JP2009157050A - Uttering verification device and uttering verification method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To reduce a verification cost of an uttering content by using an automatic speech detection technology. <P>SOLUTION: An uttering verification system includes a user interface capable of separately performing: word detection result verification operations S11 to S14 of a first stage, in which only whether a result detected by the automatic speech detection technology from a speech waveform to be detected is correct, is verified by a person; and content verification operations S15 to S18 of a second stage in which a detection audible range is determined based on the verification result, and which verifies whether a content is correct from the detection audible range. Based on the past verification result, the number of detection words to be verified is reduced. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to an utterance verification system that verifies whether a predetermined content is uttered with respect to a speech waveform.

  In recent years, a method of using voice information has been devised, such as monitoring call center operations in a company, recording a conversation between an employee and a customer, and verifying the validity of the conversation for use in complaint handling and corporate compliance.

  When a recorded utterance is heard and verified by a human, it takes an enormous amount of time according to the recorded length. Although it is conceivable to shorten the verification time by applying a conventional technique such as speech speed conversion, human listening ability is limited, and it still takes a lot of time. Although automation by computer processing is also considered, it is difficult due to the diversity of speech (the influence of speaker nature, environment, etc.), and a technology effective in a practical environment where various utterances are made has not yet been established. Therefore, at the current technical level, computerization and human verification should be combined into a system for maximum efficiency.

As such a system, for example, Japanese Patent No. 3285145 describes a method for verifying a recorded voice database. In this method, whether or not labels attached to a large amount of utterance databases are correct is automatically searched using speech recognition, and a person who does not match is audited. As a result, human verification time can be greatly reduced.
Japanese Patent No. 3285145

  The problem to be solved by the present invention is that when verifying whether the content of speech uttered at a call center conforms to a predetermined rule, the accumulated utterance data must be verified almost as it is, It takes a lot of time.

  In the prior art, utterance verification in units of words is assisted by a computer. However, utterance verification at the word level is insufficient to verify the content of the utterance. This is because the content of an utterance is usually composed of at least a sentence, that is, a plurality of words, and even when the same content is uttered, the types and order of words to be uttered are different. Therefore, even if the utterance verification can be performed in units of words, it cannot be said that the content has been verified and the problem of the present invention cannot be solved.

  Furthermore, the knowledge level required for the verification operator differs between the task of verifying the word and the task of verifying the content. When verifying whether or not the contents are correct by listening to the voice, the person to be verified must be familiar with the corresponding work contents and be familiar with the judgment criteria. When human resources with specialized knowledge are required in this way, the cost as labor costs increases. This is also a problem to be solved.

  In the present invention, a first-step “word detection result verification” operation for verifying only whether or not a result detected by an automatic speech detection technology from a speech waveform to be verified is correct, and a listening range is determined based on the verification result. Provided is an utterance verification system including a user interface that can individually determine and verify whether the content is correct from the audition range, and perform a second-stage “content verification” operation. In addition, the number of detected words to be verified based on past verification results is reduced.

  An utterance verification device according to the present invention includes a word detection unit that obtains a position including an utterance of a specified word from an input speech waveform as a word detection result, and a waveform that includes a position obtained as a word detection result from the speech waveform. Detected word verification processing unit that reproduces and stores input word detection affirmative / negative determination as a detected word verification result, a file that stores a determination range determination rule based on words, a detected word verification result, and an audition A listening range determination unit that determines the listening range of the audio waveform based on the range determination rule. In addition, a content verification unit is provided that reproduces a waveform including a listening range from the input audio waveform, and stores affirmation / negative determination as to the content verification result.

  The word detection unit obtains, as a word detection result, a score indicating the probability of the detection together with the position where the utterance of the designated word is included from the input speech waveform, and the detected word verification processing unit has previously detected the detected word verification processing The score prior distribution information is obtained from a database consisting of the detected word verification result obtained by the section and the score of the detected word, and whether the word detection is positive or negative is determined based on the score in the word detection result and the score prior distribution information. It may be. At this time, the detected word verification processing unit estimates the positive score distribution from the score of the score prior distribution information determined to be positive, and based on the positive score distribution and the predetermined allowable error rejection rate, the allowable error rejection score threshold value , And an acceptable false rejection score threshold is compared with the score of the word detection result to determine whether the word detection is positive or negative. In addition, the detected word verification processing unit estimates a negative score distribution from the score of the score prior distribution information that has been determined to be negative, and sets an allowable misacceptance score threshold based on the negative score distribution and an allowable error acceptance rate specified in advance. It is possible to determine whether the word detection is affirmative or negative by calculating and comparing the allowable false acceptance score threshold with the score of the word detection result.

