JPH0752354B2 - Voice recognizer - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a voice recognition device for recognizing a voice of an unspecified speaker, and in particular, a standard pattern having a low contribution to correct recognition is deleted to improve recognition performance. The present invention relates to an improved voice recognition device.

(Prior Art) Conventionally, a voice recognition device for an unspecified speaker receives an analog voice signal as an input, performs A / D conversion, performs frequency analysis, etc., and converts it into some feature amount. Recognize the category of the standard pattern with the maximum similarity by performing similarity calculation (pattern matching) between the input pattern converted into the amount and the standard pattern created in advance for each category to be recognized. It is organized into categories.

As a method of creating a standard pattern, for each category to be recognized, processing such as clustering and averaging is performed on a voice pattern, which is the feature amount of a plurality of speakers (generally about 1000 people) prepared in advance, per category. The standard method is to create multiple standard patterns (about 20 in practice). The voice pattern used in creating the standard pattern is hereinafter referred to as a learning pattern.

(Problems to be Solved by the Invention) However, the voice recognition device having the above configuration has the following problems.

The learning pattern used for creating the standard pattern is made up of only a limited range of speakers, and efficiently covers the characteristic distribution of the limited speakers. Therefore, when the learning pattern is recognized by using the standard pattern, the possibility of misrecognition is extremely small, and the clustering is efficiently performed in each category, so that a plurality of standard patterns can be recognized in the feature space. Since all the standard patterns have the highest similarity among all the standard patterns when the learning pattern is correctly recognized, the standard patterns are evenly distributed among the standard patterns. However, in the case of recognizing an unspecified speaker, the number of speakers to be recognized is far greater than the number of speakers constituting the learning pattern, and the variation in the feature distribution is also large in the learning pattern. It will be large compared to that.

That is, the feature distribution of the learning pattern and the feature distribution of the unlearned input pattern when recognizing an unspecified speaker are different. Therefore, when a recognition process is performed on an unlearned input pattern, the frequency at which the similarity becomes the maximum among all standard patterns during normal recognition due to the difference in feature distribution as described above Concentrate on a specific standard pattern rather than evenly between them, and among the standard patterns, the number of times that the maximum similarity among all standard patterns at the time of correct recognition is small and the contribution to recognition is small, or other category A standard pattern that tends to cause misrecognition appears.

For the above reasons, having a standard pattern that has a small contribution to correct recognition causes an ineffective similarity calculation time for recognition, resulting in an increase in recognition processing time, or recognition due to the existence of a standard pattern that easily causes erroneous recognition. There is a problem that a phenomenon that causes a decrease in the rate occurs, and in any case, the performance of the voice recognition device deteriorates.

In order to solve this problem, for example, as disclosed in Japanese Patent Laid-Open No. 59-15990, deleting a standard pattern that easily causes misrecognition also contributes to correct recognition of the deleted standard pattern. When it is large, the number of false recognitions and the number of correct recognitions both decrease, which makes it difficult to improve the overall recognition performance.

The present invention provides a speech recognition apparatus capable of eliminating the above-mentioned drawbacks and deleting a standard pattern having a small contribution to recognition to improve the recognition performance.

(Means for Solving Problems) According to the present invention, first storage means for storing in advance a standard pattern for each category created by processing a plurality of voice patterns, and the first storage means. Means for calculating a similarity between the standard pattern stored in the storage means and an input voice pattern to be recognized, and means for determining and outputting the category name of the standard pattern having the maximum similarity as a recognition result. When a recognition process is performed on the input voice pattern using the standard pattern, a counting unit that counts the number of correct recognition results and the number of false recognition results for each standard pattern, and at the time of creating the standard pattern When the recognition processing is performed on the plurality of used voice patterns using the standard pattern, the result of counting the number of correct recognition results for each standard pattern is calculated in advance. A second storage means for storing the same; and a calculation means for calculating the non-contribution degree to the correct recognition of each standard pattern based on the counting result of the counting means and the storage content of the second storage means. Accordingly, there is provided a voice recognition device in which a standard pattern having a high non-contribution degree is deleted from the first storage means based on the calculation result of the calculation means.

