JP2006084706A - Voice recognition performance estimating method, device, and program, recognition trouble word extracting method, device, and program, and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a recognition performance estimating device capable of evaluating the performance of a word list used for voice recognition and a recognition trouble word extracting device capable of extracting a word causing recognition trouble from words included in the word list. <P>SOLUTION: The voice recognition performance estimating device 100 is equipped with a word similarity calculating means 104 of inputting two words which are an object word and an opponent word to calculate the word similarity; a word misrecognition score calculating means 103 of inputting the object word and word list to calculate a word misrecognition score; a mean word misrecognition score calculating means 102 of inputting the word list and calculating the mean word misrecognition score of all the words included in the word list; and a voice recognition rate estimating means 101 of inputting three elements which are the number of the words included in the word list, the mean word misrecognition score, and a candidate number narrowing-down reference value set when a solution candidate is searched for in voice recognition processing and calculating an estimated value of a voice recognition rate. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to the technical field of speech recognition, and more particularly to the field of predicting the performance of a word list used for speech recognition and designing a recognition vocabulary.

As a conventional technique in the field of speech recognition performance estimation that estimates the performance of a speech recognition device before actually performing speech recognition, it automatically estimates the misrecognition tendency of phonemes from the parameters of the acoustic model used for speech recognition, There is a method (Non-patent Document 1) that calculates an acoustic similarity between two words, calculates a similarity between words based on the acoustic similarity, and estimates the recognition rate from the similarity with the opposing word. It was.
Conventionally, in designing a recognition vocabulary used for speech recognition, a word predicted to be pronounced by a user is designed as a recognition vocabulary.
“A Study on Model-Based Error Rate Estimation for Automatic Speech Recognition”, IEEE Transactions on Speech and Audio Processing, Volume 11, No 6, November 2003.

When the speech recognition technology is put to practical use, it is necessary to estimate the accuracy of the recognition result to some extent in advance in order to estimate the quality of the service to be provided. In order to estimate the accuracy of recognition results in advance, in general, speech in an environment similar to the speech input when a service is provided is collected as test data, and the accuracy is estimated by conducting a recognition experiment on this. . However, since collecting a sound in an environment similar to an actual service requires an appropriate economic burden and time, a technique for estimating the recognition accuracy without recording the sound is required.
In the method of Non-Patent Document 1 described in the background art, the recognition accuracy is estimated using only the information of the recognition vocabulary and the acoustic model, and the recording of the voice is unnecessary.

On the other hand, when the speech recognition technology is put to practical use, the speed of recognition is an important factor as well as the accuracy of the recognition result. For example, when constructing a voice response system that accepts voice input and outputs an appropriate response, if the result is not obtained until a few seconds have passed after the input voice is over, the response is delayed and gives an unnatural impression Usability is significantly reduced.
By the way, the currently widely used speech recognition method is to divide a speech waveform obtained as input into short units of several milliseconds, and what is the phoneme uttered in short units based on the frequency characteristics within a short interval. The word candidates including the phoneme are listed as a word list, and the word candidate that best matches the input from the word list is used as the recognition result.

  When enumerating all the words to be recognized when enumerating this word list, the amount of calculation increases and it may be difficult to obtain a speech recognition result in real time. It is known that a recognition result can be obtained in real time with a reduced amount of calculation if a candidate is searched only in a limited number of words. However, if the number of candidates is limited, since the correct recognition cannot be performed unless the correct answer is included in the candidates, the recognition accuracy decreases. That is, the recognition speed can be increased by changing the reference value of how many search candidates are narrowed down when performing speech recognition, but the disadvantage is that the recognition accuracy is reduced.

Therefore, in general, when speech recognition is put to practical use, the number of search candidates to be narrowed is changed depending on which of the recognition speed and the recognition accuracy is important.
In the conventional technology as described in Non-Patent Document 1, since the change in recognition accuracy due to the number of search candidates narrowed down cannot be estimated, for example, when using the speech recognition technology for a service that requires high-speed response output, it is sufficient in advance. The recognition accuracy could not be estimated with high reliability.
In addition, as the user gets used to speaking toward the computer system, the number of utterances such as “Ut”, “Oh no”, “Ah”, etc. increases. That is, the target of speech recognition includes not only important keywords such as names and place names, but also those that are not important but can appear in the user's utterance.