  Further, the speech verification method according to the present invention is input using an utterance verification device having a word detection unit, a detected word verification processing unit, a file storing a listening range determination rule, and a listening range determination unit. The method of determining the audition range of the voice waveform, the step of obtaining a position including the utterance of the word specified from the input voice waveform as a word detection result by the word detection unit, the detected word verification processing unit, Playing back a waveform including the position obtained as the word detection result from the speech waveform and storing the input word detection positive / negative determination result as the detection word verification result, the detection word verification by the listening range determination unit A step of determining a listening range of the audio waveform based on the result and the listening range determination rule;

  The first-stage word detection result verification work is a simple repetitive work that does not require specialized knowledge and can be performed at high speed, and only a limited part needs to be verified based on the result, so the cost of listening to the whole This makes it possible to check with high accuracy while reducing the above. Furthermore, by shortening the word detection result verification work based on the past verification results, the first-stage word detection result verification work itself can be shortened, and the work time can be reduced to approximately half or less.

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

  First, a first embodiment of the present invention will be described. The utterance verification system of the present invention includes the utterance verification apparatus shown in FIG. The utterance verification device 100 includes an arithmetic device 200, a storage device 300, an output device 400, and an input device 500. The storage device 300 stores programs and data, and holds a word detection program 311, a listening range determination program 314, a detected word verification UI (User Interface) processing program 313, and a content verification UI processing program 315 as programs. Also, as data, a speech waveform 360, a word detection result 370, a detected word verification result 380, and a content verification result 390 are held. Furthermore, a word dictionary 320, a phoneme model 330, and a listening range determination rule 340 are held. Although the present embodiment assumes a general computer operation in which a program is read from a storage device and an arithmetic device operates according to the mechanism, the present invention is described without depending on such a configuration. Applicable to any computer that operates according to the program.

  Next, the operation of the utterance verification device 100 will be described. First, as shown in FIG. 2, the word detection program 311 uses the phonetic model 330 to search the speech waveform 360 for a portion that can be determined to correspond to a word included in the word dictionary 320, and records the result as a word detection result 370. To do. In the following description, the audio waveform 360 is described as having two audio waveforms 361 and 362. The voice waveform is data including human voice, music, sound, noise, and the like stored in a format such as a linear PCM (Pulse Code Modulation) format and an MP3 (MPEG 1 Layer-3) format.

  The word dictionary 320 is a database in which words to be detected are described, and each record includes a word ID, a pronunciation string, and notation information. A word ID indicates a unique identifier of the record, a pronunciation string indicates a pronunciation method of the word, a pronunciation symbol (phoneme or phoneme of the language), and notation information indicates a character notation of the word. Note that the word ID may be replaced by a pronunciation string or notation information, or the notation information may be omitted. The pronunciation string may be omitted, and the pronunciation string may be directly obtained from the notation information according to a conversion rule defined by the language. Also good. In the following description, it is assumed that the word dictionary 320 is registered with two records having the word IDs “a” and “b” as illustrated. The phoneme model 330 is a database that represents the characteristics of a speech waveform corresponding to phonetic symbols that describe a phonetic string in a word dictionary. For example, a format such as HMM (Hidden Markov Model) is used.

  The word detection program 311 uses a known technique generally called word spotting, and uses the phoneme model 330 to search for a part that can be determined to correspond to a word included in the word dictionary 320 from the speech waveform 360. The specific method is well known in the literature in the technical field, and is omitted here. Here, it is assumed that the detection sections 501 to 509 are detected as illustrated.

  The word detection result by the word detection program 311 is stored in the word detection result 370. The word detection result 370 is a database, and each record includes a detection ID, a waveform ID, a word ID, and a position. The detection ID is a unique identifier of the record of the word detection result 370, the waveform ID is an identifier that identifies a speech waveform, the word ID is an identifier that identifies a word recorded in the word dictionary 320, and the position is a section where a word is detected (start Time and end time). Here, it is assumed that the detection IDs 501 to 509 are assigned to the detection sections 501 to 509 and the illustrated contents are obtained. For example, the record of the detection ID “501” records that the word corresponding to the word ID “a” is detected at the positions 5.02 to 5.89 of the waveform ID “361”.

