JP3523949B2 - Voice recognition device and voice recognition method - Google Patents

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 having a post-processing function of a recognition result in voice recognition and particularly applicable to the field of word voice recognition.

[0002]

2. Description of the Related Art An example of basic functions of a speech recognition apparatus is described in, for example, a published paper of Acoustical Society of Japan 3-1-8 (October 1983), and among them, a recognition method for an unspecified speaker. An example is disclosed in the proceedings of the 10th information theory and its application symposium "Unspecified speaker word speech recognition using fuzzy pattern matching". Shu Vol. J69-ANo. 1 (1986) "Unspecified speaker word speech recognition".

[0003]

With reference to FIG. 2, the operation of collating an unknown input voice with a standard pattern which is used for recognizing the unknown input voice and which is registered and stored in advance in the voice recognition apparatus is referred to. explain. The unknown input voice input from the microphone 1 has its input level optimized by the pre-processing unit 2 to remove signals in unnecessary bands, and then the feature extraction unit 3 features a pattern required for recognition of the unknown input voice. Is converted to. The converted characteristic pattern is compared and collated in the recognition processing means 5 with a plurality of standard patterns registered and stored in the standard pattern storage means 4 in advance, and the standard pattern and the characteristic pattern are compared for each standard pattern. The degree of similarity is determined. According to the similarity of each standard pattern thus obtained, candidate words for the unknown input speech are determined in descending order of similarity. Next, the post-processing means 6 outputs the candidate word to the outside through the result output means 7 when a certain criterion is satisfied based on the determined candidate word, and the candidate word when the criterion is not satisfied. Reject (also referred to as reject).

Next, a method of creating the standard pattern will be described. Many methods of creating standard patterns for a voice recognition device have been reported, but here, for example, a pattern matching method that has been conventionally used will be described. First, in the case of the specific speaker method, the specific speaker utters one or more times before using the speech recognition apparatus, and the input pattern based on this utterance is weighted averaged to create a dictionary of input patterns. . In a voice recognition device that handles unspecified speakers, input patterns based on the utterances of multiple speakers are weight-averaged to create a dictionary of input patterns.In that case, in order to make the standard patterns general, multiple evaluation input patterns are used. In many cases, the standard pattern is optimized by inputting "." An example of the method is Japanese Patent Laid-Open No. 5-31.
It is disclosed in Japanese Patent No. 3687. The standard pattern thus created is stored in the standard pattern storage means 4. At this time, there are two cases: a specific speaker and an unspecified speaker are individually recognized, and a case where both are recognized in a mixed manner.

Further, conventionally, not only the voice of the user who uses the voice recognition device, but also the ambient noise in the environment in which the voice recognition device is used and the original purpose of use of the voice recognition device. A voice other than the above, for example, a private language is input, and the voice input device reacts to these noises and a private language, which is a problem. Furthermore, when the user utters a word that is not registered in advance, the voice recognition device outputs some result because the effective rejection function is not provided although the obtained similarity is low. There were times when it ended up. Therefore, in order to prevent such an adverse effect, the conventional voice recognition device needs to have a function of rejecting (rejecting) the recognition result with low accuracy and prompting the user to input again. Was there. A voice recognition device relating to such a rejection function is disclosed in JP-A-61-73200 and JP-A-1-15680.
No. 0 publication. That is, the voice recognition of the unknown input voice is performed using this standard pattern, and the largest first
The first-rank standard pattern having the second-rank similarity and the second-rank standard pattern having the second-rank similarity are calculated, and the ratio between the first-rank similarity and the second-rank similarity is calculated. First
If it is equal to or more than the threshold value and the first-order similarity degree of the first-order standard pattern satisfies the second threshold value or more, the first-order standard pattern is output as a voice recognition result. It has the function to reject the standard pattern.

However, in this case, since the first threshold value and the second threshold value are fixed to certain values, the similarity between the unknown input voice and the standard pattern varies depending on the environment in which the unknown input voice is emitted. When the unspecified speaker and the specified speaker are mixed, the phenomenon that the standard of the similarity is different occurs. Therefore, depending on a certain unknown input voice, there is a problem that the standard pattern listed as a candidate is likely to be rejected even though it is a correct answer, or is difficult to be rejected even if it is an incorrect answer. The present invention has been made to solve the above problems,
An object of the present invention is to provide a voice recognition device and a voice recognition method that can obtain a stable recognition result for an unknown input voice.