On the other hand, the recognition accuracy of speech recognition tends to decrease when words having similar pronunciations are included in the recognized vocabulary (word list). For example, a voice uttered “Eto” is rarely recognized as “Ah”, but is often mistakenly recognized as “Eh”. If the recognition vocabulary is designed to include all vocabularies that are not important but can appear in the user's utterance, the vocabulary that closely resembles the utterance with important keywords such as names of people and places will be included in the recognition vocabulary. The rate may be significantly reduced.
In the conventional recognition vocabulary design, all the words that are expected to appear are designed as a word list. Therefore, from the viewpoint of whether the word list of important keywords is adversely affected, the words that can be realized are not important. It was not possible to determine whether to include in the word list.

  The present invention intends to propose a speech recognition performance estimation method and apparatus capable of estimating the performance of a word list used for speech recognition. For this purpose, the present invention receives two words consisting of a target word and an opposing word, and calculates a word similarity that indicates the degree to which the utterance corresponding to the target word is recognized as an opposing word when speech recognition is performed. A word indicating the degree to which the utterance corresponding to the target word is recognized as one of the words in the estimated word list when the calculation process, the target word and the estimated word list as the performance estimation target are received and speech recognition is performed A word misrecognition score calculating process for calculating a misrecognition score; an average word misrecognition score calculating process for receiving an estimated word list and calculating an average word misrecognition score of all words included in the estimated word list; The three elements of the number of words included in the estimated word list, the average word error recognition score, and the candidate number narrowing reference value set when searching for speech recognition solution candidates As a force, proposes a speech recognition performance estimation apparatus and executes the speech recognition rate estimation processing of calculating the estimated value of the speech recognition rate, a.

  In the present invention, the recognition impaired word search process and the speech recognition performance estimation method are executed, and the recognition impaired word search method performs the estimated word list for each word included in the estimated word list given by input. The speech recognition performance estimation method estimates the speech recognition performance of the failure search word list consisting of all the remaining words excluding the word from, and extracts a certain number of failure search word lists in descending order of speech recognition performance Then, a recognition failure word extraction method is proposed in which the removed words corresponding to each failure search word list are output as recognition failure words.

  Note that the recognition failure word extraction method similar to that proposed in the present invention can also be realized using a speech recognition performance estimation method according to the prior art. However, in order to correctly extract the recognition failure word, it is required that the accuracy of the speech recognition performance estimation method used for this is high. For this reason, the recognition failure word extraction method according to the present invention has higher reliability accuracy than the recognition failure word extraction method using the prior art in which the recognition accuracy may not be estimated with sufficient reliability. Since the speech recognition performance estimation method according to the present invention is used, the recognition failure word extraction method is clearly excellent.

According to the speech recognition performance estimation method of the present invention, when the speech recognition rate is estimated, the speech recognition rate is calculated using the number of words, the average word misrecognition score, and the candidate narrowing reference value.
Thereby, for example, the reference value is set as the maximum number of candidates when narrowing down candidates, and the estimated recognition rate increases if the maximum number of candidates is large, and the estimated recognition rate decreases if the maximum number of candidates is small. If the speech recognition rate estimation process is performed as described above, it is possible to estimate the speech recognition rate according to the candidate narrowing reference value.
In addition, when narrowing down candidates, for example, the reference value is set so that the voice recognition score such as the acoustic score and language score of the candidate must be higher than the reference value, or the highest voice recognition score among the conflicting candidates If it is set that the difference in the speech recognition score with the candidate must be below the reference value, it is possible to estimate the change in the speech recognition rate according to the reference value, as in the case of the maximum number of candidates. .

  Furthermore, according to the recognition failure word extraction method of the present invention, words that adversely affect speech recognition performance can be extracted for words other than the extraction-prevented words designated as important keywords. For this reason, it is possible to design a recognition vocabulary such as deleting words that adversely affect speech recognition from words that may appear but are not important, while taking into consideration the effect on the recognition performance of important keywords.

The speech recognition performance estimation device and the recognition failure word extraction device according to the present invention mainly install a speech recognition performance estimation program and a recognition failure word extraction program proposed in the present invention in a computer and cause the computer to execute these programs. Implementation of a computer to function as a speech recognition performance estimation device for estimating the recognition rate of a word list to be estimated for performance estimation and a recognition failure word extraction device for extracting a failure word that hinders recognition from the word list to be estimated The form is the best embodiment.
When the computer functions as a speech recognition performance estimation device, the computer includes at least a word similarity calculation means, a word error recognition score calculation means, an average word error recognition score calculation means, and a speech recognition rate estimation means. The operation of calculating the estimated value of the speech recognition rate of the estimated word list by the function of the means is executed.

