LPCC) inward 60. In order to extract the characteristics of a known continuous sound, a permanent known continuous sound database is prepared. Each known continuous sound is called the solid standard. /月明者 pronunciation-time. If you recognize an accent that is heavy or not standard, say 11 201106340

The words 'that (4) this person pronounced, _ ❻ 娜 成 into the ExP_C matrix placed in the water for a long time known continuous sound data phase. In the Yonghongzhi continuous sound database, for a known continuous sound extraction feature, first prepare a tribute library of unknown continuous sounds. 'Unknown continuous sounds and mixed riding two kinds: - more unknown continuous sounds have another - tender tear . Miscellaneous _, the average variability of each unvoiced sound. Unknown continuous sound database of the surface handle: N + of the estimated coded cepstrum (LPCC) for finding the N most recent unknown continuous around the known continuous sound with Bayesian distance 1 weighted average is used as the average of the known continuous sounds, and the number of variances of the weighted average _ mouth continuous sounds of N consecutive sounds of N consecutive sounds, the ExP average and the variance matrix are the continuous sounds The preliminary feature value 79 is placed in the continuous tone feature database. If there is no sample in the unknown treasury, in the unknown continuous sound data, find the linear prediction code of N unknown 1 and N unknown continuous sounds with the minimum absolute distance for the known continuous sound. _α(10) 2 = 1 ) a number of children. Find the weighted average of (10)) numbers as the average of the known continuous values, and find (the number of variances of the number as the known continuous sound: ^ number, this power is her initial generation of her companion = put It is known that the _axis 79. In the (10) continuous sounds of the second library, if the mean of the known continuous syllables is the same as the Bayesian distance of the Lpcc of the known continuous sounds in the permanent known continuous syllables, In special: 12 201106340 Bessie's inner branch is the smallest, then in the feature: #内库(四)氏距离找N

❻Continuous sounds. Their Bayesian turn to observe the continuous sound LPCC is N minimum. (4) The average value of the force of the known tone is used as the new average of the known continuous sound, and the weighted average of the N variances of the N known continuous sounds is used. A new variation of continuous sound is known. This method is used to repeatedly calculate the new average and variance of each known continuous sound in the feature database. Finally, the new average and the matrix of the variance are called the standard model representing the known continuous sound, and the feature data is placed. The library also uses the known continuous sounds of the known _ database to create sentences and name data. - The second ® represents an unknown sentence or name-financing process. When an unknown sentence or name 2 is input to the speech recognition method of the present invention, the splicing conversion ^ 3 is converted into a series of sound wave signal points by the group unknown continuation sound wave 11 into the human touch device 2 。. The sound wave of an unknown sentence or name is cut into D sound waves 40 of unknown 2 sounds, and then the sound wave of the voice is deleted by the first picture. Then, each unknown continuous sound wave is normalized, and the feature is extracted, and each unknown sentence of the sentence or the name has a period of _ age of the voice, and a flexible frame is formed every time period. Each continuous sound has a paste elastic frame, no filter, no overlap, and free telescopic cover all signal points. In each box, because of the location of the recorded points, Xiaoping Green is asked to purchase an estimate of the unknown coefficient. The estimated value of the most oblique square of each group is called Line I·Life Prediction Compilation (LPC) Vector 'Linear Estimation Coding (Lpc) Vector has positive 13 201106340

