JP2002268679A - Method and device for detecting error of voice recognition result and error detecting program for voice recognition result - Google Patents

Abstract

PROBLEM TO BE SOLVED: To support manual efficient correction by automatically detecting a recognition error which is inevitable when contents represented in a voice language is made into character information by using voice recognition technology. SOLUTION: Results which are outputted from a voice recognition device including recognition errors and sound information corresponding thereto are stored by correct answers, errors and sound models for discriminating whether the results are correct or incorrect, the most of which are generated from the pieces of sound information; and sound information which is newly inputted and an object of recognition is compared with those models to decide whether the recognition results are correct or incorrect.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and an apparatus for detecting an error in a speech recognition result and a program for detecting an error in a speech recognition result, and more particularly to extracting correct / incorrect information from past recognition results and discriminating between correct and wrong. In a method for creating a correct / incorrect discrimination model to be used and a recognition error correction support device that points out recognition errors using the correct / incorrect discrimination model, subtitles for news audio in broadcasting, subtitle services for hearing-impaired persons in lectures, magazines and the like The present invention relates to a method and an apparatus for detecting an error in a speech recognition result and a program for detecting an error in a speech recognition result for promoting the efficiency of proofreading work for newspaper inter-articles.

[0002]

2. Description of the Related Art Recently, a speech recognition technique for supporting the transcription of speech has been actively introduced, but at present, there is no speech recognition apparatus having a recognition rate of 100%. In particular, for unspecified speakers and those for continuous speech recognition, there is a very high possibility that the recognition result contains an error.
Therefore, in the present situation, when the recognition result converted using the speech recognition device is actually used for some purpose,
Some corrective measures are needed.

As a correction means, it is a common practice that a corrector listens to a voice to be recognized, reviews the result of a character string from the voice, and finds an error.

[0004] As described above, when a person visually recognizes the recognition result output from the speech recognition system and detects an error while listening to the actually uttered voice with his ear, the person takes one of the sentences. Since there is a tendency to focus on the flow of sentences rather than focusing on one character,
It often misses local typographical errors and omissions. In particular, since the two processes of simultaneously correcting the sentence while listening to the actually uttered voice are performed at the same time, the probability of overlooking an error is further increased as compared with the operation of simply correcting a sentence from only characters.

[0005] Also, in the online recognition error correction work in which the recognition result is used immediately, such as converting a continuously uttered voice into a character in real time and converting the sound into a subtitle in a live TV program, a normal speech speed is used. , The recognition result is corrected while listening to the voice that cannot be heard again, so the conditions become even more severe, imposing a considerable burden on the corrector, and also increasing the probability of missing errors It is inevitable that it will be high.

As described above, in order to put the speech recognition system into practical use, it is necessary to construct a system that takes into account correction of recognition errors. When considering the practical use of the system, it is indispensable to improve not only the efficiency of character conversion but also the efficiency of correction.However, in the correction part, correction by hand is more reliable at present. In order to reduce the burden at the time of correction and the uncorrected portion, the overall system will be improved.

[0007]

SUMMARY OF THE INVENTION The present invention has been made in view of the above circumstances, and automatically extracts an erroneous portion in a recognition result output from a speech recognition system, and is effective for a corrector. It is therefore an object of the present invention to provide a method and an apparatus for detecting an error in a speech recognition result and a program for detecting an error in a speech recognition result, which can reduce the error and the burden of the correction work by presenting the error.

[0008]

Means for Solving the Problems The present invention (claim 1) provides:
When converting the content represented by a speech language into character information by speech recognition, in a speech recognition result error detection method for detecting an error in the speech recognition result, a time domain based on a past speech recognition result is used. And the correct acoustic information and the error acoustic information discriminated as correct / error are stored for each minimum unit, and the learning processing is performed on the correct acoustic information and the error acoustic information. By performing the above, to generate a correct / false discrimination model that is an acoustic model that best discriminates the correct sound information and the error sound information,
An error is detected from a newly input speech recognition result and acoustic information corresponding to the recognition result using a true / false discrimination model.

According to the present invention (claim 2), the correct sound information and the error sound information are extracted from the correct answer included in the past speech recognition result and the speech signal corresponding to the error by acoustic analysis, and the extracted correct answer sound is extracted. Information and error acoustic information are stored, a representative point in the stored correct answer information and error acoustic information is obtained, the representative point is set as an initial value, and a characteristic amount included in the initial value and the representative point are obtained. An optimal acoustic model is generated by sequentially updating the initial model parameters.