  When the computer functions as a recognition failure word extraction device, the computer includes a recognition failure word search means in addition to the configuration of the speech recognition performance estimation device described above. A certain number of word lists are extracted in descending order of the speech recognition rate using the speech recognition rate of the estimated word list, and an operation of outputting the removed words corresponding to the respective word lists as recognition failure words is executed.

An embodiment of a speech recognition performance estimation apparatus according to the present invention will be described in detail below with reference to FIGS.
FIG. 1 shows the overall configuration of the speech recognition performance estimation apparatus 100. The speech recognition performance estimation apparatus 100 includes a speech recognition rate estimation unit 101, an average word error recognition rate score calculation unit 102, a word error recognition score calculation unit 103, and a word similarity calculation unit 104. The speech recognition rate r is output with the estimated word list as input.
The speech recognition rate estimation means 101 receives the candidate number narrowing reference value and the estimated word list L _input, passes the estimated word list L _input to the average word error recognition score calculation means 102, and receives the average word error recognition score as a result. take.

In the speech recognition performance estimation apparatus according to the present invention, the candidate number narrowing reference value is set to the maximum number of candidate solutions listed at each time point of the search, the average word misrecognition score is x, and the total number included in the estimated word list L _input When the number of words is y and the candidate number narrowing reference value is z, the speech recognition performance estimation apparatus 100 outputs a value r calculated by the following equation (1) as a speech recognition rate.
r = a ₁ x + a ₂ y + a ₃ (y / z) + a ₄ (1)
Here, the value of each a _n is or sound environment in which the speech recognition is performed, assumed to be previously adjusted in accordance with the acoustic model and speech recognition method used in speech recognition.

The average word error recognition score calculating unit 102 receives the object estimated word list L _{input The,} for each word included in the estimated word list L _{input The,} by the words and target words W _i, all words other than the word and create a new word list L _i including, passing word false recognition score calculating unit 103 receives the words misrecognized score SCA as a result. In the present invention, the average value of the word error recognition score SCA obtained for each word is output as the average word error recognition score x.
Further, in the present invention, as another method of calculating the average word misrecognition score x, the sum of values obtained by multiplying the value of the word misrecognition score SCA obtained for each word by the occurrence probability of each word is used as the average word misrecognition. A method of outputting as a score x is proposed.

Words misrecognized score calculating means 103, input important conflict word extraction section 201 as shown in FIG. 2, the word distance calculation unit 202, it if Toka score conversion unit 203, the target word W _i and word list L _i As a result, a word error recognition score SCA is output.
The word misrecognition score calculation means 103 receives the target word W _i and the word list L _i, and for each word included in the word list L _i , sets this as the opposite word W _k , and sets the target word W _i as a pair. As a result, the word similarity degree J is received. Important conflict word extraction section 201 of all the words contained in the word list L _i is the one word similarity J is large to a certain number extracted from the upper, it as important conflicting word list L _Imsb, with target words W _i Delivered to the inter-word distance calculation unit 202.

Word distance calculation unit 202 calculates the logarithm of word similarity between the target word own target word W _i, for each word included in the important conflicting word list L _Imsb, word similarity between the target word W _i The value obtained by subtracting the logarithmic value of the word similarity from the logarithmic value of the word itself is passed to the score conversion unit 203 as the distance between words.
The score conversion unit 203 converts the inter-word distance d of each word included in the important confrontation list L _imsb into a larger value as d is smaller, and a smaller value as d is larger, The sum of the subsequent values is taken and output as a word error recognition score SCA. This is a process for reflecting that the closer the distance between words is, the higher the importance of the target word as a partner to be erroneously recognized. Various functions can be considered as the function to be converted. For example, a SIGMOID function shown in Expression (2) can be used.

ここで、定数α及びβは、音声認識を行う音環境などに応じてあらかじめ調整されているものとする。
単語類似度算出手段１０４は、図３に示すように音素列生成部３０１と、音素継続時間付与部３０２と、音素アライメント部３０３と、辞書３０４と、音素継続時間長データ３０５と、音素コンフュージョンマトリクス３０６とからなり、対象単語Ｗ_ｉと対立単語Ｗ_ｋを入力として、単語類似度Ｊを出力する。

Here, it is assumed that the constants α and β are adjusted in advance according to a sound environment for performing speech recognition.
As shown in FIG. 3, the word similarity calculation unit 104 includes a phoneme string generation unit 301, a phoneme duration adding unit 302, a phoneme alignment unit 303, a dictionary 304, phoneme duration time data 305, and phoneme confusion. It consists of a matrix 306, receives the target word _Wi and the confronting word _Wk, and outputs the word similarity J.

