TWI395200B - A speech recognition method for all languages without using samples - Google Patents

Description

An identification method that can identify all languages without using samples

A continuous tone contains one or more syllables (monophonic). The present invention can recognize all languages without using samples of continuous sound.

The invention uses 12 elastic frames (windows), equal length, no filter, no overlap, and converts sound waves of continuous sounds of different lengths into a matrix of 12×12 linear predictive coding cepstrum (LPCC), an unknown The continuous tone is represented by a matrix of 12×12 linear predictive coding cepstrum. A 12 x 12 matrix is considered to be a vector of a 144 degree 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 unknown continuous sound (LPCC).

The invention comprises 12 elastic frames normalizing a continuous sound sound wave, and the Bayesian comparison method finds a known continuous sound for the unknown continuous sound of the speaker in the database, and divides an unknown sentence of a speaker into D sounds. Unknown continuous sound, and a window screening method, screening a known sentence as a speaker's unknown sentence.

When a continuous sound is made, its pronunciation is expressed by sound waves. A sound wave is a system that changes nonlinearly with time. A continuous sound wave contains a dynamic characteristic and also changes nonlinearly with time. When the same continuous sound is pronounced, there are a series of identical dynamic characteristics, which are nonlinearly stretched and contracted with time, but the same dynamic characteristics are arranged in the same order according to time, 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. Similar continuous sound Tedo, making identification more difficult.

A computerized language recognition system first extracts sound waves related to language information, that is, dynamic characteristics, filtering and language-independent noises, such as human voices, tones, psychology, physical and emotional speech and speech recognition. The same features of the same continuous tone are then arranged at the same time position. This series of features is represented by a series of equal-length feature vectors, called a continuous-tone feature model. At present, it is too complicated and time-consuming for a speech recognition system to produce a feature model of uniform size, because the same features of the same continuous sound are difficult to be arranged at the same time position, especially in English, which makes the identification difficult.

The general sentence or name recognition method has the following five main tasks: unknown sentence or name cut into D unknown continuous sounds, extracted features, feature normalization (feature model size is consistent, and the same features of the same continuous sound are arranged at the same time position) ), unknown continuous tone recognition, and find a suitable sentence or name in the sentence or name database. A continuous sound wave feature is commonly used in the following ways: energy, zero crossings, extreme count, formants, linear predictive coding cepstrum (LPCC) and The Mel Frequency Cepstrum (MFCC), in which linear predictive coding cepstrum (LPCC) and Mel frequency cepstrum (MFCC) are most effective and commonly used. Linear Predictive Cepstrum (LPCC) is the most reliable, stable, and accurate linguistic feature that represents a continuous tone. It uses a linear regression model to represent continuous sound waves, calculates the regression coefficients by the least squares estimation method, and converts the estimated values into cepstrums, which becomes a linear prediction. Coded Cepstrum (LPCC). The Mel Frequency Cepstrum (MFCC) converts sound waves into frequencies using the Fourier transform method. The auditory system is estimated based on the ratio of the frequency of the Mel. According to the scholar SBDavis and P. Mermelstein published in IEEE Transactions on Acoustics, Speech Signal Processing, Vol.28, No. 4, the use of dynamic time warping method in the paper of Comparatively parametric representations for monosyllabic word recognition in the continuous sentences (DTW), the Mel Frequency Cepstrum (MFCC) feature is higher than the Linear Predictive Cepstrum (LPCC) feature recognition rate. However, after many speech recognition experiments (including my previous invention), the Bayesian classification method, the linear prediction coding cepstrum (LPCC) feature recognition rate is higher than the Mel frequency cepstrum (MFCC) feature, and saves time.

As for language recognition, there are many ways to adopt it. There are dynamic time-warping, 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. The recognition rate will be good, but pulling the same feature to the same position is difficult and the distortion time is too long to apply. Vector quantization, such as identifying a large number of continuous sounds, is not only inaccurate, but also time consuming. Recently, the hidden Markov mode method (HMM) identification method is good, but the method is complicated, too many unknown parameters need to be estimated, and the estimated value and the recognition time are calculated. Recently T.F.Li (Li Zifen) was published in Pattern Recognition in 2003. The paper published in vol.36 uses the Bayesian classification method to compress various long and short series of LPCC vectors into the same size classification model with the same database. The recognition result is better than YKChen, CYLiu, GHChiang, MTLin published in the Proceedings of Telecommunication Symposium in 1990, and the paper published in Taiwan, The recognition of mandarin monosyl lables based on the discrete hidden Markov model, is better with the hidden Markov mode HMM method. However, the compression process is complicated and time consuming, and it is difficult for the same continuous sound to compress the same feature to the same time position, which is difficult to recognize for similar continuous sounds.

The speech recognition method of the present invention is directed to the above-mentioned shortcomings. From the academic point of view, according to the sound characteristics of the sound waves, nonlinear changes over time, a set of methods for extracting speech features is naturally derived. A continuous sound wave is normalized first and then converted into an equal-sized feature model sufficient to represent the continuous sound, and the same continuous sound has the same characteristics at the same time position within their feature models. There is no need to artificially or experimentally adjust the unknown parameters and thresholds within the present invention. Using the simple Bayesian classification method, the unknown continuous sound classification model can be compared with the known continuous sound standard model in the continuous sound feature database, without recompressing, distorting or finding the same features for comparison. Therefore, the speech recognition method of the present invention can quickly perform feature extraction, feature normalization and recognition.

(1) The most important object of the present invention is to simulate and calculate the characteristics of any known continuous sound in any language using features of a plurality of unknown continuous sounds, It is possible to establish the characteristics of any continuous tone of any language without using samples, that is, the present invention can correctly recognize various languages without using samples. In detail, the present invention simulates and calculates the known continuous tone by finding a sequence of N unknown continuous sounds in a space of 144 degrees using a Bayesian distance for any known continuous sound of any language, so as to achieve no known continuous sound. 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 method. It can remove non-verbal sound waves.