According to the present invention (claim 3), when an error is detected, a speech signal output from a speech recognition device is cut out from continuous speech, and an acoustic feature amount of each recognition unit is obtained. , A correct / false discrimination model, and correct / incorrect is determined based on which of the correct model and the error model of the correct / false discrimination model is similar.

[0011] The present invention (claim 4) is a speech recognition result error detection apparatus for detecting an error in the speech recognition result when converting the contents expressed in a speech language into character information by speech recognition. For past speech recognition results,
It has a feature quantity analyzed in the time domain and the frequency domain, and accumulates correct sound information and error sound information discriminated as correct / error for each minimum unit, and stores the correct sound information and the error sound information in the correct sound information and the error sound information. A learning unit for performing a learning process for generating a correct / false discrimination model which is a sound model for best discriminating correct sound information and error sound information, and a newly input speech recognition result and sound information corresponding to the recognition result. And detection means for detecting an error using a true / false discrimination model.

According to the present invention (claim 5), the learning means sorts the correct answer and the error included in the past speech recognition result into the correct answer and the error, and the sorted correct answer and the error voice signal are subjected to the acoustic analysis to obtain the correct sound information and the error. True / false discriminating means for extracting sound information and accumulating the extracted correct sound information and error sound information, and sound for obtaining a representative point in the stored correct sound information and error sound information as an initial value of the sound model. Model initial value generating means, and discriminative learning means for generating an optimal acoustic model by successively updating a feature value included in the initial value and an initial model parameter obtained from the representative point using the representative point as an initial value. And

According to the present invention (claim 6), in the detection means, a speech signal extraction means for extracting a speech signal output from the speech recognition device from continuous speech, and a feature extraction means for obtaining an acoustic feature amount of each recognition unit. A true / false model comparing unit that compares the acoustic feature quantity with the true / false discrimination model and determines whether the true / false discrimination model is similar to the correct model or the error model.

The present invention (claim 7) is a speech recognition result error detection program for detecting an error in the speech recognition result when converting the content expressed in a speech language into character information by speech recognition. For the past speech recognition results, it has a feature amount analyzed in a time domain and a frequency domain, and accumulates correct sound information and error sound information discriminated as correct / error for each minimum unit. A learning process of performing a learning process on the correct sound information and the error sound information to generate a correct / false discrimination model that is an acoustic model that discriminates the correct sound information and the error sound information best; A detection process for detecting an error from the recognition result and acoustic information corresponding to the recognition result using a true / false discrimination model.

According to the present invention (claim 8), in the learning process of the error detection program of the speech recognition result, the speech signal is sorted into the correct answer and the error included in the past speech recognition result, and the sorted correct answer and the error speech signal are extracted. A correct / error discrimination process for extracting correct sound information and error sound information by sound analysis and accumulating the extracted correct sound information and error sound information, and representing a representative point in the stored correct sound information and error sound information. The optimal acoustic model initial value generation process obtained as the initial value of the acoustic model, and the characteristic value included in the initial value and the initial model parameters obtained from the representative point are successively updated by setting the representative point as the initial value. Discriminative learning process for generating an acoustic model.

According to a ninth aspect of the present invention, in the process of detecting an error detection program for a speech recognition result, a speech signal extracting process for extracting a speech signal output from a speech recognition device from continuous speech, and a sound for each recognition unit. A feature extraction process for determining a feature amount, a sound feature amount, a comparison between a true / false discrimination model, and a true / false model matching process for judging correctness based on which one of the correct model and the error model of the true / false discrimination model is similar. Having.

As described above, according to the present invention, a result containing a recognition error output from a speech recognition apparatus and acoustic information corresponding to the result are stored for each correct answer and for each error, and the correctness / error is best determined from these acoustic information. An acoustic model to be discriminated is generated, and the acoustic information to be recognized, which is newly input from the speech recognition device, is compared with these acoustic models (correct / incorrect discriminating models) to determine whether the recognition result is correct or incorrect. By determining whether or not there is, there is a recognition error that cannot be avoided when the contents expressed in the speech language are converted into character information (automatic captioning by voice, transcription, etc.) using speech recognition technology, It is possible to support efficient correction.

