JP2006030915A - Method and device for speech recognition - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method and device for speech recognition that decides that an unnecessary word is an incorrect answer with high probability without using a dictionary of unnecessary word models, reduces malfunction, and has high recognition efficiency with a small-capacity memory by replacing an old threshold with a new threshold suitable to change of environment of a speech recognition device for deciding a recognition result outputted through speech recognition processing. <P>SOLUTION: A feature quantity of an inputted speech is analyzed to obtain a score (19) showing the feature quantity and retrieval from a database (21) is carried out; when the value of the recognition score (19) is not close to a threshold already stored in the database (21) (23), a new threshold adapted to the recognition score is calculated (25) to replace the old threshold stored in the database (36). It is decided that an unnecessary word is an incorrect answer with high probability without using a dictionary of unnecessary word models and malfunction is reduced to provide the method and device for speech recognition which has high recognition efficiency with the small-capacity memory. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a speech recognition method and apparatus. Specifically, it is intended to provide a novel method and apparatus for determining an input voice.

    Voice recognition devices have been developed as information transmission means using human generated voice. In such a speech recognition device, an acoustic model and a word dictionary are registered (stored) in advance, and when a person generates a phrase, the phrase of the speech signal is recognized with reference to the registered acoustic model and the word dictionary. I have to.

  In actual speech recognition, there are cases where words that are not related to the content of recognition (hereinafter referred to as “unnecessary words”) such as “motto” and “um” or cough may be included. This unnecessary word is likely to cause misrecognition, and in that case, a person who uses the speech recognition apparatus is requested to re-utter. As a method in consideration of dealing with such unnecessary words, a method has been proposed in which an unnecessary word model is created, registered as a dictionary, referenced in speech recognition, and unnecessary words are removed from the result.

  In addition, in a voice interaction device that performs voice interaction processing, a candidate for determining a voice input from a telephone and a recognition score (score) for the candidate are acquired, and it is determined whether or not the obtained candidate is correct. As a method, a threshold set for each word predicted to be input is stored in the database in advance, and branch judgment is executed by comparing it with the recognition score. A method for determining an appropriate response to a user has been proposed.

In the method of determining the appropriate response to the user,
1. Reference number for each word
2. Each word
3. Stores fixed data including thresholds for each word. The stored contents are input in advance and are not rewritten during the operation of the apparatus.

  However, in the method of removing unnecessary words, it is necessary to create a model of unnecessary words according to any environment in advance, which is impossible in practice. On the other hand, in the method of determining an appropriate response to the user, the threshold setting method is important. For example, the threshold set in a certain environment (time, installation location, user, etc.) When using in other environments, the effects of environmental changes are inevitable. Furthermore, in order to determine a threshold value suitable for each environment of the speech recognition apparatus, enormous data for each environment and a time for creating it are required. In addition, the threshold set by summing up the recognition scores of all words registered in the recognition dictionary covers all the variations in recognition scores for each word. As a result, there is a problem that the recognition accuracy deteriorates.

  Patent Document 1 discloses a speech recognition apparatus that recognizes an unknown input voice by comparing and collating an unknown input voice with a standard pattern (data base) prepared in advance, and has the highest similarity to the unknown input voice. A reference similarity of the pattern is generated for each standard pattern, and it is determined based on the reference similarity whether or not to use the standard pattern selected in recognition of unknown input speech as a recognition result.

  In Patent Document 2, standard information is stored in a data base. In this case, the standard information includes, for every word, a feature pattern and a continuous duration obtained based on the standard speech interval time of each word. The first similarity is obtained. When the maximum similarity is obtained at the second time for the second word or the combined word, the sending means 2 at the time when the word having the maximum similarity from the second time goes up by the second continuous long time. When the corresponding determination time is later than the first time, both the first similarity and the maximum similarity are transmitted.

