JPH09212191A - Speech recognition system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the efficiency of the system by increasing the precision of key word extraction, narrowing down the range of collation and making the collation fast, and reducing the storage capacity that standard feature patterns original require. SOLUTION: A prediction type syntax analysis part 9 consists of a prediction part 9e which predicts termination symbols in order, a grammar part 9a which stores a context free grammar, a chart part 9c which stores the termination symbols and nontermination symbols, an operation table part 9d which stores an LR table and information needed to control the process procedure of recognition according to the table, a state list part 9b which stores the predicted termination symbols by position numbers indicating positions in an input speech, and an analysis part 9f which takes an analysis according to analysis algorithm. The prediction type syntax analysis part 9 selects standard feature patterns corresponding to the predicted termination symbols out of a standard feature pattern storage part 4, and a collation part 5 collates the input speech feature pattern with the selected standard feature patterns to extract the most likelihood word.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition system, and more particularly to a voice recognition system suitable for extracting and recognizing important words in continuous voice.

[0002]

2. Description of the Related Art In the conventional voice recognition method, as a problem in extracting and recognizing a keyword (important word) from an input voice, an unregistered word portion is extracted as one of registered words at the time of utterance. It may spring out, or the registered word part may drop out without being extracted as a registered word.

First, the problem of springing out will be described with an example. For example, suppose the word you want to extract has / red square /. There are various ways of speaking by the speaker, such as / red square /
Some people say / red square / (_: some silence)
Some people say. If / red / and / square / are registered here, the recognition system may return the correct result in the case of the former utterance. However, in the latter case, there is a special unregistered word called silence, so /
After extracting the word red /, it is possible to accidentally extract other words other than the / square / that are similar to the silent feature pattern. To deal with such a problem, Japanese Patent Laid-Open No. 6-2
As disclosed in Japanese Patent No. 66386, there is a method using a garbage model. In this method, a model corresponding to a word other than the registered word is created, and the vocalized portion other than the keyword is absorbed by the model, thereby suppressing the springing.

Next, the problem of dropout will be described with an example. For example, suppose that the word / red_square / is registered and you say / red_square /. In this case, the characteristic pattern of the uttered silent portion is different from the registered pattern, and therefore the character may be dropped without being recognized. It can be dealt with by additionally registering a standard feature pattern corresponding to / red_square /, but since there are many such combinations, the amount of memory required to register the pattern increases. To address this issue, “Proceedings of the Acoustical Society of Japan, pp.141-142, (1991-0)
3) ”, there is a method of creating a keyword lattice. In this method, first, every matching section in the input speech feature pattern is compared with the standard feature pattern,
Generate a plurality of keyword candidates, and then use a parser to recognize the most likely candidate as a recognition result from the resulting lattice using a parser.

[0005]

Among the conventional speech recognition methods described above, in the method using the garbage model, since the models other than the registered words are treated as one model, the words to be extracted are also attracted to the garbage model. , May be absorbed. Therefore, model parameters must be carefully controlled. It is also possible to increase the number of garbage models in order to avoid unnecessary absorption, but in this case, the memory capacity of the models increases.

Further, in the method using the keyword lattice, since the lattice structure is created, the number of collations with the standard feature pattern increases and the amount of calculation increases. In addition, the lattice is an intermediate output for obtaining the final result, and since a parser must be used, the system efficiency is low.

The present invention has been made in view of the above circumstances, and a predictive parser having a simple mechanism and a small memory capacity is directly used for standard pattern selection to narrow the collation range of speech recognition. However, the object of the present invention is to provide a speech recognition method that speeds up the matching and enables highly accurate word extraction.

[0008]

According to a first aspect of the present invention, there is provided means for creating a characteristic pattern of an input voice, and a predictive syntax for sequentially predicting each terminal symbol corresponding to one or more words appearing in the voice. A voice including an analyzing means and a means for storing a standard feature pattern corresponding to the terminal symbol, and comparing the standard feature pattern with a feature pattern of an input voice to extract a word uttered in the input voice. In the recognition method, the predictive-type syntactic analysis unit stores a prediction unit that sequentially predicts terminal symbols, a grammar unit that stores a context-free grammar, a chart unit that stores terminal symbols or nonterminal symbols, and an operation that stores an LR table. It has a table part and a state list part for storing predicted terminal symbols for each position number, and compares the standard feature pattern corresponding to the predicted terminal symbol with the input voice feature pattern. Characterized in that so as to extract a word seems also have to speed matching to narrow the scope of the verification, in which so as to improve the efficiency of the system.