2 with, county secret age code αρ〇 vector Na more stable linear pre = touch spectrum coffee) vector 6Q. - Unknown continuous (10) linear: Encoding inverted lining (_vector__, called classification model! lgit known continuous sound standard model large - like a sentence - total stagnation summary D unknown continuous sound 9G, if - It is known that the continuous sound is the average of the known continuous sound 'its standard _ is closest to the unknown continuous sound classification f (10) predictive compilation turn (Qing. Therefore, this is the classification of unknown continuous sounds and The continuous sound database is compared to the standard model of every 2 known continuous sounds. (1) If there is a known continuous sound 2 unknown continuous age 'for the slant, the linear axis of the mosquito axis _ sound is composed of all the linear predictions (靡) Cong) has independent normal distribution, which = average and variance are calculated by the average and frequency of the known continuous sound standard model. The simple Bayesian method is to calculate the linear predictive code of the unknown continuous tone. Inversion = (_ and known The distance of the average of the continuous sounds is then adjusted by the known continuous = odd number, and the obtained value represents the unknown continuous sound with a known continuous 2 similarity. Selecting the unknown similarity with the unknown continuous sound F the highest known continuous sound generation Knowing the voice, so an unknown sentence or The name _ _ is known to connect two to mean no. - An unknown sentence or name is cut into d unknown contiguous: 'It is difficult to cut into just - an unknown sentence or the continuous sound contained in the name and ^ 'sometimes - consecutive The sound is cut into two, sometimes after two consecutive sounds, 'computer cut into one'. Therefore, _unknown continuous sound is not necessarily a speech 14 201106340 j positive continuous sound of the 'received' column F Knowing that her continuous sound does not include the continuous sound of the speaker. In identifying an unknown sentence in the traitor and name database 85, testing each known sentence and name, in the test H sentence, the name is more than the speaker _ _ _, The sentence or name gimmick - a known continuous tone is compared to the front and back of the DxF matrix similar continuous sound, with a continuous sound (when the silk-a pair can only match the last two consecutive similar sounds), then move the 3xF window (There are three known similar continuous tones in the first three columns) 12〇 Find the second known continuous tone of the sentence until the test sentence is all known to be continuous. In the sentence and name, the sentence or name is the speaker with the highest probability of the sentence or name. Sentence or name % (test sentence or name already The number of consecutive sounds in the 3xF window divided by the test sentence or the number of consecutive sounds) (10). Of course, you can select sentences or names that are approximately equal in length to the unknown sentence or name (D unknown continuous sounds) in the sentence and name database. Save time. If the sentence or name= can be identified, use the Bayesian classification method to improve the slightly-sounding features of the N most similar continuous sounds (10) in the feature data, and it will succeed. This is detailed in the following: (1) After a continuous sound input into the human language reading financial method, the continuous sound continuous sound wave is converted into a series of digitized sound wave signal points (signal Qing - she (6). Then delete the mouth sound wave signal point. The present invention provides Two methods: - is to calculate the number of variances of signal points in one hour. The second is to calculate the sum of adjacent two letters = point distances in the correction segment. Theoretically, the first method is better, because the variation of the money points is greater than The odd number of the day is y, indicating that there is voice. However, in the case of the invention, the recognition rate of the two methods is the same, but the second time is saved. (2) After the #f sound iron point is deleted, the remaining money points represent - a continuous sound full tie number point A normalizes the sound wave and extracts the feature, and divides all the signal points into the time period of 'the time interval'. . - A continuous sound with a total length of flexible frames without filters, no overlap, freely telescopic, covering all signal points. The elastic frame mm time is not a miscellaneous change, and the mathematical model table is ^ ^ ^ J. Markhoul ^ 1975 ^ ^ Proceedings of IEEE,

Vol. 63, No. 4 Published in Linear Prediction: A tutOTial review shows that the signal point has a linear relationship with the previous signal point, and the signal point of this nonlinear change can be estimated by a regression model that changes linearly with time. The signal point W can be estimated from other surface signal points, and its estimated value is expressed by the following regression modes. (1) Lu in (1), ~ 'seven=1, .." corpse, is the estimated value of the regression unknown coefficient. , the household is a letter

Number of points. L. Rabiner and B.H. 1! Qing published in 1993 Fundamentals of Speech Recognition, Prentice Hall PTR

In Englewood Cliffs, New Jersey, Durbin's cyclic formula finds s small squared estimates, which are called linear predictive coding (Lpc) vectors. The linear predictive coding of the _ inner signal point (LP 〇 vector method is detailed as follows) to represent the sum of the squared differences between the signal point ^(«) and its estimated value (8): 16 201106340 £1=|;[5(η )-2α,5(η-λ:)]2 (2) η=〇灰=1 Find the regression coefficient to minimize the sum of squares. For each unknown coefficient of refinement α, _, ί = 1,.·· , Ρ, find the partial differential of (2), and divide the partial into 〇 ' to get the normal equation of the household:

^a^Sin- k)S(n - ί) = Σ5(η)5(η-〇, 1 < ^ ^ fc=l η «Expand after (2), substituting (3), the minimum Total squared difference α E„ =X52(n)-^at£5(n)5(n-)t) nk=\ η(3) and (4) are converted to (3) (4) $atR(i - fc) = R(i), l<i^P k=i (5)

