TWI460613B - A speech recognition method to input chinese characters using any language - Google Patents

Description

a method of inputting Chinese characters in any language

The invention can input 14100 Chinese in any language, and divide the 409 known tones into 82 sentences first, and it takes only 20-30 minutes to read 82 sentences in one language for any first time. In detail, the present invention can be used in any language, such as Mandarin, Taiwanese, Hakka, Cantonese, etc., as long as 20-30 minutes to read eighty-two sentences (including 409 sounds), you can enter 14100 different Chinese characters . The recognition time is less than one second.

First, the 14108 text is divided into 409 groups with a single sound (no matter five sounds), and the first sound is represented by 409 single sounds. Each word in a group is divided into two types: common words and unusable words. The characteristics of a single tone are represented by an equal ExP=12x12 matrix. First, send a thousand different sounds once, except for the noise, and the rest are all sound waves with sound. Then use the E=12 equal length elastic frame, no filter, no overlap, and convert all the sound waves into the ExP=12x12 linear predictive coding cepstrum matrix. There are a total of one thousand different ExP linear predictive coding cepstrum matrices representing one thousand different databases.

First divide the 409 known tones into 82 sentences, first read a sentence in any language, then divide the sentence into known tones, remove the noise, and the rest are sound waves with sound. Then, using E equal length elastic frames, a sound wave of all known single sounds is converted into an ExP linear predictive coding inverse spectrum matrix. Using the distance, the tone is divided into the nearest L databases of the ExP linear predictive coding cepstral matrix of the 1000 database. 409 means that the 14100 words are all placed in the 1000 database, and each tone can be placed in L databases.

When the first voice is recognized in the same language for the first time, it is first converted into the ExP linear predictive coding cepstrum matrix, and then the distance between the unknown monophonic ExP matrix and the thousand database ExP matrix is from one thousand. In the database, find the F closest database. The single tone is known from the F closest database, and the unknown tone is found by distance. All single-word pleads represented by known tones recognized as unknown tones appear

With Visual Basic, you can quickly identify the word you want in less than a second. The method is simple, no need for samples, no training, mathematical calculation and identification, anyone can use it in time, the pronunciation is not standard or the wrong voice can be. Tested in Mandarin, Taiwanese, and pronunciation, the speed is fast and the accuracy is high. If not accurate, the present invention can correct the tone characteristics in time to achieve a recognition rate of 100.

A continuous acoustic feature is commonly used in the following ways: energy, zero crossings, extreme count, formants, linear predictive coding cepstrum (LPCC) and plum Frequency Frequency Cepstrum (MFCC), in which linear predictive coding cepstrum (LPCC) and Mel Frequency Cepstrum (MFCC) are the most efficient 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 the linear predictive coding cepstrum (LPCC). The Mel Frequency Cepstrum (MFCC) converts sound waves into frequencies using the Fourier transform method. The auditory system is estimated based on the ratio of the frequency of the Mel. According to the scholar SB Davis 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 tones, 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. Speech by TF Li (Li Zifen) published in 2003 in Pattern Recognition, vol. 36 Speech recognition of mandarin monosyllables, Li, Tze Fen (Li Zifen) in 1997 in the US Patent Certificate, Apparatus and Method for Normalizing and Categorizing Linear Prediction Code Vectors using Bayesian Categorization Technique, USA Patent No. 5,704,004, Dec. 30,1997, Li Zifen in 2008 in the Republic of China Patent Certificate I 297487 (2008, 6, 1) name speech recognition method and Li Zifen in 2009 in the Republic of China Patent Certificate No. I 310543 (2009, 6, 1) Name A method of identifying a similar Mandarin single tone by a continuous quadratic Bayesian classification method. Using the Bayesian classification method, the same database will be used for different lengths and lengths of a series of LPCC vectors. Various methods are compressed into feature models of the same size, and the recognition results are compared to YK Chen, CYLiu, GH Chiang, MT Lin, published in Proceedings of Telecommunication Symposium, 1990. The recognition of mandarin monosyllables based on the discrete hidden Markov model It is better to use the hidden Markov mode HMM method. However, the compression process is complicated and time consuming, and it is difficult for the same tone to compress the same feature to the same time position, which is difficult to recognize for similar tones.