[0007]

According to one aspect of the present invention, a voice recognition apparatus is characterized in that the standard pattern storage means for storing a standard pattern and the unknown input voice are compared and collated with each other. A voice recognition device comprising a recognition means for recognizing an unknown input voice, wherein the reference similarity stores a reference similarity provided for each standard pattern of the standard patterns selected for the unknown input voice. The storage means and the standard pattern selected for the unknown input speech
It becomes a standard to reject the adoption of turns and stores the above standard similarity
Discard generated based on the above-mentioned reference similarity stored in the means
A rejection threshold value storage means for storing a rejection threshold value, wherein the recognition means determines whether or not the selected standard pattern is a recognition result for the standard pattern selected in the comparison and collation based on the reference similarity. When making a decision ,
Calculate the similarity between the unknown input voice and each standard pattern
First-ranking standard putter with first-ranking similarity
And a second-order standard pattern having a second-rank similarity degree.
In addition, the ratio between the above-mentioned first-rank similarity and second-rank similarity is
1 threshold value or more, and the first similarity is the second threshold value or more
Regarding the certain first standard pattern, the recognizing means is
The 1st rank similarity of the 1st rank standard pattern is the rejection
If the number is above the threshold, the first standard pattern is adopted.
Characterized in that that.

[0008]

According to such a voice recognition apparatus, the standard similarity is added to each standard pattern, and the standard pattern selected and listed as a candidate based on the standard similarity is adopted as the voice recognition result. It is determined whether or not. Therefore, the voice recognition is performed not only based on the similarity between the unknown input voice and the standard pattern, but also based on the reference similarity, so that the correct answer rate in the voice recognition result is improved. You can As described above, the matters specifying the invention described in the above-described aspect act so that a stable recognition result can be obtained for the input voice.

According to another aspect of the present invention, the voice recognition method stores a standard pattern and recognizes the unknown input voice by comparing and collating the unknown input voice with the standard pattern. The reference similarity provided for each standard pattern of the standard pattern having the highest similarity to the unknown input voice is stored, and is generated based on the reference similarity and selected for the unknown input voice. the adoption of the standard pattern storing the rejection threshold for rejecting the upper
The above standard pattern selected by comparison
Select whether or not to use the selected standard pattern as the recognition result.
When making a decision based on the reference similarity,
The highest similarity is calculated by calculating the similarity between the voice and each standard pattern.
Threshold 1st rank standard pattern and 2nd rank with similarity
2nd standard pattern with similarity of
Note: The ratio between the first-rank similarity and the second-rank similarity is greater than or equal to the first threshold value.
And the first similarity is greater than or equal to a second threshold value.
Regarding the 1st standard pattern,
When the first-rank similarity is equal to or greater than the rejection threshold, the above
The first standard pattern is adopted .

[0011]

[0012]

BEST MODE FOR CARRYING OUT THE INVENTION A speech recognition apparatus according to an embodiment of the present invention will be described below with reference to the drawings. This voice recognition device has a post-processing function in voice recognition, when a voice of an unknown input that is a word that is not registered in advance, that is, an unknown input voice is uttered to the voice recognition device,
The function of effectively rejecting the unknown input voice is provided to prevent erroneous recognition of voice recognition. In FIG. 1, the same components as those shown in FIG. 2 are designated by the same reference numerals, and the description thereof will be omitted. Further, in order to facilitate the description, the following description will be made by taking an unspecified speaker recognition method as an example. In the present embodiment, the output side of the feature extraction means 3 is connected to the standard pattern generation means 11 at the time of creating the standard pattern, and the recognition processing included in the recognition means 15 at the time of recognizing the unknown input voice after the standard pattern is created. Connected to the means 5. The output side of the standard pattern generation means 11 is connected to the standard pattern storage means 4. The standard pattern generating means 11 is supplied with a voice pattern obtained by converting the standard pattern generating input voice into a pattern required for recognition by the feature extracting means 3. In the case of an unspecified speaker, the standard pattern generation means 11 is a known device that creates a standard pattern, and uses a voice pattern obtained from a standard pattern generation input voice uttered by a plurality of persons for each word to generate each standard pattern. Generate one standard pattern for each word. In the case of a specific speaker, a standard pattern corresponding to each word is generated one by one using a voice pattern obtained by one person uttering a plurality of times for each word. The output side of the standard pattern storage means 4 is connected to the recognition processing means 5, and the standard pattern stored in the standard pattern storage means 4 is sent to the recognition processing means 5 during the recognition operation of the unknown input voice. Further, the output side of the recognition processing means 5 is connected to the reference similarity generation means 12 at the time of generating the reference similarity, which will be described later, and is connected to the rejection determination means 14 included in the recognition means 5 at the time of the recognition operation. Connected.