In both Patent Document 1 and Patent Document 2, a standard pattern prepared in advance is registered and stored in a data base, and the similarity of the input speech with these standard patterns is recognized. In this regard, Patent Document 1 and Patent Document 2 belong to a method for determining an appropriate response to the above-described user.
JP-A-9-152888 Japanese Unexamined Patent Publication No. 9-16995

  In the conventional technology, an unnecessary word model corresponding to every environment must be created in advance, and there is a problem that it is impossible to solve in reality. Furthermore, it is unavoidable that the influence of environmental changes is unavoidable, and in order to determine a threshold value suitable for each environment of the speech recognition apparatus, a huge amount of data and time are required, and the recognition accuracy deteriorates. There was a problem that was not possible.

  The present invention replaces a new threshold value suitable for a change in the environment of the speech recognition apparatus with an old threshold value in the determination of the recognition result output in the speech recognition processing, thereby avoiding the use of an unnecessary word model dictionary. Another object of the present invention is to provide a speech recognition method and apparatus that recognizes an unnecessary word as an incorrect answer with high probability, reduces malfunctions, and has high recognition efficiency with a small memory.

  In order to determine the input speech, the feature amount is analyzed by referring to the acoustic model and the word dictionary, and the likelihood indicating the feature amount is obtained as a recognition score, and the feature amount of many words stored in the word dictionary And a database containing the thresholds corresponding to them, scoring reliability for at least one feature data to obtain a recognition score, and obtaining feature data with a recognition score as a search result, It is determined whether the answer is correct or incorrect from the threshold value of the search result, and the correct recognition score is calculated by calculating the average and variance of the correct recognition scores for each predetermined short period. If it is not close to the threshold already stored in the database, a new threshold that adapts to this correct recognition score is calculated to the old threshold stored in the database. Replace .

  According to the present invention, for each of many words stored in the word dictionary, a new threshold value that is optimal for various changing environments is always calculated, and the old and new stored in the database as necessary. Since threshold values are replaced with threshold values, the threshold values stored in the database are always optimal for the environment, input speech is correctly determined regardless of environmental changes, and unnecessary words have a high probability. Therefore, it is possible to discriminate incorrect answers, thereby reducing malfunctions and enabling efficient speech recognition with a data base having a small memory capacity.

  In order to determine the input speech, the feature amount is analyzed by referring to the acoustic model and the word dictionary, and the likelihood indicating the feature amount is obtained as a recognition score, and the feature amount of many words stored in the word dictionary And a data base storing threshold values corresponding to them, the reliability is scored for at least one feature data as a recognition score, and feature data with a recognition score is obtained as a search result. A correct answer recognition score is created by determining whether the answer is correct or incorrect from the recognition score of the search result, and calculating the average and variance of the recognition scores taken as correct answers every predetermined short period.

  If the value of the correct recognition score is not close to the threshold already stored in the database, a new threshold that adapts to the correct recognition score is calculated and stored in the database. Replace with the old threshold value. For many words contained in the word dictionary, always calculate a new threshold that best suits the changing environment and replace it with the old threshold stored in the database as needed. Therefore, the threshold value stored in the database is always optimal for the environment, and the input speech can be determined correctly regardless of environmental changes, and unnecessary words can be determined as incorrect answers with high probability. A speech recognition method and device could be realized.

  FIG. 1 is a circuit configuration diagram showing an embodiment of the present invention. A recognition processing unit 10 generally called a recognition engine includes an acoustic model 11, a word dictionary 12, and a feature amount analyzer 13. When the audio signal 18 is input to the feature quantity analyzer 13, a spectral feature quantity (feature vector) is extracted. The feature vector of the speech signal 18 is checked against the acoustic model 11 and the word dictionary 12, and a plurality of acoustic models are selected in order from the one with the highest likelihood (yudo) indicating the possibility of matching, A recognition score x (0 ≦ x ≦ 1) indicating each likelihood value is attached to each selected acoustic model and output as feature data 19.