Phoneme string generation unit 301 inputs the conflict word W _k and target words W _i, from each of the sound information, to create a series of corresponding phoneme, phoneme sequence corresponding to the target word W _i W _ion phoneme duration To the assigning unit 302, the phoneme string W _kon corresponding to the conflict word W _k is output to the phoneme alignment unit 303.
Phoneme duration applying unit 302 receives the phoneme string W _ion of the target word W _i, to impart an average phoneme duration of each phoneme given by the phoneme duration data 305 to each phoneme of a phoneme sequence W _ion, Output to phoneme alignment unit 303 as phoneme string W _iont with duration information.

Phoneme alignment unit 303, and the phoneme string _{W iont} with duration information corresponding to the target word _{W i,} receives the phoneme string _{W kon} corresponding to the confrontation word _{W k,} of the phoneme string _{W iont} corresponding to the target word _{W i} At each time point when each phoneme continues, any phoneme in the phoneme string W _kon corresponding to the conflict word W _k is associated.
An example of association is shown in FIG. Correspondence, as phonemes side corresponding to the target word W _i is the probability to be recognized as a phoneme of the side corresponding to the conflict word, the value obtained by integrating over phoneme-set at all time points becomes largest, the dynamic This is done using genetic programming. The probability that a phoneme is mistakenly recognized by another phoneme is given to the phoneme confusion matrix 306. The phoneme confusion matrix 306 holds the probability of being recognized by the other phoneme when one phoneme is uttered for any combination of phonemes and phonemes. It is prepared in advance according to the acoustic model used.

  The phoneme alignment unit 303 normalizes the probability values accumulated over the phoneme pairs at all time points after the phoneme association is performed, and outputs the normalized value as the word similarity.

FIG. 5 shows an embodiment of the recognition failure word extraction apparatus according to the present invention. 501 shown in FIG. 5 shows a recognition failure word extraction apparatus. The recognition failure word extraction device 501 according to the present invention includes a recognition failure word search device 502 and a speech recognition performance estimation device 100.
The recognition failure word search device 502 receives a candidate number narrowing reference value, an estimated word list L _input, and an extraction prevention word list L _off, and is included in the estimated word list L _input from these inputs. For each word W _i , a new word list L _i obtained by removing each word W _i from the estimated word list L _input and a candidate number narrowing reference value are passed to the speech recognition performance estimation device 100, and the speech recognition performance estimation device 100 The speech recognition rate r _i of the new word list L _i is received from this, the recognition failure word is searched using this speech recognition rate r _i as a clue, and the recognition failure word W _off is output. Here, the extraction prevention word list L _off means a word list that lists important words in the words included in the estimated word list L _input , and is necessary for the execution of various services using speech recognition such as place names and personal names. This is a word list of important keywords. By preventing the words listed in this word list from being extracted, the word list can be adjusted without impeding the execution of the service.

The operation of the recognition failure word search device 501 will be described below using the operation flow shown in FIG.
When the recognition failure word extraction device 501 receives the candidate number narrowing reference value and the estimated word list L _input consisting of N words and the extraction blocked word list L _off consisting of M words in step 601, in step 602, the estimated word list L target words each word other than the extraction blocking words included in the _{input W i (i = 0~N-} M-1) to start the loop.
In the loop inside the step 603, creates a word except the target word _{W i} from the estimated word list _{L input The} list _{L i} (Fig. 9), the speech recognition in step 604, the reference value narrowing down word list _{L i} and the number of candidates The speech recognition rate r is obtained by inputting the performance estimation device 100.

When the loop is finished, in step 606, selects a certain number of word list L _i in descending order of speech recognition performance, in step 607, cognitive impairment word target word W _i corresponding to the selected word list L _i Output as.
Hereinafter, specific operation examples of the respective units will be described with reference to the drawings.
FIG. 7 is an example of the estimated word list L _input input to the speech recognition performance estimation apparatus 100 according to the present invention. FIG. 8 is an example of a probability prepared in the phoneme confusion matrix 306. Originally, there are about 40 phonemes in Japanese, but here, for the sake of simplicity, only the phonemes included in the words in FIG. 7 and pauses meaning silent intervals are considered.

Also, in the following example, the candidate number narrowing reference value is set as the maximum number of candidates when narrowing candidates, and the value is set to 6.
The average word error recognition score calculating unit 106 (FIG. 1), the new receives an object to be estimated word list _{L input The,} each of the words included in the estimated word list _{L input The} intended word _{W i,} the other word deliver to word erroneously recognized score calculation device 103 as a word list _{L i.} Each line of FIG. 9 shows an example of a pair of target words W _i and word list L _i. Because the word list L _i includes four words, a pair of target words W _i and word list L _i of quadruplicate in this case is created.

