TWI310543B - A method for classifying similar mandarin syllables using two consecutive bayesian decision rules - Google Patents

Description

1310543 IX. Description of invention: 1. [Technical field to which the invention belongs] There are many identical or similar monotones in the country, which makes it difficult to identify single tones, especially the physiology and rationality of the pronunciation and pronunciation of men and women, the same - monophonic, Sounds, the difference is very different 'there is the tone of the provincial and local people's pronunciation, even if the same - single tone will sound like a single tone. The present invention is to find a truly unknown tone in a similar known tone by using a similar known sound with an unknown single tone. According to the detailed description, the sound wave system of the county has a flexible frame with the same length. There is no fortune, no overlap, and the monophonic sounds of different lengths are framed. The new one is normalized and _ into a green chowder. LPCC) vector. First use the (four) Bayesian classification (Bayesdecisic) n rule) to find the most known single tone of the unknown tone. The precision Bayesian classification finds the ίί tone from the paste known to her mono. Normalization of sound waves and extraction features are simple and time-saving. The identification is fast and accurate, and the recognition range is wide. 2. [Prior Art] When the early sound is pronounced, its pronunciation is expressed by sound waves. Sound waves are a kind of system that changes linearly with time. - A single sound wave contains the characteristics of the perceptual state, and it also changes nonlinearly with time. When the same single tone is pronounced, there are a series of identical sorrow characteristics, which are nonlinearly stretched and contracted with time, but the same dynamic characteristics are arranged in the same order, but the time is different. When the same single tone is pronounced, it is very difficult to arrange the same dynamic characteristics at the same time position. It is more difficult to identify because of the similarity of the single tone. As early as 6 1310543 - a computerized language predicate system, we must first extract the sound waves related to the heart ^, that is, the dynamic characteristics, the language is not _ murmur, such as the human voice ° color shell = 'speaking mental and physical and emotional and voice Identifying irrelevant predecessors. The same features of the same tone are then arranged in the phase_time position. This - even (4) features are represented by the "equal length series of feature vectors, called a single tone (10) sign model. At present, the speech recognition system has to generate a large-scale feature model too complex.

Miscellaneous and time consuming. The same features of the same tone are difficult to align in the same-time position, making alignment difficult. - General #音识别(四) There are three charms to defend: the characteristics of the training, the normalization of the features (the size of the special type - the same feature of the _ single tone arranged at the same time position) and the unknown tone to identify a single sound wave feature commonly used There are the following types: energy, zero crossing points (_(10) where"

Extreme number (extreme c〇unt), _ (f〇r_s), linear prediction coded cepstrum (LPCC) and Mel frequency cepstrum (MFCC), where Lpc(:

And mfcc is the most effective 'and is commonly used. The LPCC is the most reliable and stable word for a single tone. Its sinuous singularity represents a single-sound wave, which is returned to the turtle with a least squares estimate, and its estimate is converted into a cepstrum, which becomes the LPCC. The MFCC converts the sound waves into frequencies using the Fourier transform method. The auditory system is estimated based on the ratio of the frequency of the Mel. According to scholars s. B. Davis and Ρ Mermelstein published in IEEE in 198