The voice pattern converted from the feature extracting means 3 at the time of generating the reference similarity is supplied to the reference similarity generating means 12 through the recognition processing means 5. The reference similarity generation means 12 has the reference similarity generation means 12 in advance.
Is calculated or the similarity between the evaluation pattern and the voice pattern is calculated for each of the evaluation patterns, based on the plurality of evaluation patterns stored in or externally and the voice pattern supplied from the feature extracting means 3. The reference similarity is calculated based on this calculation result. What is the evaluation pattern?
Dozens or more obtained when a plurality of persons (for example, 100 persons) different from a plurality of persons who uttered to create the above-described standard pattern uttered the reference similarity generation input voice for the word "pencil", for example. Hundreds of voice patterns. More specifically, the reference similarity generation unit 12 resembles the voice pattern corresponding to the “pencil” supplied from the feature extraction unit 3 with, for example, the tens to hundreds of the above “pencil” voice patterns. Calculate each degree. Therefore, several tens to several hundreds of similarities can be obtained for the word “pencil”, and the reference similarity generating means 12
Statistical processing is performed on these tens or hundreds of similarities. In this embodiment, the average value is calculated as the statistical processing, but the present invention is not limited to this, and a value such as variance may be calculated. The average value thus obtained, for example, becomes the reference similarity. Therefore, a large reference similarity means that a large number of people utter a voice pattern that is similar to the voice pattern of an unknown input voice. That is, the reference similarity is the degree to which the standard pattern listed as a candidate for the unknown input voice and the unknown input voice match,
In other words, it represents the selection correct answer rate of the selected standard pattern. Then, the generated reference similarity is stored in the reference similarity storage means 16 in association with each standard pattern. In this way, the reference similarity is set for each standard pattern.

The recognition means 15 connected to the feature extraction means 3 at the time of recognition operation has a recognition processing means 5 and a rejection decision means 14. The recognition processing unit 5 calculates the similarity between the voice pattern converted by the feature extraction unit 3 for the unknown input voice and the standard pattern read from the standard pattern storage unit 4, and the maximum similarity is calculated based on the calculated similarity. The first standard pattern having the first similarity and the second standard pattern having the second similarity are determined, and these are sequentially transmitted. On the other hand, the output of the reference similarity storage means 16 is connected to the recognition processing means 5 and the rejection threshold value generation means 13. The rejection threshold generation means 13 generates a rejection threshold by multiplying each reference similarity stored in the reference similarity storage means 16 by a coefficient having a value of 1 or less, and the rejection threshold determination means 14 determines the rejection threshold. Alternatively, it is sent to the rejection threshold value storage means 17. The output side of the rejection threshold value storage means 17 is rejection determination means 1
4 and the rejection threshold value storage unit 17 stores the rejection threshold value, and during the voice recognition operation, the rejection threshold value storage unit 1
The rejection threshold value read from 7 is sent to rejection determination means 14.