Ep = R(0) (6) jfc=l In equations (5) and (6), use # to indicate the number of signal points in the frame, W—i ^(0 = + 0) ϊ ^ 0 (7) /»=〇 Use the Durbin loop to quickly calculate the linear predictive coding (Lpc) vector as follows: E0 =/?(〇) 勾=_ 一客〇·-·7·)]/£μ (8) (9) αί° = of -kta^, E { =(l-^j2)^j_i (10) (ID (12) 17 201106340 (8-12) Formula Cyclic Calculation 'Get the least squares estimate of the regression coefficient α Strict 1,...'' C Linear Estimation Coding (LPC) The vector) is as follows: (13) Pre-coded cepstrum

Aj = ajP). \<j<P Then use the following formula to convert the LPC vector to a more stable linear (LPCC) vector<,y = 1," household, ;=1 i Μ (14) (15) - elasticity The reduction-rotation prediction code reversing vehicle (service) vector: '.··0. According to the method of identification of the local county, the handsome, because the final linear prediction coding cepstrum (LPCC) is almost 0. A continuous The sound is recorded by the linear prediction code. The cut-off feature contains a linear Z-coded cepstrum (the matrix of _ is a continuous tone... a continuous tone contains one or more syllables. , =) The method calculates the cepstrum of the unknown continuous sound wave by the equation (8-(5)) (sense vector, the matrix of the LPCC with the same size, called the classification model of the unknown continuous sound. 2 second wealth, Speech recognition _, __ classification of unknown continuous sounds A matrix of fine LPCC. ❹.卞"chuan=1' household, q unknown continuous sound classification model. In and with a known continuous tone d heart, all continuation Total) 'In comparison, in order to quickly calculate the alignment value, assume (5) 201106340

There is an independent normal allocation of the households, and its average and variation are estimated by the mean and variance within the known continuous sound standard model. The conditional density function of x is expressed in eight xlc). The decision theory in Speech Rec〇gniti〇n 〇f mandarin monosyllables published by TF Li (Li Zifen) in 2003 in Pattern Recognition, Vol. 36 shows that the Bayesian classification is as follows: Assume that the feature database has a total of known A standard model of continuous sound. The heart '=1'.._, "the probability of the occurrence of the table discontinuity ^ = 1,%, that is, the previous probability, ** § 3-1. The heart represents a decision method. Define __ a simple loss function (loss function) ‘that is, the probability of error in d (misclass is as in (5) probabiUty): If the decision maker knows that an unknown continuous syllable c, then the loss function, he = 1. If, for example, an unknown continuous tone coffee ^, there is no loss L (Ci, 々 x)) = 0. The identification method is D, ., , and Guang. With Γ·_, the matrix value belongs to the range of the known continuous tone Ci.汝 also px in η, ^ / sentence unknown continuous sound belongs to the known continuous sound Ci error average probability ΚΜ) = Ρ > κ, ά (χ)) η χ ^ ^ m (16) In (16) 'r = (Uj, r, jr, 疋^乂 outside the range. The D-form method is recognized, that is, the method of dividing the melon is known to be green and not deductible to take 4) the range of construction and shellfish. Find an identification method in D. The average probability of error recognition (10) reaches a minimum of 201106340 (Π) R (T^-minR(T>d) satisfies the formula (17) of the method of financing, called "related to the previous probability" The classification method can be expressed as follows: (18) ''The range of known continuous sounds is the same in (18),;·=1,

all〕}. As with all known continuous tones, the probability of occurrence is the same as the "maximum probability" method. Bayesian classification (18) identifies the conditional density function of all z/(X|C,.), i=l,...,w, f (x\ c.)= Π n (19) In (19), ^'^乂 knows the total number of continuous sounds ^ For the convenience of calculation, '(10)) is taken as a logarithm, and the constant is deleted, and the Bayesian distance %) = / = 1,.. .,m. (10)