(Operation) The first storage means stores in advance a standard pattern for each category created by processing a plurality of voice patterns (learning patterns). On the other hand, the second storage means stores the result of counting the number of correct recognitions (contribution to correct recognition) for each standard pattern when the recognition process is performed on these learning patterns using the standard pattern. It is stored in advance. The counting means corrects recognition frequency (contribution to correct recognition) and misrecognition frequency (contribution to misrecognition) when recognition processing is performed on the input voice pattern to be recognized by using a standard pattern. Is counted for each standard pattern. The degree of non-contribution to each standard pattern considers not only the number of correct recognitions and the number of false recognitions counted by the counting means but also the number of correct recognitions stored in the second storage means (for example, from the number of false recognitions It is calculated (by subtracting both positive recognition times). Since the standard pattern is deleted according to the non-contribution calculated in this way, people who perform utterances far from the utterance pattern (learning pattern) of the standard generator used when creating the standard pattern are concentrated in the book. When a voice recognition device is used (for example, when the device is used in an area with a specific dialect), the recognition contribution is high for the original standard utterance, but the recognition contribution is limited only in that area. It is possible to prevent the inconvenience that the standard pattern is determined to be low and is deleted by mistake. As a result, the recognition performance can be improved.

(Embodiment) FIG. 1 is a block diagram showing an embodiment of the present invention. In the figure, 1 is a similarity calculator that calculates the similarity between the input pattern A and the standard pattern B, and 2 is the standard pattern B.
A standard pattern storage unit for storing the standard device, 3 a control unit for controlling the entire apparatus, 4 a learning pattern positive recognition contribution ratio storage unit for storing the contribution of learning patterns for each standard pattern to correct recognition, and 5 a standard pattern for each standard pattern. A positive recognition counting unit that counts the contribution to correct recognition, 6 is a false recognition counting unit that counts the contribution to false recognition for each standard pattern, and 7 is a learning pattern correct recognition contribution storage unit 4 and a correct recognition count. It is a non-contribution degree calculation unit that calculates the non-contribution degree in the recognition processing for each standard pattern from the output signals of the unit 5 and the misrecognition counting unit 6 and outputs the number of the standard pattern to be deleted to the standard pattern storage unit 2.

Next, the operation will be described with reference to the operation flowchart of FIG. First, the control unit 3 outputs a recognition start command signal 3a to the similarity calculation unit 1, the correct recognition counting unit 5, and the false recognition counting unit 6. The similarity calculation unit 1 sequentially inputs the input pattern A from the time when the control unit 3 inputs the recognition start command signal 3a (step). At this time, the correct category name is also entered. The similarity between the input pattern A and all the standard patterns B stored in the standard pattern storage unit 2 is calculated (step) and compared with the correct answer category name to determine whether the recognition is correct or erroneous. (Step).
As a result of this determination, if the recognition is correct, the correct recognition standard pattern number 1a, which is the standard pattern number having the highest degree of similarity, is output to the correct recognition counting unit 5, and if the recognition is incorrect, the recognition error counting unit 6 is performed. The operation of outputting the erroneously recognized standard pattern number 1b, which is the standard pattern number with the highest similarity, is repeated until the input of the input pattern A to be recognized is completed. In order to confirm the effect of the present invention, it is desirable that the number of times the input pattern is input is 10,000 or more. After the input of the input pattern 1 is completed, the similarity calculation unit 1 outputs the recognition end signal 1c to the control unit 3. The control unit 3 is the similarity calculation unit 1
After inputting the recognition end signal 1c, the counting result output command signal 3b is output to the correct recognition counting unit 5 and the false recognition counting unit 6.

The positive recognition counting unit 5 has a function of calculating the positive recognition contribution for each standard pattern by the following means. The similarity calculation between one unlearned input pattern and each standard pattern (similarity calculation unit 1) is performed, and a score is given to the standard pattern that gives the maximum similarity among all the standard patterns in the case of correct recognition. Add 1 (step). This operation starts from the time when the recognition start command signal 3a is input from the control unit 3 and is performed for each voice input until the time when the count result output command signal 3b is input from the control unit 3. The score obtained for each standard pattern in the above operation is defined as the correct recognition contribution rate, and is used as a parameter indicating the magnitude of the contribution to the correct recognition for an unlearned input pattern.

The misrecognition counting unit 6 has a function of calculating the misrecognition contribution rate for each standard pattern by the following means. A similarity calculation (similarity calculation unit 1) between a certain unlearned input pattern and each standard pattern is performed, and a score is given to the standard pattern that gives the maximum similarity among all the standard patterns when misrecognized. Add 1 (step). This operation starts from the time when the recognition start command signal 3a is input from the control unit 3 and is performed for each voice input until the time when the count result output command signal 3b is input from the control unit 3. The score obtained for each standard pattern in the above operation is defined as the misrecognition contribution rate, and is used as a parameter representing the magnitude of the contribution to the misrecognition for an unlearned input pattern.

