JP2002297181A - Speech recognition vocabulary registration determination method and speech recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To prevent the registration of vocabularies liable to give rise to erroneous recognition. SOLUTION: The numerical value to evaluate the easy tendency to mingling of the new vocabulary and the already registered vocabulary is calculated by evaluation value calculating means 5. Deciding means 6 sets the threshold to the numerical value given from the evaluation value calculating means 5 and decides whether the registration of the new vocabulary is permitted or prohibited.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION The present invention relates to a method for preventing a decrease in recognition performance by selecting a vocabulary (word or phrase, or a combination thereof) that is likely to cause erroneous recognition when selecting or registering a vocabulary to be subjected to speech recognition. Regarding the detection technology, it is useful for a user of the speech recognition system to register a new vocabulary in the system, and for a developer of the speech recognition system to select a recognition target vocabulary to be registered in the system.

[0002]

2. Description of the Related Art In a task-limited voice recognition device, a personal registration type voice portal device, a center-provided voice portal device, and the like, such as an airline ticket ticket reservation system and an automatic telephone reception system, a voice input type word processor is used. As compared with the recognition system, the input speech is of poor quality due to telephone input and the number of tasks is limited, so that the number of vocabulary words to be recognized is often limited to prevent a reduction in recognition performance. That is, only a limited vocabulary corresponding to the task is recognized.

In this type of speech recognition system, usually, a system developer empirically selects a limited vocabulary. For example, system developers have known from experience that the vocabulary "Uncle" and the vocabulary "Grandfather" are easily misunderstood, so if "Uncle" is registered, "Grandfather" is unsuitable. It is excluded and is not registered. However, since it depends on experience, the efficiency of selecting a vocabulary for speech recognition is not good.

On the other hand, it is conceivable that a user of the system can customize the vocabulary to be recognized. However, since the user has little experience, there is a risk that a new vocabulary that is easily confused with a registered vocabulary (confusing) is registered, and the recognition performance is reduced. For example, even if “Uncle” has already been registered, it is conceivable that “Uncle” is additionally registered as a new vocabulary unconsciously. Therefore, it is practically difficult for the user to increase the vocabulary to be recognized.

[0005]

SUMMARY OF THE INVENTION It is an object of the present invention to detect a vocabulary which is likely to cause erroneous recognition when selecting or registering a vocabulary to be recognized, and to prevent the vocabulary from being registered. As a result, for system developers,
The recognition target vocabulary can be efficiently selected. Further, for the user, it is possible to customize the vocabulary to be recognized, for example, without increasing the recognition performance.

[0006]

The invention according to claim 1 is a speech recognition vocabulary registration / judgment method, which quantifies the ease of confusion between a new vocabulary and a registered vocabulary for a speech recognition device,
A threshold value is set for the obtained numerical value (hereinafter, evaluation value), and it is determined whether registration of a new vocabulary is permitted or prohibited.

According to a second aspect of the present invention, in the first aspect, characters or character strings of the new vocabulary and the registered vocabulary are decomposed into recognition units such as phonemes and syllables to correspond to the newly registered vocabulary. Speech recognition characterized by obtaining a first recognition unit sequence and a second recognition unit sequence corresponding to a registered vocabulary, and obtaining a similarity between the first and second recognition unit sequences as the evaluation value. This is a vocabulary registration determination method.

A third aspect of the present invention is the speech recognition vocabulary registration determination according to the second aspect of the present invention, wherein the similarity is obtained by taking a sum or a product of scales representing distances between recognition units. Method.

According to a fourth aspect of the present invention, there is provided the voice recognition vocabulary registration according to the second aspect, wherein the similarity is obtained by using a previously obtained confusion matrix between the recognition units. This is a determination method.

The invention according to claim 5 is the invention according to claim 2, wherein the recognition unit is HMM (Hidden Markov Model).
Is a speech recognition vocabulary registration / judgment method characterized in that the similarity is obtained by using a distribution in an acoustic parameter space expressing the expression.

According to a sixth aspect of the present invention, in the first aspect of the invention, a pseudo acoustic parameter sequence corresponding to a character or a character string of a new vocabulary is generated, and the generated acoustic parameter sequence is subjected to a speech recognition process. And determining the evaluation value from the obtained likelihood.

The invention according to claim 7 is the invention according to claim 6, wherein the speech recognition processing includes a first speech recognition processing using a grammar for accepting a new vocabulary, and a grammar for accepting only registered vocabulary. Perform a second speech recognition process to be used,
A speech recognition vocabulary registration / determination method characterized in that a difference between a likelihood obtained in a first speech recognition process and a likelihood obtained in a second speech recognition process is used as the evaluation value.