Jl ^ Ji σιμ

The Bayesian classification (18) becomes for each known continuous tone Ci, calculating your ;) value (2〇), generation C, ·) also known as the unknown continuous sound and the known continuous sound ^ similarity, or shell Baysian distance. In (20), x = u;i} , _/=ι,...,£·, hi,···,/1, is the linear predictive coding cepstrum in the unknown continuous sound classification model (LPCC) The value '(~, station} is estimated by the mean and the number of variations in the standard model of known continuous tones. The most important contribution of the present invention is that there is no sample in the known continuous tone feature database for each known continuous tone C Find the center point of each other's stability 201106340 〇 = (~} and the range of clear non-overlapping here is called ~} is to indicate the range of known continuous sounds (10) like Lpcc matrix. (5) Extract - a known continuous tone ·, first - A database of unknown continuous sounds. There are two types of unknown continuous sound data. One is that there are samples of unknown continuous sounds, and the other is no samples. There is a database of samples, and the average value of each unknown continuous sound is first sought. The number of variances. In the unknown continuous sound (4) library with samples, find the N most recent unknown continuous sounds around the known continuous sound with Bayesian distance. Then find N unnumbered N averages and the known continuous sounds. Miscellaneous predictive coding cepstrum (N+1 weighted averages of to as the average of known continuous tones, and two N The continuous sound of N_Wei's domain flat touches as a continuation of the sound of the 2 biliary t-shirts _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ In the middle, use the minimum absolute value of 5 around the "unknown continuous sound. The known continuous sound and N unknown connected 2' mispredicted coding spectrum (_ as (4)) numbers. Find: 1 weighted average as The average value of the known continuous sounds, and the seeking (the number of sub-variables is known as the moment of the loss of 2, such as the maid_touch, the heart of the library 79. Name P heart, reading * u 曰贝曰In the known continuation feature database, the average of the performance scores (four) is known as tearing one = 21 201106340 to be in the shape of the snake is not the smallest in the characteristic axis, then the _ sound is known in the shape It is the IV counties of the five-year-old continuous sounds of the smoked Bayes, and the early average letters are taken from the smaller ones. Find the N new averages of the N known continuous sounds and use N known continuous Sound: 'sound

For the far known new variations of the continuous sound. In this way, repeat the new 5 lbs of each known continuous sound in the library. ==: _ 嶋 — 音 - (8) (4) Touch - an unknown continuous tone e-touch error, this 剌 provides Nazheng old features To make the continuous sound recognizable correctly; Eight ω is found in Teh by Bayesian classification (10)

Phase-continuation—,...,—: value

Use {/V 〜 = ?^=Σ: ijt ;}, £=1,...,ρ , (22) Penalty ruth, representing the unknown continuous tone of the new standard grip continuous tone feature database, and then Test the sound must be a work force, (l) ii kiss scarf, the most similar known continuous sound mean and the cont of the unknown continuous sound, 1, the weight of the ^ value of the most similar continuous sound The average value is used as the weighted average of the not-only dice and the unknown, and the estimated cross-spectrum (LPCC) is used as the weighted average of the unknown continuous sound. { J =X:.,E, I = l, ... 5p 'represents the 22 201106340 new standard model of unknown continuous sound. (7) In order to confirm the original language and the same language, this (4) performs a two-person speech recognition experiment. (a) First establish an unknown continuous stream f library. This monophonic database was purchased from the Central Research Institute in Taiwan. The database has a total of single-sentences (three figures), all of which are female pronunciations. The samples range from 6 to 99, and the pronunciation is almost the same. (6) Convert all samples from the (2) method to the Lpa office, with a total of 12,400 matrices. (c) In the 388 Mandarin singles, use the sample to average and the number of variances. (4) Blindly mix 3_monophonic' to make the single tone with the sample mean and the variance number into a 388 unknown continuous sound database (one Mandarin single tone and one syllable continuous sound). ^ (e) Re-find - male-female 654 Mandarin mono, 154 vocabulary, one German, one Japanese and three Taiwanese pronunciations to create two 813 permanent known continuous sound databases, each _continued (2) In the 813 known continuous tones of the permanent known continuous sound database, for each known continuous sound, find the 388 unknown continuous sound with Bayesian distance (10). And 15 unknown continuous sounds, the linear predictive coding cepstrum (LPCC) of the known continuous sound and the sample mean of the N unknown continuous sounds are n+i weighted averages 23 201106340 Miles power is not renewed The sample variation kernel H is known to be the threat of continuous sound. This average and the L and number matrix are called the preliminary features of the known continuous sound. 79 The known continuous tone feature database. That is, the feature database contains (10) The mean and the variance matrix 80. (g) In the feature data, if the average of the known continuous sounds and the LPCC of the known continuous sound in the Yong 3 continuous database are not the same Take small. In the 813 continuous tone feature data, use the Bayesian distance to find N=15 known connection: N N flat with N consecutive performances The mean and the unweighted average of the known continuous sounds are listened to the new average of the known continuous sounds. The weighted average of the variances of the n known continuous turns is the new variance of the known continuous sound. Repeated calculation Xinping and Wei have been repeated many times. The final 12-can average and variable-number matrix is called the fresh model, indicating the known continuous sound feature, and there are 8〇 in the known continuous tone feature database. Benming performs the following continuous tone recognition. The recognition rate depends on the person. Because there are too many similarities, the top three are even right. · 1 Identify 384 Mandarin singles, 1 German, 1 Japanese, 2 Taiwanese (3rd) (very good recognition rate) 2 Identify 154 English, 1 German (4th) (very good recognition rate) 3 Identify 154 English and 388 Mandarin, 1 German, 1 Japanese, 2 Taiwanese 24 201106340 (The recognition rate is very good) 4 Identify 654 Mandarin Monophonic, 1 German, 1 Japanese, 3 languages (the third and fifth pictures have good recognition rate, no good for the first three). The sentence of the speaker is said or not. 'We first establish - English and country = sentence wide And the name of the database 'each sentence or name _ continuous sound all 1 by the joint holding In the cargo warehouse (3_4), English and Mandarin are known to be arbitrarily composed = adult and name, and 384 Mandarin characters are composed of 407 national sentences and %, (different six figures). The identification method is as follows: u J卞汉

