JP5091202B2 - Identification method that can identify any language without using samples - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an identification method correctly identifying all language sentences by improving a feature of continuous sound, and thereby, identifying all languages such as Taiwan-Chinese, English, Japanese, German, French, Korean, Russian, Cantonese and Taiwanese, without using samples. <P>SOLUTION: The continuous sound (word) includes one or more single short duration sounds, and a feature of the continuous sound of all languages is extracted from unknown continuous sound of all languages. The unknown continuous sound is displayed by using a matrix value, and dispersed to 144-dimensional space, and the feature of known continuous sound of all languages is dispersed to the 144-dimensional space, and simulated and calculated by the feature of the unknown continuous sound around the known continuous sound. The continuous sound including 12 elastic frames with the same length, without filters and without overlapping, is converted to sound wave 12&times;12 matrix with various lengths, and compared and identified by Bayes' identifying method. <P>COPYRIGHT: (C)2011,JPO&amp;INPIT

Description

The present invention relates to an identification method that can identify any language without using a sample. In particular, a continuous tone includes one or many syllables (single notes), and can identify all languages without using a continuous tone sample. Converts continuous sound waves of equal length, no filters, no overlap, and varying lengths into a matrix of 12x12 linear predictive cepstrum coefficients (LPCCs) using one stretchable frame An unknown word or continuous tone is displayed using a matrix of 12 × 12 linear prediction cepstrum coefficients, and one 12 × 12 matrix is represented as one vector in one 144-dimensional space. Recognize that many unknown words or vectors of continuous sounds are scattered in a 144-dimensional space, and when a speaker emits one known continuous sound, the features of the known continuous sound are Simulated and calculated by leaf or continuous sound features (LPCC), including 12 stretchable frames, normalizing certain continuous sound waves, Bayesian comparison is performed in the raw database Search for one known continuous sound for an unknown word or continuous sound, split one unknown sentence of one speaker into D unknown words or continuous sound, and one The window screening relates to an identification method capable of identifying any language without using a sample for screening one known sentence as an unknown sentence of a speaker.

When a certain continuous sound is emitted, the pronunciation is displayed by sound waves. A sound wave is a kind of system that performs a non-linear change according to time. A sound wave of a certain continuous sound includes a kind of dynamic characteristic and performs a non-linear continuous change according to time. When a homologous continuous tone is emitted, it has a series of homologous dynamic properties and performs non-linear extension and contraction over time. However, homologous dynamic characteristics are arranged in the same order based on time but are different in time. When homologous continuous sounds are emitted, it is very difficult to align homologous dynamic properties on the same time position. Furthermore, since there are a lot of similar continuous sounds, the identification is made more difficult.
( However, hereinafter, “in-phase” means “same language” (for example, “Japanese”) .)

In some computerized language identification systems, it is first necessary to extract sonic-related language information, i.e., dynamic characteristics, and filter noise that is unrelated to the language. For example, the timbre of the human voice, the tone of the sound, the psychology at the time of speaking, the physiology, and the emotion are irrelevant to the voice identification, and are deleted first. Subsequently, the homologous features of the homologous continuous sounds are arranged on the homologous time position. This series of features is displayed using a sequence feature vector having the same length, and is called a feature pattern of a continuous sound. In current speech identification systems, the generation of feature patterns with matching sizes is too complicated and time consuming. This is because the homologous features of homologous continuous sounds are very difficult to line up at the same time position, and in particular, English is more difficult to identify.

There are the following five main tasks in a general sentence or name identification method. Divide an unknown sentence or name into D unknown words or continuous sounds, extract features, normalize the features (characteristic pattern sizes match, and homologous features of homologous words or continuous sounds Are identified at the same time position), and unknown words or continuous sounds are identified, and the sentence or name database is searched for a matching sentence or name. The characteristics of certain continuous sound waves often include energy, zero crossings, extreme count, formants, linear prediction cepstrum coefficient (LPCC), and mel frequency cepstrum coefficient (MFCC). Expressed.

Of these, the linear prediction cepstrum coefficient (LPCC) and the mel frequency cepstrum coefficient (MFCC) are the most effective and widely used. Linear predictive cepstrum coefficients (LPCC) represent the most reliable, stable and accurate language features of a continuous tone. It uses a linear regression method and represents a continuous sound wave, and calculates a regression coefficient by a minimum square estimation method. When the estimated value is further converted into a cepstrum, a linear prediction cepstrum coefficient (LPCC) is obtained.

The Mel frequency cepstrum coefficient (MFCC) converts sound waves into frequencies using a Fourier transform method. Furthermore, the auditory system is estimated based on the mel frequency proportionality. SB Davis and P. Mermelstein publish a paper "Comparison of parametric representations for monosyllabic word recognition in continuously spoken" published in "IEEE Transactions on Acoustics, Speech Signal Processing, Vol. 28, No. 4" published in 1980. According to sentences, the Mel frequency cepstrum coefficient (MFCC) feature using the dynamic time warping method (DTW) is higher than the identification rate of the linear prediction cepstrum coefficient (LPCC) feature. However, in several speech recognition experiments (including the inventors' previous invention), the linear predictive cepstrum coefficient (LPCC) feature identification rate using the Bayes classifier is more than the mel frequency cepstrum coefficient (MFCC) feature. It is also expensive and saves time.

