JPH09212189A - Method and device for speech recognition - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the precision of speech recognition by giving a threshold (what is called a moving threshold value) which can reflect a difference in the similarity of a 1st candidate in detail according to the difference in the similarity of the 1st candidate. SOLUTION: For example, the distance of a 1st candidate is defined as a variable and the distance difference between 1st and 2nd candidates is defined as a function of the distance of the 1st candidate. This function can be given a pattern of an increase trend which increases the distance difference as, for example, the distance of the 1st candidate increases. Then the concrete distance value (a) of the 1st candidate is substituted in the function (threshold value function) and then a function threshold value f(a) characteristic of the distance value 8a) is obtained. When the actual distance difference between the 1st and 2nd candidates is in a lower area (B) on the basis of the function threshold value f(a), the 1st candidate is judged to be misrecognized (rejected) and when the difference in an upper area (C), the 1st candidate is judged to be recognized correctly.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is based on a comparison of the similarity between an input voice and a plurality of predetermined voice candidates, and recognizes the voice as matching with any of the plurality of candidates. The present invention relates to a recognition method and its device.

[0002]

2. Description of the Related Art Conventionally, as this type of voice recognition device,
As a similarity criterion, for example, distance calculation is performed on the pattern matching between the actual input voice and a plurality of candidates, the candidate with the smallest distance having the highest similarity is set as the first candidate, and the first candidate is recognized. There is one that determines whether or not. Then, as a reject determination method in which the first candidate cannot be used as a recognition result, for example, one of the following methods or a combination thereof is known.

If the distance a of the first candidate is larger than the threshold value, it is rejected. The difference between the distance a of the first candidate and the distance b of the second candidate (a-
If b) is smaller than the threshold value, it is rejected. Ratio of distance a of the first candidate and distance b of the second candidate (a / b)
However, if is smaller than the threshold value, it is rejected. Here, is the case where the recognition result cannot be obtained because the similarity of the first candidate is absolutely low, and or is the first similarity because the similarity of the first candidate and the similarity of the second candidate are close to each other. This is the case where the candidate cannot be determined as the recognition result. As a combination of and, there is JP-A-61-114299 and the like, and a related one is JP-B-3-44316 and the like.

[0004]

However, further improvement is desired in view of the demand for surely eliminating erroneous recognition and the like even by these methods. In particular, the recognition accuracy is greatly affected by how to determine the threshold value that serves as a criterion for determination such as erroneous recognition. If the threshold is set to a strict criterion, the recognition accuracy is improved, but the first candidate is rejected at a high rate, which is inconvenient for the user. On the other hand, if the threshold value is made gentle, the rate at which the first candidate is rejected decreases, but a voice candidate not intended by the user is presented, which causes a decrease in reliability.

An object of the present invention is not to be uniform according to the difference in the similarity of the first candidate (for example, the magnitude of the distance), but a threshold value (so to speak, move so that the difference in the similarity is finely reflected). By assigning a threshold value), erroneous recognition is eliminated as much as possible, and what is to be judged as correct recognition is prevented (rescue) from being carelessly rejected.

[0006]

A speech recognition method according to the present invention compares an input voice with a plurality of predetermined candidates and calculates the similarity between them to determine that the input voice has a plurality of candidates. Assuming that one of the first candidate and the second candidate, which is ranked from the one with the highest similarity, is used as a function of the similarity of the first candidate, on the assumption that which one corresponds to the first candidate and the second candidate. By defining / correlating in advance and applying a concrete value of the similarity of the first candidate to the function, the corresponding function value is used as a function threshold, and the first candidate is adopted as a recognition result based on this function threshold. It is characterized by determining whether or not to do.