  The data base 21 stores threshold values s that can be rewritten corresponding to each word stored in the word dictionary 12. The data search unit 22 acquires, as search information 31, the threshold value s of the word to which the acoustic model closest to the applied feature data 19 belongs, and outputs it as a search result 32. The threshold value determination unit 23 receives the feature data 19 including the recognition score x and the search result 32 representing the threshold value s, and compares them, so that the correct value is obtained from the recognition score x and the threshold value s corresponding to each word. Judge whether it is recognized as a thing. When the recognition score x is larger than the threshold value s, it is recognized as being correct with a high probability, and the recognition score x is output as the correct answer recognition score 33. If the recognition score x is smaller than the threshold value s, the incorrect answer recognition score 34 is output as an incorrect one. When the incorrect answer recognition score 34 is output, the user is requested to input voice again, or the subsequent operation is stopped.

  The post-processing unit 24 receives the correct answer recognition score 33 and executes a voice recognition application for providing information that the user wants to know in order to proceed to the next step. In the voice recognition application, for example, information prompting the user to input voice is provided by screen output, voice guidance, or a responder (operator). For screen output and voice guidance, it is necessary to transfer necessary information to a display device or voice device. In addition, in order for the operator to respond, it is necessary to make a call to the operator or transfer information.

  When the necessary post-processing is completed, the post-processing unit 24 outputs the correct answer recognition score 33 as the recognition score 35. The threshold calculation processing unit 25 receives the recognition score 35, calculates the average and variance of the correct recognition scores x corresponding to each word every short period, and the new threshold value obtained as a result is the current database 21. If it is not close to the threshold value in use, the new threshold value is output as threshold update data 36, and the threshold value of the data base 21 is updated. By updating the threshold value, the database 21 is adapted to the new environment, so that the speech recognition rate for the same word in the subsequent speech input is increased.

  FIG. 2 is an internal configuration diagram of the threshold value calculation processing unit 25 that plays an important role among the components shown in FIG. The threshold calculation processing unit 25 includes a threshold calculation pre-processor 41, a frequency determination unit 42, a threshold calculator 43, and a threshold update unit 44. The threshold calculation pre-processor 41 receives the recognition score 35 (value is x) determined to be correct, and uses this recognition score x to calculate the average and variance necessary for threshold calculation for each word. Pre-processing for calculation is performed and a pre-processing result 46 is output. The number determination unit 42 receives the preprocessing result 46, determines whether or not the threshold calculation preprocessing has been executed a preset number of times, and outputs the number determination result 47.

  Upon receipt of the number determination result 47, the threshold calculator 43 calculates the average value, variance, and standard deviation of the correct recognition score x using the preprocessing result 46, and obtains the threshold calculation result 48. The threshold updater 44 receives the threshold calculation result 48 and determines that the difference between the previous standard deviation and the current standard deviation (new and old standard deviation value difference) is stable. Is larger than the margin α, the average value obtained by the threshold calculator 43 and the subtraction value of the standard deviation are stored in the data base 21 as the new threshold s, and the difference between the old and new standard deviation values is smaller than the margin α. In this case, it is determined that the difference between the old and new standard deviation values has converged, and the update of the new threshold value s in the database 21 is stopped.

The calculation process in the threshold value calculation preprocessor 41 will be described in detail. The threshold calculation preprocessor 41 receives the recognition score 35 determined to be correct and performs preprocessing for calculating the average and variance necessary for threshold calculation for each word using the recognition score x. . An average added value E _TN (x) ′ necessary for calculating the average value and an average square added value E _TN (x ² ) ′ required for calculating the variance are calculated. First, the weight m is added to the recognition score x determined to be the correct recognition score 33, and a predetermined number of times N are added to determine and store the average value E _N (x) of equation (1). Here, the weight m is Σ _n = 1, and this Σ represents the accumulation of n = 1 to N.
E _N (x) = Σm _n x _n (1)
However, Σ in the formula (1) represents the accumulation of n = 1 to N.