Many methods have already been adopted for language identification. For example, there are a dynamic time-warping method, a vector quantization method, and a hidden Markov model method (HMM). If the homologous pronunciation is different in the change over time, the homologous feature is extended to the same time position while comparing. Although this discrimination rate is very high, it is very difficult to extend homologous features to the same position, the warp time is too long, and it is not applicable. Vector quantization is not only inaccurate but also time consuming when identifying large numbers of continuous sounds. The hidden Markov model method (HMM) is a good discrimination method, but it is complicated and requires too many unknown parameters to be estimated, so it takes time to calculate and identify the estimated values.

In the paper “Speech recognition of mandarin monosyllables” published in “Pattern Recognition, vol. 36” published in 2003, TF Li uses a Bayesian classifier and uses a homologous database. A series of LPCC vectors was compressed into a classification pattern of homologous size. The recognition results are based on the paper “The recognition of mandarin monosyllables based on the discrete” published in 1990 by “Proceedings of Telecommunication Symposium, Taiwan” published by YK Chen, CYLiu, GH Chiang, and MT Lin. In "Hidden Markov model", it is better than using the HMM method of hidden Markov model. However, the compression process is complex, time consuming, and it is very difficult to compress the homologous features of homologous continuous sounds to homologous time positions, and it is extremely difficult to identify similar continuous sounds.

The speech identification method of the present invention naturally derives a method for extracting speech features by performing non-linear changes according to time based on speech features with sound waves, from the viewpoint of science, in view of the above drawbacks. A sound wave of a certain continuous sound is first normalized, and then converted into a feature pattern having a size sufficient to represent the continuous sound. Moreover, homologous continuous sounds have homologous features at the time positions of homology within their feature patterns, and it is not necessary to adjust unknown parameters and reference values in the present invention by human or experiment. Use simple Bayesian classifiers to compare unknown word or continuous sound classification patterns with known continuous sound standard patterns in the continuous sound feature database, without having to look for and compare recompressed, warped, or homologous features . Therefore, the voice identification method of the present invention can quickly complete feature extraction, feature normalization, and identification.

The problem to be solved by the present invention is to provide an identification method capable of identifying any language without using a sample.

In order to solve the above problems, the present invention provides an identification method capable of identifying any language without using the following sample.
Since the most important object of the present invention is to simulate and calculate the characteristics of any one known continuous sound in any language using a number of unknown words or continuous sound characteristics, the present invention It is possible to construct features of continuous sounds in any language without using a sample, that is, various languages can be accurately identified without using a sample of the present invention. Specifically, the present invention uses Bayesian distances for any one known continuous sound in any language to search for N unknown words or continuous sound matrices in a 144-dimensional space, Continuous sounds can be simulated and calculated, thus constructing any known continuous sound feature without using a sample of known continuous sounds. Thus, any language can be identified.
The present invention provides a language identification method, which can eliminate speech waves that do not have a language.
The present invention provides a sonic normalization and feature extraction method for continuous sounds. It uses E mutually equal stretchable frames, does not overlap, has no filter, can freely adjust all wavelengths based on the length of some continuous sound wave, and according to time within the sound wave of continuous sound A series of dynamic characteristics with nonlinear changes is converted into a feature pattern of equal magnitude, and the homologous continuous sound wave feature pattern shows a homologous feature at the homologous time position. Have. Immediate identification is possible, and a computer immediate identification effect can be achieved.
The present invention provides a method for identifying unknown words or continuous sounds by a simple and effective Bayesian method, minimizing the probability of identification error, reducing calculation, quick identification, and high discrimination rate.
The present invention provides a method for extracting features of a continuous sound, and the sound wave of a continuous sound has a dynamic characteristic that performs a non-linear change according to a kind of time. The present invention uses a sound wave that performs a non-linear change according to a regression model estimation time that performs a linear change according to time, and generates a minimum square estimate value (LPC vector) of a regression unknown coefficient.
The present invention uses sound waves (sound sample points) with all sounds. A smaller number E = 12 mutually equal stretchable frames is used, there is no filter, no overlap, and all sample point features are included. If a continuous sound wave is too short, the continuous sound is not deleted, and if it is too long, some sample points are not deleted or compressed. If human hearing can distinguish this continuous sound, the present invention can extract the characteristics of the continuous sound. Therefore, the speech identification method of the present invention can extract speech features as much as possible by applying sample points each having one speech. Since E = 12 stretchable frames do not overlap and the number of frames is small, the time to calculate feature extraction and linear prediction cepstrum coefficients (LPCC) can be greatly reduced.
The identification method of the present invention can identify continuous sounds that are too fast to speak or too slow to speak. When talking too fast, some continuous sound waves are very short. If the length of the stretchable frame is shortened, the present invention can cover short waves using E equal length stretchable frames of homologous number, and E linear prediction cepstrum coefficients. (LPCC) vector is generated. The E linear prediction cepstrum coefficient (LPCC) vectors can effectively represent the characteristic pattern of the short sound as long as the short sound can be discriminated by humans. The continuous sound waves emitted when speaking too late are longer, the stretchable frame stretches, and the generated E linear predictive cepstrum coefficient (LPCC) vectors can effectively represent the long sound.
The present invention provides a method for stabilizing and adjusting all known continuous tone features in a database, whereby all continuous tone features occupy their position and space relative to each other in a 144-dimensional space. Thus, accurate identification can be performed.
When identifying one sentence or name, first, the unknown sentence or name is divided into D unknown continuous sounds, and the present invention uses Bayesian methods to separate each unknown word or continuous sound, In the continuous sound feature database, select F known continuous sounds that are most similar. One sentence is displayed by D × F known continuous sounds and is difficult to cut, so it is divided into a relatively large or relatively small number of unknown words or continuous sounds. Compares one known continuous sound in a sentence or name by F similar known continuous sounds in three rows before and after an unknown word or continuous sound, and each sentence or name in the sentence and name database In contrast, a method for screening one known continuous sound using known similar continuous sounds in a 3 × F window and searching for the most likely sentence or name from the sentence and name database. Is simple and has a very high success rate (identifies 70 English sentences and names and 407 Taiwanese sentences and names).
The present invention provides two techniques to modify the characteristics of continuous sounds, thereby successfully identifying unknown words or continuous sounds and unknown sentences or names.
In the present invention, one Taiwanese Chinese single sound is a continuous sound of only one syllable, and all Chinese and foreign language features are displayed by a matrix of the same sample size. Therefore, the present invention can simultaneously identify various languages.