Further, the speech recognition apparatus according to the present invention suitable for carrying out such a method analyzes the following requirements, that is, candidate storage means for storing a plurality of candidates, and characteristics of input speech. A voice analysis means; a similarity calculation means for calculating the similarity between the analysis result and the plurality of candidates; and a similarity difference between the first candidate and the second candidate, which is a reference from the higher similarity, or the similarity. Relative value calculating means for calculating the ratio, the similarity difference or similarity ratio between the first candidate and the second candidate in advance,
Reference function storing means for defining / associating as a function of similarity of the first candidate and storing the relationship as a reference function for recognition suitability determination, applying the similarity of the first candidate as a variable value to the reference function, and corresponding The first candidate is recognized on the basis of the function threshold value determining means having the function value as a threshold value, and the magnitude difference between the similarity difference or the similarity ratio between the first and second candidates and the function threshold value. As a result, it includes a suitability judging means for judging whether it is suitable or not.

FIG. 2B shows a comparative example with respect to the present invention. Distance as the similarity of the first candidate (for the input voice)
If the distance of the first candidate is larger than the threshold value P, the first candidate is determined to be erroneous recognition (area A ′), and if it is smaller than the threshold value P, the distance of the first candidate is further decreased. And whether the difference between the distances of the second candidate and the second candidate is larger than the uniformly fixed threshold value Q. Below the threshold Q (B ')
In the area of (1), it is determined that the first candidate is erroneous recognition because the priority of the first candidate with respect to the second candidate is not clear, and in the area above the threshold value Q (C ′), the first candidate is determined to be correct recognition.

Here, when the distance value of the first candidate is small (the similarity with the input voice is high), even below the threshold value Q,
Actually, the first candidate may be a correct recognition (marked with O), and conversely, when the distance value of the first candidate is large (the similarity with the input voice is low), even if the upper side of the threshold value Q is exceeded, Is the first
The candidate may be misrecognized (x mark). In other words, those that should not be judged (rejected) as misrecognition (circle)
Should be rejected, and conversely should be rejected (×
Mark) results in being recognized as correct, which leads to a decrease in recognition accuracy.

On the other hand, in the present invention, an example thereof is shown in FIG.
As shown in (a), not a fixed threshold value but a threshold value as a function is defined (reference function). That is,
For example, the distance of the first candidate is used as a variable, and the distance difference between the first and second candidates is taken as a function of the distance of the first candidate. This function can be, for example, a pattern of increasing tendency in which the distance difference increases as the distance of the first candidate increases. Then, when the specific distance value a of the first candidate is substituted into the function (which can be called a threshold function), the distance value a
To obtain a threshold value f (a) peculiar to This is a function threshold in the sense that it is given by a function. Further, with the function threshold value f (a) as a reference, the first and second
If the actual distance difference of the candidate is in the lower (B) area, the first
The candidate is judged (rejected) as false recognition, and if it is in the upper area (C), the first candidate is judged as correct recognition.

As a result, when the distance of the first candidate is small, the threshold value of the distance difference between the first and second candidates can be made gentle, and it should be judged that the recognition is proper (O).
Mark) is prevented from being rejected. So to speak, it is a relief from rejection. On the other hand, when the distance of the first candidate is large, it is possible to make the threshold value of the distance difference between the first and second candidates more strict, so that what is originally rejected (x mark) is correctly recognized. Prevent being judged. As a result, it is possible to improve the recognition accuracy while suppressing the number of rejects from being unnecessarily increased.

Note that, in FIG. 2A, similarly to FIG. 2B, a foot cut threshold P is set for the distance of the first candidate, and an area (A) deviating from the foot cut threshold P is set. With regard to the above, it is possible to judge misrecognition without considering the function threshold value, and for the region within the cutoff threshold value P, the recognition judgment using the function threshold value can be performed as the next step. . However, reverse the order of the recognition steps,
After the recognition judgment by the function threshold, the cutoff threshold P
May be added, or the recognition threshold may be independently determined based on the function threshold value without using the cutoff threshold value P in the first place.