Transactions on Acoustics, Speech Signal Processing, 7 1310543 vol. 28, No. 4 Published by parametric representations for monosyllabic word recognition in the continuous sentences Sentence Time Warping (DTW) ' MFCC features than LPCC feature recognition rate high. However, after many speech recognition experiments (including my previous invention), the Bayesian classification method, the recognition rate of Lpcc special gift is higher than the MFCC feature, and saves time. As for the five-word recognition, there are many ways to adopt it. There are dynamic time-wai*ping ' vector quantization and hidden Markov mode (hmm). If the same pronunciation changes in time, one side compares the same feature to the same time position. It is good to recognize the s tolerance, but it is difficult and too long to pull the same feature to the same position and cannot be applied. Vector quantization, such as recognizing a large number of monotones, is not only inaccurate, but also time consuming. Recently, the HMM identification method is good, but the method is complicated, too many unknown parameters need to be estimated, and the estimated value and the recognition time are calculated. Recently, TF Li published in Paper Reocgnition, vol. 36 in 2003, using the Bayesian classification method in Paper Recognition of mandarin monosyllables, to compress various length and length series of LPCC vectors into the same size classification model with the same database. The result is more than γ.Κ. Chen, CY Liu, GH Chiang, MT Lin published in 1990 in Proceedings of Telecommimication Symposium, Taiwan. The recognition of mandarin monosyllables 8 based on the discrete _en Mark〇v Markov's model sister (臓) square silk, but Wei touched the lion _, and the same single tone is difficult to squint to ride the Lai steel (four) position, for similar tones, it is difficult to identify. The speech discriminating method of the present invention is directed to the shortcomings of the description, from the financial aspect, according to the sound characteristics of the sound waves, and the non-linear changes over time, and the self-extracting speech special wealth method. The monophonic sound waves are normalized first and then converted into an equal-sized feature model sufficient to represent the tones, and the same tones have the same characteristics at the same time position within their feature models. There is no need to manually or experimentally adjust the unknown parameters and gates. Using the simple Bayesian classification method, the unknown early classification model can be compared with the known mono standard model in the database, without recompression, distortion or phase _ feature comparison. Therefore, the speech recognition system of the present invention can quickly perform feature extraction, feature normalization and recognition. In order to improve the tone recognition rate, the continuous use of the second Bayesian classification method will make it difficult to identify t-like early-sounding ridges, greatly improving the overall target tone. Recognition rate 2' Wide application, the use of the elastic frame can identify very short or very long monophonic sound waves, and is also effective for the recognition of very short syllables in English. [Summary of the Invention] (1) The present invention provides a Linyin identification wire. It does not have a voice sound wave removed. (2) This (4) provides the method of normalizing and extracting features of fresh sound waves. It uses A equal elastic frame, no overlap, no filter, can be free to cover all wavelengths according to the single 1310543 2 tone wavelength, and can convert the dynamic miscellaneous changes of the series p object into a uniform size. The feature model, and the feature models of the same monophonic sound waves have the same features at the same time position. It can be identified in time to achieve instant recognition of the computer. (3) The present invention provides a simple and effective method for recognizing unknown tones. Sentences, • The wrong rate is minimized, the calculation is small, the recognition is fast, and the recognition rate is high. ~ (4) The inventor's invention provides a method for extracting monophonic features. A single sound wave has a dynamic characteristic that changes nonlinearly with time. The present invention estimates a sound wave that changes nonlinearly over time using a regression model that varies linearly with time. The least squared estimate of the regression unknown coefficient (LPC vector) produced. (5) This hair secret ride has a voice sound wave (sound wave rhyme). With fewer equal elastic frames, all signal point features are covered without overlapping. | Suppose it is a single tone, delete the material, and not because it is too long, delete or compress some signal points. As long as the person can discern the single 曰', the present invention can extract the single tone. According to the voice identification of the invention, the system applies each signal point with speech, and can extract speech features as much as possible. Because (4) the two miscellaneous frames do not overlap, the number of frames is small, which greatly reduces feature extraction and calculates LPCC time. (6) The speech discriminating system of the present invention can clarify the monosyllabic speech too much. When the speech is too fast, the single tone is very short, especially for the English syllable 10 pronunciation, the length of the elastic frame of the present invention can be reduced, and the same number of π equal length elastic frames are still used to cover the short sound wave. Generate an LPCC vector. As long as the short human is discernible, the f LPCC vectors can effectively represent the feature model of the short tone. Speaking too slowly, the single tone is longer. The bullet _ will elongate. The generated box LPCC vector can also effectively generate long sounds. (7) The present invention includes a set of voice recognition secrets. First, multiple people emit the same known tone, producing a variety of sound waves that are not long or short. Weaving the same elastic frame by I's no filter, no overlap, covering all sound waves, producing 10,000 Lpcc vectors for a known tone. Multiple samples, that is, multiple feature models representing the known tone. Multiple feature models are internally identical, with the same feature vector set, ie, the paste (10) for each sample is approximately the same as the sequential LPCC vector. Then calculate the LPCC average and the number of variances for the sample of the special type, and obtain a matrix with the same size and the average of the undertones. The standard model called the known tone has a database order. In the same way, an unknown monophonic sound wave is used with an equal-length elastic frame, without a chopper, and the entire sound wave is not overlapped to produce a paste LPCC vector. The "solid Lpcc vector is called a classification model of unknown tones. When comparing the unknown tones and the known tones in the database with the simple Bayesian classification, all Lpccs in the classification model are assumed to be independent. Normally assigned random variables whose mean and variance are replaced by the sample mean and sample variability in the known monophonic standard model. The average and unknown in each known mono standard model are calculated in the database. The LPCC distance in the classification model of the monophony is adjusted by the variation of the known tones. The most similar known singles are selected with the unknown tones #, and each similar known is calculated by the second Bayesian classification. The miss-categorization risk of the £ sample closest to the unknown tone. Then find a similar known tone from the total Bayesian distance of #similar known tones, its total The distance is the smallest and is judged as an unknown single tone. [Summary of the Invention] The invention execution program is illustrated by Figure- and Figure 2. The diagram _ is the process of establishing the database. The surface of the material library has a fresh surface with a known tone. Know the characteristics of the tone. - Known The sound is entered into the receiver 2G in the form of a continuous sound wave. The digitizer 3 converts the continuous sound wave into a signal point of the serial sound wave number. The previous processor 4 has two methods of deleting: (1) calculation - The number of _ points in the hour and the variation of the general murmur. The smaller than the latter, the ship tree segment is not green, should be. (2) Calculation - the hourly segment of the gamma-family and the general ride, such as the former Less than the latter, the _ segment is from the voice, _. After the previous processing 40, the 'paid-sequence has the known tone signal point. The sound wave is normalized to take _, and the entire money point of the known tone is divided into ^ Waiting time period. Each time period consists of - box. - One tone - Common · Isometric frame 1310543 5〇 'No test n 'Do not overlap, according to the length of all signal points of the single tone, with the solid frame length from the whole cover Signal point. So the box is called elastic=, the length is free to expand and contract, but the length of the frame is the same. Unlike the Ha ing 固 solid 'filter' waver' semi-overlapping, fixed length, can not be freely adjusted with the wavelength. __ non-heterogeneous changes, sound waves contain - a phonological kinetic energy feature, also over time Changes. Because not overlap, (iv) present