The identification method that can identify any language without using the sample of the present invention improves the characteristics of a certain continuous sound, thereby correctly identifying any language sentence, and thus, without using the sample, Taiwanese Chinese, English, Japanese All languages such as German, French, Korean, Russian, Cantonese, Taiwanese can be identified.

It is a flowchart which shows the construction process of a known continuous sound permanent database, a known continuous sound feature database, a sentence and a name database. It is a flowchart which shows the process of the identification method of one unknown sentence or a name. It is a figure which shows the identification method of 384 Taiwanese Chinese single sound, 1 German, 1 Japanese, and 2 Taiwanese. It is a figure which shows the identification method of 154 English and one German. FIG. 3 is a diagram of a method for identifying 269 Taiwanese Chinese phonetics and 3 Taiwanese words. The sentence and name database is a diagram showing that there are 70 English sentences and 407 Chinese sentences and names. It is a Visual Basic identification diagram showing a method for simultaneously identifying English and Taiwanese Chinese sentences and names. It is a Visual Basic identification diagram showing a method for simultaneously identifying English and Taiwanese Chinese sentences and names.

The best mode for carrying out the present invention will be described in detail below with reference to the drawings.

1 and 2 illustrate the enforcement process of the present invention.
FIG. 1 shows the construction process of three databases: a known continuous sound permanent database, a known continuous sound feature database, and a sentence and name database.
The continuous sound feature database includes a standard pattern of all known continuous sounds and indicates the characteristics of known continuous sounds.
First, one known continuous sound or one sentence or name 1 is input (the sentence or name is divided into a number of continuous sounds), and the receiver 20 is entered in a certain continuous sound wave 10 format.
The digital converter 30 converts continuous sound waves into sequence sound wave digital sample points.

The preprocessor 45 has the following two types of deletion methods.
A variance value of sample points and a variance value of general noise within a certain time frame are calculated. If the former is smaller than the latter, the certain time frame has no audio and should be deleted.
The sum of the distances between two consecutive sample points within a certain time frame and the sum of the general noise are calculated. If the former is smaller than the latter, the certain time frame has no audio and should be deleted.

As the preprocessor 45 passes, the sequence comprises the known continuous tone sample points.
First, normalize sound waves, then extract features, and divide all sample points of known continuous sound into E equal time frames.
Each time frame constitutes one frame.
A certain continuous sound has a total of E isometric frames 50, has no filter, and does not overlap.
Based on the length of all sample points in the continuous sound, the length of E frames is freely adjusted to cover all sample points.
Therefore, the frame is called a stretchable frame, and the length can be freely stretched, but the lengths of the E stretchable frames are the same.
Unlike a Hamming window, it has filters, is half-overlapped, has a fixed length, and cannot be freely adjusted according to wavelength.

A certain continuous sound changes nonlinearly according to the sound wave time, and the sound wave includes one voice dynamic feature and changes nonlinearly according to time. Since there is no overlap, the present invention uses relatively few (E = 12) stretchable frames and covers all continuous sound waves. Since the sample point can be estimated from the previous sample point, the linear change regression method is performed according to the required time, the non-linear change sound wave is closely estimated, and the regression unknown coefficient is estimated using the minimum flat method. Within each frame, a set of unknown coefficient minimum square estimates is generated, which is referred to as a linear prediction code (LPC) vector.
Further, the linear prediction code (LPC) vector is converted into a relatively stable linear prediction cepstrum coefficient (LPCC). A sound wave of a continuous tone includes a speech dynamic feature that changes nonlinearly according to a sequence time. In the present invention, the sound wave is converted into E linear prediction cepstrum coefficient (LPCC) vectors 60 having the same magnitude. .

In order to extract the characteristics of one known continuous sound, first, one permanent known continuous sound database is prepared. Each known continuous tone is pronounced once by a speaker with a standard and pronounced pronunciation. When identifying poor or non-standard utterances, they are pronounced by such a speaker and all known continuous sounds are converted into E × P LPCC matrices in a permanently known continuous sound database. Incorporate In order to extract one known continuous sound feature in a permanently known continuous sound database, an unknown word or continuous sound database is first prepared.

There are two types of unknown words or continuous sound databases. One type is an unknown word or sample with continuous sound, and the other type has no standard. An unprocessed database with samples first determines the mean and variance of each unknown word or continuous sound. In a database of unknown words or continuous sounds with a sample, use the Bayesian distance to find the N nearest unknown words or continuous sounds around that known continuous sound. Further, N average values of N unknown sounds and N + 1 weighted average values of the linear prediction cepstrum coefficients (LPCC) of the known continuous sounds are obtained, and the average values of the known continuous sounds are obtained. The weighted average value of N variance values of the continuous tone is set as the known variance value of the continuous tone. This E × P mean value and variance value matrix is the initial feature value 79 of the known continuous tone and is incorporated into the continuous tone feature database.