The function (standard function) that defines the threshold value is not limited to a linear function (simple straight line), but may be a stepwise function as shown in FIG. It can be appropriately determined according to the target recognition standard, such as a curved shape like (b) and (c), a polygonal shape like (d) and (e).

Further, the above function defining the threshold is
The distance ratio (a / b) between the distance a of the first candidate and the distance b of the second candidate is not limited to the similarity difference such as the distance difference between the first and second candidates, and the distance of the first candidate is used as a variable. It can also be defined by the similarity ratio of. Even in this case, the degree to which the similarity of the first candidate is more significant than that of the second candidate is reflected, and particularly when the distance ratio (a / b) is adopted, the smaller the value, the first candidate. Of the second candidate is higher than that of the second candidate, and in that sense, the distance difference (ba)
It is the opposite relationship. Therefore, in order to obtain the same effect as described above in FIG. 2A, the function that is based on the distance ratio (a / b) has an inclination opposite to the distance difference (ba) (downward in the figure). Will be the one.

[0015]

BEST MODE FOR CARRYING OUT THE INVENTION Based on the embodiment shown in the drawings,
An embodiment of the present invention will be described. FIG. 1 is a block diagram showing an example of a device configuration. A voice analysis unit 2 extracts a characteristic pattern of a voice input from a voice input unit such as a microphone 1, and the feature pattern is temporarily stored in an input voice buffer 3. . On the other hand, the standard voice pattern dictionary unit 4 stores a standard voice pattern for each word, for example, as a candidate storage unit. The CPU 5 determines the characteristic pattern of the input voice as
By comparing with each standard voice pattern that is multiple candidates and performing distance calculation, which is an example of similarity calculation, the distance between each candidate and the characteristic pattern of the input voice is calculated, and the result is the candidate rank and distance. It is temporarily stored in the buffer 6. Here, the first candidate, the second candidate, and so on are in order of decreasing distance (higher similarity).

The reference function memory 7 as the reference function storage means defines the distance difference between the first and second candidates as a function of the distance (variable) of the first candidate, as shown in FIG.
A related reference function (also called a threshold function) is stored. This is defined in advance as a function equation, or the distance between the first candidate and the distance difference between the first and second candidates are associated and stored as a storage table such as a conversion table or a correspondence map. Further, when the cutoff threshold P for the distance of the first candidate in FIG. 2A is stored, a cutoff threshold memory 8 for that is added.

The CPU 5 reads the distance value a of the first candidate from the candidate rank / distance buffer 6 and uses this distance value a as shown in FIG.
The function threshold value f (a) is obtained by applying it to the reference function of (a), and this is temporarily stored in the function threshold value buffer 9. The CPU 5 also uses the candidate rank / distance buffer 6
The distance difference (b−a) between the first candidate distance “a” and the second candidate distance “b” is calculated based on the above data, and the calculated value (b−a) is calculated.
By comparing whether or not a) is larger than the function threshold f (a), it is determined whether or not the first candidate may be a recognition result. That is, it is determined whether the first candidate is correct recognition or incorrect recognition, and the recognition result is output from the I / F unit 10. In this embodiment, the CPU 5 also serves as the similarity calculation means, the relative value calculation means, the function threshold value determination means, and the suitability determination means.

Next, the flow of voice recognition as described above will be described with reference to FIG. The voice input is captured in step S1, and the characteristic pattern is extracted in step S2. In S3, distance calculation is performed between the characteristic pattern and each standard pattern (candidate) to obtain each distance.

In S4, the distance of the first candidate is shown in FIG.
It is compared whether it is larger than the cutoff threshold P of (a), and if it is larger, the similarity of the first candidate is insufficient,
Reject in S8, and the corresponding process (for example, a message that the user has difficulty recognizing voice or a request for re-input) is performed in S9.