According to a second aspect of the present invention, in the voice recognition system according to the first aspect, the score storage unit is provided, and the standard feature pattern and the input voice feature pattern are collated each time a terminal symbol is predicted. It is characterized in that the plausible words are extracted in order while storing the accumulated score at the time of matching in the score storage unit, and thus the accuracy of the system is improved by digitizing the criterion of the word extraction. It is intended to improve.

According to a third aspect of the present invention, in the voice recognition system according to the second aspect, a position storage unit is provided, and the collation start position of the standard feature pattern corresponding to the subsequent terminal symbol is the start end of the input voice feature pattern or later. And set the collation start position to be near the end position of the word section extracted by collation of the standard feature pattern and the input feature pattern corresponding to the terminal symbol, thereby accurately determining the collation position. , The range of collation is narrowed, so that the collation is speeded up and the efficiency of the system is improved.

According to a fourth aspect of the present invention, in the voice recognition system according to the second or third aspect, a pruning unit is provided, and the standard feature pattern and the input voice feature pattern are collated each time a terminal symbol is predicted. The number of plausible candidates for the cumulative score at the time of matching is left in the pruning unit to trim the remaining candidates. Is to speed up the system and improve the efficiency of the system.

According to a fifth aspect of the present invention, in the speech recognition system according to any one of the first to fourth aspects, the standard feature pattern corresponding to the predicted terminal symbol has a unit pattern smaller than the terminal symbol unit. It is characterized in that the standard feature patterns are created by connecting them, so that the storage amount originally required by the standard feature pattern is reduced and the efficiency of the system is improved.

According to a sixth aspect of the present invention, in the speech recognition system according to the fifth aspect, the standard feature pattern corresponding to the predicted terminal symbol has a unit smaller than the terminal symbol unit.
In addition, it is characterized in that the characteristic patterns depending on the acoustic environment of the preceding and succeeding ones are connected and created, so that the memory amount originally required by the standard characteristic pattern is reduced,
In addition, the accuracy of collation is improved and the efficiency of the system is improved.

According to a seventh aspect of the present invention, in the voice recognition system according to the fifth or sixth aspect, a standard feature pattern corresponding to a terminal symbol is created in advance with a feature pattern having a unit smaller than the unit of the terminal symbol and collation is performed. The feature is that sometimes the standard feature pattern is selected by referring to the terminal symbol number, and the input feature pattern is matched, so that the memory amount originally required by the standard feature pattern is saved, and in addition, the matching procedure is performed. Is simplified to enable high-speed matching and improve the efficiency of the system.

[0015]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a schematic block diagram of an embodiment of the present invention. In FIG. 1, the input voice 1 is described in detail by Niimi, "Voice Recognition", Kyoritsu Shuppan (1979) and the like. The feature pattern creating unit 2 converts the feature pattern into a feature pattern by the analysis method as described above. Here, the 10-dimensional cepstrum parameter, the power, and the first-order difference in the time direction of the power are extracted and used as the feature pattern. However, as the analysis conditions, the sampling frequency is 16 kHz, the windowing by the Hamming window (window width 16 ms), and the linear prediction coefficient analysis order 14 are used. The shift width per frame is 10 msec. The analysis method is not limited to the above method, and any acoustic analysis method such as frequency analysis may be used.

Next, a standard feature pattern to be compared with the input voice feature pattern is selected in association with the terminal symbol predicted by the predictive syntax analysis unit 9. FIG. 2 is a block diagram of the predictive syntax analysis unit. The grammar shown in Table 1 is stored in the grammar unit 9a. Terminal symbols start with '*' and all other symbols are non-terminal symbols. Table 1 is briefly described to show the embodiment, but the grammar section 9a can make a more complicated description if the notation is based on the context-free grammar.