(!) cutting - an unknown sentence or name becomes D = neighbor: the sum of the signal points, such as too small, the 'for the Qiuqiu cut into D buckles, check 繂 θ should be 疋 two,, g The boundary should be cut, one;; lpIc f; tt Λ;: ί r50 60 ^90 ^ and the Mandarin feature database + select the most similar Bayesian continuation [2〇); and the Mandarin (Figure)) - unknown sentences or The name is in the continuation package ii (b) in the sentence and name database to find the speaking sentence and the name, the length of the selection is (10) j in 47 words and the same length (D unknown continuous sound f and 4 words of the sentence or The name matches the sentence or the name of the 5 known continuous sounds and both contain - a comparison sentence or the name inside the known ^ bei 曰 曰 throwing as the mother column similar continuous sound = such as m (four) material name is a D (d if _ Library_sentence and name 冉.

Cc) recognizes that the correct continuous sound is not hard, the number 疋 D-] or D+] or the comparison sentence or name (in the database), 'the invention is screened with the secret window, in the first; Continuous tone, using the return matrix 25 201106340 = the last three columns are similar to known continuous sounds (ie the first

How many DxF matrices are selected than the known large sentences in the sentence or the name or the name of the female secret library - the probability is the most likely if a sentence or name _ acknowledgment error - is in one or more of them Among the similar known continuous sounds, there is a known continuous sound in (155+384). The first Ν=15 is known to be connected to ^1 (2= similar continuous sound and Lpa of the unknown continuous sound: plus ^ = = make continuous sounds in their F similar known continuous tones

T column English and national sentences and name recognition, identification is almost all correct, 1 identify 70 English sentences and names (very good). 2 Distinguish s 407 national sentence and name (very good) 3 Identify 70 English sentences and names and sentences and names (very good) * Attach a Visual Basic identification diagram (the first sentence έ έ s s 及 及 及 及 子 子, the name method. The seventh figure of the seventh figure shows the simultaneous [simplified description of the drawing] The first figure and the second figure illustrate the invention execution program. The first figure shows the known continuous sound permanent database, the continuous sound feature database and the three databases of sentences and names: rogue, the first figure shows an unknown sentence or name identification method flow. The third picture shows the identification of 384 Mandarin singles! German, } Japanese, 2 Taiwanese. The fourth picture shows the recognition of 154 English and 1 German. The fifth picture shows the recognition of 269 Mandarin singles and 3 words. The sixth picture shows that the sentence and name database has 7 English sentences and Chinese sentences 26 201106340 and the name. The seventh and eighth diagrams illustrate the Visual Basic recognition diagram showing the simultaneous recognition of English and national sentences and name methods. [Explanation of main component symbols] (1) Establish a permanent database of known continuous sounds, pronounce a continuous sound or a sentence' sentence and divide it into several known continuous sounds. (10) continuous sound continuous sound wave (20) receiver (30) sound wave digital converter (45) remove noise (50) E elastic frame normalized sound wave (60) least square method calculate linear predictive coding cepstrum ((10)) Vector, distance (absolute distance), for each known continuous sound (permanent database), find the nearest unknown continuous sound in the unknown continuous sound database. (79) A known continuous sound (permanent database) for the mother, with an unknown continuous sound and a weighted average of the succumbs of the sequel, for the known continuous sound, a preliminary feature. Then, the feature database is used to find 1^ known perpetual scores and the known continuous tone LPCC to obtain a weighted average value, which is calculated multiple times. The last ^ flat sentence value (ΕχΡ mean and variance) represents the nominal model of the known continuous tone. τ Γ ^ ^^曰Characteristic database contains all the averages and the number of variants of the standard model. The continuous tone of the continuous tone feature database is created to identify the sentence and name. 27 201106340 sentence and name database (2) input - not (four) Or name (11) a group of unknown continuous sound waves (40) will be - a sentence or name cut into an unknown continuous sound (9 〇) sound, qing _ generation fine solid head continuous sound classification model (100) use classification Compare each known continuous tone standard and unknown classification model 110(110)-sentence or name towel to find the closest known 曰 consecutive 曰 for each unknown continuous sound, one sentence or name has a total of Dxp Knowing the most similar continuous tone representation (120) In the sentence and name database, use the similar known continuous tones of the 3xF window to filter all sentences and each known continuous sound in the name (130) in the sentence and name database. One of the most likely sentences or names

