JP2003316377A - Device and method for voice recognition - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To determine correctness or incorrectness of a result of voice recognition in a highly precise manner employing a small amount of processing. <P>SOLUTION: Discrimination is made to determine the correctness or incorrectness of a recognition result RCG by collating an acoustic model HMMsb and a feature vector V(n) of uttered voice, obtaining a recognition result RCG that indicates an acoustic model through which maximum likelihood is obtained, a first score FSCR which indicates the value of the maximum likelihood and a second score SSCR which indicates a next likelihood value and comparing an evaluation value FSCRX (FSCR-SSCR) based on the scores FSCR and SSCR with respect to a beforehand set threshold value THD. When it is determined that the result RCG is correct, a speaker adaptation is conducted for the model HMMsb. When it is determined that the result RCG is incorrect, no speaker adaptation is performed for the model HMMsb so as to realize improvement in the precision of the speaker adaptation or the like. <P>COPYRIGHT: (C)2004,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device and a voice recognition method for performing voice recognition by speaker adaptation, for example.

[0002]

2. Description of the Related Art As a main cause of difficulty in voice recognition,
Generally, it is said that the uttered voice of a speaker has individual differences such as a vocal organ and a habit of utterance.

Conventionally, a MLLR (Maximum Likelihood Linear Regres) has been used as a robust (robust) speech recognition algorithm for uttered speech which is affected by such individual differences.
A speaker adaptation method for increasing the recognition rate for a specific speaker by performing speaker adaptation using a sion) or MAP (Maximum a Posteriori) estimation method or the like.

In this speaker adaptation method, a large number of acoustic models (initial acoustic models) for each phoneme or word are created from the voices of an unspecified number of speakers, and the speech of the speaker to be speaker-adapted. Those acoustic models are adapted to the speaker according to the feature quantity of.

[0005]

By the way, in order to perform an appropriate speaker adaptation, it is necessary to apply the speaker adaptation to the phoneme or word-based acoustic model in correspondence with the speech recognition result.

[0006] In other words, to give a concrete example, when the speaker utters "Tokyo", the acoustic model of "Tokyo" is selected only when the speech is correctly recognized as "Tokyo". It is necessary to adapt the speaker, and if the acoustic model of "Tokyo" is adapted to the speaker according to the result of erroneous recognition, incorrect speaker adaptation will be performed.

If the speaker utters "Tokyo", the uttered voice is erroneously recognized as "Kyoto" and the speaker uses the feature amount of the uttered voice of "Tokyo" with respect to the acoustic model of "Kyoto". If adaptation is performed, appropriate speaker adaptation cannot be performed, and the voice recognition rate after speaker adaptation may be reduced.

Therefore, as a precondition for performing speaker adaptation, it is important to surely determine whether the uttered voice matches the result of voice recognition, that is, whether the result of voice recognition is correct. ing.

However, according to the conventional speaker adaptation method,
Since the amount of processing for determining whether or not the result of voice recognition is correct is enormous, it is possible to perform new and rapid speaker adaptation without causing stress to the speaker. The development of a judgment processing method has been an important issue.

The present invention has been made in view of these problems, and an object of the present invention is to provide a voice recognition device and a voice recognition method that enable quick and appropriate speaker adaptation.

[0011]

In order to achieve the above-mentioned object, the invention according to claim 1 is a speech recognition apparatus which applies speaker adaptation to an acoustic model according to a feature vector of uttered speech,
Collate the above acoustic model with the feature vector of the speech,
A speech recognition unit that outputs a recognition result indicating an acoustic model for which the maximum likelihood is obtained, a first score indicating the maximum likelihood value, and a second score indicating the next likelihood value, and the first score. And an evaluation value based on the second score and a threshold value set in advance, and when the evaluation value has a predetermined relationship with the threshold value, a determination unit that determines the recognition result as correct, If it is determined that the above recognition result is correct,
And a speaker adaptation processing means for applying speaker adaptation to the acoustic model.

According to the invention of claim 5, in a voice recognition method for applying speaker adaptation to an acoustic model by a feature vector of a spoken voice, the acoustic model and a feature vector of a spoken voice are collated to determine a maximum likelihood. A recognition result indicating the obtained acoustic model, and a first score indicating the value of the maximum likelihood,
A first step of obtaining a second score indicating the next likelihood value;
A second step of comparing the evaluation value based on the first score and the second score with a preset threshold value, and determining the recognition result to be correct when the evaluation value has a predetermined relationship with the threshold value. And a third step of applying speaker adaptation to the acoustic model when the recognition result is determined to be correct in the second step.