[0017]

[Table 1]

Next, the action table obtained from the grammar is used as the action table section.
It is stored in 9d in the form shown in Table 2. The contents of Table 2 are AV Aho et al., “Compilers-Principles, Te
chniques, and Tools ", Addison-Wesley (1986), which is the same as the LR analysis table, which is composed of an ACTION part and a GOTO part. Actions include state transition, grammar application, acceptance, There are four types of error: In the ACTION part, the one that starts with's 'is the state transition action, and the one that starts with'r' is the grammar application action. In Table 2, in the ACTION part state 0
The state transition behaviors of [s3] and [s4] are written in / * red / and / * blue / respectively, but after predicting / * red /, go to state 3 and / * blue / It indicates that after the prediction, the process proceeds to the state 4. In state 5, '* $' is given the action of applying the grammar [r1]. This indicates that the first grammar in Table 1 applies if it is the end of the symbol. '* $' Is set to the last position of the terminal symbol string predicted by the terminal symbol representing the end.

[0019]

[Table 2]

Next, in the analyzing unit 9f and the predicting unit 9e, the predicted terminal symbols are taken out one by one from the head, and Table 3 and Table 4 are taken.
Then, the algorithm shown in Table 5 is applied, and the result is recorded in the chart section 9c and the status list section 9b in the form shown in Table 6. The sentence will be correct if the acceptance operation is finally performed. Further, all the phrase structures defined as correct in the grammar are recorded in the chart portion 9c.

[0021]

[Table 3]

[0022]

[Table 4]

[0023]

[Table 5]

The process flows (1) to (9) will be described below with reference to Tables 3, 4, and 5. (1) PARSE: The start state number is stored in the state list of position number 1. The start state number is 0 because it is the beginning of the sentence. Predicting terminal symbols that can be transitioned from this state number is / * red / and / * blue /. Also, the start position number and end position number are 1 and 2, respectively, as shown in the GOTO section of Table 2, so (* red, 1,2) and (* blue, 1 ,,
2) and are memorized. Here, the start position number represents the start position of the terminal symbol. The end position number similarly represents the end position. In this way, ADDARC (* red,
1, 2) and ADDARC (* blue, 1, 2) occur, and the processing proceeds in parallel.

(2) ADDARC (* red, 1,2): (* red, 1,2) is not yet stored in the chart part, so C
Run HECKARC (* red, 1,2). Hereinafter, A
Since the processing after DDARC (* blue, 1, 2) is the same, it is omitted.

(3) CHECKARC (* red, 1, 2): When the status list of position number 1 is taken out, it is [0]. Next, the state transition destination of the terminal symbol / * red / in state 0 is examined from the operation table of Table 2. Since it is known that 3 is the transition destination state number, the list L = [3]. For the state number 3 in the list L, the state transition destination is checked from the operation table of Table 2.
Since there is nothing beginning with's' in the ACTION part of the state number 3 column in Table 2 and there is only reduction operation, state transition is impossible. Therefore, the procedure proceeds to the procedure (4) in Table 5, and the operation is stored as [r4], so RL = 4. When the grammar number 4 in the list RL is taken out from the grammar section of Table 1, (COLO
R, * red). _R'h is empty, except for the top-right elements of the right-hand side
[] Therefore, ADDARC (COLOR, 1,2)
Execute

(4) ADDARC (COLOR, 1,2): Since (COLOR, 1,2) is not yet stored in the chart part, CHECKARC (COLOR, 1,2) is executed.

(5) CHECKARC (COLOR, 1,
2): As shown in the GOTO section of Table 2, it is [0] when the status list of position number 1 is taken out. Next, the state transition destination of the non-terminal symbol / COLOR / in state 0 is checked from the operation table of Table 2. Since it can be seen that 2 is the transition destination number as shown in the GOTO section of Table 2, L = [2]. For state number 2 in L, the state transition destination is checked from the operation table of Table 2.
As shown in the ACTION section of Table 2, since the transition operation is possible in the state 2, the state number 2 is recorded in the state list portion 9b of the position number 2. When the applicable grammar number for the state number 2 of L is checked from the operation table of Table 2, there is nothing that starts with'r 'indicating the applicable grammar number, and there is no reduction operation. Therefore, in the process flow (3) As mentioned, R in procedure (4) of Table 5
L becomes empty []. Since the variable Ret = 1 so far, CHECKARC (COLOR, 1,2) is successfully completed, and ADDARC (COLOR, 1,2) is returned to.

(6) ADDARC (COLOR, 1, 2):
(COLOR, 1, 2) is stored in the chart portion 9c, the process ends successfully, and the process returns to CHECKARC (* red, 1, 2).