  Next, the detected word verification UI processing program 313 creates a detected word verification UI as shown in FIG. 3 based on the word detection result and presents it to the word detection result determination operator. The detected word verification UI includes a word detection result information presentation unit 610, a speech waveform reproduction instruction unit 620, an affirmative determination unit 630, and a negative determination unit 640.

  The word detection result information presentation unit 610 includes information related to the word detection result, for example, a pronunciation string or notation information corresponding to the word ID of the current word detection result record, information indicating the remaining process of work, and the like. included. When this is designated, the speech waveform reproduction instructing unit 620 presents the speech waveform designated by the waveform ID and position of the current word detection result record to the word detection result determination operator by the output means. Note that the presented speech waveform may be only the speech waveform specified by the waveform ID and the position, or a section including several seconds before and after the position, or the entire speech waveform specified by the waveform ID.

  When this is specified, the affirmative determination unit 630 and the negative determination unit 640 store in the detected word verification result 380 that positive and negative determinations have been made for the currently detected word detection result record, respectively. The detected word verification result 380 is composed of a detection ID and a verification result, as shown in FIG.

  The detected word verification UI processing program 313 repeats the above processing for all records of the word detection result 370 to create a detected word verification result 380. Here, it is assumed that the verification result as shown in FIG. 4 is instructed by the word detection result determination operator. The order of the records to be presented is arbitrary, but if records with the same word ID are presented together, the word detection result discriminating operator can continuously discriminate the same word, so that an increase in work efficiency can be expected.

  Next, the listening range determination program 314 determines the listening range based on the word detection result 370, the detected word verification result 380, and the listening range determination rule 340, as shown in FIG. Store. The inspection range determination rule 340 includes a content ID and a condition. The content ID is an identifier for designating the content defined by the rule, and the condition is a condition indicating the situation in which the detected word verification result is suitable for the rule. Here, it is assumed that the condition “detected word verification result of word ID = a and word ID = b exists in a positional relationship within 10 seconds” is specified as the content ID “901”.

  In general, when performing the condition determination, the result determined to be negative by the detected word verification result 380 is excluded from the condition determination. Further, it is the entries detected within the same waveform ID that make the condition determination.

  The audition range includes a content ID, a waveform ID, and a section. The content ID has the same meaning as the content ID in the audition range determination rule. The waveform ID is an identifier for designating a speech waveform. The section is used when a condition is hit in a range to be audited, for example, in the audition range determination rule. Represents the minimum continuous section including all the detected entry sections. From the word detection result shown in FIG. 2, the detected word verification result shown in FIG. 4, and the audition range determination rule shown in FIG. 5, as shown in FIG. 11.42-14.31 "are obtained. Here, since the verification result of detection 502 is negative in the detection word verification result 380 shown in FIG. 4, the audition range determination program determines that it does not match the rule, and excludes it from the content determination processing in the subsequent stage. Such a case is a typical example in which the work cost for content determination can be reduced by the result of detection word determination in the present invention.

The following listening range determination rules can be considered.
(1) Among the entries in the word detection result, when the detection ID is affirmed by the detection word verification result, if all of the word ID group exist within a certain time, the range of the word ID group is audited. Range.
(2) Among the entries in the word detection result, when the detection ID is affirmed by the detection word verification result, if all of the word ID groups exist in a predetermined order within a certain time, the word ID group Is set as the audition range.
(3) The above-described listening range is expanded to a voice break. The voice segmentation is an arbitrary position in a section where it is determined that there is no voice based on an index based on voice power or voice quality. This facilitates later content determination work.

  Next, the content verification UI processing program 315 creates a content verification UI as shown in FIG. 6 based on the listening range 385 and presents it to the content verification operator. The content verification UI includes an audition range information presentation unit 710, a speech waveform reproduction instruction unit 720, an affirmative determination unit 730, and a negative determination unit 740.

  The audition range information presentation unit 710 includes information related to the audition range, for example, information explaining the contents to be verified of the target audition range, information indicating the remaining steps of the work, and the like. When this is designated, the voice waveform reproduction instructing unit 720 presents the voice waveform designated by the waveform ID and the section of the current audition range record to be presented to the content verification operator by the output means. The presented speech waveform may be only the speech waveform specified by the section, the section including several seconds before and after the section, or the entire speech waveform specified by the waveform ID.

  When this is specified, the affirmative determination unit 730 and the negative determination unit 740 store in the content verification result 390 that affirmative and negative determinations have been made for the audition range record that is the current target. As shown in FIG. 7, the content verification result 390 includes a content ID and a verification result.