In words misrecognized score calculating means important conflict word contained in 103 extraction unit 201 (FIG. 2) receives the target word W _i and word list L _i for each row in FIG. 9, each of words contained in the word list L _i for calculates word similarity between the target word W _i using the word similarity degree calculating unit 104, extracts a having a large word similarity between the target word W _i fixed number as a key conflict word list _L imsb. Each row of Figure 10 shows the target words _{W i,} and important conflict word list _{L Imsb,} an example of a word similarity between target words _{W i.} In this example, the number of important conflicting words extracted is 2, and the word similarity is expressed as a logarithm of probability values, and thus takes a negative value. However, the larger the value (the closer it is to 0). which means that close to the target word W _i. Also, numbers are entered in the column of the target word W _i represents the similarity between the target word own target word W _i.

In a word distance calculation unit 202 (FIG. 2) receives the target word _{W i} and important conflict word list _{L Imsb,} for each word included in the important conflicting word list _{L Imsb,} word directed to the logarithm of the word similarity _{W i} The inter-word distance is calculated by subtracting from the logarithmic value of the word similarity with the target word itself. Each row of FIG. 11 shows an example of the distance between words of the important conflicting word. For example, important words between the distance of the confrontation word "morning" when the target word _{W i} is "red", has become a -0.21--0.57 = 0.36. In this way, the value of the word distance between, has become the smaller the value the closer to the target word W _i.

In the score conversion unit 203, the distance between words of each important confrontation word is converted into a larger value as it is smaller, and into a smaller value as it is larger, using a conversion function. Each row of FIG. 12 shows an example of the converted inter-word distance of the important conflict word list L _imsb . In this example, the function to be converted is calculated by substituting α = 10 and β = 0.4 into the SIGMOID function given in Equation (2). The values of α and β are not absolute values and should be adjusted according to the environment in which speech recognition is used. Further taking the sum of these values, a word erroneous recognition score SCA corresponding to each target word W _i is calculated. Each row of Figure 13 shows an example of a word misrecognition score SCA corresponding to the target word W _i.

The average word error recognition score calculation means 102 takes the average of the word error recognition scores SCA included in each line of FIG. 13 and outputs this as the average word error recognition score. In this example, 0.8 + 0.6 + 0.17 + 0.0024 is divided by 4, and 0.394 becomes the average word error recognition score.
The speech recognition rate estimation apparatus 100 calculates the speech recognition rate using Equation (1) using the obtained average word error recognition score 0.394, the number of words 4, and the candidate number narrowing reference value 6. The value of _{a n} of formula (1), for example _{a 1 = -1.5, a 2 =} -0.2, a 3 = 0.8, When a 4 = 80,
−15 × 0.394−0.2 × 4 + 0.8 × (4/6) + 80 = 73.8 Equation (3)
Thus, the speech recognition rate is estimated as 73.8%.

The value of these a _n are not absolute, the value should be adjusted according to the environment in which the speech recognition is used.
In the word similarity calculation means 104 (FIG. 3), the target word _Wi and the conflict word _Wk are received, and the phoneme string generation unit 301 first generates a phoneme string corresponding to the word. In the example shown in FIG. 14, “red” and “morning” are received, and the corresponding phoneme sequences using the dictionary are / a /, / k /, / a / and / a /, / s /, / a /. , The ones with front and back pose sections are generated.
Next, the phoneme string _{W ion} corresponding to the target word _{W i,} to impart phoneme duration in the phoneme duration applying unit 302 (FIG. 3). In the example of FIG. 15, since the target word _Wi is “red”, the duration is given to / pause /, / a /, / k /, / a /, / pause /.

Next, the phoneme alignment unit 303 (FIG. 3), for each time point each phoneme of the phoneme sequence _{W ion} corresponding to the target word _{W i} is continued, either _{W kon} phoneme string corresponding to the conflicting word _{W k} Match any phoneme. An example associated with FIG. 16 is shown. There are various combinations of associations. For each associated phoneme, association is performed using dynamic programming so that the value obtained by integrating the probability values given on the phoneme confusion matrix 306 is maximized. To decide.
The probability of each phoneme to be associated will be described taking the phoneme correspondence 1601 in FIG. 16 as an example. In phonemes corresponding 1601, phonemes corresponding to the target word and W _i / a / a, Conflict word W corresponding phoneme _k is / s / So phoneme confusion matrix 306 on at of / a / line given in FIG. 8 , / K / column value 0.03 is the probability value of each phoneme associated.