(7) CHECKARC (* red, 1, 2):
Since the variable Ret = 1, CHECKARC (* red, 1,
2) Succeeded and finished, ADDARC (* red, 1, 2)
Return to

(8) ADDARC (* red, 1,2): (* red, 1,2) is stored in the chart portion 9c, and the process ends successfully.
Return to PARSE.

(9) PARSE: ADDARC (* red,
Since 1,2) and ADDARC (* blue, 1,2) have been analyzed successfully, the index numbers corresponding to / * red / and / * blue / are sent to the standard feature pattern storage unit 4. When the analysis operation as described above is repeated, finally, as shown in Table 6, 6
One sentence is predicted as a candidate. Keyword extraction is
This can be realized by verifying these prediction sequences with standard feature patterns corresponding to the predicted terminal symbols.

[0033]

[Table 6]

By the operation of the predictive syntax analysis unit 9 described above, the index number of the terminal symbol is sent to the standard feature pattern storage unit 4 in the order of the input voice.
The terminal symbol corresponds to the standard feature pattern stored in advance in the standard feature pattern storage unit 4, and the collation unit 5 collates with the input voice feature pattern.

[0035]

[Equation 1]

In the following description, the symbol in column A of Table 7 is used, but since it is the symbol in column B corresponding to it, it should be read as the symbol in column B.

[0037]

[Table 7]

Here, P _{q (l)} is an index of the standard feature pattern corresponding to the l (lowercase L) (1 ≦ l ≦ L) th in the sequence, and has V standard feature patterns. In the embodiment, W [(s)] in Expression 1 corresponds to the terminal symbol sequence accepted by the predictive syntax analysis unit 9. Table 6
As shown in the example of
In each case, L = 2 since it is composed of two terminal symbols. In this case, the number of standard feature patterns is the terminal symbol.
Since it is equal to the number of (/ * red /, / * blue /, / * △ /, / * □ /, / * ○ /), V = 5. Similarly, the input voice feature pattern X is expressed as follows. _{X = {x 1, ...,} x m, ..., x M} and (Expression 2) subintervals of the verification unit 5, the utterance and X of m ₁ frame of speech feature patterns were up to m ₂ frame , And a score with the standard feature pattern sequence associated with the partial section are obtained by the following Expression 3.

[0039]

[Equation 2]

M ₁ [q (l)] and m ₂ [q (l)] are model q (l)
The start point and the end point are respectively represented by the end points of the extraction section of the word corresponding to. In addition, M [q (l)] = (m ₂ [q (l)] − m ₁ [q
(l)]) is the length of the extraction section. Equation 4 calculates the score for the extraction interval and the matching function r. The matching function is a function that represents a matching path, and can be obtained by well-known dynamic programming or the like. The matching function associates the m-th frame input speech feature amount with the r (m) -th frame standard feature amount. The scores associated in this way are defined as follows.

[0041]

(Equation 3)

H is a positive value used for the threshold operation of the score. According to Equation 5, the larger the score, the more likely the candidate is as a recognition result. However, the method of measuring the score is not limited to this. When the score becomes small like the criterion based on the Euclidean distance, it becomes more likely. May be designed to Every time l (L) in Expression 3 increases, D [(s)] is calculated in the accumulated score storage unit 6. The comparison unit 7 arranges D [(s)] for the current candidates in descending order of score. Then, the pruning unit 8 selects and saves J candidates (for example, 5 candidates) from the candidates with a large score, and the pruning unit 8 prunes the remaining candidates. With respect to the saved candidate, l (ell) is incremented by 1 and the prediction-type parsing unit 9 predicts the next terminal symbol again. This process is repeated, and when all the candidates are predicted by the predictive syntax analysis unit 9, the candidate with the highest score in D [(s)] of Expression 3 is output as the recognition result. In the embodiment of FIG. 1, the matching time can be shortened by introducing the following condition in Expression 4.

[0043]

(Equation 4)

This equation 6 is (l) in the input voice feature pattern.
This indicates that the collation of the next standard feature pattern, that is, the l-th standard feature pattern is started from the frame within the section detected using the (L) -1) th standard feature pattern. m * ₁ [q (l-1)] and m * ₂ [q (l-1)] are the start point and end point of the already detected section. FIG. 5 shows the matching start point when / □ / is extracted after / red / is detected. In the method according to FIG. 5, the verification of / □ / is started from the point'α 'before the end point m * ₂ [q (l)] of / red /.