[0018]

FIG. 1 shows the configuration of a recognition error detection system according to the present invention.

The recognition error detecting system 1 shown in FIG. 1 comprises an error detecting device 4 and an error presenting unit 7, and the error detecting device 4 receives a past speech recognition result 2 as a learning unit 5.
And a detection unit 6 that receives a speech recognition result from the speech recognition device 3.

The voice recognition device 3 used in the present invention
Does not require a recognizer of special specifications for the system, but uses a general-purpose recognizer.

FIG. 2 is a flowchart showing an outline of the error detection processing of the speech recognition result of the present invention.

The operation of the components shown in FIG. 1 will be described with reference to FIG.

Step 101) The past speech recognition result (recognized character string and speech signal) 2 is input to the learning unit 5. It is assumed that past speech recognition results are stored in a database or the like.

Step 102) The learning section 5 extracts features (correct / false acoustic information) by acoustic analysis of a speech signal corresponding to a correct answer and an error contained in the input past speech recognition result.

Step 103) A true / false discrimination model that best discriminates the extracted true / false acoustic information is generated by discrimination learning and output to the detection unit 6.

Step 104) A new speech recognition result (recognized character string and speech signal) is input from the speech recognition device 3 to the detection unit 6.

Step 105) The detecting unit 6 sets the acoustic information corresponding to the new input speech recognition result and the learning unit 5
Compare the correct / false discrimination model obtained from
It is determined whether or not the error is an error, and the result is output to the error presenting unit 7.

Step 106) The error presenting section 7 displays an error in the recognition result obtained from the detecting section 6.

Thus, by viewing the result presented by the error presenting section 7, the corrector can easily narrow down a portion to be corrected, and improvement in correction work efficiency can be expected. In particular, in the correction that requires real-time properties, the corrector can concentrate on listening to the sound only by paying attention to the point where the error is presented in the recognition result, and the load at the time of the correction is reduced. At the same time, the probability of missing an error can be reduced. Further, since the correction time is shortened, the real-time property is further improved.

[0030]

Embodiments of the present invention will be described below with reference to the drawings.

First, the configuration of the error detection device will be described with reference to FIG.

The recognition error detection system 1 converts the correct and erroneous character strings output from the past speech recognition device 3 and the past speech recognition result 2 which is the corresponding speech signal, and the unknown input speech into character information. A speech recognition device 3 for conversion, an error detection device 4 for detecting an error with respect to an input from the speech recognition device 3 using a true / false discrimination model generated by a learning process, and a detection result from the error detection device 4 corrected. And a presentation unit 7 for presenting to a person.

The error detection device 4 calculates the past speech recognition result 2
And a learning unit 5 that generates a true / false discrimination model from the correct / error speech signals included in the speech recognition unit 3, selection of a true / false discrimination model corresponding to the recognition result output from the speech recognition device 3, and a speech signal corresponding to the recognition result. The detection unit 6 performs a comparison operation with the selected true / false discrimination model.

The past speech recognition result 2 is a speech signal subjected to speech recognition and a character string of a recognition result corresponding to the speech signal, and uses information input and output by a normal speech recognition device. And

Most of the voice recognition device 3 is applicable as long as it can obtain a recognition result, and the recognition target and the recognition method are not particularly limited.

The learning unit 5 generates a true / false discrimination model to be used for the right / false comparison from a voice signal obtained from the past voice recognition result 2 and a character string including an error output corresponding thereto. I do.

Hereinafter, the learning section 5 will be described in detail.

FIG. 3 shows the configuration of the learning section according to one embodiment of the present invention.

The learning section 5 shown in FIG.
It is composed of a correct / error database unit 52, an error database unit 53, a correct answer acoustic model initial value generator 54, an error acoustic model initial value generator 55, a discrimination learning unit 56, and a correct / false discriminant model 57 having a correct answer acoustic model and an error acoustic model. .

FIG. 4 is a flowchart of a process in the learning section according to one embodiment of the present invention.

The operation of the learning section 5 shown in FIG. 3 will be described below with reference to the flowchart of FIG. Step 201) The learning unit 5 describes a recognition result including an error obtained from the past speech recognition result 2 as a minimum unit (word, morpheme, etc., hereinafter a recognition unit, etc.) output by the speech recognition device by the true / false discrimination unit 51. ), An operation of sorting as a correct answer and an error is performed.