If there are no samples in the unknown raw database, the unknown word or continuous sound database is searched for N unknown words or continuous sounds around the known continuous sound using the minimum absolute distance. The linear prediction cepstrum coefficient (LPCC) of the known continuous sound and N unknown words or continuous sounds is defined as (N + 1) numbers. Find the weighted average value of (N + 1) numbers and make it the average value of the known continuous sounds, and find the variance value of (N + 1) numbers and make the variance value of the known continuous sounds, This matrix of E × P mean and variance values represents the initial features of the known sound and is incorporated 79 into a known continuous sound feature database.

In the known continuous sound feature database, if the average value of one known continuous sound and the Bayesian distance of a similar known continuous sound LPCC in the permanent known continuous sound database are If not, use the Bayesian distance in the feature database and look for N known continuous notes, and the LPCC for that known continuous note in those Bayesian matrices is the N smallest. N known continuous sounds are obtained, N average values and LPCC weighted average values of the known single notes are set as new average values of the known continuous sounds, and N variances of the N known continuous sounds are obtained. Use the weighted average of the values as the new variance value for that known continuous sound. This method is repeated several times to calculate new average values and variance values for each known continuous tone in the feature database. Finally, the new E × P mean and variance matrix is referred to as the standard pattern, representing its known continuous sound, and incorporated into the feature database 80. In addition, a sentence and name database is constructed 85 using known continuous sounds from a known feature database.

FIG. 2 shows a procedure for identifying one unknown sentence or name. After entering one unknown sentence or name 2 into the speech recognition method of the present invention, the receiver 20 is entered by a set of unknown words or continuous sound waves 11. The digital converter 30 converts to a series of sonic sample points. A sound wave of one unknown sentence or name is divided into D unknown words or sound waves 40 of continuous sounds. Further, the pre-processor 45 shown in FIG. Next, normalize the sound wave of each unknown word or continuous sound, extract features, and divide all sample points comprising the sentence or name of each unknown word or continuous sound into E equal time frames. Each time frame forms a stretchable frame 50. Each continuous tone has a total of E stretchable frames, no filters, no overlap, stretches freely and covers all sample points.

Since each sample point can be estimated from the previous signal within each frame, the estimated value of the regression unknown coefficient is obtained using the minimum flat method. The set of least square estimates that occur within each frame is referred to as a linear prediction code (LPC) vector. The linear prediction code (LPC) vector is successfully distributed and further transforms the linear prediction code (LPC) vector into a relatively stable linear prediction cepstrum coefficient (LPCC) vector 60. One unknown word or sequence The sound has E linear prediction cepstrum coefficient (LPCC) vectors as feature patterns and is called a classification pattern 90 and is the same size as a known continuous sound standard pattern. A sentence has a total of D classification patterns and represents 90 unknown words or continuous sounds. If one known continuous sound is this unknown word or continuous sound, the average value of the standard pattern is the linear prediction cepstrum coefficient (LPCC) closest to the unknown word or continuous sound classification pattern. Thus, the simplified Bayesian identification method of the present invention compares the standard pattern of each known continuous sound 100 with the unknown word or continuous sound classification pattern and the continuous sound database 80.

If one known continuous sound is that unknown word or continuous sound, all linear prediction cepstrum coefficients (LPCC) in the unknown word or continuous sound classification pattern to save computation time Are independent normal distributions, and their average number and variance are estimated by the average and variance within a known continuous tone standard pattern. The simplified Bayesian method calculates the distance between the unknown word or the linear prediction cepstrum coefficient (LPCC) of continuous sounds and the average number of known continuous sounds. Further, the value obtained by adjusting by the known continuous sound dispersion value represents the similarity between the unknown word or continuous sound and one known continuous sound. An unknown word or continuous sound and a known continuous sound with the highest F similarity are selected and set as an unknown word or continuous sound. Thus, one unknown sentence or name is displayed 110 using D × F known continuous sounds.

After dividing one unknown sentence or name into D unknown words or continuous sounds, it is difficult to just divide the continuous sound and number included in one unknown sentence or name. At times, a continuous sound is divided into two parts. At other times, the two continuous sounds are pronounced very similar, and the computer divides them into one. Thus, the D unknown words or continuous sounds are not necessarily the true number of continuous sounds of the speaker. Therefore, the F continuous sounds similar to the known F in a line do not necessarily include the continuous sound of the speaker. When identifying an unknown sentence or name, each sentence and name database 85 tests each known sentence and name. Test whether a sentence or name is a speaker's sentence or name, and the sentence or name from the first known continuous sound before or after the continuous sound with similar D × F matrix Compare with similar sounds in three rows (naturally, the first comparison can only compare similar sounds in two rows in the middle and the back). Next, move to the 3 × F window (three similar rows of similar similar continuous sounds) 120 and look for the second known continuous sound of the sentence. In this way, all known continuous sounds of the sentence are tested.

In the sentence and name database, the sentence or name with the highest probability is the sentence or name of the speaker (the number of known continuous sounds in the tested sentence or name in the 3 × F window is the number in the tested sentence or name. Divide by the number of continuous notes 130). Of course, in the sentence and name database, unknown sentences or names (D unknown words or continuous sounds) can be selected and compared with sentences or names whose lengths are approximately homologous, thereby saving time. If the sentence or name cannot be identified, the Bayes classifier is used to search the feature database for the N most similar continuous sounds79, and if the continuous sound features in the sentence are improved, the identification is always successful. To do.