According to an eighth aspect of the present invention, there is provided the speech recognition vocabulary registration determination method according to the seventh aspect, wherein the acoustic parameter sequence is generated using an HMM (Hidden Markov Model).

According to a ninth aspect of the present invention, in the invention according to the eighth aspect, the acoustic parameter sequence is generated by applying perturbation using the variance of the HMM or connecting acoustic parameters of the noise HMM at random. This is a speech recognition vocabulary registration / judgment method.

A tenth aspect of the present invention is characterized in that the speech recognition vocabulary registration / judgment method according to the second aspect and the speech recognition vocabulary registration / judgment method according to the sixth aspect are used in combination. This is a speech recognition vocabulary registration determination method.

According to an eleventh aspect of the present invention, there is provided a speech recognition apparatus, comprising a speech input unit, a vocabulary registration unit, and a speech recognition unit, wherein a speech input from the speech input unit is registered in the vocabulary registration unit. In a speech recognition apparatus that recognizes words based on vocabulary (hereinafter, registered vocabulary), a character input unit for inputting a character, and the input character or character string (hereinafter, new vocabulary) is confused with a registered vocabulary. Evaluation value calculating means for calculating a numerical value (hereinafter, evaluation value) for evaluating ease of use, and setting a threshold value for the obtained evaluation value to determine whether to permit or prohibit registration of a new vocabulary It is characterized by having means.

According to a twelfth aspect of the present invention, in the invention according to the eleventh aspect, first conversion means for decomposing the new vocabulary into recognition units and generating a first recognition unit sequence corresponding to the new vocabulary, Second conversion means for decomposing the registered vocabulary into recognition units and generating a second recognition unit sequence corresponding to the registered vocabulary, wherein the evaluation value calculation means includes a first evaluation value Calculating a similarity between the first recognition unit sequence obtained by the conversion means and the second recognition unit sequence obtained by the second conversion means. It is.

According to a thirteenth aspect of the present invention, in the twelfth aspect, the evaluation value calculating means calculates the similarity by calculating a sum or a product of scales representing distances between recognition units. It is a speech recognition device that is a feature.

According to a fourteenth aspect, in the twelfth aspect, the evaluation value calculating means calculates the similarity using a confusion matrix between recognition units obtained in advance. This is a voice recognition device characterized by the following.

According to a fifteenth aspect of the present invention, in the twelfth aspect of the present invention, the evaluation value calculating means includes the recognition unit H
A speech recognition apparatus characterized in that the similarity is calculated using a distribution in an acoustic parameter space representing a MM (Hidden Markov Model).

According to a sixteenth aspect of the present invention, in the invention according to the eleventh aspect, there is provided an acoustic parameter sequence generating means for generating a pseudo acoustic parameter sequence corresponding to a new vocabulary and providing the generated sequence to the speech recognition means. The evaluation value calculating means calculates the evaluation value using the likelihood obtained from the voice recognition means.

According to a seventeenth aspect of the present invention, in the invention according to the sixteenth aspect, the speech recognition means performs a first speech recognition process using a grammar for accepting a new vocabulary, and a grammar for accepting only a registered vocabulary. Performing a second speech recognition process to be used, wherein the evaluation value calculating means calculates, as the evaluation value, a difference between a likelihood obtained in the first speech recognition process and a likelihood obtained in the second speech recognition process. Is calculated.

According to an eighteenth aspect of the present invention, in the invention according to the seventeenth aspect, the acoustic parameter sequence generating means includes:
This is a speech recognition device that generates an acoustic parameter sequence using an HMM (Hidden Markov Model).

According to a nineteenth aspect, in the invention according to the eighteenth aspect, the acoustic parameter sequence generating means includes:
A speech recognition apparatus characterized in that a perturbation is given by using the variance of the HMM, or an acoustic parameter sequence is generated by randomly connecting acoustic parameters of a noise HMM.

The invention according to claim 20 is the invention according to claims 1 to 1
0 is a computer program for causing a computer to execute the speech recognition vocabulary registration determination method according to any one of the inventions.

According to a twenty-first aspect of the present invention, there is provided a recording medium storing the computer program according to the twenty-first aspect.

[0027]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of the present invention will be described.

Referring to FIGS. 1 and 2, a basic speech recognition apparatus and a speech recognition vocabulary registration / judgment method of the present invention will be described.
FIG. 1 shows a configuration example of a speech recognition device. This apparatus has a voice input unit 1, a vocabulary registration unit 2, a voice recognition unit 3, a character input unit 4, an evaluation value calculation unit 5, and a determination unit 6, and mainly includes a computer program and a computer program for executing the computer program. Computer.