Up to now, there is no way to input Chinese characters. Recently, Chinese has been entered with sentences (word breaks or broken words). The range of Chinese characters is small, it is not useful, it is not accurate, and it takes a long time to train. It may take dozens of times to train. The speech recognition method of the present invention is directed to the above shortcomings. From the academic point of view, according to the sound characteristics of the sound wave, the nonlinear change with time, naturally derives a set of extracted speech feature methods, and all the monophonic sounds of any language are equal to the ExP=12x12 matrix. Indicates that the tone recognition rate is improved. If the recognition is unsuccessful, the present invention provides three techniques for correcting the monophonic features so that the recognition is correct. You can enter 14100 Chinese characters in voice using any language.

(1) The object of the present invention is to input 14100 Chinese characters by voice using any language.

(2) The object of the present invention is to have a high recognition rate for the Chinese characters of the 14100, preferably up to 100%.

(3) The purpose of the present invention is to identify the time is short, the input speed is fast, and it is better to be faster than any current one. Any person only needs 20-30 minutes to practice and can input 14100 Chinese characters. .

(4) In order to achieve the above three objectives, the present invention applies a single-tone sound wave normalization and extraction feature method. It uses a small number of E=12 equal elastic frames, equal length, no overlap, no filter, can be normalized according to a single tone, short wavelength, freely adjust all the monophonic wavelengths covered, can all the single tones that can be identified Converted to an equal ExP=12x12 linear predictive coding cepstrum matrix. A series of dynamic characteristics that change nonlinearly with time are converted into an equal-sized ExP linear predictive coding cepstrum matrix, and the same monotone feature model has the same characteristics at the same time position. It can be compared in time to achieve instant recognition and input of the computer.

(5) The present invention applies one thousand different databases, can recognize a large number of single tones, has a high speed, and the accuracy is greatly improved. It mainly distributes all known tones in the database closest to the sound of one thousand databases. When identifying unknown tones, first find the closest database with the unknown monophonic sounds, and then the closest from F. Know the single tone in the database to find the unknown tone to be recognized. There are not many known single tones in the F closest database, and it is easy to identify. The present invention does not use samples, does not require statistical calculations and recognition, and is identified by mathematical calculations and by using the distance between the inverted-spectrum ExP matrices of the linear predictive coding of the tones.

(6) The identification method of the present invention can recognize a single tone that is too fast or too slow to talk. When the pronunciation is too fast, a single sound wave is very short, and the length of the E=12 equal length 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. A single tone is longer when the pronunciation is too slow. The length of the E=12 equal length elastic frame will be elongated. The same number of E linear predictive coding cepstrum (LPCC) vectors are generated to effectively represent the long tone.

(7) If the recognition is unsuccessful, the present invention provides three techniques for correcting the monophonic features so that the recognition is correct.

The invention execution program will be described using the first diagram and the second diagram. The first picture shows the establishment of M=1000 databases, each with similar known tones. The second figure shows that the user recognizes the unknown tone and inputs the Chinese execution program.

In the first figure, there are M = 1000 different sounds 1, and a sound sound wave is converted into a digitized signal point 10 to remove the noise or mute 20. First, the sound sound wave is normalized and the feature is extracted, and all the signal points of one sound sound wave are divided into E=12 and the like, and each time frame constitutes a frame. A sound has a total of equal length frames 30, no filters, no overlap, according to the length of all signal points of the sound, E equal frame length freely adjust all signal points with cover. Therefore, the frame is called an isometric elastic frame, and the length is free to expand and contract, but the E "elastic frames are the same length". Unlike the Hamming window, there are filters, semi-overlaps, fixed lengths, and no freedom to adjust with wavelength. Because a sound wave changes nonlinearly with time, the sound wave contains a dynamic feature of the voice, which also changes nonlinearly with time. Since there is no overlap, the present invention uses less (E=12) equal-length elastic frames covering all sound waves, since the signal points can be estimated from the previous signal points, and the nonlinear changes are closely estimated using a regression pattern that varies linearly with time. The sound wave is estimated by the least square method of the regression unknown coefficient. In each equal-length elastic frame, the P=12 linear predictive coding cepstrums 40 are calculated by the least square method, and one sound is represented by the ExP linear predictive coding cepstrum (LPCC) matrix, and the sound of the ExP linear predictive coding is inverted. The spectrum matrix represents a database with a total of one thousand databases 50. First divide the 409 known tones into 82 sentences, first read a sentence in any language, and then divide the sentence into known tones, removing mute and noise. The known tones to be recognized are converted into an ExP Linear Predictive Coded Cepstrum (LPCC) matrix 60 using E equal elastic frames. The known tones are divided into the L closest databases by distance 70. All 409 known tones identified to be identified are assigned to M = 1000 different databases. There are M=1000 databases, each database contains similar known tones, and 409 stands for 14100 words in the 1000 database. Each tone can be placed in L databases.