(3) The present invention provides a method for normalizing and extracting continuous sound waves. It uses E equal elastic frames, does not overlap, has no filter, can freely adjust the entire wavelength of the cover according to a continuous sound wavelength, and can convert a series of dynamic characteristics in a continuous sound wave with a nonlinear change with time into an equal size. The feature model, and the feature models of the same continuous sound waves have the same features at the same time position. It can be identified in time to achieve instant recognition of the computer.

(4) The present invention provides a simple and effective method for identifying unknown continuous sounds, which has the minimum probability of recognition, low calculation, fast recognition and high recognition rate.

(5) The present invention provides a method for extracting continuous sound features, and the continuous sound waves have a dynamic characteristic that changes nonlinearly with time. The present invention estimates a least squares estimate of the regression unknown coefficients (LPC vectors) using a regression model that varies linearly with time.

(6) The present invention uses all voice sound waves (sound wave signal points). Use less The number E = 12 equal elastic frames, without filters, does not overlap all signal point features with cover. It is not because a continuous sound wave is too short to delete the continuous sound, nor because it is too long, to delete or compress part of the signal point. As long as human hearing can discern this continuous sound, the present invention can extract features from the continuous sound. Therefore, the speech recognition method of the present invention applies each signal point having speech, and can extract speech features as much as possible. Since E=12 elastic frames do not overlap, the number of frames is small, which greatly reduces feature extraction and calculates linear predictive coding cepstrum (LPCC) time.

(7) The identification method of the present invention can recognize continuous sounds that are too fast or too slow to speak. When the speech is too fast, a continuous sound wave is very short, and the length of the elastic frame of the present invention can be reduced, and the same number of E equal length elastic frames are still used to cover the short sound wave. Generates E linear predictive coding cepstrum (LPCC) vectors. The E linear predictive coding cepstrum (LPCC) vector can effectively represent the feature model of the short tone as long as the short human is discernible. The continuous sound waves emitted by the speaker are too slow. The elastic frame will stretch. The generated E-Linear Estimated Coded Cepstrum (LPCC) vectors are also effective to represent the long tone.

(8) The present invention provides a method for stabilizing and adjusting the characteristics of all known continuous tones in a database, so that the features of all consecutive tones occupy their own positions and spaces in a space of 144 degrees for identification.

(9) When identifying a sentence or name, first cut the unknown sentence or name into D unknown continuous sounds. The present invention uses the Bayesian method in the continuous sound feature database to select the most similar F known. Continuous sound. One sentence with D ×F known continuous sounds indicate that the number of unknown continuous sounds may be cut into more or less due to the difficulty of cutting, and the present invention uses three similarly known continuous pairs of sounds in the sentence or name before and after each unknown continuous sound. A known continuous sound, that is, in the sentence and name database, filter a known continuous sound for each sentence or name with a known similar continuous sound of a 3×F window, and then find the most from the sentence and name database. Possible sentences or names, simple method, high success rate (identification of 70 English sentences and names and 407 national sentences and names).

(10) The present invention provides two techniques for modifying the characteristics of continuous sounds such that unknown continuous sounds and unknown sentences or names are successfully recognized.

(11) The present invention treats a single tone as having a syllable continuous sound, and the features of both Chinese and foreign languages are represented by the same size matrix.

Therefore, the present invention can recognize various languages at the same time.

The invention execution program will be described using the first diagram and the second diagram. The first figure is a flow chart showing the establishment of a known continuous sound permanent database, a known continuous sound feature database and a sentence and name database. The continuous tone feature library contains a standard model of all known continuous tones, representing the characteristics of known continuous tones. Entering a known continuous tone or a sentence or name (sentence or name will be cut into multiple consecutive tones) 1 enters the receiver 20 in the form of a continuous stream of sound 10. The digitizer 30 converts the continuous sound waves into signal points of a sequence of sound waves. The previous processor 45 has two methods of deleting: (1) calculating the number of variations of the signal points in one hour and one The number of murmur variations. If the former is smaller than the latter, the hour segment has no voice and should be deleted. (2) Calculate the sum of the distance between the two consecutive signal points and the general noise in one hour. If the former is smaller than the latter, the hour segment has no voice and should be deleted. After passing through the previous processor 45, a sequence of such known continuous tone signal points is obtained. The sound wave is normalized and then the feature is extracted, and all the signal points of the known continuous sound are divided into E and other time periods, and each time frame constitutes a frame. A continuous sound has a total of E equal length frames 50, no filters, no overlap, and all the signal points of the cover are freely adjusted according to the length of all the signal points of the continuous sound. Therefore, the frame is called a flexible frame, and the length is free to expand and contract, but the length of the E elastic frames is the same. Unlike the Hamming window, there are filters, semi-overlaps, fixed lengths, and no freedom to adjust with wavelength. Since a continuous sound wave changes nonlinearly with time, the sound wave contains a dynamic feature of the voice, and also changes nonlinearly with time. Since there is no overlap, the present invention uses less (E=12) elastic frames covering all continuous sound waves, since the signal points can be estimated from the previous signal points, and the regression pattern that varies linearly with time is used to closely estimate the nonlinear changes. Sound waves, using the least squares method to estimate the regression unknown coefficients. A set of least squares estimates of unknown coefficients is generated in each box, called Linear Predictive Coding (LPC Vector). The linear predictive coding (LPC) vector is then converted to a more stable linear predictive coding cepstrum (LPCC). A continuous sound wave contains a sequence of speech dynamics that change nonlinearly with time, and is converted into equal-sized E linear predictive coding cepstrum (LPCC) vectors 60 within the present invention. In order to extract the characteristics of a known continuous tone, first prepare a permanent known Continuous sound database. Each known continuous tone is pronounced once by a standard, clearer. If a tone or non-standard speech is recognized, then the person is pronounced, converting all known continuous tones into E x P LPCC matrices in a permanently known continuous tone database. In the permanent known continuous sound database, for a known continuous sound extraction feature, a database of unknown continuous sounds is prepared first, and there are two unknown continuous sound data bases: one is that the unknown continuous sound has samples, the other is no standard. With a sample database, first find the average and variation of each unknown continuous sound. Find the N most recent unknown continuous sounds around the known continuous sound with a Bayesian distance in the unknown continuous sound database with samples. The N average values of the N unknown sounds and the N+1 weighted averages of the linear predictive coding cepstrum (LPCC) of the known continuous sound are taken as the average of the known continuous sounds, and N consecutive sounds are obtained. The weighted average of the N variances is used as the variance of the known continuous tone, and the E×P mean and the variance matrix are the preliminary feature values 79 of the known continuous tone placed in the continuous tone feature database. If there is no sample in the unknown database, in the unknown continuous sound database, find N unknown continuous sounds around the known continuous sound with the minimum absolute distance. The linear predictive coding cepstrum (LPCC) of the known continuous tone and the N unknown continuous tones is regarded as (N+1) numbers. Find the weighted average of (N+1) numbers as the average of the known continuous sounds, and find the (N+1) number of variances as the variance of the known continuous sound, this E×P average And the matrix of the variogram represents the preliminary feature of the known sound, and is placed in the known continuous sound feature database 79. In a known continuous tone feature database, if a known continuous tone The mean value and the Bayesian distance of the LPCC of the same known continuous sound in the permanent known continuous sound database are not the smallest in the feature database, then find N known continuous points in the feature database with Bayesian distance. Tones, their Bayesian matrix is N minimum for the LPCC of known continuous tones. Finding N average values of N known continuous sounds and the LPCC weighted average of the known single sounds as the new average of the known continuous sounds, and using the weighted average of the N variances of the N known continuous sounds as the 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 E×P new average and the variance matrix are called standard models representing the known continuous sounds. In the database 80, a sentence and name database 85 is created using the known continuous sounds of the known feature database.