FIG. 3 is a schematic block diagram showing another embodiment. The phoneme feature pattern storage unit 11 of FIG. 3 stores a standard pattern in units of phonemes smaller than the standard feature pattern storage unit 4 of FIG. Therefore, when the terminal symbol "/ red /" is predicted by the predictive parsing unit 9, the feature patterns "/ a /, / k /, / i /" of the phoneme feature pattern storage unit 11 are linked by the linking unit 12. Create a standard feature pattern called / akai /. The grammar section is created as shown in Table 8 so that the phoneme feature pattern can be selected when the terminal symbol is predicted.
Since the terminal symbol of the grammar section in Table 8 is written as a phoneme symbol, the feature pattern created for each phoneme can be directly selected.

[0046]

[Table 8]

Further, in order to improve the extraction accuracy, it is also possible to use a feature pattern considering the acoustic environment before and after the phoneme. At this time, the phoneme feature pattern storage unit 11 stores phoneme feature patterns such as / -ak /, / aka /, / kai /, and / ai- / in advance in a similar example. For example, the feature pattern of / aka / represents a phoneme of / k / where both the preceding and following phonemes are'a '. The'- 'indicates that any phoneme may precede or follow. When such a feature pattern is used, the grammar shown in Table 9 is used.

[0048]

[Table 9]

FIG. 4 is a schematic block diagram showing another embodiment of the present invention. In the present embodiment, phoneme feature patterns are connected to create a word feature pattern corresponding to a terminal symbol in advance. First, switch S ₂
Is put in A and the switch S ₁ is turned off to create a word feature pattern. The terminal symbol-phoneme symbol correspondence list 13 has Table 10
Data such as is stored. Based on this table 10,
The feature patterns corresponding to the terminal symbols are created by connecting the phoneme feature patterns by the connecting unit 12 and stored in the word feature pattern storage unit 14. The word feature pattern created at this time is given the same number as the terminal symbol number.

[0050]

[Table 10]

Next, switch S ₂ is switched from A to B,
Further, by turning on the switch S ₁ , recognition can be performed. At the time of recognition, the predictive syntax analysis unit 9 predicts a terminal symbol, and the terminal symbol number is used as the word feature pattern storage unit.
If sent to 14, the word feature pattern can be selected and can be matched with the input voice feature pattern. Subsequent processing is the same as that of the above-described embodiment, and will be omitted. In addition,
In the present embodiment, the phoneme symbol and the terminal symbol can be associated with each other in the form as shown in Table 10 even for the characteristic patterns in which the preceding and subsequent phoneme environments are taken into consideration.

[0052]

According to the invention of claim 1, 3, 4 or 7, it is possible to narrow the collation range of the voice recognition and increase the collation speed with a simple mechanism. According to the invention of claim 2, it is possible to quantify the criteria for word extraction and improve the accuracy.
According to the invention of claim 5, the storage amount originally required by the standard feature pattern group can be reduced. According to the invention of claim 6, in addition to the memory saving which is the effect of the invention of claim 5, the accuracy of word extraction can be increased. Thus, according to the present invention, it is possible to improve the speed and accuracy of matching by extracting only the necessary information from the continuous speech that is relatively freely uttered like a spoken word. Can be applied to.

[0053]

[Brief description of drawings]

FIG. 1 is a schematic block diagram of continuous speech recognition according to an embodiment of the present invention.

FIG. 2 is a block diagram showing an embodiment of a predictive syntax analysis unit.

FIG. 3 is a schematic block diagram of another embodiment of the present invention.

FIG. 4 is a schematic block diagram of another embodiment of the present invention.

FIG. 5 is a diagram illustrating an embodiment of claim 3;

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Speech, 2 ... Feature pattern creation part, 4 ... Standard feature pattern storage part, 5 ... Collation part, 6 ... Cumulative score storage part, 7 ... Comparison part, 8 ... Pruning part, 9 ... Predictive syntax analysis part, 9a ... Grammar part, 9b ... State list part, 9c ... Chart part, 9d ... Action table part, 9e ... Prediction part, 9f ... Analysis part, 10 ... Recognition result, 11 ... Phoneme feature pattern storage part, 12 ... Connection part, Thirteen
... Terminal symbol-phoneme symbol correspondence list, 14 ... Word feature pattern storage unit, A ... B changeover switch, B ... A changeover switch, C ... On / Off changeover switch.