According to the inventions of the speech recognition apparatus according to claim 1 and the speech recognition method according to claim 5, at the time of speaker adaptation processing, by comparing the acoustic model and the feature vector of the spoken voice, the maximum Based on the recognition result indicating the acoustic model for which the likelihood is obtained, the first score indicating the maximum likelihood value, and the second score indicating the next likelihood value, and further based on the first score and the second score. The evaluation value is compared with a preset threshold value. Then, when the evaluation value has a predetermined relationship with the threshold value, the recognition result is determined to be correct, and speaker adaptation is performed on the acoustic model. As a result, appropriate speaker adaptation is performed based on the correct recognition result.

According to a second aspect of the invention, in the first aspect of the invention, when the evaluation value does not have a predetermined relationship with the threshold value, the determining means determines the recognition result as an error, The speaker adaptation processing means does not perform speaker adaptation on the acoustic model when the recognition result is determined to be an error.

The invention according to claim 6 is the invention according to claim 5, wherein in the second step, the recognition result is determined to be an error when the evaluation value does not have a predetermined relationship with a threshold value. However, in the third step, when the recognition result is determined to be an error, speaker adaptation is not performed on the acoustic model.

According to the inventions of the voice recognition device and the voice recognition method according to the sixth aspect, when the recognition result is determined to be an error, speaker adaptation is not performed on the acoustic model. As a result, the speaker adaptation is not performed based on the erroneous recognition result, so that the deterioration of the voice recognition accuracy after the speaker adaptation is prevented.

According to a third aspect of the invention, in the invention of the speech recognition apparatus according to the first aspect, the evaluation value is calculated by a difference value between the first score and the second score. To do.

According to a seventh aspect of the invention, in the invention of the speech recognition method according to the fifth aspect, the evaluation value is calculated by a difference value between the first score and the second score. To do.

According to the inventions of the voice recognizing device and the voice recognizing method according to claim 7, the evaluation value is calculated by the difference value between the first score and the second score. When the obtained evaluation value has a predetermined relationship with the threshold value, the recognition result is determined to be correct, and when the evaluation value does not have the predetermined relationship with the threshold value, the recognition result is determined to be incorrect. Achieves improved judgment accuracy.

According to a fourth aspect of the invention, in the invention according to the second aspect, when the determination means determines that the recognition result is an error, the recognition result is prohibited from being output, and
It is characterized in that it comprises means for presenting that the recognition result is incorrect.

Further, in the invention according to claim 8, in the invention according to claim 6, when the recognition result is determined as an error in the second step, the recognition result is prohibited from being output and the recognition result is The feature is that an error is presented.

According to the invention of the voice recognition device according to claim 4 and the voice recognition method according to claim 8, effective information such as whether or not a proper speaker adaptation is made is presented to the user or the like.

[0023]

BEST MODE FOR CARRYING OUT THE INVENTION Preferred embodiments of the present invention will be described below with reference to the drawings.

(First Embodiment) A first embodiment of the present invention will be described with reference to FIGS. Note that FIG.
FIG. 3 is a block diagram showing a configuration of a voice recognition device of this embodiment.

In FIG. 1, the speech recognition apparatus is an HMM.
(Hidden Markov Model) is used for voice recognition, and a voice input signal v (t) output from a microphone (not shown) as a sound pickup means is used as a feature vector sequence V (n) in the cepstrum region. An acoustic analysis unit 1 for converting the output into a speech output, a speech recognition processing unit 2, a word model generation unit 3, a word dictionary 4, an acoustic model storage unit 5, a correctness determination unit 6,
And a speaker adaptation processing unit 7.

The acoustic model storage unit 5 stores an acoustic model (standard phoneme HMM) in units of subwords such as phonemes generated from the voice of an unspecified speaker.

As will be described later, the acoustic model storage unit 5 stores an acoustic model (standard phoneme HMM) obtained in advance from the voice of an unspecified speaker as an initial acoustic model. After that, when the speaker adaptation is applied, the initial acoustic model is updated to the speaker adapted acoustic model. Therefore, if the speaker adaptation is continued or repeated, the speaker adaptation is performed on the speaker adaptation acoustic model stored in the acoustic model storage unit 5, and further the speaker adaptation acoustic model is further performed. Since the speaker adaptation is performed and the iterative processing is performed, the acoustic model storage unit 5 stores the repeatedly updated speaker-adaptive acoustic model. By thus updating the speaker-adaptive acoustic model, the recognition performance for the speaker's voice is improved.