System, more than (2) box, including a single tone, because the signal point can be estimated by the #面彳号&丨' with a linear regression pattern with time to closely match the __ changing sound wave. Win], the flat method estimates the regression unknowns generated within each box - the least squares estimate of the group unknown coefficients is called linear predictive coding (LPC vector). The Lpc vector is then converted to a more stable defeated spectrum (LPGG). - The voice dynamics containing the phase-to-linear variation in the sound-sounding sound are replaced by the equal-sized LPCC vector 6G in the present invention. In order to establish a dragon library, a known single tone is pronounced by a plurality of people, and the multi-person pronunciation sound is compiled into a plurality of samples of f equal-sized LPCC vectors by the _transfer method. The _ Lpcc vector of multiple samples represents an equally known tone, so the fixed LPCC vectors are sequential, and the LPCC vectors of multiple samples should be approximately the same, that is, at the same-time position, multiple sample LPCC vectors. Then calculate the average value of the multiple samples Lp (:c and «number. This (10) LPGG average f-surface fascia - a standard general of the known tones (four), called the standard model 70 of the known tones. 1310543 Fresh model with sample mean and variance represents a known single tone, stored in the database 80. I —π 不00 - an unknown tone recognition method flow. When entering an un

After knowing the tone 2 to the speech recognition method of the present invention, the unknown tone continuous sound wave 11 enters the receiver 21, and the digital machine 3 turns into a series of sound signal points. ^ The image-previous processor 40 deletes the sound without speech. The sound wave is positively banded into the 'extraction feature, and the signal of the unknown single tone having all the voices is squared into a square, and the elastic frame 5G is formed every time period. - There are a total of $ elastic frames ‘no «waves, no overlap, free telescopic cover all signal points. In the mother frame, since the signal point can be estimated by the previous signal, the least square method is used to obtain the unknown coefficient and the value. The set of least squares estimates produced by each frame is called the LPC vector. The lpc vector has a normal allocation. The loss vector is then converted to a more stable LPC cepstrum, which is the (10) vector. An unknown single 代表 represents the feature model in an LPCC, called the classification model (10), and the size of the known single tone model. If a known tone is this unknown tone, the average of its standard model is closest to the LPCC of the unknown tone classification model. Therefore, the simple Bayesian identification method of the present invention is compared with the classification of unknown unknowns _ and Longxian - a known single tone (9). If a known tone is the unknown tone, in order to save time, it is assumed that all Lpccs in the facet of the unknown tone have independent normal assignment, and their mean and & The sample mean and the sample a31〇543 variation are estimated to be 60. First, a simple Bayesian method is used to calculate the loss of an unknown single tone [ 〃 α the distance of the mean of a single day and then adjust by the known tone variation, the resulting _ generation unknown unknown single reading - a known monophonic phase city . Select the highest known single tone 11〇 with the unknown early sound #. A fine Bayesian classification is then used to calculate the total Bayesian distance (mis_categorizati〇n risk) of each sample that is known to be the closest to the unknown tone. From • a similar known single tone to find a similar known tone, its λ sample has the smallest total Bayesian distance and is judged as an unknown single tone of 13〇. 5·[Embodiment] (1) After a single 3-input incense recognition system, the single-tone continuous sound wave is converted into a series of sampled signal k-points (signal sampled p〇ints). Then delete the point that does not have a voice signal. The present invention provides two methods: one is to calculate the number of variations of signal points within one hour. The second is to calculate the sum of the distances of adjacent signal points in the time period. In theory, the first method is better, because the number of variations of the I signal point is greater than the number of noise variations, indicating that there is speech. However, when the present invention recognizes a single tone, the two methods recognize the rate, but the second saves time. (2) After the voice signal point is deleted, the remaining signal points represent a single tone signal point. The sound waves are normalized and the features are extracted, and all the signal points are divided into equal periods, and a frame is formed every time period. A single tone has a flexible frame of equal length, no filter, no overlap, free stretch, and covers 15 1310543 all signal points. The signal points in the elastic frame change nonlinearly with time, which is difficult to represent with a mathematical model. Because J. Markhoul was published in 1975.