The distribution is performed by a method of discriminating by comparing the voice signal to be recognized with the recognition unit character string, or by preparing a correct character string in advance and comparing the correct character string with the recognition unit character string. There are many methods such as discrimination, but the discrimination method is not particularly limited.

Step 202) The correct answer and the incorrect speech signal that have been sorted out by the correct / error discriminating section 51 are subjected to feature extraction by acoustic analysis, and the correct answer is stored in the correct database 52 and the error is stored in the error database 53 for each recognition unit. Store,
Accuracy information for each recognition unit, which is present in large quantities in the past recognition results, and associated acoustic information are accumulated in a database. The data is a pair of character information and acoustic information for each recognition unit, and the acoustic information is time-series data having various feature amounts analyzed in a time domain and a frequency domain.

Step 203) The correct answer acoustic model initial value generation section 54 obtains a representative point of data in the acoustic information for each recognition unit included in the correct answer database section 52, and the identification learning section 56 uses it as an initial value. There are various methods for obtaining a representative point, such as a clustering method in which a distortion measure is defined based on a distance value between data and a point that minimizes the distortion is used as a representative point, but the method is not particularly limited. Also, the number of representative points can be set arbitrarily according to the distribution state of data.

Step 204) Similarly, the error acoustic model initial value generation unit 55 obtains a representative point based on the acoustic information for each recognition unit included in the error database unit 53.

Step 205) In the identification learning section 56,
In order to find a representative point that best discriminates all correct and error data included in an arbitrary recognition unit, it is included in the initial values generated by the correct answer acoustic model initial value generator 54 and the error acoustic model initial value generator 55. An optimal acoustic model is generated by sequentially updating the feature amounts and the initial model parameters obtained from the representative points. There are many data updating methods, such as a gradient method that defines an identification error measure and changes a feature amount or a model parameter in a direction to reduce the identification error measure, but a learning method is not particularly limited.

Step 206) The identification learning unit 56 accumulates the correct and incorrect model data obtained by the learning as the correct and incorrect discrimination model 57 for each recognition unit in the correct and wrong acoustic models, respectively.

Step 207) Each model data of the true / false discrimination model 57 is sent to the detection unit 6.

Here, a clustering method will be described as an example of obtaining a representative point in acoustic information for each recognition unit in step 203 in the flowchart of FIG.

First, in the correct answer acoustic model initial value generating section 54 and the error acoustic model initial value generating section 55, as initial values for discriminating learning, clusters that effectively represent correct answer and error characteristics are extracted. Correct Answer Database 5
2. The data collected for each error database unit 53 is subjected to clustering by the k-means algorithm to determine a representative point. Since the samples used for clustering are time-series data, the distance dif (i, f: sample number) between the voice patterns is equal to the DP distance d (Xi
, Xj). However, Xi and Xj are different time-series audio samples. The centroid calculation is
In order to avoid data deformation as much as possible, select a mimi-max center y (ω) between samples as shown in the following equation (1).
It is a representative point.

[0051]

(Equation 1) Here, ω represents an arbitrary cluster. The n-th intra-cluster total distortion Δnk for convergence determination is ^obtained by the following equation (2).

[0052]

(Equation 2) The convergence determination is performed by applying an arbitrary threshold value Δth to the following equation (3).

[0053]

(Equation 3) Here, L is the number of samples in a cluster, k is the number of repetitions, and M is the number of clusters according to the number of repetitions.

Next, as an example of generating an optimal acoustic model by successively updating the initial model parameters obtained from the representative points in the above step 205, for example, the correctness and falseness existing in each morpheme sample From the data,
A correct / incorrect judgment template is generated by identification learning based on the judgment error minimization criterion. First, learning is performed by defining a classification error scale that also takes into account the classification weight from the initial template, and sequentially updating each parameter in the direction of the effect of the classification error scale using the gradient method.

Next, a newly input speech recognition result,
The operation of the detection unit 6 that detects an error from a speech signal corresponding to a recognition result using a true / false discrimination model will be described.

The detection unit 6 detects an error using acoustic information corresponding to a recognition unit output from the speech recognition device 3 and a true / false discrimination model corresponding to the recognition result character string obtained by the learning unit 5. .