This will be described in detail below.
After a continuous sound is input to the speech identification method, the continuous sound wave is converted into a series of signal sampled points. In addition, sample points that do not have audio sound waves are deleted. The present invention provides two methods. First, the variance of sample points within a certain time frame is calculated. The second calculates the sum of the distances between two adjacent sample points within the time frame. Theoretically, the first method is better, but the variance value of the sample points is larger than the noise variance value, indicating that speech is present. However, when the present invention identifies continuous sounds, the two methods have the same identification rate, but the second method can save time.

After deleting sample points that do not have sound, the remaining sample points represent all sample points of a continuous tone. First, normalize the sound waves, then extract features, and divide all sample points into E equal time frames. Each time frame forms one frame. A continuous tone has a total of E equal length stretchable frames, has no filters, does not overlap, stretches freely, and covers all sample points. The sample points in the stretchable frame change nonlinearly with time and are difficult to represent with a mathematical model. Because J. Markhoul published a paper “Linear Prediction: A tutorial review” in “Proceedings of IEEE, Vol. 63, No. 4” published in 1975. This is because there is a linear relationship between the previous sampling point and the previous sampling point, and it is explained that the sampling point of this nonlinear change can be estimated using a regression model that changes linearly according to time.





Then, according to the last linear prediction cepstrum coefficient (LPCC), it approximates to 0. A continuous tone has E linear prediction cepstrum coefficient (LPCC) vector display features, that is, contains a continuous sound with a matrix display of one E × P linear prediction cepstrum coefficient (LPCC), and one continuous tone is one Or contains many syllables.

(3) In the same way, when E linear prediction cepstrum coefficient (LPCC) vectors of one unknown word or continuous sound wave are calculated by equation (8-15), E × P of the same size A matrix of LPCCs is provided, which is called an unknown word or continuous tone classification pattern.

(5) In order to extract the characteristics of one known continuous sound, first, a database of unknown words or continuous sounds is prepared. There are two types of unknown words or continuous sound databases. One type is a sample of unknown words or continuous sounds, and the other type is no sample. In an unprocessed database with samples, first, an average value and a variance value of each unknown word or continuous sound are obtained. In a database of unknown words or continuous sounds with a sample, use the Bayesian distance to find the N nearest unknown words or continuous sounds around that known continuous sound. Further, N average values of N unknown sounds and N + 1 weighted average values of the linear prediction cepstrum coefficients (LPCC) of the known continuous sounds are obtained, and the average value of the known continuous sounds is obtained. The weighted average value of the N variance values of the continuous tone is set as the variance value of the known continuous tone. This E × P mean value and variance value matrix is the initial feature value 79 of the known continuous tone and is incorporated into the continuous tone feature database. If there are no samples in the unknown word or continuous sound database, search for N unknown words or continuous sounds around the known continuous sound in the unknown word or continuous sound database using the minimum absolute distance. . The linear prediction cepstrum coefficient (LPCC) of the known continuous sound and N unknown words or continuous sounds is defined as (N + 1) numbers. Find the weighted average value of (N + 1) numbers and use it as the average value of its known continuous sounds, and find the variance value of (N + 1) numbers and use it as the variance value of the known continuous sounds . This matrix of E × P mean and variance values represents the initial features of the known continuous tone and is incorporated 79 into the known continuous tone feature database. In the known continuous sound feature database, if the average value of one known continuous sound and the Bayesian distance of a similar known continuous sound LPCC in the permanent known continuous sound database are If not, use the Bayes distance in the feature database and look for N known continuous sounds. The LPCC for that known continuous tone of their Bayesian matrix is the N smallest. N known continuous sounds are obtained, and N average values and LPCC weighted average values of the known continuous sounds are set as new average values of the known continuous sounds. The weighted average value of the variance values is used as the new variance value of the known continuous sound. This method is repeated several times to calculate new average values and variance values for each known continuous tone in the feature database. Finally, the new average and variance matrix of E × P is referred to as the standard pattern and represents its known continuous sound, incorporated into the feature database 80, using the known continuous sound of the known feature database, Build sentence and name database 85.