28

Claims (1)

201106340 VII. Patent application scope: A method for identifying various languages without using samples. The steps include: (1) an unknown continuous sound database (with or without samples); by test (2) - a known continuous sound permanent database By a clear standard speaker, each known continuous sound is pronounced once, such as the test population sound is very heavy; (3) - a sacred ϋ (ρΐΈ-ρπχ^_) 林 语音 voice voice sound point points (sampled points) or murmur; ° (4) - a continuous sound wave normalization and extraction feature method: with e elastic box will The sound waves are normalized and converted into equal-sized linear predictive coding cepstrum (LPCC) ExP feature matrix; (5): an unknown continuous tone red f- and pumping (four) method: normalize the sound wave and convert it into size and Knowing the continuous phonetic scale (represented by a matrix of (10) averages and variances), the feature matrix of equal size, called unknown continuous-segment facet, contains linear predictive coding cepstrum (LPCC); (8)- Simplify the Bayesian classification: compare the unknown continuous sound classification model with all known continuous sound standard models of known continuous sound feature databases (by a list of ExP averages and variances)* Continuous sound is known, which is the least conveyed to the unknown continuous sound, and is recognized as an unknown continuous sound; (7) in the known permanent database of sounds, for each known continuous sound, there are 29 201106340 ^ object, sound database f, (four) turn to the nearest °, _ sound 'if unknown, please do not have money, use absolute distance to find N recently unknown continuous sound, · (8) two stocks have samples The N+1 weighted averages of 'calculating the N most recent unknown continuous tones and the linear estimates of the known continuous tones' are the average continuous sounds, n unknown continuous tones (10) The weighted average of the variances is the variation of the known continuous sound. The ExP average and the variance matrix are called the preliminary features of the known continuous sound, and are placed in the known continuous sound feature database. (9) If the continuous continuous There is no sample in the sound database. N is a linear estimate of the unknown continuous sound. The horse's cepstrum (off) and the linear predictive code of the known continuous sound are cepstrum (4) ··+1), and the (10) number is weighted. The mean and the number of variances, the mean value of the Εχρ and the matrix of the variance are called the preliminary features of the known continuous sound, and there is a library of known continuous sound features; (10)- Repeat calculation and stabilization of each known continuous sound feature Method for making known continuous sound features Each known continuous tone has a mutual stability feature (represented by a matrix of mean values and variances), a standard model called the known continuous tone, placed in a known database; '曰(11) will An unknown sentence or name is cut into D unknown continuous sound methods; (12) - A simplified shell classification method in D unknown continuous sounds, for each - 30 201106340 Unknown continuous sounds in known continuous sound features (four) library t , select the F most similar known continuous sounds - an unknown sentence or the name is represented by the known continuous sound of the _ matrix; (10) using the known continuous sound of DXF_, comparing the sentence and the name database with all sentences and names, Look for - the most likely known sentence or name; (10) - a method of correcting a continuous tone feature, so that the speaker's sentence or name is identified as being read. 2 • According to the identification method of the patent application scope 1, the method of the steps: the item mentioned in the item - the miscellaneous words recognize various languages (3) delete the sound waves or murmurs without speech, including the signal points in the second a small segment 'Calculating the number of variances of signal points and the variation of the general murmurs' as recorded _ the number of variograms is smaller than the murmur, and the coffee goes to that time; (^-hours are the points of the number, the sum of the peaks of the two recorded points and the general miscellaneous _ The age of the 辑 跋 辑 Μ 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者 后者The matrix, the steps are as follows: (= 2 aliquots - a continuous sound wave recording method 1 (10) Sexual changes closely estimate the nonlinear changes of the sound waves, the full length of the sound waves into the transition ^ 'mother time period formation - the rotation box, one Continuous sound has a paste frame, no tender 11 (FUW stack, can be freely stretched and covered 201106340, Long sound wave 'is not a fixed length Hamming (Hari (8) window; (b) inside the mother box, with - linearly varying sound waves over time; the mode of rotation is estimated over time (c) with Durbin's loop mode R ( i) = ^S(n)S(n + i\ ,·>〇η=〇E〇=R(〇) Yes, + =_-乞〇_刀]/κ }-\ 卜1 α,ω Et = (1-^2) £m aj=aT^ l<j<P Find regression coefficient, square estimate #W, call the silky estimate mother (LPC) vector, reuse | 1-1 / α '«·=α.