Next, control unit 3 determines whether or not the input of the input pattern is completed (step). If the processing has not been completed, the above processing is repeated, and if the processing has been completed, the next processing is performed.

The learning pattern normal recognition contribution storage unit 4 has a function of storing the learning pattern normal recognition contribution for each standard pattern defined by the following description. The similarity between a certain learning pattern and each standard pattern is calculated, and a score of 1 is added to the standard pattern that gives the maximum similarity among all the standard patterns when the correct recognition is performed. The above operation is performed for all learning patterns in advance and the score obtained for each standard pattern is defined as the learning pattern positive recognition contribution 4a,
It is stored as a parameter indicating the magnitude of contribution to correct recognition for the learning pattern.

The correct recognition counting unit 5 and the false recognition counting unit 6 input the count result output command signal 3b from the control unit 3, and then the former recognizes the positive recognition contribution rate.
5a, the latter outputs the misrecognition contribution 6a to the non-contribution calculation unit 7. The non-contribution calculation unit 7 calculates the contribution 5a of correct recognition and the contribution 5 of false recognition.
6a and the learning pattern correct recognition contribution 4a are input to calculate the non-contribution in the recognition process for each standard pattern as described below (step). The standard pattern number is SN
O, positive recognition contribution REC _(i) , i = 1 to SNO, false recognition contribution E
RR _(i) , i = 1 to SNO, learning pattern positive recognition contribution is IR
When E _(i) , i = 1 to SNO and non-contribution degree is NDEL _(i) , i = 1 to SNO, NDEL _(i) = CE ・ ERR _(i) -CR ・ REC _(i) -CI・ IRE _(i) Define in. That is, the misrecognition contribution rate 6a is large, and the correct recognition contribution rate is 6a.
The smaller the 5a and the contribution 4a of the learning pattern positive recognition, the larger the non-contribution.

After calculating the non-contribution degree, the non-contribution degree calculating unit 7 deletes the standard pattern number 7a from the standard pattern having the largest non-contribution degree in order to delete the standard patterns of the predetermined number of deletion patterns. Output to 6. The standard pattern storage unit 2 deletes the standard pattern corresponding to the input deleted standard pattern number 7a (step).

As described above, in the present embodiment, by calculating the non-contribution degree during the recognition process by using the three types of parameters including the correct recognition contribution rate, the learning pattern correct recognition contribution rate, and the incorrect recognition contribution rate, for example, the incorrect recognition contribution rate Is large and the contribution to correct recognition is large, the standard pattern with a low contribution to recognition is avoided by avoiding the deletion of standard patterns that reduce the number of false recognitions but also reduce the number of recognitions. Since I deleted
The recognition processing speed can be increased by reducing the number of times of pattern matching, and the recognition accuracy can be improved by decreasing the number of false recognitions.

(Effects of the Invention) As described above, according to the present invention, a standard pattern that originally should have a high contribution to a recognition result is erroneous with respect to a recognition result by a speaker having a voice quality greatly deviating from a standard utterance. Since it is possible to avoid the risk of being deleted in advance and to delete only the standard pattern that has a low contribution to the true recognition, it is possible to speed up the recognition processing speed and improve the recognition accuracy by reducing the number of times of pattern matching.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, and FIG. 2 is a flow chart showing the operation of the embodiment of FIG. 1 ... Similarity calculation unit, 2 ... Standard pattern storage unit, 3 ...
Control unit, 4 Learning pattern correct recognition contribution storage unit, 5
…… Correct recognition counting unit, 6 …… False recognition counting unit, 7 …… Non-contribution degree calculation unit.

Claims (1)

[Claims] 1. A first storage means for storing in advance a standard pattern for each category created by processing a plurality of voice patterns, and the standard pattern stored in the first storage means. A means for calculating a similarity with an input voice pattern to be recognized, a means for determining and outputting a category name of a standard pattern having the maximum similarity as a recognition result, and a standard pattern for the input voice pattern. When performing recognition processing using, the counting means for respectively counting the number of correct recognition results and the number of erroneous recognition results for each standard pattern, and for the plurality of voice patterns used when creating the standard pattern A second storage unit that stores in advance the result of counting the number of correct recognition results for each standard pattern when the recognition process is performed using the standard pattern; The non-contribution means calculates the non-contribution degree of each standard pattern to the correct recognition based on the counting result of the number means and the stored content of the second storage means. A voice recognition device, characterized in that a standard pattern having a high degree of contribution is deleted from the first storage means.