The second diagram shows the flow of unknown tones and input texts that are first recognized in the same language. First pronounce the first sound in the same language with a clear unknown sound. The unknown single tone is digitized into signal point 10, removing silence and noise 20. E equal-length elastic frames normalize the sound waves, extract features, and divide the signal points of the unknown single sounds to be recognized into E, etc., and form a flexible frame 30 every time period. There are a total of E equal length elastic frames, no filters, no overlap, and free telescopic covers all signal points. Within each frame, the estimated value of the regression unknown coefficient is obtained by the least squares method because the signal point can be estimated from the previous signal. The P=12 least squares estimate produced in each box is called the linear predictive coding (LPC) vector, and the linear predictive coding (LPC) vector is converted to the more stable linear predictive coding cepstrum (LPCC) vector, an unknown single. The tone represents 41 with an ExP linear predictive coding cepstrum matrix. The present invention finds the F closest database by using the distance or weighted distance of the ExP linear predictive coding cepstrum matrix of the unknown tone and the M=1000 database 80 ExP linear predictive coding cepstrum matrix, that is, the The F databases have F minimum distances 84 from the linear predictive coding cepstrum matrix of the unknown tones. Then use the distance or weighted distance to find the known single tone in the F closest to the database, and find the unknown single tone. All the single characters of the group represented by the known single tone that is recognized as the unknown single tone will appear 90.

The invention is described in detail later:

(1) After a sound (monophonic) is pronounced clearly, the sound (monophonic) sound wave is converted into a series of signal sampled points 10. Then delete the point that does not have a voice signal, and delete all the mute and noise 20 before the sound (single tone) and after the sound (single tone). After the voice signal point is deleted, the remaining signal points represent all the signal points of a single voice (single tone). The sound wave is normalized and then the feature is extracted, and all the signal points are divided into E=12 and the like, and a frame is formed every time period. A sound (monophonic) has a total of E "equal length" elastic frames, without filters, without overlapping, freely telescopic, covering all signal points 30. In each isometric elastic frame, the signal points change nonlinearly with time and are difficult to represent in mathematical models. Because J. Markhoul published in Proceedings of IEEE, Vol.63, No. 4 in 1975, Linear Prediction: A tutorial review and Li, Tze Fen (Li Zifen) in 1997 in the US Patent Certificate, Apparatus and Method for Normalizing and Categorizing Linear Prediction Code Vectors using Bayesian Categorization Technique, USA Patent No. 5,704,004, Dec. 30,1997, which shows that the signal point has a linear relationship with the previous signal point, and the nonlinear change signal can be estimated by a regression model that changes linearly with time. point. 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. In 1993, L. Rabiner and BH Juang published Fundamentals of Speech Recognition, Prentice Hall PTR, Englewood Cliffs, New Jersey and Li, Tze Fen (Li Zifen) in 1997 in the US Patent Certificate, Apparatus and Method for Normalizing and Categorizing Linear Prediction Code Vectors using Bayesian Categorization Technique, USA Patent No. 5, 704, 004, Dec. 30, 1997 The calculation formula of Durbin's cycle finds the least squares estimate. 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 box produces a linear predictive coding cepstrum (LPCC) vector ( a ' ₁ ,..., a ' _P ) 40. According to the speech recognition method of the present invention, P = 12, since the final linear prediction coding cepstrum (LPCC) is almost zero. A sound (monophonic) feature is represented by E linear predictive coding cepstral (LPCC) vectors, that is, a matrix containing E×P=12x12 linear predictive coding cepstrums (LPCC) represents a sound, a sound The linear predictive coding cepstrum ExP matrix represents a database with a total of one thousand databases 50.

(2) Divide the 409 known tones into 82 sentences, first read a sentence in any language, and then divide the sentence into known tones, remove the mute and noise, and use the formula (8-15) to know the order. The sound is converted into a linear predictive coding cepstrum (LPCC) ExP matrix 60. Find the L closest database with the distance or weighted distance between the known single tone linear predictive coding cepstrum (LPCC) ExP matrix and the linear predictive coding cepstrum ExP matrix of all M=1000 different sounds. Know the single tone into the nearest 70 database. There are M=1000 databases, each database contains similar known tones, and 409 stands for 14100 words in the 1000 database. Each tone can be placed in L databases.