FIG. 5 shows the configuration of the detection unit according to one embodiment of the present invention.

The detection section 6 shown in FIG.
1, a feature extraction unit 62, a true / false model collating unit 63, and a true / false discrimination model 64.

FIG. 6 is a flowchart of the processing in the detection unit according to one embodiment of the present invention.

Step 301) The speech signal extraction unit 61 uses the speech signal output from the speech recognition device 3 to
A speech signal corresponding to each recognition unit is cut out from continuous speech. Step 302) The feature extraction unit 62 calculates an acoustic feature amount of each recognition unit. The acoustic features use the same features as those used at the time of creating the true / false discrimination model, and there are various acoustic features such as a feature in the frequency domain and a feature in the time domain. The method for extracting the feature and the like are not particularly limited.

Step 303) True / false model collating unit 63
Then, a correct answer is obtained by comparing the acoustic feature amount obtained by the feature amount extraction unit 63 with a model corresponding to the recognition character string output by the speech recognition device 3 in the true / false discrimination model 64 sent from the learning unit 5. And whether it is an error. In the determination of correctness, the acoustic information corresponding to the input recognition unit is matched with the correct answer model and the error model, and a determination is made based on which model is similar. If the model form can be expressed as a parameter distance, the closer the distance is, if the model can be expressed as a probability value, the higher the probability likelihood is the determination result.

Step 304) The model matching unit 63
The judgment result is output to the error presenting unit 7.

Thus, the error presenting section 7 presents an error to the corrector based on the determination result output from the true / false model collating section 63. In the presentation of the error, various character attributes such as color, size, and spacing are added to the character information of the recognition result and displayed, thereby clearly indicating a correction portion and prompting a corrector to urge the corrector to correct.

In the above embodiment, the operation of the learning section 5 has been described with reference to FIG. 4 and the operation of the detection section 6 has been described with reference to FIG. 6. However, the operations shown in FIGS.

It is also possible to install the constructed program in a CPU of a computer used as an error detection device, or to distribute the program via a network.

It should be noted that the present invention is not limited to the above-described embodiment, but can be variously modified and applied within the scope of the claims.

[0067]

When the current acoustic recognition system is used, since the recognition result always contains an error, some correction means is required. Usually, the correction is performed manually by a corrector. However, according to the present invention, an error can be automatically detected by using a true / false discrimination model, and the error can be effectively presented to a corrector. This can reduce not only the burden on the corrector but also the number of uncorrected portions.

Further, the present invention can be applied as long as a recognition result can be obtained without depending on the internal structure of the speech recognition system, so that the speech recognition system does not need to be modified and the application range is expanded. .

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a recognition error detection system of the present invention.

FIG. 2 is a flowchart showing an outline of an error detection process of a speech recognition result according to the present invention.

FIG. 3 is a configuration diagram of a learning unit according to an embodiment of the present invention.

FIG. 4 is a flowchart of a process in a learning unit according to one embodiment of the present invention.

FIG. 5 is a configuration diagram of a detection unit according to an embodiment of the present invention.

FIG. 6 is a flowchart of a process in a detection unit according to one embodiment of the present invention.

[Explanation of symbols]

 Reference Signs List 1 recognition error detection system 2 past speech recognition result 3 speech recognition device 4 error detection device 5 learning unit 6 detection unit 7 error presentation unit 51 correct / error discrimination unit 52 correct answer database unit 53 error database unit 54 correct answer acoustic model initial value generation unit 55 Error acoustic model initial value generation unit 56 Discrimination learning unit 57 True / false discrimination model 61 Audio signal extraction unit 62 Feature extraction unit 63 True / false model matching unit 64 True / false discrimination model

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Atsushi Imai 1-10-11 Kinuta, Setagaya-ku, Tokyo Japan Broadcasting Corporation Broadcasting Research Institute (72) Inventor Toru Toki 1-1-10 Kinuta, Setagaya-ku, Tokyo 11 Japan Broadcasting Corporation Broadcasting Research Institute F term (reference) 5D015 GG01 GG03 GG04 LL04

Claims (9)