The rejection decision means 14 calculates the first-order similarity degree of the first-order standard pattern supplied from the recognition processing means 5 during the recognition operation from the reference similarity degree of the first-order standard pattern. If it is less than the rejection threshold, the adoption of the first standard pattern is rejected. On the other hand, if the second-rank similarity of the second-rank standard pattern is equal to or more than the rejection threshold calculated from the reference similarity of the second-rank standard pattern, the second-rank standard pattern is set as an unknown input voice. It is adopted as the recognition result of. The output of the rejection decision means 14 is connected to the result output means 7, and the result output means 7 sends the standard pattern adopted as the recognition result to the outside.
The voice recognition device according to the above-described embodiment is provided with the standard pattern generation unit 11, the reference similarity generation unit 12, and the rejection threshold generation unit 13, as shown in FIG. However, in the case of a type of speech recognition device that does not require changes in the standard pattern, the standard similarity and the rejection threshold,
The standard pattern generation unit 11, the reference similarity generation unit 12, and the rejection threshold generation unit 13 are not provided. In such a voice recognition device, the standard pattern, the reference similarity, and the rejection threshold are stored in advance corresponding to the standard pattern storage unit 4, the reference similarity storage unit 16, and the rejection threshold storage unit 17, respectively.

The operation of the speech recognition apparatus configured as above will be described below. Before the voice recognition processing is executed, first, a standard pattern is generated, a reference similarity and a rejection threshold are generated. The standard pattern is generated by, for example, 50 people uttering a "pencil", and based on, for example, 50 voice patterns transmitted from the feature extracting means 3, a "pencil" pattern having the highest recognition rate is generated. Is generated. In this way, a standard pattern is generated for each word. Further, as described above, the reference similarity is generated for each standard pattern from the calculation result of the similarity between the evaluation pattern and the voice pattern of the unknown input voice. Each generated reference similarity is stored in the reference similarity storage means 16 in association with each standard pattern. The rejection threshold is generated for each reference similarity by multiplying the reference similarity by a coefficient.
The generated rejection threshold value may be stored in the rejection threshold value storage means 17 and sent from the rejection threshold value storage means 17 to the rejection decision means 14, but may be sent directly from the rejection threshold value generation means 13 to the rejection decision means 14. May be. In the following description, the case where the rejection threshold value generation unit 13 directly sends the rejection threshold value determination unit 14 will be described.

Next, the voice recognition operation will be described. The unknown input voice input from the microphone 1 is converted into a voice pattern required for recognition by the feature extraction unit 3 after the input level is optimized by the preprocessing unit 2 to remove unnecessary bands. The voice pattern is sent to the recognition processing means 5, and the recognition processing means 5 compares and collates each standard pattern read from the standard pattern storage means 4 to obtain the degree of similarity with each standard pattern. Further, the reference processing unit 5 is supplied with the reference similarity corresponding to each standard pattern read from the standard pattern storage unit 4 from the reference similarity storage unit 16. The recognition processing means 5 determines candidate words such as a first standard pattern, a second standard pattern, ... In the comparison and collation operation with the standard pattern, for example, it is possible to recognize the distinction between male and female, and to define a plurality of standard patterns in the same word to correspond to the dialect.

The rejection decision means 14 is supplied from the recognition processing means 5 with standard patterns which are candidate words in the order of, for example, the first standard pattern, the second standard pattern ,.
Further, each rejection threshold based on each reference similarity corresponding to these standard patterns is supplied from the rejection threshold generation means 13. The rejection determination unit 14 determines whether the similarity in each standard pattern supplied from the recognition processing unit 5 is less than or equal to each rejection threshold supplied from the rejection threshold value generation unit 13 as the first standard pattern and the second rank. Judge in the order of standard patterns. Then, when the similarity is less than the rejection threshold, the rejection determining unit 14 regards the standard pattern having the similarity as an incorrect answer and rejects it. On the contrary, if the similarity is equal to or more than the rejection threshold, the standard pattern having the similarity is regarded as a correct answer and is sent to the result output means 7, and the result output means 7 sends the standard pattern to the outside.

The rejection operation described above will be described more concretely, taking the case where an unknown input voice "telephone" is input as an example. As a result of comparison and matching between the unknown input voice of “telephone” and each standard pattern in the recognition processing means 5, the similarity between the unknown input voice and the standard pattern A is 100, and the similarity with the standard pattern B is 100. Is 70
If the similarity to the standard pattern C is 80, the recognition processing means 5 uses the standard pattern A as the first standard pattern, the standard pattern C as the second standard pattern, and the third standard pattern. The standard pattern B is ranked as. The standard pattern A has a standard similarity of 15
0 is added, and the standard pattern B has a standard similarity of 20.
0 is added, and the standard pattern C has a standard similarity of 60.
Is attached. If the coefficient for determining the rejection threshold is 0.8, for example, the rejection threshold for standard pattern A is 120 and the rejection threshold for standard pattern B is 16
The rejection threshold for 0 and the standard pattern C is 48. Therefore, in the rejection decision means 14, for the standard pattern A, the similarity 100 of the standard pattern A is the standard pattern A.
Since it is less than 120, which is the rejection threshold of, the standard pattern A is rejected although it is ranked first in the similarity. The second standard pattern C and the third standard pattern B are not rejected because each similarity is equal to or more than each rejection threshold, and the standard pattern C having the highest similarity among these results as a recognition result. It is sent from the output means 7.