Proceedings of IEEE, Vol. 63, No. 3 Published paper Linear

Prediction: A tutorial review states that the signal point has a linear relationship with the previous k-point. The model of the regression that varies linearly with time estimates the signal point of this nonlinear change. The signal point (8) can be estimated from the previous signal point, and its estimated value S, (8) is represented by the following regression pattern: «>〇(1) k=l '' In the formula (1), Α, λ = 1,._.,Λ Is the regression unknown coefficient estimate, /> is the number of previous signal points. In L. Rabiner and B. H. Juang, 1993, FundamentalsofSpeechRecognition, Pren-tice Hall PTR, Englewood Cliffs, New Jersey

Durbin's loop formula finds the least squares estimate, and this set of estimates is called the LPC vector. The Lpc vector method for finding the signal points in the frame is detailed as follows: To sum the sum of the squared differences between the signal points and their estimated values y(8): (2) «=0 k=\ Find the regression coefficient to minimize the sum of squares. For each unknown regression coefficient β'々ι, _.", ask for the partial differential of (2), and divide the partial differential into 〇, and get the normal equation of the corpse: Ρ Σ ~Σ --明〇二Σ 5(8)5(w - 〇, \<ι<ρ (3) 16 1310543 After expanding (2), substitute (3) to get the smallest total squared difference &=2>2 (8)- ΧΣ ) 抑 - 幻 4 4 ) rt k=\ η (3) and (4) are converted to - k) two R(i,, \<i<P (5) k=\ k=l

In equations (5) and (6), use # to indicate the number of signal points in the frame, Ni R(i) = ^S(n)S(n + i), />0 (7) w=0 with Durbin's The loop quickly calculates the LPC vector as follows: E0=R(〇) (8) k, ])]/Ε^ (9) ./=1 α,(ί) = ki (10)

αΚ-Ι^α”, (11) £,=(1-^)^, (12) (8-12) The formula is cyclically calculated to obtain the least squares estimate of the regression coefficient, j· = 1,..., eight (LPC) The vector) is as follows: α,=αΓ, \<j<P (13) The following formula converts the LPC vector to a more stable LPC cepstrum (LPCC) vector fl'.,., / = 1,__" corpse, 17 ( 14)

- an elastic frame produces - hunger (Χ vector κ, .. _, α,; according to the speech recognition system of the present invention, with / Μ 2, because the final LPCC is almost 〇. A single tone to represent the features of the solid LPCC vector 'Also _ a matrix containing nanosolid Lpcc represents a monophonic feature. (3) A known tone is pronounced by multiple people, producing a matrix of multiple samples, representing the known tone, and multiple The average value and the number of variances of the LPCC sample are obtained - the financial matrix contains LP (the average value of the X sample and the sample variation. This matrix is called the standard feature of the known tone, or the standard model. (4) The same as ( 8-15) Counting f of Yang single sound waves

The LPCC vector has a matrix of LPCCs of the same size, called a classification model of unknown tones. In the figure, the speech recognizer 1〇〇 receives a classification of the unknown tone, a matrix of ^^LPCC. Use 1 = 匕 y = 1, 』, "1, ·..', to represent the unknown tone feature model. When comparing with a known tone q, , in order to quickly calculate the alignment value, assume that there is a household , independent normal distribution 'its average and variation, estimated by the sample mean and sample variation within the known mono standard model. The conditional density function of Z is represented by /(Xk). 18 1310543 in Pattern Recognition, Vol. 36 The decision theory in Speech recognition of mandarin monosyllables shows that the Bayesian classification is as follows: Assume that the database has a standard model of known monotones. K, w, for singles The sound c, called, w, the probability of occurrence, that is, the previous probability, then |^, = 1. To represent a decision method. Definition