The word dictionary 4 stores a large number of words and sentence dictionary data (also referred to as text data) in advance.

The word model generation unit 3 is stored in the acoustic model storage unit 5 in accordance with the data of each word or sentence stored in the word dictionary 4 (hereinafter, these data are collectively referred to as “word data”) WRD. Acoustic model HM of each phoneme
An acoustic model (hereinafter, referred to as "word model") HM corresponding to word data WRD by combining Msb
Generate Mw.

The voice recognition processing section 2 includes a word model generation section 3
The word model HMMw generated in (1) and the feature vector V (n) of the uttered voice supplied from the acoustic analysis unit 1 are collated in real time. Then, as a result of matching all the word models HMMw with the feature vector V (n), the recognition result RCG indicating the word model HMMw having the maximum likelihood is output.

Further, the voice recognition processing section 2 recognizes the recognition result RCG.
In addition to outputting the above-mentioned maximum likelihood value as the first score FS
The CR and the next largest likelihood value (the second largest likelihood value with respect to the maximum likelihood) are output as the second score SSCR.

For example, when the speaker utters "Tokyo", the speech recognition processing unit 2 will generate the feature vector series V (n) and all the word models HMMw. As a result of the matching, it is assumed that the likelihood for the word model of “Tokyo” has the maximum value and the likelihood for the word model of another word “Kyoto” has the second largest value. , Its maximum likelihood value is the first score FSCR, the second largest likelihood value is the second score SS
Output as CR. In the case of such an example, since the uttered voice of “Tokyo” and the word model of “Tokyo” corresponding to the first score FSCR match, it corresponds to the case where the recognition result RCG is correct. To do.

On the other hand, when the recognition result RCG is erroneous recognition, for example, when the speaker utters "Tokyo", the result of the voice recognition by the voice recognition processing unit 2, for example, the word "Kyoto". ) ”Has the maximum likelihood, and the likelihood of the“ Tokyo ”word model has the second largest likelihood, the maximum likelihood value is the first score FSCR, The second largest likelihood value is output as the second score SSCR. in this way,
Speech of "Tokyo" and first score FSCR
The recognition result RCG is not correct because the word model of “Kyoto” corresponding to is not matched.

The correctness determination unit 6 outputs the first score FSC output when the voice recognition processing unit 2 voice-recognizes a speech.
R and the second score SSCR are applied to the score evaluation function represented by the following expression (1), and the score evaluation value G (L) obtained thereby is compared with a predetermined threshold value THD.

G (L) = FSCR × (FSCR−SSCR) (1) The variable FSCR shown in the right side of the above formula (1) is
First score FSCR value (maximum likelihood value), variable SSCR
Is the value of the second score FSCR (second likelihood value), and the score evaluation value G (L) is a variable related to the likelihood L obtained by the evaluation operation on the right side.

Then, the score evaluation value G (L) is the threshold value THD.
Greater than (that is, when G (L)> THD)
, The recognition result RCG is determined to be correct, and when the score evaluation value G (L) is smaller than the threshold value THD (that is, G (L)
≦ THD), the recognition result RCG is determined as an error,
Each judgment result RSLT is output.

Here, the score evaluation function and the threshold value THD of the above equation (1), and the determination principle of the correctness determination unit 6 will be described in detail.

The score evaluation function of the above equation (1) and the threshold value THD
Is experimentally determined in advance by the statistical method described below.

First, an arbitrary speaker is made to speak a predetermined number N of words and sentences, and the word dictionary 4, the word model generation unit 3 and the voice recognition processing unit 2 are caused to perform voice recognition processing, and the total number N is reached.
The first score FSCR and the second score SSCR output for each word or sentence of are measured experimentally.

Then, a word or sentence when the uttered voice and the recognition result RCG match (that is, a word or sentence when correctly recognized) and a word or sentence when the uttered voice and the recognition result RCG do not match. Sentences (that is, words and sentences when misrecognized).

For example, the total number N of words or sentences spoken as an experiment target is 500, the number X of correctly recognized words or sentences is 400, and the number Y of incorrectly recognized words or sentences is 100. If the number is 500, the 500 words or sentences are classified into the numbers X and Y, respectively.