(7) In order to prove that the present invention can discriminate all languages at the same time, the present invention conducted a voice discrimination experiment of two people.
(a) First, an unknown word or continuous sound database is constructed. This phone database was purchased from the Central Research Institute in Taiwan. There are a total of 388 Taiwanese Chinese phonetics in the database (Fig. 3), all of which are pronounced by women, and there are 6 to 99 samples, and the pronunciation of many singles is almost the same.
(b) From the method in section (2), when all samples are converted to an E × P LPCC matrix, it has a total of 12400 matrices.
(c) In 388 Taiwanese Chinese single notes, the average value and the variance value are obtained using samples.
(D) At random, 388 Taiwanese Chinese single notes are mixed, and 388 samples have average and variance single notes as 388 unknown words or continuous sound database (1 Taiwan Chinese A single word is a continuous sound with only one syllable).
(e) Next, one male and one female each pronounces 654 Taiwanese Chinese sounds, 154 English, 1 German, 1 Japanese, and 3 Taiwanese, Construct two 813 permanently known continuous tone databases. Each continuous tone is displayed by a linear prediction cepstrum coefficient (LPCC) E × P matrix.
(f) In 813 known continuous sounds in a permanent known continuous sound database, for each one known continuous sound, using Bayesian distance 20, in 388 unknown words or continuous sounds, Search for N = 15 unknown words or continuous sounds. The linear prediction cepstrum coefficient (LPCC) of the known continuous sound and the sample average value of N unknown words or continuous sounds are obtained as N + 1 weighted average values, and the average value of the known continuous sounds is defined as N A weighted average value of the sample variance values of the unknown words or continuous sounds is obtained and set as the dispersion value of the known continuous sounds. This average and variance 12 × 12 matrix is referred to as the initial feature of the known continuous tone 79 and is present in the known continuous feature database. That is, the feature database includes 813 12 × 12 average and variance value matrices 80.
(g) In the feature database, if the average value of one known continuous sound is the same as in the permanent continuous sound database, the Bayes distance of the LPCC of the known continuous sound is not minimum. N = 15 known continuous sounds are searched using 813 continuous sound feature Bayes distances. A weighted average value is obtained using N average values of N continuous sounds and LPCC of the known continuous sounds, and set as a new average value of the known continuous sounds. A weighted average value is obtained for the dispersion values of N known continuous sounds, and set as a new dispersion value of the known continuous sounds. New average and variance values are calculated several times. The last 12 × 12 mean and variance matrix fails with the standard pattern and represents its known continuous tone features and is present in 80 in the known continuous tone feature database.
The present invention performed the following continuous tone identification. The identification rate is determined by the person and there are too many similarities, so the top three are correct.
Identify 384 Taiwanese Chinese singles, 1 German, 1 Japanese, 2 Taiwanese (see Figure 3) (very high recognition rate)
Identify 154 English, 1 German (see Figure 4) (very high identification rate)
154 English and 388 Taiwanese Chinese, 1 German, 1 Japanese, and 2 Taiwanese at the same time (very high identification rate)
(4) Identify 654 Taiwanese Chinese singles, 1 German, 1 Japanese, 3 Taiwanese (see Fig. 5) (identification rate is high, but not as high as the above three examples)

(8) In identifying a speaker's sentence or name, we first built one English and Taiwanese sentence and name database. All continuous sounds in each sentence or name are arbitrarily composed of known English and Taiwanese Chinese in the continuous sound feature database (384 + 154). 154 English words compose 70 English sentences and names, and 384 Taiwanese Chinese words compose 407 Taiwan Chinese sentences and names (see FIG. 6).
The identification method is as follows.
(a) One unknown sentence or name is divided into D unknown words or continuous sounds, and each unit time frame calculates the sum of two sample drop distances adjacent to each other. If it is too small, the time frame is noisy or stuttering, and there is too much accumulation of adjacent unit timeframes (no more than two continuous syllable times) with no audio signal, and all are noise or stuttering. It should be divided at the boundary between two continuous sounds, and divided into a total of D unknown words or continuous sounds. Next, the 45, 50, 60, and 90 processes of FIG. 2 are used to convert to an E × P LPCC matrix. For each unknown word or series of sounds, the Bayes classifier 20 is used to select the most similar F known series of sounds in the feature database of English and Taiwanese (at the same time, English And may include Taiwanese Chinese (Figure)). An unknown sentence or name is displayed by a known continuous sound that is most similar to D × F.
(b) Look up the sentence or name of the speaker in the sentence and name database. Among the 477 English and Taiwanese sentences and names, (D ± 1) known continuous sound sentences and names Select.
(c) If the database selection is of length equal to the sentence or name to be compared and the sentence or name of the speaker, if there are D unknown words or continuous sounds, then each of D columns F D known series of sentences or names to be compared with similar known continuous sounds of F are compared in order, and F similar continuous sounds are known continuous sounds in the sentence or name to be compared. See if. If all the similar continuous sounds in each row include one continuous sentence or a known continuous sound in the name, the correct continuous sound is identified as D. That is, the comparison sentence or name is the sentence or name of the speaker.
(d) If the database comparison sentence and the number of known continuous sounds in the name are D-1 or D + 1, or (c) the number of identified accurate continuous sounds is not D, the present invention uses a 3 × F window. To screen. In the comparison sentence or name (in the database), the i-th known continuous sound is the similar known continuous sound of the front and rear three rows in the D × F matrix (that is, the i−1, i, i + 1th columns). Used to compare the i th known continuous sounds and calculate how many comparison sentences or known continuous sounds in the name are in the D × F matrix. Next, the probability of the comparison sentence or name is obtained by dividing by the total number D, and the sentence or name having the maximum probability in the database is selected as the pronunciation of the speaker.
(e) If the identification of a sentence or name is an error, there will always be one or many of the D unknown words or continuous sounds, and not those F similar known continuous sounds . The Bayes classifier 20 is used to search for (N = 15) known continuous sounds from among (155 + 384) known continuous sounds, and to search for N similar continuous sounds and their unknown words or continuous sounds. Find the LPCC weighted average and improve the unknown word or continuous sound. Thus, the D unknown words or series of sounds are within those F similar known series and the second test will always succeed.
In the present invention, the following English and Taiwanese Chinese sentences and names were identified. Almost everything is correct, but it varies from person to person.
(1) Identify 70 English sentences and names (very good).
(2) Identify 407 Taiwanese Chinese sentences and names (very good)
(3) Identify 70 English sentences and names and 407 Taiwan Chinese sentences and names (very good).

The present invention has been confirmed to be able to achieve the expected purpose after many tests. Moreover, its functions are outstanding, it has not been seen in public publications before application, and there is no fact of public use. Therefore, the present invention has novelty that is a requirement of claims and is sufficient compared to conventional similar products. It has great progress, is highly practical, meets social needs, and has a very high industrial utility value.