The second diagram shows the flow of an unknown sentence or name recognition method. When an unknown sentence or name 2 is input to the speech recognition method of the present invention, a set of unknown continuous sound waves 11 enters the receiver 20, and the digital converter 30 converts to a series of sound signal points. The sound wave of an unknown sentence or name is cut into sound waves 40 of D unknown continuous sounds, and the sound waves without speech are deleted by the previous processor 45 of the first figure. Then, each unknown continuous sound wave is normalized, and features are extracted, and the signal points of each sentence or name each of the unknown continuous sounds having speech are divided into E and the like, and a flexible frame 50 is formed every time period. There are a total of E elastic frames for each continuous sound, no filters, no overlap, and free telescopic covers all signal points. In each frame, because the signal point can be estimated from the previous signal, the least square method is used to estimate the regression unknown coefficient. Valuation. The set of least squares estimates produced in each box is called the linear predictive coding (LPC) vector, the linear predictive coding (LPC) vector is normally allocated, and the linear predictive coding (LPC) vector is converted to a more stable linear estimate. Coded Cepstrum (LPCC) vector 60. An unknown continuous tone represents a feature model with E linear predictive coding cepstrum (LPCC) vectors, called the classification model 90, which is the same size as the known continuous tone standard model. There are a total of D classification models representing D unknown continuous sounds 90. If a known continuous sound is this unknown continuous sound, the average of its standard model is closest to the linear predictive coding cepstrum of the unknown continuous sound classification model ( LPCC). Therefore, the simple Bayesian recognition method of the present invention compares 100 with a classification model of unknown continuous sounds and a standard model of each known continuous sound of the continuous sound database 80. If a known continuous sound is the unknown continuous sound, in order to calculate the time-saving, it is assumed that all linear predictive coding cepstrums (LPCC) in the classification model of unknown continuous sound have independent normal allocation, and their mean and variance are known. Estimation of mean and variance in the continuous sound standard model. The simple Bayesian method is to calculate the distance between the linear predictive coding cepstrum (LPCC) of the unknown continuous sound and the average of the known continuous sounds, and then adjust the known continuous sound variation, and the obtained value represents the unknown continuous sound and one. Continuous sound similarity is known. Selecting the F-similar similarity to the unknown continuous sound The highest known continuous sound is represented as an unknown continuous sound, so an unknown sentence or name is represented 110 by D x F known continuous sounds. After an unknown sentence or name is cut into D unknown continuous sounds, it is difficult to cut into an unknown sentence or the continuous sound and number contained in the name. Sometimes a continuous sound is cut into two. Sometimes two consecutive notes are very close, and the computer is cut into one. Therefore, D unknown continuous sounds are not necessarily the number of true continuous sounds of the speaker, so a certain F known similar continuous sounds do not necessarily contain The continuous sound of the speaker. When identifying an unknown sentence or name, in the sentence and name database 85, test each known sentence and name, test whether a sentence or name is the speaker's sentence or name, and the sentence or name is known from the first one. The continuous tonal ratio is similar to the continuous three-column continuous continuous sound of the D×F matrix similar continuous sound (of course, the first comparison can only compare the last two consecutive similar sounds), and then move the 3×F window (the front and rear three columns are known) Similar continuous sound) 120 finds the second known continuous sound of the sentence until the test sentence is all known to be continuous. In the sentence and name database, the highest probability sentence or name is the speaker's sentence or name (the number of known continuous sounds in the test sentence or name is divided by the number of 3×F windows divided by the number of consecutive words in the test sentence or name) 130 . Of course, it is possible to save time by selecting sentences or names that are approximately equal in length to the unknown sentence or name (D unknown continuous tones) in the sentence and name database. If the sentence or name is unrecognizable, use the Bayesian classification to find the N most similar continuous sounds (79) in the feature database to improve the continuous sound features in the sentence. The invention is described in detail later:

(1) After a continuous tone input speech recognition method, the continuous sound continuous sound wave is converted into a series of signal sampled points. 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 two signal points in the time period. In theory, the first method is better because of the variation of signal points. The number is greater than the number of noise variations, indicating that there is speech. However, when the present invention recognizes continuous sounds, the two methods have the same recognition rate, but the second saves time.