A simple loss function, that is, the probability of error (mi sc 1 ass ifi cat i〇n pr〇babi 1 i ty ) is as follows: If the decision method j judges an unknown single tone, the loss function i ( c, rf (function... If ^ is judged against an unknown tone, there is no loss with c,, i/(x)) = o. The recognition method is as follows: [hUw, indicating that the matrix value belongs to the known tone c, The range of I. In the case of I, the unknown single tone belongs to the known single tone c, and the average probability of error is (16) '-=l

'(υ) in t(16), Γ7 is Γ, outside the range. All methods of the speech recognition method are also indicated by D, which also divides the range of π known tones. Find the average probability of error in D in the identification method (minimum, in %, W (17) R{t, dr) two minR (t, d) del) meet the (Π) formula identification method d do With the previous probability of Bayesian classification. It can be expressed as follows: , knife 19 1310543 dT(x) = Ci if dif{x\c)>ejf{x\cj) (18) In (18), ; = 'is also known as a single tone ^ The range is for all corpses /, rf = {χΚ / (small. If all known tones have the same probability of occurrence - then Bayesian classification and maximum probability method - Bayesian classification (18) identifies an unknown tone) When calculating the conditional density function / 〇c|c,), / = l,...,w,

Π 4ΐπ\ σ, ijt _1Σ(νΐ% 2 ./ί σίβ (19) In (19) ' /4,.,·^, (the total number of known tones). For the convenience of calculation, take (19) Logarithm, and delete the constant, get β

The Bayesian classification (10) becomes a (10) value (2〇) for each known single tone, and %) is also called an unknown tone and a known tone, or a Bayesian distance plus a 1_ break. In the formula (10), the name of the LPCC value in the unknown single-tone classification model, the average number of samples and the sample variation of the county model axis with known tones and the first-to-four (four) classification method are called:赖 库々's early knives 叱) value up to #min, a, A, 匕 音. From the similar known single tones unknown singles #similar known single from similar known tones find ^L twice The Bayesian classification method produces an unknown tone. (4), /=1,...to 20 1310543

...' Μ戈表, a (10) matrix of Lpcc for the second sample of a similar known tone. If the first, · a similar single tone, sample samples), in the second Bayesian classification towel, the unknown single tone off matrix U and its first / similar single tone, the sample matrix off the series Then, in this classification, 6 is the average. ^The unknown single tone is W belongs to L-like tone, the average value of J is the continuation sample <, matrix, the variance of the shot is IJC / ii6 = V tw *VJ (21) in (2)) In the 'AH her single tone off (four). However, the condition of this I becomes Wei (8) and cannot be calculated. Therefore, (4), a known single tone is changed 7'^ in (21) where '% is the same as =Γ but Shanpin. Therefore unknown single tone j and the first, similar

When a sample heart {Peach pair, the variation of J is still the number in the database 8〇, · _, % of a single tone. Use the second: know the single audio recording _ material tone mail Na (four) when the distance Orns- her gQrizatlQn such as) (10) then the following expression, β 丄if ίβ ^ (22) used, "value, sigh" may be Negative. In the ruler, a request for "the most money, the ring and the _ eye seems to have the distance of the hat (also known as the total 'ah. And t unknown early sounds similar to the similarity of the known single tone), the smaller the distance In the classification of 21 1310543 h, the total Bayesian distance of the #similar similar singles is the smallest, and the greater the similarity is judged. The second is to find a similar known tone, which is the same as the known single tone. Unknown early 9. The second Bayesian classification is a bit like 々-/?/? classifier, but the invention uses Bayesian distance (2)). In the second classification, each of her is as simple as this experiment. The county is close to (Iberian distance) unknown tone -, _ similar to single use too many samples 'may be caught - some are not very good samples (can not represent this similar