1. Build one known continuous sound permanent database, pronounce a continuous sound or a sentence, and divide the sentence into a number of known continuous sounds.
10 continuous sound wave 20 receiver 30 sound wave digital converter 45 noise removal 50 E stretchable frame normalized sound wave 60 calculate linear prediction cepstrum coefficient (LPCC) vector by minimum flat method 70 using Bayesian distance (absolute value distance), For each known continuous sound (permanent database), look for the N newest unknown words or continuous sounds in the unknown word or continuous sound database.
79 For each known continuous sound (permanent database), a weighted average value is obtained by using the surrounding N unknown words or continuous sounds and LPCC of the known continuous sounds. The initial features of the known continuous sound are incorporated into a feature database. Further, using the Bayes distance in the feature database, N known continuous sounds and the known continuous sound LPCC weighted average value are obtained, and calculation is performed several times. The last weighted average value (E × P average value and variance value) represents the standard pattern of the known continuous sound.
80 The known continuous sound feature database contains a standard pattern of all mean and variance values.
85 Build a sentence and name database of sentences and names to be identified using the continuous sounds of a known continuous sound feature database.
2 Enter an unknown sentence or name.
11 A set of unknown words or continuous sound waves 40 A sentence or name is divided into D unknown words or continuous sounds.
A 90 linear prediction cepstrum coefficient (LPCC) matrix of D unknown words or series represents D unknown words or series classification patterns.
A 100 Bayes classifier is used to compare each one known continuous tone standard pattern with an unknown word or continuous tone classification pattern.
110 Search for F known continuous sounds closest to each unknown word or continuous sound in a sentence or name, and the sentence or name is based on a total of D × F known similar sounds. expressed.
In the 120 sentence and name database, each known continuous sound in all sentences and names is screened using similar known continuous sounds in a 3 × F window.
In the sentence and name database, look for one most likely sentence or name.

Claims (9)

A method for identifying utterances in any language, with the following steps:
(1) An unprocessed database comprising a plurality of samples of an arbitrary language and consisting of unknown words or continuous sounds; or an unprocessed database consisting of an unknown word or continuous sounds without having a sample of arbitrary languages; With
The plurality of samples, said emitted by an unknown word or continuous sound the same speaker, step consists of at least a plurality of words or continuous sound,
(2) providing a permanent database of known words pronounced by a speaker with a standard, clear and clear utterance or by a subject;
(3) using a processor to delete noise and a time frame without a speech signal from the speech waveform;
(4) Normalize the total length of the utterance waveform of one word or continuous sound, and use E = 12 stretchable frames without filters and without overlap, Transforming to ExP = 12 × 12 identically sized matrices of linear prediction cepstrum coefficients (LPCC);
(5) calculating an average value and a variance value of LPCC of samples from a plurality of samples in an unprocessed database having the plurality of samples;
(6) From the unprocessed database having the plurality of samples, N samples having the average value and the variance value of the LPCC of the samples and using a simple Bayes classifier are closest to the known words in the permanent database. Find N unknown words with a Bayesian distance of
Locating N unknown words from an unprocessed database without the sample with N absolute distances closest to known words in the permanent database;
(7) N unknown words having an average value and variance value of LPCC of the samples and having N Bayes distances closest to the known words in the raw database having the plurality of samples. From the (N + 1) data of the N LPCCs of the known words and the LPCCs of the known words in the permanent database, an average value and a variance value of the LPCCs of the known words are calculated,
Display an ExP = 12 × 12 matrix of LPCC mean and variance values of the known language as features of known words called standard patterns, and other known words of several different languages And storing a standard pattern of the known words in a word database,
And creating necessary sentences and names from known words in the word database and storing them in the sentences and name database;
(8) If the unknown word or continuous sound in the raw database does not have a sample, it has the N absolute distances closest to the known words in the permanent database, does not have the sample Consider N LPCCs of N unknown words in the raw database and LPCCs of known words in the permanent database as (N + 1) data;
Calculate an average value and a variance value of the (N + 1) data,
And storing in the word database an Exp = 12 × 12 matrix of mean values and variance values of the LPCC as known word features, referred to as standard putters;
(9) Normalize the input unknown word or continuous sound waveform length using E = 12 extendable frames without filter and without overlap,
And converting the entire waveform length into an ExP = 12 × 12 identically sized matrix of LPCCs, referred to as the unknown word classification pattern;
(10) The standard pattern of each known word in the word database is matched with the inputted classification pattern of the unknown word,
And using a simplified Bayesian classifier, searching the word database for a known word having a Bayes distance closest to the unknown word;
(11) dividing one unknown sentence or name into D unknown words;
(12) Using a Bayes classifier, search for F known words most similar to the unknown word from the word database;
And displaying the unknown sentence or name by a DxF matrix of similar known words in several languages,
(13) matching the DxF matrix of similar known words displaying the unknown sentence or name with all known sentences and names in the sentence and name database;
And searching the sentence and name database for a known sentence or name most likely to be the unknown sentence or name,
(14) improving the characteristics of unknown words in the input unknown sentence or name so as to ensure that the input unknown sentence or name is correctly identified. A method of identifying utterances in any language characterized by The step (3) further includes:
(A) within a unit time frame, calculate the variance value of the sample points of the speech signal and the variance value of the sample points of the noise, and if the variance value of the sample points of the speech signal is smaller than the variance value of the noise sample points Delete the unit time frame,
(B) calculating a sum of absolute distances between two adjacent utterance signal sample points and a sum of absolute distances between two adjacent noise sample points within a unit time frame; The method of claim 1, further comprising the step of deleting the time frame if the sum of absolute distances between sample points of the signal is smaller than the sum of absolute distances between sample points of the noise. To identify utterances. The step (4) further includes:
(A) E = 12 equal length stretchable frames to divide the full waveform length of one word or continuous sound into E = 12 equal sections and cover the full waveform length Each section as a stretchable frame without filters and without overlapping so that they can touch and stretch them,
(B) Use a linear regression model with P = 12 regression coefficients to estimate a nonlinear time-varying waveform within each stretchable frame, and use the least squares method to calculate P = 12 Generate linear predictive code coefficients (LPC);
(C) Using Durbin's recurrence equation with N points in each frame,