(2) After the voice signal point is deleted, the remaining signal points represent all the signal points of a continuous tone. The sound wave is normalized and then the feature is extracted, and all the signal points are divided into E and other time periods, and a frame is formed every time period. A continuous sound has a total of E equal length elastic frames, no filters, no overlap, free stretch, covering all signal points. The signal points in the elastic frame change nonlinearly with time and are difficult to express by mathematical models. Because J. Markhoul published in Proceedings of IEEE, Vol. 63, No. 4 published in 1975, Linear Prediction: A tutorial review shows that the signal point has a linear relationship with the previous signal point, and the regression model can be used to change linearly with time. Estimate the signal point of this nonlinear change. The signal point S ( n ) can be estimated from the previous signal point, and its estimated value S '( n ) is represented by the following regression mode: In equation (1), a _k , k =1,..., P are the estimated values of the regression unknown coefficients, and P is the number of previous signal points. The least squares estimate is obtained using the cyclic formula of Durbin in L. Rabiner and BHJuang in 1993, Fundamentals of Speech Recognition, Prentice Hall PTR, Englewood Cliffs, New Jersey. This set of estimates is called a linear predictive coding (LPC) vector. The linear predictive coding (LPC) vector method for finding the signal points in the frame is detailed as follows: the sum of the squared differences between the signal point S ( n ) and its estimated value S '( n ) is represented by E ₁ : Find the regression coefficient to minimize the sum of squares E ₁ . For each unknown regression coefficient a _i , i =1,..., P , find the partial differential of (2) and divide the partial differential into 0 to get the normal equation of P group: After expanding (2), substitute (3) to get the minimum total squared difference E _P (3) and (4) are converted to

In equations (5) and (6), N is used to indicate the number of signal points in the frame. The linear predictive coding (LPC) vector is quickly calculated using Durbin's loop as follows: E ₀ = R (0) (8)

(8-12) Formula loop calculation to obtain the regression coefficient least squares estimate a _j , j =1,..., P , (linear predictive coding (LPC) vector) as follows: Then use the following formula to convert the LPC vector to a more stable linear predictive coding cepstrum (LPCC) vector a ' _j , j =1,..., P ,

An elastic frame produces a linear predictive coding cepstrum (LPCC) vector ( a ' ₁ ,..., a ' _P ). According to the speech recognition method of the present invention, P = 12, since the final linear prediction coding cepstrum (LPCC) is almost zero. A continuous tone is characterized by E linear predictive coding cepstral (LPCC) vectors, that is, a matrix containing E×P linear predictive coding cepstrums (LPCC) represents a continuous tone, and one continuous tone includes one or more syllable.

(3) The same method calculates the E linear predictive coding cepstrum (LPCC) vectors of an unknown continuous sound wave with the formula (8-15), and the matrix of the same size E × P LPCC, called the classification of unknown continuous sounds model.

(4) In the second figure, the speech recognizer 100 receives a classification model of an unknown continuous sound, a matrix of E × P LPCC. Use X = { X _jl }, j =1,..., E , l =1,..., P to represent the unknown continuous sound classification model. In comparison with a known continuous sound c _i , i =1,..., m ( m represents the total number of all consecutive sounds), in order to quickly calculate the alignment value, it is assumed that { X _jl } has E × P independent Normal allocation, its mean and variation ( μ _ijl , ), estimated by the mean and variance in the known continuous tone standard model. The conditional density function of X is denoted by f ( x | c _i ). The decision theory in TFLi (Li Zifen) published in 2003 in Pattern Recognition, Vol.36 published in Speech recognition of mandarin monosyllables shows that the Bayesian classification is as follows: It is assumed that the feature database has a standard model of m known continuous sounds. _Let θ _i , i =1,..., m denote the continuous sound c _i , i =1,..., m , the probability of occurrence, that is, the previous probability, then . A decision method is represented by d . Define a simple loss function, that is, the misclassification probability of d : if the decision method d judges an unknown continuous sound d ( x ) ≠ c _i , then the loss function L ( c _i , d ( x ))=1. If d is judged to an unknown continuous sound d ( x ) = c _i , there is no loss L ( c _i , d ( x )) = 0. The recognition method is as follows: Γ _i , i =1,..., m , indicating that the X = x matrix value belongs to the range of the known continuous sound c _i . That is, X is Γ _i , d and the unknown continuous sound belongs to the known continuous sound c _i . d the average probability of error In (16), τ = ( θ ₁ ,..., θ _m ), It is outside the range of Γ _i . All methods of speech recognition, that is, all methods of dividing the range of m known continuous tones, are represented by D. Find a recognition method d _τ in D to minimize its average error rate (16), expressed as R ( τ , d _τ ) The recognition method d _τ satisfying the formula (17) is called the Bayesian classification method related to the previous probability τ . The following expressions can be used: d _τ ( x )= c _i if θ _i f ( x | c _i )> θ _j f ( x | c _j ) (18) In the formula (18), j =1,..., m , j ≠ i , that is, the range belonging to the known continuous sound c _i is for all j ≠ i , Γ _i ={ x | θ _i f ( x | c _i )> θ _j f ( x | c _j )} . As with all known continuous tones, the Bayesian classification is the same as the maximum probability method.

When the Bayesian classification method (18) recognizes an unknown continuous sound, it first calculates the conditional density function f ( x | c _i ) of all X , i =1,..., m , In (19), i = 1, ..., m , (the total number of continuous tones is known). For the convenience of calculation, take the logarithm of (19) and delete the constant to get the Bayesian distance. The Bayesian classification (18) becomes for each known continuous sound c _i , and the l ( c _i ) value (20) is calculated, and l ( c _i ) is also called the similarity between the unknown continuous sound and the known continuous sound c _i . , or Baysian distance. In (20), x = { x _jl }, j =1,..., E , l =1,..., P , is the linear predictive coding cepstrum (LPCC) in the unknown continuous sound classification model. Value, { μ _ijl , } Estimate the mean and variance within the standard model of known continuous tones. The most important contribution of the present invention is to find a mutually stable center point c _i ={ μ _ijl } and a clearly non-overlapping range Γ _i ={ x for each known continuous sound c _i without using a sample in a known continuous tone feature database. θ _i f ( x | c _i )> θ _j f ( x | c _i )}, j ≠ i . (21) where x = { x _ijl } is an E × P LPCC matrix representing a known continuous sound c _i range.