  The content verification UI processing program 315 repeats the above processing for all the records in the listening range 385 and creates a content verification result 390. Here, it is assumed that the verification result as shown in FIG. 9 is instructed by the content verification operator. The order of the records to be presented is arbitrary. However, if records with the same content ID are presented together, the content verification operator can continuously determine the same content, so that an increase in work efficiency can be expected.

  By the operation of the utterance verification device described above, the result of content verification is obtained from the speech waveform, and the object of the present invention is achieved.

  In the present embodiment, it is assumed that the utterance verification device is one device, but it may be composed of two or more devices. For example, the function of the word detection program and the detected word verification UI processing program is the first utterance verification device, the function of the audition range determination program and the content verification UI processing program is the second utterance verification device, and necessary data is a network, etc. It is conceivable to realize an utterance verification system by sharing through the network.

  In this embodiment, the detected word discrimination UI and the content discrimination UI are assumed to be a GUI (Graphical User Interface), but may be replaced with a CUI (Character User Interface) or a VUI (Voice User Interface) having equivalent functions. Good. As an implementation example of the VUI, it is conceivable that the user can instruct affirmation, denial, and re-listening by voice input or a button at hand after explaining the word to be discriminated and the contents to be discriminated by voice.

  FIG. 12 is a flowchart showing the flow of processing according to the first embodiment of the present invention. First, in step 11, word detection is performed by the word detection program 311 on the input speech waveform. The detection result is recorded as a word detection result together with the word ID of the detected word and positional information in the speech waveform. Next, the processing from step 12 to step 13 is performed on all word detection results X by the detected word verification UI processing program 313. In step 12, the detected word verification UI for X is displayed, the detected word portion of the input speech waveform is reproduced, and the input device 500 waits for a determination regarding word detection to be input. In step 13, the input determination result is stored in the detected word verification result. In step 14, it is determined whether the word detection result X remains. If it remains, the process returns to step 12 to repeat the process, and if not, the process proceeds to step 15.

  In step 15, the audition range is determined by the audition range determination program 314 based on the detected word detection result and the audition range determination rule. Next, proceeding to step 16, the content verification UI program 315 displays the content verification UI for the listening range Y, reproduces the waveform specified by the listening range Y among the audio waveforms, and inputs the playback content. It waits for a positive or negative determination to be input from the device 500. In step 17, the input determination result is stored as a content verification result. In step 18, it is determined whether the verification range Y remains. If it remains, the process returns to step 16 to repeat the process, and if not, the process ends.

  next. A second embodiment of the present invention will be described. The second embodiment uses the score of each detected word output from the word detection program in the first embodiment, and reduces the cost of the detected word verification based on the score. An utterance verification apparatus 101 of this embodiment is shown in FIG. The difference from the utterance verification device 100 shown in FIG. 1 is that the storage device 300 includes a threshold discrimination database 350 and further includes a scored word detection result 371 instead of the word detection result. Further, the operations of the word detection program 311 and the detected word verification UI processing program 313 are different as follows.

  The operation of the word detection program 311 is in accordance with the contents described in the first embodiment (FIG. 2), except that a word detection result with score 371 shown in FIG. 9 is output instead of the word detection result. Information of the score of the detected word indicated by each record (indicating how likely the detection result is) is added to the scored word detection result. For calculating the score, a method based on an acoustic likelihood difference from an acoustic model is generally known. The specific method is well known in the literature in the technical field, and is omitted here.

  The operation of the detected word verification UI processing program 313 conforms to the content described in the first embodiment, but selects a target detection result based on the score information of the scored word detection result and the threshold discrimination database.

  First, details of the threshold discrimination database are shown in FIG. Each record in the threshold discrimination database is composed of a word ID, a score, and a verification result. The threshold discrimination database stores the discrimination results of the detected words discriminated in the past by this system.

First, the detected word verification UI processing program 313 obtains a positive score distribution and a negative score distribution of a certain word ID from the threshold discrimination database 350. Assuming that the positive score distribution and the negative score distribution are normal distributions, the sample mean and sample variance are obtained. Next, based on the predetermined allowable error acceptance rate A and the allowable error rejection rate B, the allowable error acceptance score threshold value α and the allowable error rejection score threshold value β are obtained. If the probability distribution functions of the positive score distribution and the negative score distribution are set as f ₁ (X) and f ₂ (X), respectively, f ₂ (X = α) = 1−A, f ₁ (X = β) = Α and β satisfying B. These are illustrated in FIG.