When the first tone is used to recognize an unknown single tone in the same language, the user first pronounces an unknown single tone that is to be recognized. The unknown single-tone sound is digitized into signal point 10, and the mute and noise 20 are removed, and all mute and noise are deleted before the unknown single tone and after the unknown single tone. E equal-length elastic frames normalize the sound waves, extract features, and divide the signal points of the unknown single sounds with speech to be equal to E, and form a flexible frame every time period. There are a total of E equal length elastic frames, no filters, no overlap, and free telescopic covers all signal points 30. Within each frame, the estimated value of the regression unknown coefficient is obtained by the least squares method because the signal point can be estimated from the previous signal. Each box is converted to a linear predictive coding cepstrum (LPCC) vector using the (8-15) formula, and an unknown single tone is represented by an ExP linear predictive coding cepstrum matrix. The present invention also finds the F closest database 84 using the distance or weighted distance between the unknown tone linear predictive coding cepstrum ExP matrix and all linear predictive coded cepstrum ExP matrices 80 of M = 1000 different sounds. From the F closest database, use the unknown tone linear prediction to encode the distance between the cepstrum ExP matrix and the linear estimate of the known single tone in the F database. The weighted distance is used to find the unknown single tone that the user wants, and all the single-words of the group represented by the known tones that are recognized as unknown tones appear 90.

(3) In order to confirm that the present invention can input Chinese characters quickly and accurately. The 82 sentences are all composed of 409 tones, each with 5 different monos. After the inventor pronounced the first sentence with 82 sentences, the sentence was cut into single tones, and each single tone was divided into L=5 nearest databases. Each database consists of two groups, one group storing common words and the other group storing less common words. With Visual Basic, the total training time is 26 minutes.

(4) After testing 409 tones, 90% of the tones are ranked first. After the correction, all appeared in the top four, because the similar sound is too much, there are six groups of the same sound, the inventor is also unclear. The recognition and input time is less than 1 second.

(5) Figure 3 uses the software (Visual Basic) to input the Chinese part of the manual. The fourth figure shows the three commonly used words.

(1). . . First M=1000 different sounds

(10). . . Sound wave digitization

(20). . . Remove noise and silent time

(30). . . E equal length elastic frames normalize all sound waves

(40). . . Calculate P linear predictive coding cepstrums by the least square method in each equal-length elastic frame

(50). . . A linear predictive coding of a sound Cepstrum ExP matrix represents a database with a total of one thousand databases

(60). . . First divide the 409 known tones into 82 sentences, first read a sentence in any language, then divide the sentence into known tones, remove the mute and noise, and convert it into linear predictive coding cepstrum (LPCC) ExP matrix

(70). . . Use the distance to classify the known single-tone linear predictive coding cepstrum (LPCC) ExP matrix into the L closest database

(80). . . There are M=1000 databases, each database contains similar known tones, and 409 stands for 14100 words in the 1000 database. Each tone can be placed in L databases.

(2). . . In the same language, the first sound is clearly pronounced to identify the unknown single tone.

(41). . . In each equal-length elastic frame, P linear predictive coding cepstrums are calculated by the least square method, and an unknown single tone is represented by a linear predictive coding cepstrum ExP matrix.

(84). . . Use the distance in M=1000 databases to find F and the closest database to identify unknown tones.

(90). . . Similar known tones in the F closest database, using the distance to find the unknown tones that are to be recognized, all the words that are recognized by the known tones of the unknown tones appear in the group.

The first and second figures illustrate the invention execution procedure. The first figure shows the establishment of M=1000 different databases, each containing similar known tones. The second figure shows the process of identifying unknown tones and inputting Chinese. The third figure shows the input of the Chinese part of the invention using Visual Basic software. The fourth picture shows the third picture of the less common word.

(2). . . In the same language, the first sound is clearly pronounced to identify the unknown single tone.

(10). . . Sound wave digitization

(20). . . Remove noise and silent time

(30). . . E equal length elastic frames normalize all sound waves

(41). . . In each equal-length elastic frame, P linear predictive coding cepstrums are calculated by the least square method, and an unknown single tone is represented by a linear predictive coding cepstrum ExP matrix.

(80). . . There are M=1000 databases, each database contains similar known tones, and 409 stands for 14100 words in the 1000 database. Each tone can be placed in L databases.