Therefore, the weighted average added value E _TN (x) ′ of the N times × T times (N and T are predetermined integers) recognition scores x is calculated using Equation (2).
E _TN (x) ′ = Σm _n x _n (2)
However, Σ in equation (2) represents the sum of n = (T−1) N + 1 to TN (N and T are predetermined integers).
On the other hand, the variance σx ² is calculated using equation (3) using the nature of the variance.
σx ² = E (x ² ) −E (x) ² (3)

Further, the threshold calculation pre-processor 41 calculates an average square addition value E _TN (x ² ) ′ of the recognition score x. For this purpose, the recognition score x is squared using formula (4), weight p is added, and N times are added. Here, the weight p is Σp _n = 1, and Σ represents the sum of _n = 1 to N.
E _N (x ² ) = Σp _n x _n ² (4)
However, Σ in equation (4) represents a sum of n = 1 to N.
Therefore, the weighted average square addition value E _TN (x ² ) ′ of the N times × T times (N and T are predetermined integers) recognition scores x is calculated using Equation (5).
E _TN (x ² ) ′ = Σp _n x _n ² (5)
However, Σ in equation (5) represents the sum of n = (T−1) N + 1 to TN (N and T are predetermined integers).

The number determination unit 42 determines whether or not the number of executions of the threshold value calculation preprocessing unit 41 has been reached, that is, N times and T times. The threshold calculator 43 receives the preprocessing result 46 and the frequency determination result 47, and every N times, the average E _TN (x) of the equation (6), the variance σ _TN x ² of the equation (8), and the equation (9) The standard deviation σ _TN x is obtained, and the threshold value s _TN ′ of equation (10) is calculated. The average E _TN (x) of the equation (6) up to N times T times (N and T are predetermined integers) is the average addition value E _TN of the equation (2) calculated by the threshold calculation pre-processor 41. Calculate using (x) ′.
E _TN (x) = (1 / T) {(T-1) E _{(T-1) N} (x) + E _TN (x) '} (6)

The variance σ _TN x ² up to N times T times (N and T are predetermined integers) is the weighted mean square addition value E _TN (x ² ) 'is used to calculate an average square addition value E _TN (x ² ) by N times by the number N using Equation (7). From the calculation results of Equation (6) and Equation (7), Obtained using 8).
^{_{E TN (x 2) = (}} 1 / T) {(T-1) E (T-1) N (x 2) + E TN (x 2) '} (7)
σ _TN x ² = E _TN (x ² ) −E _TN (x) ² (8)

The standard deviation σ _TN x up to N × T times (N and T are predetermined integers) is calculated by Equation (9) taking the square root of Equation (8).
σ _TN x = (σ _TN x ² ) ^1/2 (9)
The threshold value s _TN ′ is a value of Expression (10) obtained by subtracting the standard deviation λσ _TN x from the average E _TN (x) obtained by Expression (6). Here, λ is a value set in advance according to the speech recognition apparatus.
s _TN '= E _TN (x) -λσ _TN x (10)

The threshold updater 44 determines whether or not the threshold value s _TN ′ calculated by the threshold value calculator 43 should be replaced with the threshold value s stored in the data base 21 by the formula ( 11). Here, a margin α for determining that the fluctuation of σ is stable and converged is used. That is, when the absolute value of the difference between the previous standard deviation σ _(T−1) (x) and the standard deviation σ _TN x obtained by the equation (9) is larger than the value of the margin α (equation (11)) , Replace the threshold value s stored in the database 21 with the threshold value s _TN ′ obtained this time (formula (12)).
| σ _(T-1) (x) -σ _TN x | ≧ α (11)
s = s _TN '(12)
If the absolute value of the difference between the previous standard deviation σ _(T-1) (x) and the standard deviation σ _TN x is smaller than the margin α (equation (13)), the standard deviation σ converges. The threshold value adaptation (update) is stopped.
| σ _(T-1) (x) -σ _TN x | <α (13)

Database 21 shows how to determine the appropriate response to users of background technology
1. Reference number for each word
2. Each word
3. Stores rewritable data including thresholds for each word. Here, what is different from the background art is that the data in 3. is rewritable data that was previously fixed. To make it rewritable
4. Number of correct answers for each word
5. Average addition value E _TN (x) '(Formula (2))
6. Mean square addition value E _TN (x ² ) '(Formula (5))
Are stored in the data base 21, and when the threshold value s is rewritten, these values are simultaneously updated.