Further, based on this result, each record of the word ID of the scored word detection result is processed as follows according to the value of the score S.
(1) When S> α, the verification result of the record is accepted.
(2) If S <β, reject the verification result of the record.
(3) In other cases, a word detection result verification UI for the record is created and presented to the word detection result verification operator. When an affirmative discrimination unit or a negative discrimination unit is designated, the result is stored in the detected word verification result 380 and simultaneously stored in the threshold discrimination database 350.

Note that the above-described score determination method is merely an example, and other methods such as the following are conceivable.
1. The permissible false acceptance score threshold value α and the permissible false rejection score threshold value β are obtained from the threshold discrimination database for each detection ID, but at regular intervals or at arbitrary timing.
2. If false rejection is not acceptable, only α is used instead of β. This is equivalent to setting the allowable error rejection rate B = 0 and the allowable error rejection score threshold β = −∞. This is useful for tasks that cannot be rejected.

On the other hand, even if a task is not allowed to be accepted incorrectly, it will be rejected at the subsequent content determination even if it is erroneously accepted at the time of word determination. . Still, if erroneous acceptance is not allowed at the time of word determination, similarly, only β is used instead of α. This is equivalent to setting the allowable error acceptance rate A = 1 and the allowable error acceptance score threshold α = ∞.
3. A set of identifiers that can be assumed to have the same score distribution, such as speakers and environments, is separately assigned to the scored word detection result and the threshold discrimination database, and only those having the same identifier are used to allow the acceptable false acceptance score threshold α A rejection score threshold β is obtained. This method is effective when it is known that the score distribution varies depending on the speaker and the environment.

  As a result of the above processing, there is a record in which the verification result is automatically determined theoretically as in (1) and (2) above, so that the word detection result determination work is shortened while maintaining a certain detection accuracy. it can.

Moreover, you may use a score as follows for the audition range determination rule in 1st embodiment.
(1) Of the entries in the word detection result, when the detection ID is affirmed by the detection word verification result, if there is a certain word ID group within a certain time, and the total score of those entries Alternatively, when the average value exceeds a predetermined value, the range of the word ID group is set as the audition range.
(2) Of the entries in the word detection result, when the detection ID is affirmed by the detection word verification result, if there is a certain word ID group in an order determined within a certain time, and those entries When the total value or average value of the scores exceeds a predetermined value, the range of the word ID group is set as the audition range.

  FIG. 13 is a flowchart showing the flow of processing according to the second embodiment of the present invention. First, in step 21, word detection is performed by the word detection program 311 on the input speech waveform. The detection result is recorded as the word detection result together with the word ID of the detected word, the position information in the speech waveform, and the score. Next, the processing from step 22 to step 25 is performed on all word detection results X by the detected word verification UI processing program 313. In step 22, X determines whether discrimination based on the score threshold is possible. If possible, the process proceeds to step 23, and the automatic determination result is stored as a detected word verification result. If discrimination based on the score threshold is impossible, the process proceeds to step 24, where the detected word verification UI for X is displayed, the detected word portion of the input speech waveform is reproduced, and the determination for word detection from the input device 500 is performed. Wait for input. In step 25, the input determination result is stored in the detected word verification result. Next, it progresses to step 26 and updates a score threshold value. The score threshold is updated by obtaining an affirmative score distribution and a negative score distribution of a certain word ID based on the detection word verification results so far, and accepting false acceptance based on a predetermined acceptable false acceptance rate A and acceptable false rejection rate B. This is performed by obtaining the score threshold α and the allowable error rejection score threshold β. In step 27, it is determined whether the word detection result X remains. If it remains, the process returns to step 22 to repeat the process, and if not, the process proceeds to step 28.

  In step 28, the audition range is determined by the audition range determination program 314 based on the detected word detection result and the audition range determination rule. Next, the process proceeds to step 29, where the content verification UI program 315 displays the content verification UI for the listening range Y, reproduces the waveform specified by the listening range Y among the audio waveforms, and inputs the playback content. It waits for a positive or negative determination to be input from the device 500. In step 30, the input determination result is stored as a content verification result. In step 31, it is determined whether the verification range Y remains. If it remains, the process returns to step 29 to repeat the process, and if not, the process ends.

The figure which shows the structural example of the speech verification apparatus by this invention. The figure explaining the detail of an audio | voice waveform, a word detection result, and operation | movement of a word detection program. The figure which shows detected word verification UI. Explanatory drawing of a detection word verification result. The figure explaining the detail of the audition range and operation | movement of the audition range determination program. The figure which shows content verification UI. Explanatory drawing of a content verification result. The figure which shows the structural example of the speech verification apparatus by this invention. Explanatory drawing of the word detection result with a score. Explanatory drawing of a threshold discrimination database. Explanatory drawing of positive score distribution and negative score distribution. The flowchart which shows the flow of a process. The flowchart which shows the flow of a process.