The voice input means 1 is for giving a voice signal to the voice recognition means 3, and for example, a voice signal input means for inputting a voice signal from outside, such as a telephone, and a microphone which converts voice itself into a signal can be used. It is.

The vocabulary registration means 2 registers a vocabulary to be recognized.

The voice recognition means 3 recognizes the voice input from the voice input means 1 based on the registered vocabulary of the vocabulary registration means 2.

The character input means 4 provides a new vocabulary as a character-based character or character string to the evaluation value calculating means 5 and the vocabulary registration means 2, and includes, for example, a CUI type input means such as a keyboard, a character recognition means, a GUI. Type input means can be used. If the new vocabulary is a kanji, an alphanumeric character, a symbol, or the like, its reading is given to the evaluation value calculation means 5.

The evaluation value calculation means 5 calculates a numerical value (evaluation value) for evaluating the degree of confusion between the new vocabulary given from the character input means 4 and the vocabulary registered in the vocabulary registration means 2 (confusingness). Then, the obtained evaluation value is given to the judgment means 6. This quantifies the ease of confusion.

The judgment means 6 sets a threshold value for the evaluation value given by the evaluation value calculation means 5 and judges whether registration of a new vocabulary is permitted or prohibited. This determination result is given to the vocabulary registration means 2. The threshold value can be set by obtaining it in advance through experiments or the like.

The vocabulary registration means 2 registers the new vocabulary given from the character input means 4 as the recognition target vocabulary if the result of the determination is permission, and does not register the new vocabulary if the registration is prohibited. If the registration is prohibited, this is notified by warning means 7 such as visual or auditory means.

Therefore, in the vocabulary registration in the speech recognition apparatus shown in FIG. 1, as shown in FIG. 2, the ease of confusion between a new vocabulary and a registered vocabulary is first quantified (step S1). A threshold value is set for the obtained evaluation value, and it is determined whether registration of a new vocabulary is permitted or prohibited (step S2). If registration is permitted, a new vocabulary is registered (step S3), and if registration is prohibited, a warning is issued (step S4).

The computer of the speech recognition apparatus reads the computer program and executes such a procedure. In other words, the procedure to be executed by the computer is described in the computer program. The computer can read the computer program recorded on an appropriate recording medium, or can read it through an appropriate transmission means such as a network.

The numerical value of the degree of confusion in the evaluation value calculation means 5 can be achieved, for example, by performing DP matching or pseudo speech recognition.

In the case of DP matching, the characters or character strings of the new vocabulary and the registered vocabulary are decomposed into recognition units such as phonemes and syllables, respectively. , And a similarity between the first and second recognition unit sequences is obtained. This similarity is an evaluation value of the degree of confusion between the new vocabulary and the registered vocabulary.

In a simple example, matching between recognition units, replacement,
Scores (scales or distance scales) representing distances for deletion and insertion are set as fixed values irrespective of combinations of recognition units, and the matching score (usually 0), replacement score, deletion score, and insertion score are used. DP matching is performed between the two recognition unit sequences. For example, a path having the minimum total score (sum of scores) is searched, and the total score of the path is set as the similarity between the two recognition unit sequences. When none of the match score, the replacement score, the deletion score, and the insertion score are 0, the minimum product score (product of the scores) can be used as the similarity between the two recognition unit sequences. FIG. 3 shows a configuration example of a voice recognition device that realizes this method. However, only the parts necessary for the evaluation value calculation are shown, and the voice input means and the voice recognition means are not shown. The computer of the speech recognition device also reads the computer program and executes the above procedure. The computer program describes a procedure to be executed by the computer.

The voice recognition device shown in FIG.
A first conversion unit 9 and a second conversion unit 10 are additionally provided. The storage unit 8 stores a match score, a replacement score, a deletion score, and an insertion score between recognition units as fixed values.

The first conversion means 9 obtains a new vocabulary from the character input means 4, breaks it down into recognition units, generates a first recognition unit sequence corresponding to the new vocabulary, and gives it to the evaluation value calculation means 5. . The second conversion unit 10 acquires the registered vocabulary from the vocabulary registration unit 2 and decomposes the vocabulary into recognition units, generates a second recognition unit sequence corresponding to the registered vocabulary, and generates an evaluation value calculation unit 5.
Give to. The evaluation value calculation means 5 uses the fixed matching score, replacement score, deletion score, and insertion score between the recognition units stored in the storage means 8 to obtain the first recognition unit obtained from the first conversion means 9. By calculating the DP matching between the sequence and the second recognition unit sequence obtained from the second conversion means 10, the minimum sum score or the minimum product score is calculated by the similarity between these two recognition unit sequences. (Evaluation value of ease of confusion between new vocabulary and registered vocabulary) is calculated. In this case, when the evaluation value (similarity) is larger than the set threshold value, the determination means 6 permits the registration of a new vocabulary,
If it is smaller, registration will be prohibited.