3 and 4 are flowcharts showing the flow of the threshold value s determining operation in the threshold value calculation processing unit 25 of FIGS. When the threshold value calculation process is started, the average value E _N (x) of the equation (1) is obtained and stored, and N times × T times (T = 0, 1, 2,...) The weighted average addition value E _TN (x) ′ of the recognition score x is calculated using the equation (2) (S1, FIG. 3). Therefore, the variance sigma] x ² calculated by Equation (3), followed by Formula (4) to calculate a weighted average sum of x ² E _{N (x} ^2), the square of the average which was based on The added value E _TN (x ² ) ′ is calculated using equation (5) (S2).

Therefore, n is incremented to n = n + 1 (S3). As a result of the increment, it is determined whether or not n exceeds a predetermined number N (S4). If N does not exceed N (N> n) (S4N), the process returns to step S1 and the operations of S1 to S4 are repeated. When N is exceeded (N ≦ n) (S4Y), the average E _TN (x) of equation (6) is calculated (S5). Next, an average square addition value E _TN (x ² ) is calculated by the equation (7) (S6), and a variance σ _TN x ² up to N × T times is calculated by using the equation (8) ( S7). Therefore, the standard deviation σ _TN x is calculated by the equation (9) obtained by taking the square root of the equation (8) (S8, FIG. 4).

Using the margin α for determining that the fluctuation of σ is stable and converged, the previous standard deviation σ _(T−1) (x) and the standard deviation σ _TN x obtained by the equation (9) If the absolute value of the difference is larger than the value of the margin α (equation (11), S9Y), an equation to replace the threshold value s stored in the data base 21 with the threshold value s _TN ′ obtained this time. The threshold value s _TN ′ is calculated from (10) (S10). Therefore, the threshold value s stored in the database 21 is replaced with the threshold value s _TN ′ obtained this time (formulas (12), S11). n = 0, E _TN (x) ′ = 0, E _TN (x ² ) ′ = 0 is reset (S12), T is incremented to T + 1 (S13), and the operation returns to step S1. This operation ends when T reaches a predetermined value.

  5 to 8 are flowcharts showing the overall operation flow in the information guidance system including the voice recognition apparatus of FIG. The illustrated information guidance system is installed in a place where information is provided, has guidance by pressing a button, recognizes a user's utterance content via a microphone, and uses the obtained recognition result to create a database 21. The target information data file is obtained from the server, and after confirming with the user whether or not the recognition result is correct, a service for providing voice guidance of the target information is provided.

  At the start of information guidance by voice recognition, the information guidance button is pressed (S51, FIG. 5). The information guidance guidance is output (S52), and the recognition signal A is input from the microphone by the user while the recognition processing unit 10 refers to the acoustic model 11 and the word dictionary 12 to recognize the first candidate and the second candidate recognition result A. And B and the respective recognition scores x1 and x2 are acquired (S53). The recognition result A and B obtained as the feature data 19 and the respective recognition scores x1 and x2 are obtained by searching the data base 21 and closest to them, and the information data file A ′ of the first candidate A And the threshold value s1, the information data file B 'of the second candidate B, and the threshold value s2 are extracted (S54).

  If the recognition score x1 of the first candidate A is less than or equal to the threshold s1 of the first candidate A (x1 ≦ s1) (S55N, FIG. 6), the incorrect answer flag is set to TRUE (S56TRUE) and the recognition was not possible (error (Recognition) is notified to the user and re-input guidance is provided (S57), the incorrect answer flag is set to TRUE, and the process returns to the recognition process in Step S53 (FIG. 5) (S58). When the recognition score x1 of the first candidate A is greater than or equal to the threshold value s1 of the first candidate A (x1 ≧ s1) (S55Y, FIG. 6), it is determined that the word “A” uttered by the user is correct with a high probability. The confirmation guidance is output (S59), and the recognized threshold value s1 is acquired as the confirmation processing result (S60).