(5) Extracting the characteristics of a known continuous sound, first preparing a database of unknown continuous sounds, there are two kinds of unknown continuous sound database: one is that there are samples of unknown continuous sound, and the other is no sample. With a sample database, first find the average and variation of each unknown continuous sound. Find the N most recent unknown continuous sounds around the known continuous sound with a Bayesian distance in the unknown continuous sound database with samples. The N average values of the N unknown sounds and the N+1 weighted averages of the linear predictive coding cepstrum (LPCC) of the known continuous sound are taken as the average of the known continuous sounds, and N consecutive sounds are obtained. The weighted average of the N variances is used as the variance of the known continuous tone, and the E×P mean and the variance matrix are the preliminary feature values 79 of the known continuous tone placed in the continuous tone feature database. If there is no sample in the unknown continuous sound database, in the unknown continuous sound database, find N unknown continuous sounds around the known continuous sound with the minimum absolute distance. The linear predictive coding cepstrum (LPCC) of the known continuous tone and the N unknown continuous tones is regarded as (N+1) numbers. Find the weighted average of (N+1) numbers as the average of the known continuous sounds, and find the (N+1) number of variances as the variance of the known continuous sound, this E×P average And the matrix of the variance number represents the preliminary feature of the known continuous tone, and is placed in the known continuous tone feature database 79. In the known continuous tone feature database, if the average of a known continuous tone and the Bayesian distance of the LPCC of the same known continuous tone in the permanently known continuous tone database is not the smallest in the feature database, Then in the feature capital N known continuous tones are found in the library by Bayesian distance, and their Bayesian matrix is N minimum for the LPCC of the known continuous tone. Find N average values of N known continuous sounds and LPCC weighted average of the known continuous sounds as the new average of the known continuous sounds, and use the weighted average of the N variances of the N known continuous sounds as the 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 E×P new average and the variance matrix are called standard models representing the known continuous sounds. The database 80 then uses the known continuous tones of the known feature database to create a sentence and name database 85.

(6) If an unknown continuous sound c is recognized as an error, the present invention provides two techniques for correcting the old feature so that the continuous sound is recognized correctly;

(a) Using the Bayesian classification (20), find N consecutive sounds { μ _ijl that are most similar to the continuous sound c in the feature database. }, i =1,2,..., N , then average (or weighted average) With { μ _jl , }, j =1,..., E , l =1,..., P , represents the new standard model of the unknown continuous sound, exists in the continuous sound feature database, and then tests the sound to be successful.

(b) in (a), the N most similar known continuous sound mean values and the linear predictive coding cepstrum (LPCC) of the unknown continuous sound are weighted averages as the new average of the unknown continuous sounds, to N The weighted average of the variance of the most similar continuous sounds as the new variance of the unknown continuous sound With { μ _jl , }, j =1,..., E , l =1,..., P , represents the new standard model of the unknown continuous sound.

(7) In order to confirm that the present invention can recognize any language at the same time, the present invention performs a 2-person speech recognition experiment.

(a) First establish an unknown continuous sound database. This monophonic database was purchased from the Central Research Institute of Taiwan. There are 388 Mandarin singles in the database (the third picture), all of which are female pronunciations, ranging from 6 to 99 samples, and many monophonic pronunciations are almost the same.

(b) Convert all samples into E×P LPCC matrices from the method in (2), a total of 12400 matrices.

(c) In the 388 Mandarin singles, use the sample to average and the number of variances.

(d) Blindly blending 388 Mandarin singles, making 388 samples with sample mean and variance numbers into 388 unknown continuous sound databases (one Mandarin single tone is also a one-syllable continuous sound).

(e) Find another man and a woman to pronounce 654 Mandarin singles, 154 words, 1 German, 1 Japanese and 3 Taiwanese pronunciations to create two 813 permanent known continuous sound databases, each of which is Linear Predictive Coded Cepstrum (LPCC) E x P matrix representation.

(f) In the 813 known continuous tones of the permanent known continuous sound database, for each known continuous sound, find N=15 unknown continuous sounds in 388 unknown continuous sounds with Bayesian distance (20). It is known that the linear predictive coding cepstrum (LPCC) of continuous sound and the sample mean of N unknown continuous sounds are N+1 weighted averages for the known continuous sound mean, and samples of N unknown continuous sounds are obtained. Variant number The weighted average is the number of variations of the known continuous tone. This average and variance 12 x 12 matrix is called the preliminary feature 79 of the known continuous tone, and there is a library of known continuous tone features. That is, the feature database contains 813 12×12 averages and a variance matrix 80.

(g) In the feature database, if the average of a known continuous tone and the LPCC of the known continuous tone in the permanent continuous tone database are not the smallest. In the 813 continuous tone feature data, N = 15 known continuous sounds were found using the Bayesian distance. The weighted average of the N averages of the N consecutive tones and the LPCC of the known continuous tones is the new average of the known continuous tones. The weighted average of the variances of the N known continuous tones is the new variation of the known continuous tones. Repeat the calculation of the new average and the number of variations multiple times. The final 12×12 mean and variance matrix is called the standard model and represents the known continuous tone feature, which exists in the known continuous tone feature database.