Further, before performing the comparison using the above-mentioned reference similarity, the rejection decision method conventionally used may be first performed. On the contrary, the rejection determination method conventionally used may be applied to the standard pattern having the reference similarity or the similarity exceeding the rejection threshold. That is, for example, the first
The second standard pattern and the second standard pattern are obtained, and the ratio of the first similarity, which is the similarity in the first standard pattern, to the second similarity, which is the similarity in the second standard pattern, is first. For the first-rank standard pattern that is equal to or higher than a threshold and the first-rank similarity is equal to or higher than a second threshold, when the first-rank similarity or reference similarity of the first-rank standard pattern is equal to or higher than a rejection threshold. A method of adopting the first standard pattern as a recognition result may be adopted. By adopting such a configuration, it is possible to obtain a voice recognition result with higher accuracy than in the case of performing voice recognition based on the reference similarity.

Further, it may be configured as follows.
Depending on the environment in which the voice recognition device is used, an unknown input voice pattern different from the voice sample used when creating the standard pattern may occur. In order to be able to handle such cases as well, when creating a standard pattern, input in various environments, create a standard pattern for each environment according to each environment, and further for each standard pattern in such a different environment. An environment-specific reference similarity corresponding to the reference similarity is generated. In this way, a plurality of environment-based reference similarities corresponding to the environment are held for each standard pattern, and the environment-based reference similarity to be used is switched according to the instruction of the user, and further based on the environment-based reference similarity. It is also possible to switch and use the generated rejection threshold for each environment. In addition, if the system including the voice recognition device is provided with a detection function of its installation environment and its noise type, etc., it is possible to switch the environment-based reference similarity and the environment-based rejection threshold by using them. You may instruct. By adopting such a configuration, it becomes possible to set environment-specific rejection thresholds according to various environments in which unknown input speech is emitted,
The voice recognition operation is performed according to the environment, and thus the voice recognition result can be obtained with high accuracy.

Although the above description has been made for the case of an unspecified speaker, in the case of a specified speaker, an evaluation mode of a registered word is provided separately from the time of voice registration, and the recognition result of the evaluation mode is set. The above-mentioned reference similarity may be generated based on the similarity to. Further, it may be provided with a function of recalculating the generated reference similarity at the time of actual use or performing an update calculation.

Although the above description has been made using the word-based pattern matching method, the voice recognition apparatus and the voice recognition method in this embodiment are limited to the word-based method or the pattern matching method. However, it is also possible to apply it to a recognition method using a statistical method.

As described above, the reference similarity is added to each standard pattern, and the recognition result of the unknown input voice is not obtained only from the similarity with each standard pattern, but is recognized based on the reference similarity. Since the result is calculated, it cannot be rejected by the conventional rejection function.
It is possible to suppress the occurrence of erroneous answers, and more reliably prevent erroneous recognition due to noise other than voice or private language. Especially since each word has its rejection criteria,
The recognition result is not uniformly obtained only because the similarity is high. Therefore, the deviation of the rejected word is reduced and the recognition result is stabilized due to the variation of the voice samples when the standard pattern is created. In addition, if the voice sample at the time of creating the standard pattern of each word is good, the number of rejected words decreases relatively, and in the opposite case, the number of rejected words increases. It will be possible.

[0025]

As described above in detail, according to the present invention, the standard similarity is added to each standard pattern, and whether or not the standard pattern selected and listed as the candidate is adopted as the voice recognition result. Is determined based on the reference similarity. Therefore, not only the voice recognition is performed only based on the magnitude of the similarity between the unknown input voice and the standard pattern, but the voice recognition is performed in consideration of the reference similarity.
The correct answer rate in the voice recognition result can be improved,
A stable recognition result can be obtained for an unknown input voice.