The value obtained by integrating the probability values of the associated phonemes is normalized by the total duration of the phonemes of the target word is the output value of the word similarity calculation means 104. In this example, normalization is performed by dividing the logarithm value of the integrated value by the length of time, and this value is -0.57, which is a value corresponding to "morning" in the first row of FIG. Yes.
An example of the operation of the recognition failure word extraction device 501 will be described with reference to FIG. The maximum number of candidates when narrowing candidates is set as the candidate number narrowing reference value, and the number of recognition failure words to be extracted is 1.
In the example of FIG. 17, the to-be-estimated word list L _input , the candidate number narrowing reference value, and the extraction prevention word list L _{off shown} in FIG.

In the loop starting from step 602 shown in FIG. 6, the target word _Wi is composed of words other than those included in the predetermined extraction prevention word list, and the estimated word list is shown in the first column of FIG. It becomes each word L _{input the,} as the corresponding word list _{L i} shown in the second row, a list of words, except for the target word _{W i} from the estimated word list _{L input the.}
The reference value narrowing down word list L _i and the number of candidates of speech recognition rate corresponding to the word list L _i input to the speech recognition performance estimating apparatus 100, shows the resulting speech recognition performance in the third row of FIG. 17. In this example, since the speech recognition rate of the word list of “morning bandit cow” obtained by excluding “red” is the highest, the recognition failure word corresponding to this word list is “red”. When all the words given in the estimated word list L _input are words included in the extraction prevention word list, the recognition failure word extraction operation is not executed, and the recognition failure word extraction result outputs “pear”. .

The speech recognition performance estimation device and the recognition failure word extraction device according to the present invention described above install the speech recognition performance estimation program and the recognition failure word extraction program according to the present invention in a computer, and these are installed in a central processing unit provided in the computer. This is realized by deciphering and executing the program.
The speech recognition performance estimation program and the recognition performance estimation program according to the present invention are described in a computer-readable program language, and are recorded on a recording medium such as a magnetic disk or a CD-ROM that can be read by a computer. Or it is installed in a computer through a communication line.

  The speech recognition performance estimation method, speech recognition estimation device, speech recognition estimation program, recognition failure word extraction method, recognition failure word extraction device, and recognition failure word extraction program according to the present invention are used in the design field such as an automatic guidance system using speech. Be utilized.

The block diagram for demonstrating the whole speech recognition performance estimation apparatus by this invention. The block diagram for demonstrating the structure of the word misrecognition score calculation means shown in FIG. The block diagram for demonstrating the internal structure of the word similarity calculation means shown in FIG. The figure for demonstrating a mode that the target word phoneme string and an opposing word phoneme string are matched in the phoneme alignment part shown in FIG. The block diagram for demonstrating the structure of the recognition impairment word extraction apparatus of this invention. The flowchart for demonstrating operation | movement of the recognition disorder | damage | failure word search apparatus shown in FIG. The figure for demonstrating an example of the to-be-estimated word list | wrist input into the speech recognition performance estimation apparatus of this invention. The figure for demonstrating the mode inside the phoneme confusion matrix shown in FIG. The figure for demonstrating the example of the target word and word list handed over to the word misrecognition score calculation means shown in FIG. The figure for demonstrating the list | wrist which matched the target word and important conflict word output from the important conflict word extraction part shown in FIG. The figure for demonstrating the example of the distance between words of an object word and an important conflict word. The figure for demonstrating the conversion example of the distance between words shown in FIG. The figure for demonstrating the example of the word recognition score which the score conversion part shown in FIG. 1 outputs. The figure for demonstrating operation | movement of the phoneme string production | generation part shown in FIG. The figure for demonstrating operation | movement of the phoneme continuation time provision part shown in FIG. The figure for demonstrating operation | movement of the phoneme alignment part shown in FIG. The figure for demonstrating the operation | movement condition of the recognition impairment word extraction apparatus by this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 Speech recognition performance estimation apparatus 304 Dictionary 101 Speech recognition rate estimation means 305 Phoneme duration data 102 Average word misrecognition score calculation means 306 Phoneme confusion matrix 103 Word misrecognition score calculation means 501 Recognition failure word extraction apparatus 104 Word similarity Calculation means 502 Cognitive impairment word search device 201 Important allele extraction unit 202 Interword distance calculation unit 203 Score conversion unit 301 Phoneme sequence generation unit 302 Phoneme duration addition unit 303 Phoneme alignment unit