The present invention performs the following continuous tone recognition, and the recognition rate is determined by the person. Because there are too many similarities, the top three finalists 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 (very good recognition rate)

4 Identify 654 Mandarin singles, 1 German, 1 Japanese, 3 Japanese (3rd and 5th) (good recognition rate, no good for the first three)

(8) To identify the sentence or name of a speaker, we first establish a database of English and national sentences and names. The continuous sounds in each sentence or name are all known from the continuous sound feature database (384+154). English and Mandarin are arbitrarily composed of 154 English words composed of 70 English sentences and names, and 384 Mandarin words consisting of 407 national sentences and names (Sixth). The method of identification is as follows:

(a) cutting an unknown sentence or name into D unknown continuous sounds, calculating the sum of the distances of adjacent two signal points per unit time period, such as too small, the time period is noise or mute, and the adjacent unit time period without voice signal accumulates too More (more than two consecutive syllables), indicating that it is all noise or mute, it should be cut between two consecutive sound lines, cut into D unknown continuous sounds, and then use the second figure 45, 50, 60 and The 90 process is converted into an E×P LPCC matrix. For each unknown continuous sound, use the Bayesian classification (20) to select the most similar F known continuous sounds in the English and Mandarin feature database (possibly including both English and Mandarin (picture)), an unknown sentence or name. D x F is most similar to the known continuous tone representation.

(b) Find the sentence or name of the speaker in the sentence and name database. In the 477 English and national sentences and names, select (D ± 1) known continuous sentences and names.

(c) If the choice of database is the same as the sentence or name of the speaker and the sentence or name of the speaker (D unknown continuous tones, then D each column of F similar known continuous tones and alignment sentences or names The D known continuous tones are compared in order, to see if there are any similar continuous tones in the sentence or the name in the F consecutive continuous tones. For example, each column of similar continuous tones contains a comparison sentence or within the name. A continuous sound is known, and it is recognized that the correct continuous sound is D, and the sentence or name of the comparison is the sentence or name of the speaker.

(d) If the database matches the sentence and the known continuous number of sounds in the name is D-1 or D+1 or (c), the correct continuous sound is not D, and the present invention uses a 3×F window to screen. In the comparison sentence or name (in the database), the ith known continuous sound, using the similarly continuous continuous sounds (ie, the i - th , i , i +1 columns) in the three columns of the D×F matrix For the ith known continuous sound, calculate how many ratios of the D×F matrix are compared to the known continuous sounds in the sentence or name, and divide by the total number D to get the probability of the matching sentence or name, and select the highest probability in the database. The sentence or name is the speaker's pronunciation.

(e) If a sentence or name is identified incorrectly, it must be in D unknown continuous sounds, one or more of them are not in their F similar known continuous tones, using the Bayesian classification (20) at (155+384) Knowing that the continuous Nine is the first N=15-order known continuous sound, and the N similar continuous sounds and the LPCC weighted average of the unknown continuous sound are used to improve the unknown continuous sound, so that D unknown continuous sounds are in their F Similar to known continuous sounds, the test will be successful.

The present invention implements the following English and national sentences and name recognition, and the identification is almost entirely correct and varies from person to person:

1 Identify 70 English sentences and names (very good).

2 Identify 407 country sentences and names (very good)

3 Identify 70 English sentences and names with 407 national sentences and names (very good).

* Two Visual Basic recognition diagrams (7th and 8th) indicate the simultaneous recognition of English and national sentences and names.

(1) ‧‧‧Create a permanent database of known continuous sounds, pronounce a continuous sound or a sentence, and divide the sentence into several known continuous sounds

(10)‧‧‧Continuous sound continuous sound waves

(20) Receiver

(30)‧‧‧Sonic Digital Converter

(45) ‧ ‧ remove noise

(50) ‧‧‧E elastic frames normalized sound waves

(60) ‧‧‧Last-flat method for calculating linear predictive coding cepstrum vectors

(70) ‧‧‧ Use the Bayesian distance (absolute value distance) to find N recently unknown continuous sounds in the unknown continuous sound database for each known continuous sound (permanent database).

(79) ‧ ‧ For each known continuous sound (permanent database), use the surrounding N unknown continuous sounds and the linear predictive coding cepstrum of the known continuous sound to obtain a weighted average for the known continuous sound Preliminary features are placed in the feature database. Then, in the feature database, N known continuous sounds and the known continuous sound LPCC are used to find a weighted average value, and the calculation is performed multiple times. The final weighted average (E × P mean and variance) represents the standard model of the known continuous sound

(80) ‧‧‧The known continuous tone feature database contains standard models for all mean and variance

(85) ‧ ‧ Build a sentence and name to be recognized using the continuous sound of a known continuous tone feature database Sentence and name database

(2) Enter an unknown sentence or name

(11)‧‧‧A group of unknown continuous sound waves

(40) ‧ ‧ Cut a sentence or name into D unknown continuous sounds

(90) ‧‧‧D linear predictive coding cepstrum matrix of unknown continuous sounds representing D unknown continuous sound classification models

(100)‧‧‧Compare each known continuous sound standard model and unknown continuous sound classification model by Bayesian classification

(110)‧‧‧ In the sentence or name, find the closest F known continuous sounds for each unknown continuous sound, one sentence or name has a total of D × F known most similar continuous sounds

(120) ‧‧‧ In the sentence and name database, use the similar known continuous tones of the 3×F window to filter all sentences and each known continuous sound in the name

(130) ‧‧ Find the most likely sentence or name in the sentence and name database

The first and second figures illustrate the invention execution procedure. The first figure shows the known continuous sound permanent database, the known continuous sound feature database and the three database establishment flow of sentences and names, and the second figure shows the flow of an unknown sentence or name recognition method.

The third picture shows the recognition of 384 Mandarin singles, 1 German, 1 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 70 English sentences and 407 Chinese sentences and names.

The seventh and eighth diagrams illustrate the Visual Basic recognition diagram showing the simultaneous recognition of English and national sentences and names.