[Brief description of drawings]

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

FIG. 2 is a block diagram showing a configuration of a conventional voice recognition device.

[Explanation of symbols]

2 ... Preprocessing means, 3 ... Feature extraction means, 4 ... Standard pattern storage means, 5 ... Recognition processing means, 7 ... Result output means, 11
... standard pattern generation means, 12 ... reference similarity generation means,
13 ... Rejection threshold generating means, 14 ... Rejection determining means.

Claims (6)

(57) [Claims] 1. A voice recognition device comprising: a standard pattern storing means for storing a standard pattern; and a recognizing means for recognizing the unknown input voice by comparing and collating the unknown input voice with the standard pattern. Te, and the reference similarity storage means for storing the reference similarity provided for each standard pattern of the standard patterns selected for the unknown input speech, adoption of standard patterns selected for the unknown input speech
Stored in the above-mentioned standard similarity storage means
The rejection threshold generated based on the above-mentioned standard similarity
And a rejection threshold value storage means for paid, it is the recognition means when determined based whether the recognition result a standard pattern which is the selected for the standard pattern selected in said comparison match to the reference similarity ,
In addition, the similarity between the unknown input voice and each standard pattern is calculated.
The first landmark with the highest similarity, which has the highest similarity
Second-order standard putter with second-order similarity to quasi-pattern
Of the above-mentioned first-rank similarity and second-rank similarity.
The ratio is equal to or higher than the first threshold, and the first similarity is the second threshold.
The above-mentioned 1st standard pattern which is more than the value
The means of recognition is that the above-mentioned first-rank similarity of the above-mentioned first-rank standard pattern is
The first standard pattern when it is equal to or more than the rejection threshold
A voice recognition device characterized by adopting . 2. A standard pattern generation means for generating the standard pattern and sending the generated standard pattern to the standard pattern storage means, and a standard similarity generated for generating the standard similarity to the standard similarity storage means. and a reference similarity generation means for sending, according to claim 1 Symbol placement of the speech recognition device. 3. The reference similarity has a plurality of environment-based reference similarities for one standard pattern, and the rejection threshold storage unit is generated for each of the plurality of environment-based reference similarities. stores environment by rejection threshold, the said recognition means to use for determining whether or not the recognition result a standard pattern selecting the environment-specific rejection threshold depending on the environment of the unknown input speech is generated, according to claim 1 or 2 Symbol placement of the voice recognition device. Wherein said reference similarity, for the same word one of which is generated from a plurality of standard patterns generated for the input speech
A standard pattern, determine the similarity between a plurality of evaluation patterns generated from a plurality of reference similarity generating the input speech for the same words an input voice of a different person from the above standard pattern generating input speech, The method according to claim 1 , which is generated by performing statistical processing based on the plurality of calculated similarities.
The voice recognition device according to any one of 3 above. 5. The voice recognition device according to claim 1 , wherein the standard pattern is a standard pattern for recognition of an unspecified speaker by a plurality of times of utterances by a plurality of people. 6. A voice recognition method for recognizing an unknown input voice by storing a standard pattern and comparing and collating an unknown input voice with the standard pattern, wherein the voice recognition method is most similar to the unknown input voice. The standard similarity that is set for each standard pattern of the high standard pattern is stored, and the rejection threshold for rejecting the adoption of the standard pattern that is generated based on the standard similarity and that is selected for the unknown input speech is stored. The standard pattern selected in the above comparison
Whether to use the selected standard pattern as the recognition result
When making a decision based on the above-mentioned standard similarity, unknown input
The similarity between the voice and each standard pattern is calculated, and the
1st standard pattern and 2nd with high 1st similarity
The second standard pattern having similarity of rank
The ratio between the first-rank similarity and the second-rank similarity is less than or equal to the first threshold.
Above, and the first similarity is equal to or greater than a second threshold
Regarding the first standard pattern, the above first standard pattern
If the above-mentioned first-rank similarity of is above the rejection threshold, then
A voice recognition method characterized in that a first-rank standard pattern is adopted .