In the above example, the matching score, the replacement score, the deletion score, and the insertion score between the recognition units are fixed values, but the degree of confusion between the recognition units differs depending on the combination of the recognition units. Therefore, the match score between the recognition units,
By setting the replacement score, the deletion score, and the insertion score to values according to the combination of the recognition units, the similarity (evaluation value) can be obtained more accurately.

As an example, using a confusion matrix of recognition units such as phonemes and syllables, the degree of confusion between recognition units is quantified and converted into a DP.
An example of performing matching will be described.

A confusion matrix as shown in FIG. 4 is obtained by conducting a recognition experiment on recognition units such as a phoneme recognition experiment and a syllable recognition experiment in advance.
For simplicity, the two recognition units are a and b.

In FIG. 4, the event of the recognition unit on the vertical axis indicates the number of recognition units recognized on the horizontal axis, the symbol del indicates the number of events of the recognition unit on the vertical axis deleted, and the symbol ins indicates the number of events. Indicates the number of axis recognition units inserted. For example, the value 80 indicated by the vertical axis a and the horizontal axis a indicates that the event of the recognition unit a is the recognition unit a
The value 15 indicated by the vertical axis a and the horizontal axis b is the number of events recognized in the recognition unit a recognized as the recognition unit b, that is, the number replaced. The value 5 indicated by the vertical axis a and the horizontal axis del is the number of events of the recognition unit a deleted, and the value 30 indicated by the vertical axis ins and the horizontal axis a is the number of recognition units a inserted. is there.

From this confusion matrix,
It is possible to define a score that indicates the degree of confusion between recognition units. One example is shown by the following equations (1) to (8). hit (aa) = log (80 / (80 + 15 + 5)) ... Equation (1) sub (ab) = log (15 / (80 + 15 + 5)) ... Equation (2) del (a) = log (5 / (80 + 15 + 5))… Equation (3) hit (bb) = log (70 / (20 + 70 + 10))… Equation (4) sub (ba) = log (20 / (20+ 70 + 10)) ... Equation (5) del (b) = log (10 / (20 + 70 + 10)) ... Equation (6) ins (a) = log (30 (80 + 20 + 30)) ... Equation (7) ins (b) = log (5 (15 + 70 + 5)) Equation (8) where hit (aa) is the matching score (the number of correct answers is Sub (ab) is the replacement score (the logarithm of the ratio of the number of replacements to the number of cases) at which the event of the recognition unit a is recognized as the recognition unit b, and del (a) is the recognition unit a The deletion score (the logarithm of the ratio of the number of deletions to the number of cases) at which events are deleted, and hit (bb) is the matching score at which the event of the recognition unit b is recognized as the recognition unit b (the ratio of the number of correct answers to the number of cases) , Sub (ba) is the replacement score (the logarithm of the ratio of the number of replacements to the number of cases) at which the event of the recognition unit b is recognized as the recognition unit a, and del (b) is The deletion score at which the event of the recognition unit b is deleted (the logarithm of the ratio of the number of deletions to the number of cases), and ins (a) is the insertion score at which the recognition unit a is inserted (the logarithm of the ratio of the number of insertions to the number of observations) ,
ins (b) is the insertion score at which the recognition unit b is inserted (the logarithm of the ratio of the number of insertions to the number of observations).

A match score, a replacement score, a deletion score, and an insertion score are determined by using such a confusion matrix, and DP matching is performed using the determined scores.
Similarity between two recognition unit series (evaluation value of ease of confusion)
Ask for.

However, in the examples shown in the above equations (1) to (8), the matching score, the replacement score, the deletion score and the insertion score indicate not the distance between the recognition units but the degree of confusion. In DP matching, the maximum sum score or product score is used as the evaluation value (similarity).

In order to use the minimum sum score or product score as an evaluation value (similarity), DP matching is performed using a score in which the sign is inverted as shown in the following equations (9) to (16). Good. In this case, a path having the minimum total score or product score is searched for, and the total score or product score of the path becomes the evaluation value (similarity). hit (aa) =-log (80 / (80 + 15 + 5)) ... Equation (9) sub (ab) =-log (15 / (80 + 15 + 5)) ... Equation (10) del (a) = -log (5 / (80 + 15 + 5))… Equation (11) hit (bb) =-log (70 / (20 + 70 + 10))… Equation (12) sub (ba) =-log ( 20 / (20 + 70 + 10)) ... Equation (13) del (b) = -log (10 / (20 + 70 + 10)) ... Equation (14) ins (a) = -log (30 (80+ 20 + 30))… Equation (15) ins (b) = -log (5 (15 + 70 + 5))… Equation (16)

FIG. 5 shows an example of the configuration of a speech recognition apparatus for realizing this method. However, only the parts necessary for the evaluation value calculation are shown, and voice input means, voice recognition means, warning means, and the like are not shown. The computer of the speech recognition device also reads the computer program and executes the above procedure.
The computer program describes a procedure to be executed by the computer.