·+Σ(~)α.·-;α,/·>ι</<ρ j-\ l α'.=Σ(Ζ«, p<i /=α-ρ The converted linear predictive coding (LPC) vector is a stable linear predictive coding cepstrum (LPCC) vector α', ·, ΐϋΡ; (d) a linear predictive coding cepstrum (LPCC) vector is used to represent a continuous tone 32 201106340 4 According to one of the methods described in the first paragraph of the application of the scope of the transfer, the step (5) further includes a -type method, the steps are as follows: (5) will tear the sound wave (four) wheel, round _ ^ unknown _ sound with scale length Box, no money wave g,, free telescopic cover all sounds _ point; " A sample that can be used to identify unknown continuous sounds in various languages can be identified (with a regression pattern that varies linearly with time in a flexible box). (c) Use Durbin's Cyclic mode N -im^^)S(n+iX i>() Λ=〇~U Ε0 = R(Q) Outline - , (0 Ei=(l-kf)E τ(Ρ) is JSP Calculate the regression coefficient The least squares are calculated... (LPC vector) (d) The LPC vector is further transformed into a stable linear prediction coded spectrum (Lpcc) vector (6) using the formula corpus 33 201106340 a'i=£(f)W μ The estimated code scrambling hunger (10) vector, (tearing Lpcc matrix), as the classification model of the unknown continuous sound. According to the application of the patent (4), the use of a sample does not recognize the various languages. The financial law's step (6) contains a simple Bayesian (Bay which identifies the unknown continuous sound method, the steps are as follows: (4) - The characteristics of an unknown continuous sound are classification models, using a cadence matrix 5 7 -1, ·., £ 5 ^=1 d The main one ^L· ,.·.,ρ, the table is not, in order to quickly identify, the office has an LPCC and tampers the independent random variables, there is a normal allocation, if the continuous continuous sound and A known number of filaments 9 c, ι·-1'.··'% (the calendar is the total number of known continuous sounds) a comparison of the average number and the number of variances (/w4) Knowing the continuous (quad) mean value and the estimate of the variance, the conditional density function of Lai X Π μ 4ΐπσίΗ Knowing the classification pattern of continuous sounds _ line (4) Estimating the cepstral variation lemma 4) Using the average number in the known continuous phonological model and the _ Yi Bei's classification to materialize the continuous sound characteristics and find the continuous sound ^ Most like this is the unknown continuous standing, the male 曰 X known continuous sound to the unknown even 34 201106340 Continuation χ similarity in the following formula Zhong Zhong Ci) indicates [Υ42πσΰ (\ , C) for rapid identification, with logarithm Simplify the conditional density function (b) in (b), and delete the constants that do not need to be calculated. The Bayesian distance is also called Bayesian classification. %)); ^ 2 n % (4) Calculate the Bayer distance (10) value in (c) for each known continuous tone e; ' f = 1,%; (6) Select - a known continuous sum in the feature database, its shell distance % ;) The value 疋 is the smallest and is judged as the unknown continuous sound. 6. According to the application of patent scope i, a non-sample can identify the recognition method of each language's step (10) contains - money to calculate and stabilize all known continuous tone features of the library: ω if known in continuous The sound feature #料库 has a known continuous sound, which is not the smallest in the characteristic data 2 of the same known continuous sound in the permanent database of known continuous sounds, and the mine (four) distance is shouted in the special shoe library. The linear predictive coding cepstrum of the continuous sound is also known as the permanent data (10) 〇 the most recent known continuous performance sound; (b) the estimated coded cepstrum of the most recent continuous sound (LPCC) N+1 N The average value and the linear pre-weighted average of the known continuous sound are the new average of 35 201106340 of the known continuous sound, and the weighted average of the N variances of the (four) most recent consecutive sounds is the new variance. This new feature of the new mean and new variability is placed in the feature database; (7) Repeating the butterfly a) to (6) multiple times, the last new feature, (4) the mean and the matrix of the variance 'called the known continuous sound Registration model; 'U secret Chou material library encoding a known side continuous tone spectrum (Cong) unchanged. • According to the application of the patent scope, the use of a sample can identify the method of the various languages, the steps ((1) include an unknown sentence or the name is cut into an unknown