  If the recognized confirmation processing result is “NG” (S61NG), if there is a second candidate and the determination flag = FALSE (S62Y, FIG. 7), the same as in steps S55 to S58 for the first candidate, When the recognition score x2 of the two candidates B is equal to or less than the threshold s2 of the second candidate B (x2 ≧ s2) (S63Y), a confirmation guidance for “B” (recognized correctly) is notified to the user (S64) The determination flag is set to TRUE, the fact that it has been confirmed is displayed, and the process returns to step S60 (S65). If the confirmation result in step S61 is “OK” (S61OK), the user is notified of the confirmation guidance for “A” (recognized correctly) (S66), the confirmation flag is set to TRUE (S67), and step S70 (FIG. Go to 8).

  If the incorrect answer flag is FALSE in step S56 (S56FALSE), if the judgment flag is FALSE in step 62 (S62N), or if x2 <s2 in step 63 (S63N), it is not recognized (incorrect recognition). The user apologizes and forwards the call to the operator automatically. The operator responds to the user and provides the target information verbally (S68, FIG. 7). According to the guidance, the voice input from the user to the voice recognition device is transferred to the operator (S69), and the process proceeds to step S70. Therefore, both the judgment flag and the incorrect answer flag are set to FALSE (S70), and if the confirmation flag is set to TRUE (S71Y), the threshold value calculation process shown in FIGS. 3 and 4 is performed and the process returns to step S51 ( S72). Even when the confirmation flag is not TRUE (S71Y), the process returns to step S51 (S71N).

  As is clear from the above description, for each of the many words contained in the word dictionary, a new threshold value that is optimal for various environments is constantly calculated and stored in the database as needed. Since the threshold value stored in the database is always adapted to the optimum state for the environment, the input sound is correctly determined regardless of the environment change, and is unnecessary. Since words can be identified as incorrect answers with high probability, malfunctions are reduced, and efficient speech recognition is possible with a data base having a small memory capacity.

It is the circuit block diagram which showed the Example of this invention. (Example 1) FIG. 2 is an internal configuration diagram of a threshold value calculation processing unit, which is a component of the circuit configuration of FIG. 3 is a flowchart showing a flow of operations of a threshold value calculation processing unit in FIG. FIG. 3 is a flowchart showing the operation flow of the threshold value calculation processing unit in FIG. 2 together with FIG. FIG. 2 is a flowchart showing a flow of overall operation in an information guidance system including a voice recognition device having the circuit configuration of FIG. 6 is a flowchart showing the overall operation flow in the information guidance system including the voice recognition device having the circuit configuration of FIG. 1 together with FIG. 7 is a flowchart showing an overall operation flow in the information guidance system including the voice recognition device having the circuit configuration of FIG. 1 together with FIGS. 5 and 6. FIG. 8 is a flowchart showing an overall operation flow in the information guidance system including the voice recognition device having the circuit configuration of FIG. 1 together with FIGS.

Explanation of symbols

10 Recognition processing section
11 Acoustic model
12 word dictionary
13 Feature analyzer
18 Audio signal
19 Feature data
21 Database
22 Data search part
23 Threshold judgment part
24 Post-processing section
25 Threshold calculation processing section
31 Search information
32 Results
33 Correct answer score
34 Incorrect answer recognition score
35 recognition score
36 Threshold update data
41 Threshold calculation pre-processor
42 Count detector
43 Threshold calculator
44 Threshold updater
46 Pretreatment results
47 Judgment result
48 Threshold calculation results

Claims (5)