The voice recognition device shown in FIG.
A first conversion unit 9, a second conversion unit 10, and a score calculation unit 11 are additionally provided. The score calculation means 11 uses a confusion matrix and calculates the above equations (9) to (1).
The match score, replacement score, deletion score, and insertion score are calculated as in 6), and these are stored in the storage means 8.

The first conversion means 9 obtains a new vocabulary from the character input means 4, breaks it down into recognition units, generates a first recognition unit sequence corresponding to the new vocabulary, and gives it to the evaluation value calculation means 5. . The second conversion unit 10 acquires the registered vocabulary from the vocabulary registration unit 2 and decomposes the vocabulary into recognition units, generates a second recognition unit sequence corresponding to the registered vocabulary, and generates an evaluation value calculation unit 5.
Give to. The evaluation value calculation means 5 uses the matching score, the replacement score, the deletion score, and the insertion score between the recognition units stored in the storage means 8 to obtain the first recognition unit obtained from the first conversion means 9. By calculating the DP matching between the sequence and the second recognition unit sequence obtained from the second conversion means 10, the minimum sum score or the minimum product score is calculated by the similarity between these two recognition unit sequences. (Evaluation value of ease of confusion between new vocabulary and registered vocabulary) is calculated.
In this case, the judgment unit 6 permits registration of a new vocabulary when the evaluation value (similarity) is larger than the set threshold value, and prohibits registration when the evaluation value (similarity) is smaller than the set threshold value.

As another example, a score between recognition units can be obtained using acoustic parameters of a HMM (Hidden Markov Model) representing a recognition unit, such as a phoneme model or a syllable model, instead of the confusion matrix. For example, if the HMM is a discrete HMM, the distance between codebooks can be used. If the HMM is a continuous distribution HMM, the distance between mean vectors of the distribution can be used.

Here, an example of the distance calculation for obtaining the score between the recognition units from the acoustic parameters of the HMM will be described. For example,
The recognition unit a and the recognition unit b are three-state HMMs, and the second state of each is a mixed normal distribution that outputs the probabilities P _a and P _b represented by the following equations (17) and (18). In this case, the distance D (a, b) between the two states is calculated by the following equation (19) or
Defined in (20). Here, d (a _j , b _j ) represents the distance between distributions. Equation (20) defines the minimum value of the inter-distribution distance as D (a, b), but the maximum or average value of the inter-distribution distance may be D (a, b).

[0056]

(Equation 1)

When a score between recognition units is obtained from the acoustic parameters of the HMM representing the recognition unit, DP matching is performed using this score. That is, similarly to the case based on the confusion matrix, a path having the smallest total score or product score is searched, and the total score or product score of the path is evaluated (similarity).

In this case, if a score between recognition units is obtained from the acoustic parameters of the HMM representing the recognition unit, a score for deletion and insertion of the recognition unit cannot be obtained. Therefore,
As the DP matching, a DP matching that prohibits deletion and insertion of a recognition unit and allows a stay instead is used.

FIG. 6 shows an example of the configuration of a speech recognition apparatus for realizing this method. However, only the parts necessary for the evaluation value calculation are shown, and the voice recognition means and the like are not shown. The computer of the speech recognition device also reads the computer program and executes the above procedure. The computer program describes a procedure to be executed by the computer.

The voice recognition device shown in FIG.
A first conversion unit 9, a second conversion unit 10, and a score calculation unit 11 are additionally provided. The score calculation means 11 calculates scores between recognition units from acoustic parameters of the HMM representing the recognition units, and these are stored in the storage means 8.

The first conversion means 9 obtains a new vocabulary from the character input means 4, breaks it down into recognition units, generates a first recognition unit sequence corresponding to the new vocabulary, and gives it to the evaluation value calculation means 5. . The second conversion unit 10 acquires the registered vocabulary from the vocabulary registration unit 2 and decomposes the vocabulary into recognition units, generates a second recognition unit sequence corresponding to the registered vocabulary, and generates an evaluation value calculation unit 5.
Give to.