...the individual Ping K first order is similar to the known single tone) will be judged by the first:Beibei classification. The recognition rate is not increased, but is independent. The more 彳m of each similar single tone, the better. It can even contain monos with unclear pronunciations' because the second Bayesian classification has more chances to choose ^ better samples (more, close, unknown), if too soft, Calculating the start value takes time. In addition, the second Bayesian classification is a value calculated for all similar monophonic samples. If too many similar tones (# large) are taken for unknown tones, (10) is quiet. In this real: Medium 'must = 10 is enough. In order to prove that the speech recognition H of the present invention has a high rate of _, the range is wide, and the 22 1310543 extraction feature and recognition method is fast and time-saving, the present invention performs a plurality of speech recognition experiments. First, establish a daily Mandarin monophonic database. This early tone database was purchased from Academia Sinica and was of poor quality and incomplete. There are 493 different Mandarin singles on the surface. The number of samples per tone is not - many, only -, two samples, some are empty, and even multiple samples contain two tones 'some as many as 9G samples. Pronounced by multiple people, per person • only pronounced - times. When using the Bayesian classification, each ® language pronunciation must have at least 6 samples. When any __ sample is tested, the other samples are tralningsamples (at least five samples per ensemble are averaged and the number of variations). Only 361 Mandarin singles were combined, for a total of 4155 samples. The quality is not very good. According to the quality (student listening test), there are three categories, and the best one has 233 singles (a total of 3152 samples). The second category has 257 tones (a total of 3302 samples). The third category is the original 361 tone (a total of 4155 samples). Test one sample, and other training samples as the average and variance. It is stored in the database 80. The test sample and the sample for calculating the mean and the variance are completely separated. The three types of monophonic recognition results are placed in Table 2. In Table 2, the first Bayesian classification finds the most similar known tones with the unknown tones 71/(=5, 10, 20) according to the similarity Φ,) (20). In Table 1, there are three most similar tones of three test tones. In the second Bayesian classification, samples of similar tones all use the third type of all samples. Because the sample has a better chance of finding the material than the second 13th Bayesian classification of 23 1310543, the best spelling 2 3 table - let 'the first Bayesian classification recognize the first class-like tone, recognize H 26% ( It is identified by a similar known tone. Second: owing to Bayesian _ _ each unknown single tone (the job sample) of 2 已 〇 〇 早 ’ ’ 辨 辨 辨 辨 辨 辨 辨 辨 辨 辨 94 94 94 94 94 94 94 768 768 768 768 768 768 768 768 768 768 768 768 768 768 768 There are also 2 samples, which were originally broken by the first-time division. From the country == fruit '...12 shoulders are the best, (the most time-saving and identification rate (four). Lang paste elastic frame does not overlap, the correction can fully extract an early sound feature, such as A is too large, increase the recognition ability is very small, very time consuming However, due to the quality of the central research, the quality of the capital from Sakizaki, it is advisable to take only a few (material) best samples for a similar single tone. ^ Monophonic similarity size %) The order is as follows: The 1G of the unknown tones are most similar to the known single λ. ΐ曰1 2 3 4 5 Table 1 tone. unknown

10 Zhengfu tempering quasi-Su Lu wrong group mother @ will two temples strong bird six Liao Xian anger @兰棱 for the name of Ning will wear two" and "two, should be 'as the same Mandarin single, this is by The Academia Sinica provides 24 1310543 for the database. The total number of samples of #similar tones tested by unknown method is 4155 ---- top5 toplO top20 3097/4155= 7454 3208/4155= 7721 3254/4155= 7832 max top5 max toplO max top20 3337/4155= 8031 3581/4155= 8619 3758/4155= 9045 Table 2. 361 Mono Topi 2341/4155= 5634 筮 渌 疾 八 细 啊 啊 啊 啊 啊 啊 似 似 似 似 似 似 似 似 似 似 似 似 似 似 似 似 似 似 似 似 似 似 似 似 似 似 似 似 似The first - owed to the shell's charm: wrong, but the second Bayesian classification judged that there were 1 〇 26 test samples, the brother's knife method was judged 'but the second Bayesian classification error was 113 test samples. Total number of test tone samples = 3302 top5 toplO top20 2933/3302= 8889 3025/3302= 9161 3064/3302=.9279 max top5 max toplO max top20 3006/3302= 9104 3118/3302= 9443 3179/3302= 9628 Dive, but The second Bayesian classification judged that there were 82 test samples. Clock, but the second Bayesian classification error has 29 test samples. Total number of test tone samples = 3152 top5 toplO top20 2872/3152= 9112 2947/3152=. 9350 2993/3152- 9496 max top5 max toplO max top20 2918/3152= 9258 3012/3152= 9556 3073/3152-9749 Ϊ:ΐ3 Then, the first-person Bayesian knife method was robbed, but the second Bayesian classification error was 21 test samples. 257 topi 2273/3302=. 6884 233 topi 2246/3152= 7126 25 1310543 [Simple diagram of the diagram] Figure 1 shows the database establishment process. Figure 2 shows the flow of an unknown tone recognition method. [Explanation of main component symbols] (1) A brief description of the component symbols in Figure 1: 1. Input known tones (1) 2. Known monophonic continuous sound waves (10) 3. Known monophonic receivers (20) 4 . Sonic digital converter (30) 5. Remove noise (40) 6. Elastic frame normalizes the sound wave (50) 7. Minimum square method calculates the LPCC vector (60) 8. Calculate the average value of the known single tone LPCC and Variance (70) 9. The known tone database contains the standard model of all the mean and the variance (80). (2) The simple description of the symbol of Figure 2: 1. Input unknown tone (2) 2. Unknown single Tone continuous sound wave (11) 3. Unknown monophonic receiver (21) 4. Acoustic digital converter (30) 5. Remove noise (40) 26 1310543 6. Elastic frame normalized sound wave (50) 7. Minimum flat method Calculate the LPCC vector qo) 98 • The (10) vector of the unknown tone represents the unknown tone classification model (10). The known tone database contains the standard model of all the mean and the variance.