[Claims] 1. A speech recognition result error detection method for detecting an error in a speech recognition result when converting the contents expressed by a speech language into character information by speech recognition, comprising the steps of: Based on the characteristic amounts analyzed in the time domain and the frequency domain, the correct sound information and the error sound information discriminated as correct / error are stored for each minimum unit, and the correct sound information and the error By performing a learning process on the acoustic information, a true / false discrimination model, which is an acoustic model that discriminates the correct acoustic information and the erroneous acoustic information best, is generated. An error detection method for a speech recognition result, wherein an error is detected from corresponding acoustic information using the true / false discrimination model. 2. A sound signal corresponding to a correct answer and an error included in the past speech recognition result is extracted by sound analysis to obtain correct answer sound information and error sound information, and the extracted correct answer sound information and the error sound are extracted. Information is stored, and a representative point in the stored correct sound information and the error sound information is obtained. The representative point is set as an initial value, and a feature amount included in the initial value and an initial value obtained from the representative point are obtained. 2. The method according to claim 1, wherein an optimal acoustic model is generated by sequentially updating model parameters. 3. When detecting the error, a speech signal output from a speech recognition device is cut out from continuous speech, an acoustic feature amount of each recognition unit is obtained, and the acoustic feature amount and the true / false discrimination model are determined. 2. The error detection method for a speech recognition result according to claim 1, wherein the correctness / error is determined based on which of the correct model and the error model of the correct / false discrimination model is similar. 4. A speech recognition result error detecting apparatus for detecting an error in the speech recognition result when converting the contents expressed by a speech language into character information by speech recognition, comprising: , Has a feature amount analyzed in a time domain and a frequency domain, and stores correct sound information and error sound information discriminated as correct / error for each minimum unit, and stores the correct sound information and the error sound. A learning unit that performs learning processing on the information to generate a correct / false discrimination model that is a sound model that discriminates the correct sound information and the error sound information best; Detecting means for detecting an error from the acoustic information by using the true / false discrimination model. 5. The learning means according to claim 1, wherein said learning unit assigns the correct answer and the error included in the past speech recognition result to an error, and extracts the correct answer information and the error sound information by acoustic analysis of the assigned correct answer and the error sound signal. Correct / error discriminating means for accumulating the extracted correct sound information and error sound information, and a sound model initial value for obtaining a representative point in the stored correct sound information and error sound information as an initial value of a sound model. Generation means, and identification learning means for generating an optimal acoustic model by sequentially updating the feature amount included in the initial value and the initial model parameter obtained from the representative point with the representative point as an initial value. The apparatus for detecting an error in a speech recognition result according to claim 4. 6. A sound signal extracting means for extracting a sound signal output from a speech recognition device from continuous speech, a feature extracting means for obtaining an acoustic feature amount of each recognition unit, 5. The error of the speech recognition result according to claim 4, further comprising: a correct / error model matching unit that compares the correct / false discrimination model with the correct / false discrimination model and determines whether the model is similar to the correct model or the error model. Detection device. 7. A speech recognition result error detection program for detecting an error in the result of speech recognition when converting the contents expressed by a speech language into character information by speech recognition, comprising: , Has a feature amount analyzed in a time domain and a frequency domain, and stores correct sound information and error sound information discriminated as correct / error for each minimum unit, and stores the correct sound information and the error sound. A learning process of performing a learning process on the information to generate a correct / false discrimination model that is a sound model that discriminates the correct sound information and the error sound information best. A detection process for detecting an error from the corresponding acoustic information by using the true / false discrimination model. 8. The learning process includes: distributing the correct answer and the error included in the past speech recognition result to the correct answer and the error; extracting the correct answer acoustic information and the error acoustic information by acoustic analysis of the assigned correct answer and the error sound signal. Correct / error discrimination process of accumulating the extracted correct sound information and error sound information, and a sound model initial value for obtaining a representative point in the stored correct sound information and error sound information as an initial value of a sound model. A generation process and an identification learning process of generating an optimal acoustic model by sequentially updating the feature amount included in the initial value and the initial model parameter obtained from the representative point with the representative point as an initial value. The program for detecting an error in a speech recognition result according to claim 7. 9. The detection process includes: a speech signal extraction process of extracting a speech signal output from a speech recognition device from continuous speech; a feature extraction process for obtaining an acoustic feature amount of each recognition unit; 8. The error of the speech recognition result according to claim 7, further comprising a correct / error model matching process of comparing the correct / false discrimination model with the correct / false discrimination model and determining whether the correct / error discrimination model is similar to the correct model or the error model. Detection program.