The evaluation value calculation means 5 uses the score between the recognition units stored in the storage means 8 and the first conversion means 9
The first recognition unit sequence obtained from
DP that prohibits deletion and insertion of recognition units and allows stagnation instead of the second recognition unit sequence obtained from 0
By calculating the matching, the minimum sum score or the minimum product score is obtained as the similarity between these two recognition unit sequences (evaluation value of the ease of confusion between the new vocabulary and the registered vocabulary). In this case, the judgment unit 6 permits registration of a new vocabulary when the evaluation value (similarity) is larger than the set threshold value, and prohibits registration when the evaluation value (similarity) is smaller than the set threshold value.

Next, an example will be described in which pseudo speech recognition is performed to quantify the degree of confusion between a new vocabulary and a registered vocabulary. In order to perform pseudo speech recognition, a pseudo acoustic parameter sequence corresponding to a character or character string of a new vocabulary is generated. Speech recognition processing is performed on this pseudo acoustic parameter sequence to determine likelihood. Since an evaluation value is obtained from the obtained likelihood, it is determined whether registration of a new vocabulary is permitted or prohibited by using the evaluation value. When pseudo speech recognition is performed, acoustic features can be reflected more accurately than when an evaluation value of ease of confusion is obtained using scores between recognition units.

A pseudo acoustic parameter sequence can be generated using acoustic parameters of a recognition unit such as a phoneme model or a syllable model.

For example, the acoustic parameters of the HMM representing the recognition unit are used. In this case, the new vocabulary is decomposed into recognition units such as phonemes and syllables, and an HMM corresponding to each recognition unit is decomposed.
By concatenating the acoustic parameters that express, a pseudo acoustic parameter sequence can be created. If it is a discrete HMM, a continuous distribution HM
If it is M, the average value of the distribution expressing the acoustic feature can be used.

At this time, the variation of the acoustic feature can be verified by perturbing the acoustic parameter sequence using the variance of the HMM, for example, the variance of the distribution expressed by the acoustic parameters of the HMM. Also, the noise HM
By randomly connecting the M acoustic parameters to the acoustic parameter sequence, it is possible to verify the effect of the noise mixing. As the number of frames corresponding to each recognition unit, the reciprocal of the self-transition probability of the HMM state may be used.

Referring to FIG. 7, an example of generating a pseudo acoustic parameter sequence will be described.

First, the silence model "sil" and the phoneme model "p"
phonemes (HMM state) using h1 "," ph2 ",..., noise model N1,.
Specify a series. As a result, an acoustic parameter sequence corresponding to the phoneme (HMM state) sequence is generated as shown in FIG.

In FIG. 7, the phoneme series is designated and designated.
It can also be specified in the state series. Also, as described above, perturbation can be given to the acoustic parameter sequence. Furthermore, when generating an acoustic parameter sequence from the mixture distribution HMM, the center of gravity of the mixture distribution is used, the average vector of the distribution with the largest distribution weight is used, the average vector of the distribution picked up from the mixture distribution by Dam, and the like are used. It is possible.

When the pseudo acoustic parameter sequence as described above is created, a speech recognition process is performed on the pseudo acoustic parameter sequence. In this case, (1) speech recognition processing using grammar that accepts only registered vocabulary, (2) speech recognition processing using grammar that accepts only registered vocabulary and speech recognition using grammar that accepts only new vocabulary (3) A method of performing speech recognition using a grammar that accepts both registered vocabulary and new vocabulary. In the method (1), the likelihood itself obtained by the speech recognition processing can be used as an evaluation value. If the likelihood is smaller than a set threshold, registration of a new vocabulary is permitted. Should be banned. In the method (2), the difference between the likelihood of the speech recognition process performed by the grammar that accepts only the new vocabulary and the likelihood of the speech recognition process performed by the grammar that accepts only the registered vocabulary is used as the evaluation value. If the difference between the likelihoods is larger than a set threshold, registration of a new vocabulary is permitted, and if smaller, registration is prohibited. In the method (3), the difference between the maximum likelihood obtained in the speech recognition processing and the second largest likelihood can be used as the evaluation value, and the difference between the likelihoods is determined by the set threshold value. If it is large, registration of a new vocabulary is permitted, and if it is small, registration is prohibited.

FIG. 8 shows an example of the configuration of a speech recognition apparatus for realizing the above method (2). However, only the parts necessary for the evaluation value calculation are shown, and voice input means, warning means, and the like are not shown. The computer of the speech recognition device also reads the computer program and executes the above procedure. The computer program describes a procedure to be executed by the computer.

The speech recognition apparatus shown in FIG. 8 additionally has an acoustic parameter generation unit 12 and a storage unit 13.