Further, as shown in FIG. 2A, the value of the first score FSCR (maximum likelihood) for correctly recognized words and sentences (X = 400 words and sentences in the above-mentioned example). Histogram P (FSCR) showing the distribution of the number of words or sentences with respect to (value) and histogram Q (SSCR) showing the distribution of the number of words or sentences with respect to the value of the second score SSCR (second likelihood value). create.

That is, the histograms P (FSCR) and Q (SSC
R) together creates 400 words and sentences that have been correctly recognized by voice. Furthermore, the value of the first score FSCR (maximum likelihood value) and the second score output from the voice recognition processing unit 2
The value of the score SSCR (the second likelihood value) changes variously depending on the speech recognition environment and the like. Therefore, by assigning 400 words or sentences to each value of the first score FSCR, the same value can be obtained. Histogram P (FSCR) as shown by the solid line in the figure
Create a second score SS of 400 words or sentences in the same way
By assigning each value of CR, a histogram Q (SSCR) as shown by the dotted line in the figure is created.

As shown in FIG. 2B, the value of the first score FSCR is similarly applied to the erroneously recognized words and sentences (X = 100 words and sentences in the above-mentioned example). Histogram P (FSC
R) "and a histogram Q (SSCR)" indicating the distribution of the number of words or sentences with respect to the value of the second score SSCR ".

That is, the histogram P shown in FIG.
Both (FSCR) and Q (SSCR) are created for 100 misrecognized words and sentences. Furthermore, the voice recognition processing unit 2
Even if the erroneous recognition is performed, the value of the first score FSCR (maximum likelihood value) and the value of the second score SSCR (second likelihood value)
Changes in accordance with the voice recognition environment, etc.
By assigning 0 words or sentences to each value of the first score FSCR, a histogram P (FSCR) "shown by a solid line in FIG. 2B is created, and 100 words or sentences are created. Is assigned to each value of the second score SSCR, the histogram Q (S
SCR) "is created.

When the histogram is created in this way, the histograms P (FSCR) and Q (SSCR) in FIG.
Will be distributed ubiquitously in the range of the likelihood values separated from each other. Therefore, if the recognition result RCG is correct, the first
It can be seen that the statistical characteristics of the score FSCR and the statistical characteristics of the second score SSCR appear greatly differently.

Also, the histogram P (FSC in FIG.
R) "and Q (SSCR)" are distributed in the range of almost the same likelihood value. Therefore, when the recognition result RCG is incorrect, the statistical features of the first score FSCR and the second score SSCR of It can be seen that the statistical characteristics are almost the same.

In this way, when the recognition result RCG is correct and when it is incorrect, it is shown by the relationship between the histograms P (FSCR) and Q (SSCR) and the histograms P (FSCR) "and Q (SSCR)". There is a unique statistical feature as described above, and the score evaluation function of the above formula (1) is determined as a function capable of expressing such a statistical feature.

According to the score evaluation function of the above equation (1), when the recognition result RCG is correct, as shown in FIG. 2A, the first score FSCR has a larger likelihood than the second score SSCR. Since it is unevenly distributed on the value side, the difference value (FSCR-SSCR) in the right side of the above formula (1) becomes a large value, and by further multiplying this difference value (FSCR-SSCR) by the first score FSCR, Difference value (FS
(CR-SSCR) is obtained, which allows the score evaluation value to be more manifested.

Therefore, the score evaluation function of the above equation (1) is the first score FSC when the speech recognition is correctly performed.
By applying R and the second score SSCR, the statistical characteristics when the voice recognition is correctly performed are properly reflected, and the recognition result RCG output when the voice recognition is correctly performed is correct. Can be used as a basis for determining.

On the other hand, when erroneous recognition is performed, the process shown in FIG.
As shown in (b), since the likelihood values of the first score FSCR and the second score SSCR occur in substantially the same range, the difference value (FSCR-SSC in the right side of the above equation (1) is calculated.
R) becomes a small value, and this difference value (FSCR-S
By multiplying (SCR) by the first score FSCR, a score evaluation value that makes it possible to more reveal the characteristic of the difference value (FSCR-SSCR) is obtained.