To recognize the input speech signal (18), refer to the acoustic model (11) and the word dictionary (12), analyze the feature value, and obtain feature data (19) using the feature value as the recognition score Recognition processing (10)
The database (21) storing the feature values of many words stored in the word dictionary (12) and the threshold values corresponding to them is searched to trust at least one feature data (19). A data search process (22) for scoring the degree and obtaining feature data with a recognition score as a search result (32),
A threshold value determination process (23) for determining whether the search result (32) should be a correct answer recognition score (33) or an incorrect answer recognition score (34) from the threshold value;
Receiving the correct recognition score (33) and performing a post-processing (24) to obtain a post-processed recognition score (35) by performing the necessary post-processing,
A new recognition score is created by calculating the average and variance of the post-processed recognition score (35) every predetermined short period, and the value of this new recognition score is already stored in the database (21). If the threshold value is not close to the threshold value, a threshold value is calculated to replace the old threshold value stored in the database (21) by calculating a new threshold value adapted to the recognition score. Voice recognition method that performs value calculation processing (25). The threshold value calculation process (25)
A threshold calculation pre-processing (41) for receiving the post-processed recognition score (35) and performing a pre-processing calculation necessary for calculating the new threshold to obtain a pre-processing result (46);
Number of times determination processing (42) for obtaining a number of times determination result (47) when the number of calculations in the threshold calculation preprocessing (41) reaches a predetermined value;
In response to the number determination result (47) and the preprocessing result (46), a threshold value calculation process (43) for obtaining the new threshold value as a threshold value calculation result (48),
If the threshold calculation result (48) does not approximate the threshold stored in the data base (21), the threshold calculation result (48) is used as threshold update data ( The speech recognition method according to claim 1, further comprising: a threshold value updating process (44) for replacing the old threshold value stored in the data base (21) as 36). In the threshold calculation pre-processing (41),
From the post-processed recognition score (35), an average value (E _N (x): Expression (1)) and a weighted average addition value (E _TN (x) ′) for a predetermined number N of recognition scores (35). : Equation (2)), variance (σx ² : Equation (3)), weighted square addition value (E _TN (x ² ): Equation (4)), and the N recognition scores (35) A weighted average square addition value (E _TN (x ² ) ′: Formula (5)) is calculated over a predetermined number of times T, and is used as the preprocessing result (46).
In the threshold value calculation process (43),
From the pre-processing results (46) up to N times T times, the average (E _TN (x): Formula (6)), the mean square addition value (E _TN (x ² ): Formula (7)), The variance (σ _TN x ² : Formula (8)), standard deviation (σ _TN x: Formula (9)), and threshold value (s _TN ′: Formula (10)) are obtained, and this threshold value ( The speech recognition method according to claim 2, wherein s _TN ': Expression (10) is output as the threshold calculation result (48). To recognize the input speech signal (18), refer to the acoustic model (11) and the word dictionary (12), analyze the feature value, and obtain the feature data (19) using the feature value as the recognition score Recognizing means (10),
The database (21) storing the feature values of many words stored in the word dictionary (12) and the threshold values corresponding to them is searched to trust at least one feature data (19). A data search means (22) for scoring the degree and obtaining feature data with a recognition score as a search result (32),
Threshold determination means (23) for determining whether the search result (32) should be a correct answer recognition score (33) or an incorrect answer recognition score (34) from the threshold;
A post-processing means (24) for obtaining a post-processed recognition score (35) by receiving the correct answer recognition score (33) and performing post-processing as required;
A new recognition score is created by calculating the average and variance of the post-processed recognition score (35) every predetermined short period, and the value of this new recognition score is already stored in the database (21). If the threshold value is not close to the threshold value, a threshold value is calculated to replace the old threshold value stored in the database (21) by calculating a new threshold value adapted to the recognition score. A speech recognition device including a value calculation processing means (25). The threshold value calculation processing means (25)
Threshold calculation pre-processing means (41) for receiving the post-processed recognition score (35), performing pre-processing calculation necessary for calculation of the new threshold value, and obtaining a pre-processing result (46); ,
Count determination means (42) for obtaining a count determination result (47) when the number of calculations in the threshold calculation preprocessing (41) reaches a predetermined value;
Threshold calculation means (43) for receiving the number determination result (47) and the pre-processing result (46) and obtaining the new threshold value as a threshold calculation result (48);
If the threshold calculation result (48) does not approximate the threshold stored in the data base (21), the threshold calculation result (48) is used as threshold update data ( 5. The speech recognition apparatus according to claim 4, further comprising threshold update means (44) for replacing with an old threshold stored in the data base (21) as 36).

Images