method: (4) Calculate the phase per unit period The distance between the two adjacent signal points is the sum of the distances. If it is too small, the period is muted or murmur, and there is no voice signal. (5) Silence or noise accumulates too much in the adjacent unit period (longer than the two syllables in the continuous tone), The time period should be two consecutive sound dividing lines, which should be cut, an unknown sentence or name cut into!) an unknown continuous sound; (c) a continuous sound to remove mute and noise, the elastic frame normalization, the least squares calculation linear prediction A coded cepstrum (LPCC) vector representing an unknown connection, ', 曰$ sub or name—shares D linear predictive coding cepstrum (LPCC) ExP matrix representations. According to the first item of the application for the special fiber, the method of identifying the various languages is not included. The step (12) further includes the following: 36 201106340 After an unknown sentence or name is cut into D unknown continuous sounds, each one is unknown. Continuous ice refinement of the shell classification method, in the feature database, each known continuous sound ς.Κ} and the unknown continuous sound ^ bech distance ^c- ^= Σ ln (<am) + Φ^- l^L) Find the nearest F known continuous sounds (F may contain multiple languages at the same time), and the unknown continuous sounds are represented by the F known similar continuous sounds; (8) Therefore - an unknown sentence or name has a D column f A known similar continuous tone indicates 'that is, the known similar continuous tone of the sentence or name in the DxF scale:: Very high. According to the method of claiming the scope of patent application, the method of identifying the various languages (4) is not complicated, and the step (13) further includes the following sentence and name identification methods: (a) in the sentence and name database, select A sentence or name that is approximately equal in length to the speaker's sentence or name (ie, a sentence and name with D ± 1 known continuous tones); (b) If in the sentence and name database, select the sentence or name of the comparison , its length is just as long as the speaker's sentence or name is an unknown continuous sound) then 'D then D each column F similar known continuous tones and the selected D consecutive words of the selected sentence or name According to the order comparison, see if there are any known continuous sounds in the f-like similar known continuous sounds, such as the similar continuous sounds in each column, including the matching sentences or the names in the order. 201106340 Zhilian, S will have all D unknown continuous sounds correctly identified, the matching sentence or name is the speaker's sentence or name; (5) If the sentence and name are compared to the sentence or name there are D known companies Jiyin' Comparing the speakers, the D consecutive sounds are not fully recognized correctly (not in F similar known continuous sounds) or the matching sentences or names are not D lengths, the invention is a secret crane screening, with the scales she knows the continuous sound In the middle and the bottom three similar known continuous tones, in order to compare each of the sentences or names in the sentence and name database with 'D or D ± 1 known continuous tones, in the database, in the database Select the most likely match sentence or the sentence or name of the speaker. The probability is proportional to the known continuous pitch of the sentence or name divided by the full length (D or D ± 1) in the 3xF window. 10 · According to the identification method of “not using samples” mentioned in item 1 of the scope of patent application, step (14) further includes a method of modifying the continuous tone feature to make the sentence or name identify correctly: (a) If a sentence or name Touch error, which is determined to be in one of the D unknown continuous tones, one or more of them are not in their F similar known continuous tones, assuming that c indicates that one of the unknown continuous tones is not in it? In a similar known continuous sound, N consecutive sound features are obtained using N most similar known continuous sounds {={/ν'4}'ί = ι,···'"'; = 1'. .·, £, € = 1', p, the average value (or weighted by order) ', keep 'this average one represents the new feature of the unknown continuous sound c; 38 201106340 (b) in item (a) In the linear prediction of the tester's pronunciation, the spectrum and the N average of the N most similar known continuous tones are N+u solid: the weight average is the new average of the unknown continuous sound, and the N most The weighted average of the N variances of similar known continuous tones is the new variation of the unknown continuous tone: This new average and the variation ExP matrix represents the new standard model of the unknown continuous tone; (c) Retest the unknown The sentence or name should be successful. 39