Explanation of symbols

100, 101 Utterance verification device 100 Arithmetic device 300 Storage device 400 Output device 500 Input devices 501 to 509 Word segments detected by the word detection program

Claims (11)

A word detection unit that obtains a position where an utterance of a specified word is included as a word detection result from the input speech waveform;
A detection word verification processing unit that reproduces a waveform including the position obtained as the word detection result from the speech waveform, and stores the input word detection positive / negative determination as a detection word verification result;
A file that stores rules for determining the scope of listening based on words;
A listening range determination unit that determines a listening range of the speech waveform based on the detection word verification result and the listening range determination rule;
An utterance verification device comprising:   The utterance verification device according to claim 1, further comprising: a content verification unit that reproduces a waveform including the listening range from the speech waveform, and stores affirmation / negative determination as to a content verification result with respect to the input reproduction content. A featured speech verification device.   The utterance verification apparatus according to claim 2, further comprising a storage unit that holds the speech waveform, the word detection result, the detected word verification result, the listening range, and the content verification result. The utterance verification device according to claim 1,
The word detection unit obtains, as a word detection result, a score indicating the probability of the detection together with a position including the utterance of the specified word from the input speech waveform,
The detected word verification processing unit obtains score prior distribution information from a database including the detected word verification result obtained in the past by the detected word verification processing unit and the score of the detected word, and a score in the word detection result An utterance verification device that performs affirmative / negative determination of the word detection based on the score prior distribution information. In the utterance verification device according to claim 4,
The detected word verification processing unit estimates an affirmative score distribution from the score of the score prior distribution information determined to be affirmative, and an allowable error rejection score threshold based on the positive score distribution and an allowable error rejection rate specified in advance Utterance verification device, wherein the threshold of the allowable error rejection score is compared with the score of the word detection result to determine whether the word detection is affirmative or negative. In the utterance verification device according to claim 4,
The detected word verification processing unit estimates a negative score distribution from a score of the score prior distribution information that is determined to be negative, and an allowable misacceptance score threshold based on the negative score distribution and a pre-specified allowable misacceptance rate Utterance verification apparatus, wherein the acceptance error acceptance score threshold value is compared with a word detection result score to determine whether the word detection is affirmative or negative. Using an utterance verification device having a word detection unit, a detected word verification processing unit, a file storing the audition range determination rule, and an audition range determination unit, the audition range of the input speech waveform is determined. A method,
A step of obtaining, as a word detection result, a position including the utterance of the designated word from the input speech waveform by the word detection unit;
Reproducing a waveform including a position obtained as the word detection result from the speech waveform by the detected word verification processing unit, and storing an input word detection positive / negative determination result as a detection word verification result;
A step of determining a listening range of the speech waveform based on the detection word verification result and the listening range determination rule by the listening range determination unit;
A speech verification method characterized by comprising:   8. The utterance verification method according to claim 7, further comprising a step of reproducing a waveform including the audition range from the speech waveform and storing an affirmative / negative determination result for the input reproduction content as a content verification result. Utterance verification method.   The utterance verification method according to claim 7, wherein a score representing the probability of detection is also obtained as the word detection result, and in the step of obtaining the detection word verification result, a score in advance is obtained from the detection word verification result including a score obtained in the past. An utterance verification method, wherein distribution information is obtained, and whether the word detection is positive or negative is determined based on a score in the word detection result and the score prior distribution information.   10. The utterance verification method according to claim 9, wherein an affirmative score distribution is estimated from a score of the score prior distribution information that is determined to be affirmative, and an allowable error rejection based on the affirmative score distribution and a predetermined allowable error rejection rate. An utterance verification method comprising: calculating a score threshold value, and comparing the allowable false rejection score threshold value with a score of a word detection result to determine whether the word detection is affirmative or negative.   10. The utterance verification method according to claim 9, wherein a negative score distribution is estimated from a score of the score prior distribution information that is determined to be negative, and an allowable misacceptance is based on the negative score distribution and a predetermined allowable error acceptance rate. An utterance verification method, comprising: calculating a score threshold value and comparing the allowable false acceptance score threshold value with a score of a word detection result to determine whether the word detection is positive or negative.

Images