The acoustic parameter generating means 12 receives a character string of a new vocabulary from the voice input means 1, generates a corresponding pseudo acoustic parameter sequence using the HMM as described above, and sends it to the voice recognizing means 3. Is to give. At this time, a perturbation is given using the variance of the HMM, or the acoustic parameters of the noise HMM are randomly connected to generate an acoustic parameter sequence.

The storage unit 13 stores a grammar that accepts only new vocabulary and a grammar that accepts only registered vocabulary. The speech recognition means 3 performs a first speech recognition process 3a using a grammar for accepting a new vocabulary for a pseudo acoustic parameter, and a second speech grammar for accepting only a registered vocabulary.
And the likelihood of each is given to the evaluation value calculation means 5.

The evaluation value calculating means 5 performs the first speech recognition processing 3
The difference between the likelihood obtained in a and the likelihood obtained in the second speech recognition process 3b is calculated, and this likelihood difference is evaluated. In this case, the judgment means 6 permits registration of a new vocabulary when the evaluation value is larger than the set threshold value, and prohibits registration when the evaluation value is smaller than the set threshold value.

The above-described various speech recognition vocabulary registration / judgment methods can be used alone or in any combination. For example, all or four or more of the four methods shown in the following (1) to (4) are used in combination. (1) A method of determining a score between recognition units using a confusion matrix and performing DP matching using the score to obtain an evaluation value. (2) A method in which a score between recognition units is determined using a distribution in the acoustic parameter space expressing the HMM of the recognition unit, and DP matching is performed using the score to obtain an evaluation value. (3) A method of generating a pseudo acoustic parameter sequence corresponding to a new vocabulary using an HMM, and performing a speech recognition process on the sequence to obtain an evaluation value. (4) In generating a pseudo acoustic parameter sequence by the method of (3) above, by applying perturbation using the variance of the HMM or by randomly connecting the acoustic parameters of the noise HMM, fluctuations in generation and noise immunity are generated. Technique to evaluate gender.

When combining a plurality of such methods, the methods (3) and (4) are more accurate than the methods (1) and (2), but the software processing by the computer requires (3) Methods (4) and (4) require more processing. Therefore, in such a case, (1) or (2) or both methods are performed with a strict threshold set, and only those that are judged at this time are (3) or (4) or both. It is efficient to judge by the method.

[0078]

According to the present invention, at the time of selecting or registering a vocabulary to be recognized, it is possible to prevent the registration of vocabulary such as words and phrases that are likely to cause erroneous recognition.
This allows the system developer to efficiently select the vocabulary to be recognized. Further, for the user, it is possible to customize the recognition performance such as increasing the vocabulary to be recognized while suppressing the reduction in the recognition performance. The present invention having such effects is useful for a task-restricted voice recognition device, a personal registration type voice portal device, a center-provided voice portal device, and the like.

[Brief description of the drawings]

FIG. 1 is a diagram showing a configuration example of a speech recognition device according to an embodiment of the present invention.

FIG. 2 is a diagram showing an example of a speech recognition vocabulary registration determination method according to the embodiment of the present invention.

FIG. 3 is a diagram showing a configuration example of another voice recognition device according to the embodiment of the present invention.

FIG. 4 is a diagram showing an example of a confusion matrix.

FIG. 5 is a diagram showing a configuration example of another speech recognition device according to the embodiment of the present invention.

FIG. 6 is a diagram showing a configuration example of another voice recognition device according to the embodiment of the present invention.

FIG. 7 is a diagram showing an example of generating a pseudo acoustic parameter sequence.

FIG. 8 is a diagram showing a configuration example of another voice recognition device according to the embodiment of the present invention.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 voice input means 2 vocabulary registration means 3 voice recognition means 3a first voice recognition processing 3b second voice recognition processing 4 character input means 5 evaluation value calculation means 6 determination means 7 warning means 8 storage means 9 first conversion means Reference Signs List 10 second conversion means 11 score calculation means 12 acoustic parameter generation means 13 storage means 13

 ────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Norio Higuchi 2-1-1-15 Ohara, Kamifukuoka-shi, Saitama F-term in Kaididi Research Institute Co., Ltd. 5D015 GG01 HH00

Claims (21)