When the distance value of the first candidate clears the cutoff threshold value of S4, the actual distance value a of the first candidate is substituted into the reference function of FIG. 2 (a) in S5 ( Alternatively, the correspondence relationship is read out) to obtain the function threshold f (a) for the distance difference between the first and second candidates. Further S6
Then, it is determined whether or not the actual distance difference (ba) between the first and second candidates is greater than or equal to the function threshold f (a), and if so, S7
And conclude that the first candidate is adopted as the recognition result (the first candidate is correct recognition), otherwise, the above S8,
The flow is S9. Needless to say, the above description is merely an example, and the present invention is not limited to this.

[Brief description of drawings]

FIG. 1 is a block diagram showing an example of a device configuration of the present invention.

FIG. 2 is an explanatory view showing an example of the present invention and a comparative example in comparison and contrast.

FIG. 3 is a flowchart showing an example of the flow of voice recognition according to the present invention.

FIG. 4 is a conceptual diagram showing a modified example of a threshold function (reference function) of the present invention.

[Explanation of symbols]

 1 Microphone 2 Voice analysis unit 3 Input voice buffer 4 Standard voice pattern dictionary unit 5 CPU 7 Reference function memory 9 Function threshold buffer

Claims (5)

[Claims] 1. A method of comparing an input voice with a plurality of predetermined candidates and recognizing which of the plurality of candidates the voice corresponds to by calculating their similarity, wherein the similarity is high. The similarity difference or similarity ratio between the first candidate and the second candidate, which is the rank from the other side, is defined and related in advance as a function of the similarity of the first candidate, and a specific value of the similarity of the first candidate. Is applied to the function to determine a corresponding function value as a function threshold value, and whether or not to adopt the first candidate as a recognition result is determined based on the function threshold value. Method. 2. An apparatus which compares an input voice with a plurality of predetermined candidates and recognizes which of the plurality of candidates the voice corresponds to by calculating the similarity between them. A candidate storage unit for storing, a voice analysis unit for analyzing the feature of the voice, a similarity calculation unit for calculating the similarity between the analysis result and the plurality of candidates, and a reference from the one with the higher similarity. Relative value calculating means for calculating a similarity difference or a similarity ratio between a certain first candidate and a second candidate, and a similarity difference or a similarity ratio between the first candidate and the second candidate in advance for the first candidate. Reference function storing means for defining / associating as a function of similarity and storing the relationship as a reference function for recognition suitability determination; and applying the similarity of the first candidate to the reference function as a variable value, and corresponding function The value is the threshold Whether or not the first candidate is appropriate as a recognition result based on a number threshold value determining means and a magnitude relationship between the similarity difference or similarity ratio between the first and second candidates and the function threshold value. A voice recognition device comprising: an adequacy determining unit that determines whether or not. 3. The similarity difference between the first and second candidates is defined and related as a function of the similarity of the first candidate, the reference function storage means stores the relation, and the function threshold value is stored. When the determining unit applies the specific similarity value of the first candidate to the reference function and sets the function value as the function threshold value, the suitability determining unit determines that the suitability between the first and second candidates is large. If the difference in similarity is larger than the function threshold, it is determined that the first candidate can be adopted as a recognition result, and if the difference in similarity is smaller than the function threshold, the first candidate cannot be adopted as a recognition result. The voice recognition device according to claim 2, wherein 4. The reference function defined by the similarity difference between the first and second candidates using the similarity of the first candidate as a variable shifts to a side where the similarity of the first candidate decreases. 4. The voice recognition device according to claim 3, wherein the function threshold value is set in such a manner that it increases as a whole. 5. The similarity according to claim 3 or 4, wherein the similarity of the first candidate is compared with a predetermined fixed cutoff threshold value, and the similarity of the first candidate is the cutoff threshold value. If the first candidate cannot be adopted as a recognition result, a cutoff determination means for determining a rejection indicating that the first candidate cannot be adopted is provided. A voice recognition device, characterized in that the suitability of using the first candidate as a recognition result is determined based on the function threshold value.