(2) Enter an unknown sentence or name

(11)‧‧‧A group of unknown continuous sound waves

(20)‧‧‧ Receiver

(30)‧‧‧Sonic Digital Converter

(40) ‧ ‧ Cut a sentence or name into D unknown continuous sounds

(45) ‧ ‧ remove noise

(50) ‧‧‧E elastic frames normalized sound waves

(60) ‧‧‧Last-flat method for calculating linear predictive coding cepstrum (LPCC) vectors

(90) ‧‧‧D linear predictive coding cepstrum (LPCC) matrices of unknown continuous sounds represent D unknown continuous sound classification models

(80) The ‧ ‧ continuous tone feature database contains standard models for all mean and variance

(100)‧‧‧Comparing the known continuous sound standard model with the unknown continuous sound classification model by Bayesian classification

(110)‧‧‧ In the sentence or name, find the closest F known continuous sounds for each unknown continuous sound, one sentence or name has a total of D × F known most similar continuous sounds

(85) ‧‧‧Create a sentence and name database of the sentence and name (any set of continuous sounds) to be recognized using the known continuous sounds of the continuous tone feature database

(120) ‧‧‧ In the sentence and name database, use the 3×F window to filter all sentences and each known continuous sound in the name

(130) ‧‧ Find the most likely sentence or name in the sentence and name database

Claims (9)