10. Compare each known single tone standard model with the unknown (100) monotone classification model. The first-person Bayesian classification method selects the # most similar known tones of the unknown tones (110) 12. The second time The Census method calculates the total Bayesian distance (120) of each sample that is similar to the known single tone, which is closest to the unknown tone. The knife method selects a similar known tone, and the total Bayesian distance of its I sample is up to The smallest is judged as an unknown tone. (130) 27

Claims (1)

1310543 Yu Zhizhi revised replacement page ten, the scope of application for patents: 1. - A method of continuous secondary collection and classification of similar Mandarin singles, including: (1) Delete the signal point without the voice signal (sampled p〇 Ints) or murmur; (2)--A normalized sound wave normalization and sizing size--feature transfer method: Normalize the sound wave with an elastic frame and convert it into equal-sized LPCC feature matrix, and convert the same single sound wave into features The same matrix; (3H solid standard model method for known monophonic features, stored in the database, the standard model contains the known monophonic bribe _ sample flat touch variation; (4) - an unknown monophonic sound normalization and extraction feature method: The sound wave is normalized and the y-size is similar to the known fresh-small scale (4) levy, and the ^ is the unknown mono-faceted type 'with the mispredicted copper spectrum ((10)); (5) A simple Bayesian classification method. Compare the unknown monotone classification model with all the known monophonic standard models of the poor stock, and find the known single tones of the unknown monophonic 仏u. The similarity of the unknown monotone has been: Cognac, (6)-Second sub-classification: In each of her known tones, the total Bayesian distance of the smallest κ samples is called this similar known tones. With the unknown single tone = distance 〇 nis-categ〇rizatlon risk), in #similar similar to find a single lopsided separation she known single country & a continuous quadratic Bayesian two-wave signal point Or the method of Miscellaneous ' 其 其 令 步骤 步骤 步骤 步骤 步骤 步骤 步骤 ' ' ' ' ' 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 Change Wei is less than the murmur variation number '_ go to two I:, the inner signal point, calculate the sum of the distance between the adjacent two _ points and the distance between the two signal points of the general segment. If the former is smaller than the latter, delete the time. 3. ==τ A continuous quadratic Bayesian classification method to identify similar countries = towel Jing) package f material sound wave normalization and take the size of the characteristic matrix The method, the steps are as follows: ω- a turn, know the single sound _ point nucleus, for the singular change of the regression model, the sound of the sound is difficult, the full length of the sound wave is divided into ^ time period 'per time period - the rotation box , - a single tone mixed = sex box 'no filter, no overlap, you can freely extend the full length of the covered sound waves, not a fixed length of Hamming window; (6) each frame with a linear change with time The regression model estimates the acoustic waves that change nonlinearly with time; (c) uses the Durbin's cycle method Ni ^(0 = ^ S(n)S(n + /), 7 > 〇i-\ ^ wind /) - outside · ~_)]/ι ./=1 29 1310543 Weng Zhizhi correction replacement page aj ~aj ',, k〇,, \<j< Ehf, Ei-' ^J<p Find the minimum vector of regression coefficient, reuse Squared estimate, called Linear Predictive Coding (LPC) a \ = irl Μ 1 < / < P . cover (+«, P<i-type conversion LPC to Dong is a stable Lpc cepstrum (10) c) vector; with 亀 linear predictive coding cepstrum (LPCC) vector representation— A single-tone f WLPCC feature matrix. 1 Patent sub-item - a continuous second-time Bayesian age identification of similar countries = Early method, financial step (3) contains a standard model method for known monophonic features, the steps are as follows: (a) - a known single tone, the speaker's pronunciation, the length of the sound is not - because the same It has been gone, and the sound waves of different lengths contain a nonlinear change with time but the same dynamic characteristics, but the time position is different; (6) The sound wave of the length of the monophonic sound is adjusted and converted into multiple LPCCs by using one elastic frame. For the matrix samples with the same characteristics, the Lpcc sample mean and the sample variance are obtained. The sample mean and the sample variance are represented by a matrix 30 1310543, which is called the mono standard model. 