─────────────────────────────────────────────────── ───

[Procedure amendment]

[Submission date] March 15, 1996

[Procedure amendment 1]

[Document name to be amended] Statement

[Correction target item name] 0001

[Correction method] Change

[Correction contents]

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition system, and more particularly to a voice recognition system suitable for extracting and recognizing an important word in continuous voice.

[Procedure amendment 2]

[Document name to be amended] Statement

[Correction target item name] 0009

[Correction method] Change

[Correction contents]

According to a second aspect of the present invention, in the voice recognition system according to the first aspect, the score storage unit is provided, and the standard feature pattern and the input voice feature pattern are collated each time a terminal symbol is predicted. performed, the accumulated score in matching point is characterized in that so as to extract in sequence word plausible while stored in the score storage unit, has, matching
It is designed to speed up the time and improve the efficiency of the system.

[Procedure 3]

[Document name to be amended] Statement

[Correction target item name] 0039

[Correction method] Change

[Correction contents]

[0039]

[Equation 2]

[Procedure amendment 4]

[Document name to be amended] Statement

[Correction target item name] 0052

[Correction method] Change

[Correction contents]

[0052]

Claim 1, 2 according to the present invention, the 3, 4 or seventh invention, by narrowing the comparison range of the speech recognition in a simple mechanism, Ru can increase the speed of matching. In accordance with the invention of 請 Motomeko 5,
It is possible to reduce the amount of memory originally required by the standard feature pattern group. According to the invention of claim 6, in addition to the memory saving which is the effect of the invention of claim 5, the accuracy of word extraction can be increased. Thus, according to the present invention, it is possible to improve the speed and accuracy of matching by extracting only the necessary information from the continuous speech that is relatively freely uttered like a spoken word. Can be applied to.

Claims (7)

[Claims] 1. A means for creating a characteristic pattern of an input speech, a predictive syntax analysis means for sequentially predicting each terminal symbol corresponding to one or more words appearing in the speech, and a standard corresponding to the terminal symbol. And means for storing the characteristic pattern,
In the speech recognition method for extracting the words uttered in the input voice by comparing the standard feature pattern and the feature pattern of the input voice, the predictive syntax analysis means,
A prediction unit that sequentially predicts terminal symbols, a grammar unit that stores context-free grammars, a chart unit that stores terminal symbols or non-terminal symbols, an operation table unit that stores LR tables, and position numbers of predicted terminal symbols. A voice recognition method characterized by having a state list section for storing each of them, collating a standard feature pattern corresponding to the predicted terminal symbol with an input voice feature pattern, and extracting a plausible word. 2. The voice recognition method according to claim 1, wherein the standard feature pattern and the input voice feature pattern are collated each time a terminal symbol is predicted, and a cumulative score at the time of collation is stored in a score storage unit. However, the speech recognition method is characterized by extracting plausible words in order. 3. The voice recognition method according to claim 2, further comprising a position storage unit, wherein a collation start position of a standard feature pattern corresponding to a subsequent terminal symbol is set after a start end of an input voice feature pattern, A speech recognition method characterized in that a matching start position is near an end position of a word section extracted by matching a standard feature pattern corresponding to the terminal symbol with an input feature pattern. 4. The voice recognition method according to claim 2, further comprising a pruning unit, wherein the standard feature pattern and the input voice feature pattern are collated each time a terminal symbol is predicted,
A voice recognition method, characterized in that a predetermined number of plausible candidates of cumulative points at the time of matching are left in the pruning unit and the remaining candidates are pruned. 5. The speech recognition method according to claim 1, wherein a standard feature pattern corresponding to a predicted terminal symbol is created by concatenating feature patterns having units smaller than the terminal symbol unit. A voice recognition method characterized by: 6. The speech recognition system according to claim 5, wherein the standard feature pattern corresponding to the predicted terminal symbol has a unit smaller than the unit of the terminal symbol and depends on the preceding and subsequent acoustic environments. A speech recognition method characterized by connecting and creating created characteristic patterns. 7. The voice recognition method according to claim 5, wherein a standard feature pattern corresponding to a terminal symbol is created in advance with a feature pattern having a unit smaller than the unit of the terminal symbol, and the terminal symbol number is set at the time of matching. A speech recognition method characterized by selecting a standard feature pattern with reference to matching with an input feature pattern.