A method for recognizing all languages without using a sample, the steps comprising: (1) having a database without samples, the database containing sound waves of unknown continuous sounds; and (2) a permanent database of known continuous sounds, the database The sound wave containing the user's pronunciation of the known continuous sound and the sentence and name database containing the known continuous sound; (3) using the processor to delete the sampled points or murmurs without voice sound; (4) The sample-free database has different unknown continuous sounds and permanent data banks known as continuous sounds, normalizes each successive sound wave and extracts the characteristics of each successive sound wave: normalizes and converts the sound waves into equal sizes with E elastic frames a linear predictive coding cepstrum (LPCC) ExP feature matrix; (5) normalizing the user's unknown continuous sound wave and extracting the sound signature: normalizing the sound wave and converting it into an ExP linear predictive coding cepstrum (LPCC) Matrix, expressed as an unknown continuous sound classification model; (6) Find N ExP feature matrices of N unknown continuous sounds and ExP linear predictive coding cepstrum of known continuous sounds in a sampleless unknown continuous sound database (LPCC) a linear predictive coding cepstrum (LPCC) matrix of N unknown continuous tones with a minimum matrix distance and a linear predictive coding cepstrum (LPCC) matrix of the known continuous tone as (N+1) number, calculated ( N+1) number weighting Mean and variance, this E × P mean and variance matrix is called the known continuous sound standard model, there is a library of known continuous sound features; (7) using a Bayesian classification algorithm: comparison The user's unknown continuous sound classification model and the known continuous sound standard model find that the shell distance of a known continuous sound and the user's unknown continuous sound is the smallest, and the user is determined to be unknown continuous sound; (8) Cutting the unknown sentence or name of the user into D unknown continuous sounds; (9) using a Bayesian classification method in D unknown continuous sounds, for each unknown continuous sound in the known continuous sound feature database, Select the F consecutive shortest known continuous tones, representing the F most similar known continuous tones, an unknown sentence or name expressed in the known continuous tones of the D×F matrix; (10) D with unknown sentences or names The known continuous sound of the ×F matrix, comparing all the sentences and names of the sentence and the name database, the computer rate is compared with the known continuous sound of the sentence or the name in the 3×F window, divided by D, to find a comparison Known sentence or name is the unknown sentence or name of the user ; (11) a continuous tone having a correction characteristic. According to the scope of claim 1 of the patent application, a method for recognizing all languages can be identified without using a sample, wherein step (3) deletes the sound wave or noise without speech, and further includes: (a) signal points in one hour period, calculation The number of variations in signal points and the number of variations without sound signal points, when the number of variances of signal points is less than the variation of no signal points If the number is deleted, the time period is deleted; (b) the signal point is calculated within one hour, and the sum of the distance between the adjacent two signal points and the distance between the two signal points adjacent to the sound signal point is calculated. If the current one is smaller than the latter, the time period is deleted. According to the first aspect of the patent application, a method for identifying all languages can be identified without using a sample, wherein the step (4) further comprises a continuous sound wave normalization and extracting a feature matrix of uniform size, the steps are as follows: (a) Equalization A continuous sound wave signal point: In order to closely estimate the nonlinearly varying sound wave with a linearly varying regression mode, the full length of the sound wave is divided into E equal periods, and an elastic frame is formed every time period, and a continuous sound has a total of equal length elastic frames without filtering. Filter, which does not overlap, so that it can freely expand and contract the full-length sound wave; (b) In each frame, use a regression model that changes linearly with time to estimate the acoustic wave that changes nonlinearly with time; (c) Signal point S ( n ) can be estimated from the previous signal point, and its estimated value S '( n ) is represented by the following regression modes: In equation (1), a _k , k =1,..., P are the estimated values of the regression unknown coefficients, P is the number of previous signal points, and E ₁ is the signal point S ( n ) and its estimated value S ' The sum of the squared differences between ( n ): Find the regression coefficient to minimize the sum of the squares E ₁ . For each unknown regression coefficient a _i , i =1,..., P , find the partial differential of (2) and divide the partial differential into 0 to get the normal P group. equation: After expanding (2), substitute (3) to get the minimum total squared difference E _P (3) and (4) are converted to In equations (5) and (6), N is used to indicate the number of signal points in the frame. The linear predictive coding (LPC) vector is quickly calculated using Durbin's loop as follows: E ₀ = R (0) (8) (8-12) The formula is cyclically calculated to obtain the regression coefficient least squares estimate a _j , j =1,..., P , called the linear predictive coding (LPC) vector as follows: Then use the following formula to convert the LPC vector to a more stable linear predictive coding cepstrum (LPCC) vector a ' _j , j =1,..., P , An elastic frame produces a linear predictive coding cepstrum (LPCC) vector ( a ' ₁ ,..., a ' _P ); (d) a continuous tone is characterized by E linear predictive coding cepstrum (LPCC) vectors The E linear predictive coding cepstrum (LPCC) vector is used to represent a continuous tone. According to the claim 1, the non-sample can identify the recognition method of all languages, wherein the step (5) further comprises a classification model for calculating the unknown continuous sound, the steps are as follows: (a) dividing the unknown continuous sound wave into E, etc., each period consists of a flexible frame, an unknown continuous sound has E equal length elastic frames, no filters, no overlap, freely telescopic covers all acoustic signal points; (b) within each elastic frame, with one A regression model that linearly changes over time estimates acoustic waves that change nonlinearly over time; (c) uses Durbin's circular approach E ₀ = R (0) Calculate the regression coefficient least squares estimate a _j ,1 j P , called linear predictive coding vector (LPC vector); (d) then formulate LPC vector Converted to a stable linear predictive coding cepstrum (LPCC) vector a ' _i ,1 i P ; (e) E linear predictive coding cepstrum (LPCC) vector, E × P LPCC matrix, as the classification model of the unknown continuous sound. According to the first aspect of the patent application scope, a method for recognizing all languages without using a sample, the step (7) further comprises a Bayesian algorithm for identifying unknown continuous sounds, the steps of which are as follows: (a) an unknown continuous The classification model of the characteristics of the sound, using an E × P LPCC matrix X = { X _jl }, j =1, ..., E , l =1, ..., P , means: E × P LPCC { X _Jl } is E × P independent random variables, and there is a normal allocation, when the unknown continuous sound and a known continuous sound c _i , i =1,..., m , ( m is the total number of all known continuous sounds) For the time, the average and variation of { X _jl } ( μ _ijl , Using the known continuous sound standard model mean and the present variance estimate, then the conditional density function of X is X = { X _jl } is the linear predictive coding cepstrum (LPCC) of the classification model of unknown continuous sound, but ( μ _ijl , ()) Estimate the mean and variance in the standard model of the known continuous tone c _i ; (b) Find a known continuous tone c _{i in} the known continuous tone feature database using the Bayesian classification. X , a known continuous sound c _i versus unknown continuous sound X similarity in the following expression f ( x | c _i ) (c) Using the logarithmic density function f ( x | c _i ) in logarithm simplification (b), and deleting the constants that do not have to be calculated, the Bayer distance is obtained, and the distance is calculated as follows (d) for each known continuous sound c _i , i =1,..., m , calculate the value of the shell distance l ( c _i ) in ( c ); (e) in the feature database, select one The continuous sound c ' _{i is} known, when the known continuous sound c ' _i its shell distance l ( The value is the smallest and is judged as the unknown continuous sound. According to the method of claim 1, the method for recognizing all languages can be identified without using a sample, and the step (8) further comprises an unknown sentence or a name cut into D unknown continuous sounds: (a) calculating adjacent neighbors per unit time period The sum of the two signal points, the sum of the distances, and no sound The distance between the two signal points adjacent to the wave signal point, the current one is smaller than the latter, the time period is deleted, the time period is mute or noise, and there is no voice signal; (b) when the mute or noise adjacent unit time period is accumulated to the continuous sound When the two syllables are long, the time period is judged as two continuous sound dividing lines, and the unknown sentence or the name is D unknown continuous sounds; (c) each continuous sound is removed from the mute and the murmur, so that the elastic frame is normal. The least squares E×P linear predictive coding cepstrum (LPCC) matrix represents an unknown continuous tone, which is represented by D linear predictive coding cepstrum (LPCC) E×P matrices. According to the method of claim 1, the method for recognizing all languages is not used in the sample, and the step (9) further comprises the following steps: (a) after cutting the unknown sentence or name of the user into D unknown continuous sounds, Each unknown continuous sound { X _jl } uses the Bayesian classification method to calculate each known continuous sound c _i ={ μ _ijl in the feature database. } and the unknown continuous sound { X _jl } shell distance l ( c _i ), Find the nearest F known continuous sounds, which can contain multiple languages at the same time, and use the F known similar continuous tones to represent an unknown continuous sound; (b) Make D columns F known similarly continuous The tone indicates the unknown sentence or name. According to the method of claim 1 of the patent application, a method for recognizing all languages can be identified without using a sample, and the step (10) further includes the following sentence and name recognition: (a) in the sentence and name database, select the matching sentences and names of D-1, D, D+1 known continuous sounds; (b) in the sentence and name database, select the The sentence or name, when it is just the length and the speaker's sentence or name has D known continuous tones, then D each column of F similar known continuous tones and the selected alignment sentence or name D known continuous sounds are sequentially compared, and it is judged whether there are known continuous sounds in the similar known continuous sounds in the compared sentences or names, and each of the F similar known continuous sounds includes the alignment in turn. When a sentence or a known continuous tone is in the name, it means that there are D unknown continuous sounds in the matching sentence or name, the matching sentence or name identifies the sentence or name of the speaker; (c) in the sentence and name database When the matching sentence or name has D known continuous sounds, when the speaker's D consecutive sounds are not fully recognized, or when the matching sentence or name is D-1 or D+1 known continuous sounds, then Screening using a 3×F window containing similarly known continuous tones in the three consecutive columns of similar known continuous tones in a D×F matrix Compare the sentences and the name database with D or D±1 known continuous sounds in the matching sentence or each known continuous sound in the name, and select the most likely matching sentence or name in the database. The sentence or name of the person, the probability value is the number of known continuous sounds of the matching sentence or name falling within the 3×F window divided by D. According to the method of claim 1 of the patent application, a method for identifying all languages can be recognized without a sample, and the step (11) further comprises a modified continuous sound feature: (a) when a sentence or name is recognized incorrectly, in D unknown continuous One or more of the F-like similar known continuous tones, with c indicating that one of the unknown continuous tones is not in its F similarly known continuous tones, using the N most similar known continuous tones, Find N consecutive sound features { μ _ijl , }, i =1,..., N , j =1,..., E , l =1,..., P , the average value or weighted in order, , To make the average { μ _jl , } represents a new feature of the unknown continuous sound c ; (b) in (a), the average of the linear predictive coding cepstrum (LPCC) of the tester's pronunciation and the N average of the N most similar known continuous tones The N+1 weighted averages are the new averages of the unknown continuous sounds, and the weighted average of the N variances of the N most similar known continuous sounds is the new variance of the unknown continuous sound, the new average and The variance E×P matrix represents a new standard model of the unknown continuous sound; (c) the unknown sentence or name is retested.