Therefore, the score evaluation function of the above equation (1) has the first score FSCR and the second score FSCR when erroneous recognition is performed.
By applying the score SSCR, it is possible to appropriately reflect the statistical characteristics when the misrecognition is performed, and the basis for determining the recognition result RCG output when the misrecognition is performed as an error. It becomes possible to

Next, the threshold value THD is obtained by using the score evaluation value G (L) obtained from the score evaluation function of the above equation (1) as a target.
When the recognition result RCG is correct, it is determined to be correct, and when the recognition result RCG is incorrect, it is determined as a determination criterion for determining an error.

That is, it is generally difficult to correctly determine the correct answer / incorrect answer of all recognition results RCG. In speaker adaptation, if an incorrect answer is erroneously determined to be the correct answer, the acoustic model is erroneously applied as described above, resulting in deterioration of recognition performance. Conversely, even if the correct answer is erroneously determined as an incorrect answer, the speaker adaptation is not performed, so that the recognition performance is not improved but is not deteriorated. Therefore, the threshold value THD is selected based on the principle described below so as to sufficiently secure the performance of correctly determining an incorrect answer.

First, the first score FSCR and the second score obtained for each word or sentence when the voice recognition is correctly performed.
By applying the score SSCR and the score evaluation function of the above formula (1), the score evaluation value G (L) for each word or sentence
To calculate. Furthermore, a histogram R (G) showing the distribution of the number of words or sentences corresponding to each calculated score evaluation value G (L).
(L)) is calculated.

Similarly, by applying the respective first score FSCR and second score SSCR obtained for each word or sentence at the time of misrecognition to the score evaluation function of the above formula (1), The score evaluation value G (L) for each word or sentence is calculated. Further, a histogram T (G (L)) showing the distribution of the number of words or sentences corresponding to each calculated score evaluation value G (L) is obtained.

Thus, each histogram R (G (L)) and T (G
(L)), the histogram R (G (L)) has the distribution shown by the solid line in FIG. 2 (c), and the histogram T (G (L))
Has a distribution as shown by the dotted line in FIG.

That is, the histogram R (G (L)) is a feature when the speech of a certain word or sentence is correctly recognized, and the histogram T (G (L)) is the speech of a certain word or sentence is erroneously recognized. It will show the characteristics when it is recognized.

The number of histograms T (G (L)) is 0.
When the recognition result RCG is correct, it is possible to determine that the range WR in which the score evaluation value G (L) becomes larger than the boundary when the recognition result RCG is correct. A range in which the score evaluation value G (L) becomes smaller than the score evaluation value G (L) when the number of histograms T (G (L)) becomes 0. WT is
When the recognition result RCG is incorrect, it indicates the range in which it can be determined as an error.

Therefore, the threshold value THD is the histogram T (G
It is set to a value slightly larger than the score evaluation value G (L) when the number of (L)) becomes zero.

Then, the correctness determination section 6 outputs the threshold value THD and the above equation each time the recognition result RCG for each word or sentence is output.
The score evaluation value G (L) obtained from the score evaluation function of (1) is compared to determine whether the threshold value THD and the score evaluation value G (L) have a predetermined relationship, and the determination result RSLT is Output. That is, when the score evaluation value G (L) is larger than the threshold value THD (when G (L)> THD), the recognition result R
It is determined that CG is correct, and the score evaluation value G (L) is the threshold TH.
When it is smaller than D (when G (L) ≤THD), the recognition result RCG is determined as an error, and each determination result RSLT is output.

The speaker adaptation processing unit 7 inputs the recognition result RCG and the judgment result RSLT, and performs speaker adaptation processing according to the judgment result RSLT.

That is, when the determination result RSLT indicating that the voice recognition is correctly performed is input, the speaker adaptation is performed by the feature vector V (n) of the uttered voice uttered at that time, while the misrecognition is determined. As a result, when RSLT is input, speaker adaptation is not performed.

Furthermore, the speaker adaptation described above is performed for all or part of the phoneme model HMMsb before speaker adaptation.

That is, when the determination result RSLT indicating that the voice recognition is correctly performed is input, the utterance content of the feature vector series V (n) of the utterance voice is regarded as the recognition result RCG, and the utterance content is known. MLL, subject to
Speaker adaptation is performed on the phoneme model HMMsb before speaker adaptation by a speaker adaptation method such as R or MAP estimation method to obtain a phoneme model HMMsb "after speaker adaptation. Acoustic model HM before being adapted to the speaker
It is replaced with Msb and updated and stored.