Claims (19)

On the computer,
A word similarity calculation process for calculating a word similarity indicating a degree of recognition of an utterance corresponding to the target word as the opposite word when speech recognition is performed when two words including the target word and the opposite word are received; ,
Word misrecognition score calculation that receives the target word and the word list and calculates a word misrecognition score indicating the degree to which the utterance corresponding to the target word is recognized as any word in the word list when speech recognition is performed Processing,
An average word misrecognition score calculation process for receiving the word list and calculating an average average word misrecognition score of all words included in the word list;
Estimating speech recognition rate using as input the number of words included in the word list, the average word misrecognition score, and the candidate number narrowing reference value set when searching for solution candidates in speech recognition processing A speech recognition rate estimation process for calculating a value;
The speech recognition performance estimation method characterized by performing this. The speech recognition performance estimation method according to claim 1,
In the word similarity calculation process, each phoneme of the phoneme string corresponding to the pronunciation of the target word is assumed to continue only the average duration of the phoneme, and each continuation of the phoneme string corresponding to the pronunciation of the opposite word The time is the probability that the sum of the durations of the phonemes of the opposing word is equal to the sum of the durations of the phonemes of the target word, and that the phonemes of the target word are recognized by the phonemes of the opposing words at the same time Is determined so that the probability of integration at all times becomes the highest, and the value obtained by normalizing the integrated probability value with the sum of durations of each phoneme of the target word is used as the word similarity. Speech recognition performance estimation method. In any one of the speech recognition performance estimation methods of Claim 1 or 2,
The word misrecognition score calculation process extracts a certain number of words having a high word similarity with the target word from the word list, and for each of the extracted words, the word similarity between the target word and the target word itself The difference value obtained by subtracting the logarithmic value of the word similarity between each word and the target word is calculated from the logarithm value of the degree, and this is used as the interword distance. The smaller the interword distance value, the larger the value. A speech recognition performance estimation method characterized in that a converted value is calculated and a sum of converted values corresponding to each extracted word is used as a word error recognition score. The speech recognition performance estimation method according to any one of claims 1 to 3,
The average word error recognition score calculation processing is performed for each word included in the estimated word list when the word is the target word and all remaining words included in the estimated word list are new word lists. Speech recognition characterized by calculating a recognition score and dividing the sum of the word error recognition scores of each word included in the new word list by the number of words in the estimated word list as an average word error recognition score Performance estimation method. The speech recognition performance estimation method according to any one of claims 1 to 3,
The average word misrecognition score calculation processing is a word error when each word included in the estimated word list is the target word and all the remaining words included in the estimated word list are the new word list. A speech recognition performance estimation method characterized in that a recognition score is calculated and a sum of the occurrence probabilities of each word included in the estimated word list multiplied by the word error recognition score of the word is used as an average word error recognition score . The speech recognition performance estimation method according to any one of claims 1 to 5,
The candidate number narrowing reference value set when searching for speech recognition solution candidates is set as the maximum number of solution candidates listed at each time point of the search,
The speech recognition performance estimation process is a weighted value obtained by dividing the number of words included in the estimated word list, the average word error recognition score, and the number of words included in the estimated word list by the maximum number of solution candidates. A speech recognition performance estimation method, characterized in that a linear sum plus a constant is used as an estimated speech recognition rate. On the computer,
Executing any one of the recognition failure word search processing and the speech recognition performance estimation method according to claim 1,
In the recognition failure word search process, for each word included in the estimated word list given by input, speech recognition of a new word list composed of all remaining words excluding the word from the estimated word list Estimating the performance by the speech recognition performance estimation method, extracting a certain number of the above new word lists in descending order of speech recognition performance, and outputting the removed words corresponding to each new word list as recognition impairment words A method for extracting recognition impairment words as a feature. The recognition impairment word extraction method according to claim 7,
The recognition failure word search process includes an extraction prevention word list in the input, and for each word other than the extraction prevention word in the estimated word list given by input, the remaining words obtained by removing the word from the estimation word list Estimate the speech recognition performance of a new word list composed of all words using the speech recognition performance estimation method, extract a certain number of word lists in descending order of speech recognition performance, and remove the words corresponding to each word list Is extracted as a recognition failure word. A word similarity calculating means for calculating a word similarity indicating a degree that an utterance corresponding to the target word is recognized as the opposite word when speech recognition is performed using two words of the target word and the opposite word;