5. The roots of the wealth, please use the number of ^ ^ ^ ^ ^ Bayesian cheek method to identify similar countries 5 my early sounds, the towel step (4) contains - an unsatisfied sound wave normalization = take the feature green, the surface is unknown Single thank you, the steps are as follows: The unknown single sound wave is divided into the release period, each time period consists of - an elastic frame, an unknown single-tone shirt scale long and long frame, no Wei Bo, no overlap, freely telescopic cover all sound waves Signal point; (6) In each miscellaneous frame, use a regression pattern that changes linearly with time to make nonlinear changes with time; 13 (c) Use Durbin's cycle method ΙΎ-1 w»0 R(i) = Y ,S{n)S(n + i), />〇五〇=_ <=_-Ιχ', /-))]/丨 aj:ai, ^j<P ° least squares regression coefficient least squares estimate (LPC vector); (d) then formulate LPC vector A A + 31 /-11310543 Ρ :Σ j=i-P I is converted into a stable LPC cepstrum (LPCC) vector; (e) an approximate PCT vector, (a floating PCC matrix) is used as a classification model for the unknown tone. 6.=Please touch the patent] The method of identifying the similar country two 连续 by the continuous two-follow classification method, the step (5) includes a simple Bayesian (B qing sub) classification method to identify the unknown single tone method, the steps are as follows: a) - The characteristics of an unknown single tone are the classification model, using a sand (10) matrix especially seven} '^^...^ means 'for quick identification' fresh (10) W is assumed to be an independent random variable, with normal allocation, if unknown For monophonic and - a known monophonic, the average and the number of variances of (9) are estimated using the average number of samples and the number of sample variations in the known monophonic standard model, 条件W conditional density record / (small), below Expression: Π ^σ,β σίβ In the above formula, the classification of the seven-singular unknown tone is based on the classification of the ship's predictive coding (LPCC), but (~, 4) can be used with the known monophonic & standard model. Estimate the average number of samples and the number of sample variations; (6) The simple Bayesian classification is to find a known single tone c in the database, most like this unknown single tone, a known single tone is unknown Degree to / (small,) large 32 丄 31〇 543 丨 Qi "month" 曰 correction small said, ~ ~ one ~ one ^ The constant that must be calculated is determined by the speed (10)' with the logarithmic simplification conditional density function, and deleted: (^) = Σ / η (σ〇 () + ~ Σ (Xje ii 2 β (d) For each known J (early 曰c,,, = 】,..,, calculate %) value; 6 in the poor library, select #known monophonic Ο丨, 』, it's still Is the 7th minimum known as the unknown known monophony of the unknown single tone. • = __ 峨 蝴 蝴 蝴 辨 辨 辨 辨 辨 辨 辨 辨 辨 辨 辨 辨 辨 辨 辨 辨 辨 辨 辨 辨 辨 辨 辨 辨The classification method identifies the unknown single-tone method she had known, and the steps are as follows: a In the second Bayesian classification, κ represents - #similar known tones of an unknown single tone, to * _ ' " = 1,", Rolling, "., ruler," indicates the first tamping sample similar to the known early sound, to /_ίχΜ ., Λ. ίρΓρ „ ^ —^^'^卜^' The coffee matrix representing the LPCC of the sample. If it is unknown, the target "day custody and sample 4" is aligned, and ={;丨' is assumed to belong to the first known similar single, + and variable meals θ" 2 and early as a, its mean σ Γ ..., "Yes (10) Variance, then the similarity of the single-opening Laikou is not calculated by (10) with the secret factory's no, the pair is =1 feet, £(x^) = p^)+lZ(^^y · JK ^ μ (j 7 ' (6) counts the sum of the values of her mono (10) 咐 Napin (may be negative, because it is _ m solid minimum table is not unknown mono and / 33 1310543 lost July The daily correction replaces the total Bayesian distance or similarity of a similar known single tone. The smaller the total Bayesian distance, the greater the similarity; (C) Select a similar one from the unknown known singles of the unknown tone. Knowing the single tone, it has the smallest total Bayesian distance for the unknown single tone, and the unknown single tone is judged; (d) by identifying the test result, 庐12, /M2, deducting 5, is the best, because of the elasticity The frames do not overlap, and the 12 elastic frames can fully extract a single tone feature. 34