(84). . . Use the distance in M=1000 databases to find F and the closest database to identify unknown tones.

(90). . . Similar known tones in the F closest database, using the distance to find the unknown tones that are to be recognized, all the words that are recognized by the known tones of the unknown tones appear in the group.

Claims (2)

A method of inputting Chinese characters in any language, the steps of which include: (1) M=1000 different sounds first; (2) a pre-processor deletes before the mono (sound) and the mono ( After the sound), all the mute and noise without the sampled points of the speech sound; (3) the normalization of a sound or a single sound wave and the feature extraction method: using E equal elastic frames, no filter, no Overlap, normalize a sound or monophonic sound wave and convert it into a linear predictive coding cepstrum (LPCC) E×P matrix of equal size; (4) M=1000 linear predictive coding cepstrum (LPCC) for different sounds The E×P matrix represents M=1000 different databases; (5) The 409 known tones are first divided into 82 sentences, and a sentence is first pronounced in any language, and then the sentence is divided into known tones, and the single is deleted. Before and after the tones, and all the mute and noise without the sampled points of the voice, normalize the sound waves of a known single tone with E equal elastic frames and convert them into linear preamps of equal size. Estimated coded cepstrum (LPCC) E × P matrix; (6) with known monophonic linearity Coding cepstrum estimate (LPCC) ExP all matrix M = linear predictor encoding 1000 different sounds or weighted cepstrum distance to find the distance L between the closest ExP matrix repository, has the The linear predictive coding cepstrum (LPCC) ExP matrix of the known tone is divided into the L closest databases. Similarly, the linear prediction of all known tones is coded by the inverse spectral ExP matrix using the distance or weighted distance. The linear predictive coding cepstrum ExP matrix representing and representing 1000 database sounds is located in the nearest L database, and similar known tones are placed in the same database, and each single tone can be placed in L databases. (7) When using the same language to recognize the unknown single tone, the user deletes the unknown monophonic sound before and after the unknown tone, and all the non-speech sound signal points (sampled points) Mute and murmur, normalize the sound of the unknown monophonic speech with E equal elastic frames, and convert it into a linear predictive coding cepstrum E×P matrix of equal size, and the present invention uses the unknown monophonic linearity again. Estimate the distance or weighted distance between the coded cepstrum ExP matrix and the linear predictive coded cepstrum ExP matrix of all M=1000 different sounds to find the F closest database, and then use the unknown tone linear predictive coding Spectrum ExP matrix with F most The distance or weighted distance between the censored ExP matrices of similar known monotones in the near database, looking for the unknown single tone desired by the user, is represented by the known single tone of the unknown single tone. All the words of the group (common words and unusable words) appear in the plenary; (8) if the recognition is unsuccessful, the user will pronounce the tone again, using E An equal elastic frame converts the tone into a linear predictive coded cepstrum ExP matrix, and divides the tone's ExP feature matrix into the L closest databases by distance, and then recognizes the tone. Entering a Chinese character method in any language according to item 1 of the scope of the patent application, wherein step (3) comprises using E equal elastic frames, equal length, no filter, no overlap, a sound or single To normalize the sound waves and extract the feature matrix of the same size, the steps are as follows: (a) Delete the silence and noise before the mono (sound) and after the mono (sound), and all the sampled points without the speech sounds. Using a method of equally dividing a single tone (one sound) with sound wave signal points, in order to closely estimate the nonlinearly varying sound waves with a linearly varying regression mode, the full length sound wave signal points are divided into E=12 equal periods, each equal period Form a flexible frame, a single tone (one sound) has a total of E "equal length" elastic frames, no filter (Filter), no overlap, can freely extend and cover the full length of the sound wave, not a fixed length Hamming window ; (b) each "isometric" flexible box, ready intercropping sonic estimated using a non-linear changes linearly changed at any time intercropping regression model; (c) point signal S (n) estimated by the foregoing signal point , Which is the estimated value S '(n) represented by the following regression model: (d) In the above formula, 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 the above formula and divide the partial differential into 0 to obtain the normal equation of the P group: From the last two formulas, the minimum total squared difference E _P Convert the last two to In the above two formulas, 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) The upper 5 formula is calculated cyclically 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 , A single tone (or a sound) is represented by E linear predictive coding cepstral (LPCC) vectors (a linear predictive coding cepstrum (LPCC) E x P matrix).