A word for calculating a word misrecognition score indicating the degree to which the utterance corresponding to the target word is recognized as one of the words included in the word list when the target word and the word list are input and speech recognition is performed. Misrecognition score calculation means;
The average word error recognition score calculating means for calculating the average average word error recognition score of all the words included in the word list, using the word list as input,
Estimated value of speech recognition rate using as input the number of words included in the word list, the average word error recognition score, and the candidate number narrowing reference value set when searching for solution candidates in speech recognition processing Speech recognition rate estimating means for calculating
A speech recognition performance estimation apparatus comprising: The speech recognition performance estimation apparatus according to claim 9,
The word similarity calculating means assumes that each phoneme of the phoneme string corresponding to the pronunciation of the target word lasts for an average duration of the phoneme, and each phoneme of the phoneme string corresponding to the pronunciation of the opposite word As for the duration, the sum of the durations of the phonemes of the opposing word is equal to the sum of the durations of the phonemes of the target word, and the phonemes of the target word at the same time point are recognized as phonemes of the opposing word. It is determined that the probability obtained by integrating the probabilities at all points in time is the highest, and the value obtained by normalizing the value of the integrated probability with the sum of durations of each phoneme of the target word is used as the word similarity. A speech recognition performance estimation device. In any of the speech recognition performance estimation apparatuses according to claim 9 or 10,
The word misrecognition score calculating means extracts a certain number of words having a large word similarity with the target word from the word list, and for each of the extracted words, the word similarity between the target word and the target word itself The difference value obtained by subtracting the logarithmic value of the word similarity between each word and the target word is calculated from the logarithm value of the degree, and this is used as the interword distance. The smaller the interword distance value, the larger the value. A speech recognition performance estimation apparatus characterized by calculating a converted value so that the sum of converted values corresponding to each extracted word is used as a word error recognition score. The speech recognition performance estimation device according to any one of claims 9 to 11,
The average word misrecognition score calculation means, for each word included in the estimated word list, the word when the word is the target word and all the remaining words included in the estimated word list are the new word list Speech recognition characterized by calculating a misrecognition score and dividing the sum of the word misrecognition scores of each word included in the new word list by the number of words in the estimated word list as an average word misrecognition score Performance estimation device. The speech recognition performance estimation device according to any one of claims 9 to 12,
The average word misrecognition score calculation means, for each word included in the estimated word list, the word when the word is the target word and all the remaining words included in the estimated word list are the new word list A speech recognition performance estimation device characterized by calculating a misrecognition score and using the sum of the occurrence probability of each word included in the estimated word list multiplied by the word misrecognition score of the word as an average word misrecognition score . The speech recognition performance estimation device according to any one of claims 9 to 12,
The candidate number narrowing reference value set when searching for speech recognition solution candidates is set as the maximum number of solution candidates listed at each time point of the search,
The speech recognition performance estimation device is weighted by a value obtained by dividing the number of words included in the estimated word list, the average word misrecognition score, and the number of words included in the estimated word list by the maximum number of solution candidates. An apparatus for estimating speech recognition performance, wherein an estimated speech recognition rate is obtained by adding a constant to a linear sum. On the computer,
A recognition failure word search means; and any one of the speech recognition performance estimation devices according to claims 9 to 14,
The recognition failure word search means, for each word included in the estimated word list given by input, voice recognition of a new word list composed of all remaining words excluding the word from the estimated word list A recognition failure characterized by estimating performance by a speech recognition performance estimation device, extracting a certain number of word lists in descending order of speech recognition performance, and outputting the removed words corresponding to each word list as recognition failure words Word extraction device. The recognition impairment word extracting device according to claim 15,
The recognition failure word search means includes an extraction prevention word list in the input, and excludes the word from the estimated word list for each word other than the extraction prevention word among the words included in the estimation word list given by the input. The speech recognition performance estimation device estimates the speech recognition performance of a new word list composed of all the remaining words, extracts a certain number of word lists in descending order of speech recognition performance, and removes the corresponding word lists. A recognition failure word extraction apparatus, wherein the recognition word is output as a recognition failure word.   A speech recognition performance estimation program that is described in a program language that can be read by a computer and causes the computer to function as at least the speech recognition performance estimation device according to any one of claims 9 to 14.   A recognition failure word extraction program, which is described in a computer-readable program language, and causes the computer to function as at least a recognition failure word extraction device according to claim 15.   A recording medium comprising a computer-readable recording medium and recording either the speech recognition performance estimation program according to claim 17 or the recognition failure word extraction program according to claim 18.

Images