Needless to say, the speaker adaptation process is continuously or repeatedly performed. Therefore, the speaker-adapted phoneme model updated and stored in the acoustic model storage unit 5 is
In the next speaker adaptation, it becomes a phoneme model before speaker adaptation, and when the phoneme model obtained by that speaker adaptation is updated and stored in the acoustic model storage unit 5, the updated phoneme is further stored in the next speaker adaptation. Speaker adaptation is applied to the model,
It will be repeated.

Next, the operation of the speech recognition apparatus having the above structure will be described with reference to the flow chart shown in FIG.

In the figure, when the speaker adaptation process is started, the voice input process is started in step S100.

Next, in step S102, the word dictionary 4, the word model generation unit 3, and the voice recognition processing unit 2 perform voice recognition by collating the feature vector V (n) of the uttered voice with the word model HMMw, The recognition result RCG, the first score FSCR and the second score SSCR are output.

Next, in step S104, the correctness determination unit 6 applies the first score FSCR and the second score SSCR to the score evaluation function of the above equation (1) to obtain the score evaluation value G (L).
To calculate.

Next, in step S106, the correctness / incorrectness determination section 6 further compares the score evaluation value G (L) with the threshold value THD, and when G (L)> THD (when "YES"),
By determining that the recognition result RCG is correct, step S1
08, when G (L) ≦ THD (“NO”)
In the case of), the recognition result RCG is determined to be an error, and the processing is ended without performing the speaker adaptation processing.

When the process proceeds to step S108, the speaker adaptation processing unit 7 applies the speaker adaptation to the phoneme model HMMsb in the acoustic model storage unit 5 by the feature vector V (n),
Further, in step S110, the speaker-adapted phoneme model HMMsb "is updated and stored, and then the process ends.

Although FIG. 3 shows the speaker adaptation process for the uttered voice of one uttered word or sentence for convenience, the speaker adaptation process is performed for a continuous sentence composed of a plurality of words or sentences. When performing, the process of FIG. 3 is repeated.

Then, a plurality of uttered words or sentences are sequentially processed, and the phoneme model HMMsb for the erroneously recognized words or sentences is "N" in step S106.
By determining “O”, the speaker adaptation process (steps S108 and S110) is skipped, and the word or sentence recognized correctly is “YE” in step S106.
By determining "S", the speaker adaptation process (steps S108 and S110) is performed, thereby performing appropriate speaker adaptation.

As described above, according to the speech recognition apparatus of this embodiment, the two first scores FSCR and the second score SSC are used.
Only by comparing the score evaluation value G (L) obtained by applying R to the score evaluation function with a predetermined threshold value THD, it is possible to surely and quickly determine whether the recognition result RCG is correct or incorrect. it can. That is, it is possible to significantly reduce the processing amount for determining whether or not the result of voice recognition is correct, and it is possible to obtain high determination accuracy.

Therefore, prompt and appropriate speaker adaptation can be performed without causing stress to the speaker. Further, since it is possible to significantly reduce erroneous speaker adaptation, it is possible to prevent problems such as deterioration of voice recognition accuracy after speaker adaptation from occurring.

Also, every time a user or the like who is a speaker uses the present speech recognition apparatus, an appropriate speaker adaptation is gradually applied to the phoneme model HMMsb in the acoustic model storage unit 5. As the number of times increases, the voice recognition rate can be improved.

In this embodiment, the first score FSCR is used to calculate the score evaluation value G (L) on the basis of the first score FSCR and the second score SSCR, as shown in the above equation (1). Then, the difference between the second score SSCR and the second score SSCR is obtained, and the difference value is further multiplied by the first score FSCR. However, the present invention is not limited to this.
As a modified example, the first score FSCR and the second score SSC
The difference value of R may be used as the score value G (L).

(Second Embodiment) Next, the second embodiment of the present invention will be described.
An embodiment will be described with reference to FIG. In FIG. 4, parts that are the same as or correspond to those in FIG.

When the present speech recognition apparatus is compared with the speech recognition apparatus of the first embodiment shown in FIG. 1, the present speech recognition apparatus is provided with an erroneous judgment handling section 8 and a display section 9. Has more features.

The erroneous judgment handling unit 8 inputs the judgment result RSLT from the correctness judgment unit 6 and, when the judgment result RSLT that the recognition result RCG is correct is input, outputs the recognition result RCG and the recognition result RCG is erroneous. Recognition result R indicating recognition
When SLT is input, the output of the recognition result RCG is prohibited. As a result, the recognition result RCG is output only when the voice recognition is correctly performed.