[Claims] 1. A vocabulary for confusing a new vocabulary with a registered vocabulary for a speech recognition apparatus is digitized, and a threshold is set for an obtained numerical value (hereinafter referred to as an evaluation value) to register a new vocabulary. A speech recognition vocabulary registration / judgment method characterized by judging whether to allow or prohibit. 2. The method according to claim 1, wherein the characters or character strings of the new vocabulary and the registered vocabulary are decomposed into recognition units such as phonemes and syllables, respectively. A speech recognition vocabulary registration determination method, wherein a second recognition unit sequence corresponding to a registered vocabulary is obtained, and a similarity between the first and second recognition unit sequences is obtained as the evaluation value. 3. The speech recognition vocabulary registration determination method according to claim 2, wherein the similarity is obtained by taking a sum or a product of scales representing distances between recognition units. 4. The speech recognition vocabulary registration determination method according to claim 2, wherein said similarity is obtained by using a confusion matrix between recognition units obtained in advance. 5. The HMM of claim 2, wherein:
A speech recognition vocabulary registration / judgment method characterized in that the similarity is obtained using a distribution in an acoustic parameter space expressing a (hidden Markov model). 6. The method according to claim 1, wherein a pseudo acoustic parameter sequence corresponding to a character or a character string of the new vocabulary is generated, and the generated acoustic parameter sequence is subjected to a speech recognition process. A speech recognition vocabulary registration determination method, wherein the evaluation value is obtained. 7. The speech recognition process according to claim 6, wherein a first speech recognition process using a grammar for accepting a new vocabulary and a second speech recognition process using a grammar for accepting only a registered vocabulary are performed. And a difference between a likelihood obtained in the first speech recognition process and a likelihood obtained in the second speech recognition process is used as the evaluation value. 8. The speech recognition vocabulary registration determination method according to claim 7, wherein the acoustic parameter sequence is generated using an HMM (Hidden Markov Model). 9. The speech recognition vocabulary according to claim 8, wherein the acoustic parameter sequence is generated by perturbation using the variance of the HMM or by randomly connecting acoustic parameters of a noise HMM. Registration determination method. 10. A speech recognition vocabulary registration / judgment method characterized by using a combination of the speech recognition vocabulary registration / judgment method according to claim 2 and the speech recognition vocabulary registration / judgment method according to claim 6. 11. A speech input means, a vocabulary registration means, and a speech recognition means, wherein speech input from the speech input means is based on a vocabulary registered in the vocabulary registration means (hereinafter, a registered vocabulary). In a speech recognition apparatus for recognizing, a character input means for inputting a character, a numerical value (hereinafter, an evaluation value) for evaluating the ease of confusion between an input character or character string (hereinafter, a new vocabulary) and a registered vocabulary ), And a determination unit that sets a threshold value for the obtained evaluation value and determines whether registration of a new vocabulary is permitted or prohibited. 12. The method according to claim 11, wherein the first vocabulary is decomposed into recognition units to generate a first recognition unit sequence corresponding to the new vocabulary, and the registered vocabulary is decomposed into recognition units. And a second conversion unit for generating a second recognition unit sequence corresponding to the registered vocabulary, wherein the evaluation value calculation unit determines the first value obtained by the first conversion unit as the evaluation value. A speech recognition apparatus comprising: calculating a similarity between the recognition unit sequence of (i) and the second recognition unit sequence obtained by the second conversion means. 13. The speech recognition apparatus according to claim 12, wherein said evaluation value calculation means calculates the similarity by taking a sum or a product of scales representing distances between recognition units. 14. The method according to claim 12, wherein the evaluation value calculating means is obtained in advance. 15. The apparatus according to claim 12, wherein the evaluation value calculating means calculates the similarity using a distribution in an acoustic parameter space representing an HMM (Hidden Markov Model) of a recognition unit. Voice recognition device. 16. The speech recognition unit according to claim 11, further comprising: a speech parameter sequence generation unit that generates a pseudo acoustic parameter sequence corresponding to a new vocabulary and provides the pseudo speech parameter sequence to the speech recognition unit. A speech recognition device, wherein the evaluation value is calculated using likelihood obtained from the speech recognition device. 17. The speech recognition device according to claim 16, wherein said speech recognition means performs a first speech recognition process using a grammar for accepting a new vocabulary and a second speech recognition process using a grammar for accepting only a registered vocabulary. The evaluation value calculating means may determine the likelihood obtained in the first speech recognition process as a second evaluation value as the evaluation value.
A speech recognition apparatus for calculating a difference from a likelihood obtained by the speech recognition processing of (1). 18. The speech recognition apparatus according to claim 17, wherein said acoustic parameter sequence generation means generates an acoustic parameter sequence using an HMM (Hidden Markov Model). 19. The acoustic parameter sequence generating means according to claim 18, wherein the acoustic parameter sequence generating means perturbs using the variance of the HMM or randomly connects acoustic parameters of the noise HMM to generate an acoustic parameter sequence. A speech recognition device characterized by the following. 20. A computer program for causing a computer to execute the speech recognition vocabulary registration / judgment method according to claim 1. 21. A recording medium on which the computer program according to claim 21 is recorded.