2. The sample according to claim 1, comprising the step of: (d) displaying E = 12 LPCC vectors, which are words or continuous sounds, represented by an ExP = 12 × 12 matrix of LPCCs. To identify utterances in any language without using. The step (5) further includes
(A) Divide the entire length of the waveform of unknown words or continuous sounds into E = 12 equal sections, and form each section as a stretchable frame without filters and without overlapping,
(B) To estimate a non-linear time-varying waveform, use a linear regression model with P = 12 regression coefficients in each E = 12 stretchable frame, and use the least squares method to calculate the LPC vector Produces
(C) Perform a least squares method using Durbin's recurrence equation,




(e) Calculate LPCC mean and variance of unknown words or continuous sound samples using a matrix of ExP = 12 × 12 consisting of LPCCs of unknown words or continuous sound samples with two samples And storing the mean and variance values in a raw database having the plurality of samples without using samples as claimed in claim 1. The step (6) further includes



(E) using a simple Bayesian classifier to match the known words in the permanent database with all unknown words or continuous sounds in the raw database having the plurality of samples;

(G) After calculating the logarithmic value of f (x | ω _i ) and deleting unnecessary constants,
Use Bayesian classifier to display similarity by Bayesian distance,


(H) Each unknown word ω _i , i = 1,. . . , M, calculate the Bayesian distance l (ω _i ) from the known word X to the unknown word ω _i in (g),
(I) N Bayesian distances l (ω i) closest to the known word X in the permanent database to calculate a feature value of the known word in the permanent database, referred to as a standard pattern of known words ), The N unknown words closest to the known word X in the raw database with the plurality of samples comprising the LPCC mean and variance values of the samples around the known word A method for identifying utterances in any language without using a sample characterized in that The step (11) further includes:
(A) For a speech signal and noise within a unit time frame, the sum of absolute distances between two adjacent sample points is calculated, and if the sum of absolute distances between sample points of the speech signal is a noise sample point If the unit time frame is smaller than the sum of the absolute distances between, the unit time frame is a unit time frame without speech signal,
(B) If the unit time frame without the speech signal is longer than the time between two syllables in one word, the boundary line between two unknown words in the one word Locating and dividing the unknown sentence or name into D unknown words on the boundary,
(C) Normalize the waveform of each of the D unknown words with E = 12 stretchable frames without filters and without overlapping, and within each frame, LPC vectors and D = 12 × 12 A LPCC vector that displays an unknown word by a matrix of and including a step of displaying the unknown sentence or name by means of a D = 12 × 12 matrix of LPCC How to identify language utterances. The step (12) further includes:
Find out,
(B) identifying unknown language or utterances without using samples as claimed in claim 1, including the step of displaying unknown sentences or names with a DxF matrix of similar known words belonging to different languages how to. The step (13) further includes:
(A) Select sentences and names that match (D-1), D, and (D + 1) known words in the sentence and name database;
(B) Select the matching known sentence or name with D words, and each of the D known words in the matching sentence or name and the F known words that are the most similar Each of the D columns of
(C) If each sequence of the F similar words that is most similar includes the corresponding words of the matching sentence or name in turn, the matching sentence or name is the unknown sentence or name. Determined that there was
(D) In (c), the number of correctly identified words is not D, or if the matching sentence or name is (D−1) or (D + 1) If it contains known words, then 3xF successive 3 columns of the most similar F words in the DxF matrix of known words to sort each known word in the matching sentence or name Using a sorting window,
To compare with the i th known word in the matching sentence or name, use the (i−1) th, i th, (i + 1) th column of the F most similar known words,
Use the first two columns of the F most similar words that are most similar to compare with the first known word in the matching sentence or name, and 3 × F sorting windows from the first column to the last column Move and
Calculate the number of known words in the matching sentence or name in a 3xF sorting window;
(E) the most likely match, calculated by the number of known words of the matching sentence or name in a 3xF sorting window divided by the total number of words in the matching sentence or name. The method of identifying utterances in any language without using a sample according to claim 1, comprising the step of selecting highly matching sentences or names. The step (14) further includes:
(a) If the unknown sentence or name is not correctly identified, find the word ω of the unknown sentence or name that is not in the most similar F words;
(b) N words closest to the word ω are displayed from the word database by N matrixes of LPCC mean and variance, {μ _IJl , σ ² _IJl }, i = 1 _,. To find N known words for the word ω with a Bayesian distance of
Calculate a weighted average of N matrices,
as well as,
As new feature values, weighted averages, {μ _IJl , σ ² _IJl }, i = 1,..., E, i = 1 _,.
Replacing the standard pattern of the word ω in the word database, and storing the new feature value of the word ω in the word database as a new standard pattern of the word ω,
(C) From the word database, the N values closest to the word ω are displayed by N matrixes of LPCC mean and variance values, {μ _IJl , σ ² _IJl }, i = 1 _,. To find the N known words closest to the word ω with a Bayesian distance of

The sample of claim 1, comprising: replacing the standard pattern of the word ω and storing the new standard pattern of the word ω in the word database as a new standard pattern of the word ω. A method to identify utterances in any language without using it.