Further, when the determination result RSLT indicating that the recognition result RCG is erroneous recognition is input, by giving an instruction to the display unit 9 formed of a liquid crystal display or the like,
A warning display is displayed with characters indicating that the recognition is incorrect and that the user should speak again.

As described above, according to the present speech recognition apparatus, by using the judgment result RSLT of the correctness judgment unit 6,
It is possible to present the user with information such as whether or not an appropriate speaker adaptation has been made, and whether it is better to speak again, etc., so that the convenience of the user can be improved.

[0084]

As described above, according to the voice recognition device and the voice recognition method of the present invention, in the speaker adaptation process,
By collating the acoustic model and the feature vector of the uttered speech, the recognition result indicating the acoustic model for which the maximum likelihood is obtained, the first score indicating the maximum likelihood value, and the second likelihood value indicating the next likelihood value. Since the score is obtained and the evaluation value based on the first score and the second score is compared with a preset threshold value, the correctness determination of the recognition result is performed, so the accuracy determination is highly accurate and small. It can be realized by processing amount.

[Brief description of drawings]

FIG. 1 is a diagram illustrating a configuration of a voice recognition device according to a first embodiment.

FIG. 2 is a diagram for explaining a principle for accurately determining whether a recognition result is correct or incorrect.

FIG. 3 is a flowchart showing an operation of the voice recognition device in the first exemplary embodiment.

FIG. 4 is a diagram showing a configuration of a voice recognition device according to a second embodiment.

[Explanation of symbols]

1 ... Acoustic analysis unit 2 ... Speech recognition processing unit 3 ... Word model generation unit 4 ... Word dictionary 5 ... Acoustic model storage unit 6 ... Correctness determination unit 7 ... Speaker adaptation processing unit 8 ... Misjudgment handling unit 9 ... Display

Claims (8)

[Claims] 1. A speech recognition apparatus for speaker adaptation of an acoustic model using a feature vector of a spoken voice, wherein the acoustic model and a feature vector of a spoken voice are collated,
A speech recognition unit that outputs a recognition result indicating an acoustic model with which the maximum likelihood is obtained, a first score indicating the value of the maximum likelihood, and a second score indicating the next likelihood value, and the first score. A determination unit that compares the evaluation value based on the second score with a preset threshold value, and determines the recognition result to be correct when the evaluation value has a predetermined relationship with the threshold value; A speech recognition apparatus comprising: a speaker adaptation processing unit that applies speaker adaptation to the acoustic model when the recognition result is determined to be correct. 2. The determination means determines the recognition result as an error when the evaluation value does not have a predetermined relationship with a threshold value, and the speaker adaptation processing means determines the recognition result as an error. Then, the speech recognition apparatus according to claim 1, wherein speaker adaptation is not performed on the acoustic model. 3. The voice recognition device according to claim 1, wherein the evaluation value is calculated by a difference value between the first score and the second score. 4. The means for prohibiting output of the recognition result and presenting the fact that the recognition result is incorrect when the judging means judges that the recognition result is incorrect. The voice recognition device according to 2. 5. In a voice recognition method, wherein a speaker adaptation is performed on an acoustic model by a feature vector of a spoken voice, the acoustic model and a feature vector of the spoken voice are collated,
A first step of obtaining a recognition result indicating an acoustic model for which maximum likelihood is obtained, a first score indicating the value of the maximum likelihood, and a second score indicating the next likelihood value; and the first score. And a second step of comparing an evaluation value based on the second score with a preset threshold value, and determining the recognition result to be correct when the evaluation value has a predetermined relationship with the threshold value, When it is determined that the recognition result is correct in step 2,
A third step of applying speaker adaptation to the acoustic model;
A voice recognition method comprising: 6. The second step determines the recognition result as an error when the evaluation value does not have a predetermined relationship with a threshold value, and the third step determines the recognition result as an error. Then, the speaker recognition method according to claim 5, wherein speaker adaptation is not performed on the acoustic model. 7. The voice recognition method according to claim 5, wherein the evaluation value is calculated by a difference value between the first score and the second score. 8. The method according to claim 6, wherein when the recognition result is determined to be an error in the second step, the output of the recognition result is prohibited and the fact that the recognition result is an error is presented. Speech recognition method described in.