JPH0731507B2 - Standard pattern registration method - Google Patents

Description

The present invention relates to a standard pattern registration method used for voice recognition, character recognition and the like.

(Prior Art and Its Problems) Voice recognition has many advantages over conventional means such as a keyboard as means for inputting data and control commands to a computer. Recently, a practical voice recognition device has been developed and put to practical use in various fields of industry.

In an application field where high accuracy is required, a specific speaker type voice recognition device is used. Also, in this type of device, a standard pattern is prepared for each unit, and when the operator inputs a necessary word by voice, the input pattern and the above standard pattern are compared (matching) to search for the most similar standard pattern. A pattern matching method in which recognition is performed by doing is used. By inputting / recording (that is, registering) the standard pattern by the voice of the operator himself, erroneous recognition due to individual differences is reduced. However, it is known that even a voice of the same word by the same person is accompanied by a considerable amount of pattern variation. For this reason, a measure is taken to register a standard pattern generated by a plurality of utterances for one word so that even if the input pattern changes, the standard pattern matches any one of the standard patterns. However, as a practical matter, it is difficult for a general operator to perform voicing registration with a pattern containing various variations. Certainly, as the fluctuation of the voice pattern, there are some that are described as phenomenon names such as vowel devoicing, nasalization, and articulatory obscuration. It is difficult to understand and register as such. For this reason, the operator often cannot repeatedly register a monotonous utterance and register a variation pattern that may occur by chance, and in actual recognition, there is no standard pattern that matches the input pattern, resulting in erroneous recognition. was there.

(Object of the Invention) The present invention provides a standard pattern registration method in which standard patterns of various varied shapes can be reliably prepared, the ability to adapt to individuality is high, and an economical memory amount and processing amount are sufficient. The purpose is to

(Structure of the Invention) In the standard pattern registration method according to the present invention, a plurality of standard patterns of the same category, which are uttered by a person having specialized knowledge of voice while intentionally varying, are stored in advance, When the training pattern of is input, it is compared with each of these standard patterns, the distance is calculated, the closest standard pattern is determined, and if the distance is less than a predetermined value, this standard pattern is used as the training pattern. It is characterized in that the training pattern is additionally stored as a new standard pattern when it is corrected and is equal to or more than a predetermined value.

(Principle of the Present Invention) In the above configuration, as a standard pattern prepared in advance, variously varied patterns are selected and stored in advance. That is, a utterance pattern including vowel devoicing, nasalization, voicing blunting, etc., which is known from phonetic experience, is prepared. Since these are not uttered by the operator himself, it suffices that those having specialized knowledge of voice consciously give a voice and memorize them. Alternatively, a designer having voice knowledge may select and store a sample considered to be appropriate from among a large number of voice samples.

Since these standard patterns are not the voice patterns of the operator himself, if they are used for recognition as they are, the recognition rate is naturally poor due to the influence of individual differences. Therefore, the operator registers with his own voice. The pattern input for this purpose is called a training pattern. It is better to input multiple training patterns in order to deal with pattern fluctuations, but if these are stored separately from the above standard patterns, the number of standard patterns will become large and calculation for pattern matching will be performed. The amount is also large. When the training pattern and the standard pattern prepared in advance are matched as in the above configuration, and the most similar standard pattern is determined, the standard pattern and the training pattern are patterns having the same tendency of variation. It will be. Therefore, this standard pattern is deleted and the training pattern is used as a new standard pattern.
When a plurality of training patterns are given, the same thing is repeated.

When recognition is performed by the pattern matching method using the set of standard patterns thus obtained, the following effects can be expected.

(1) For the input pattern having the same tendency of fluctuation as the standard pattern replaced by the training pattern, since the standard pattern of the person in question exists, accurate recognition is possible.

(2) Since the standard pattern is prepared in advance for the variation of the shape which is not realized by the utterance of the training pattern, the input pattern having such variation of the shape can be recognized.

(3) Generally, the variation of the shape realized by the training pattern frequently occurs as the input pattern even during the actual recognition. These are recognized with high accuracy by (1). In the case of (2), it can be said that the probability that the input pattern actually occurs in the form of fluctuation that was not realized by the training standard pattern is considerably low. Since the standard pattern that should match such an input pattern is different from the person's pattern, the recognition rate is expected to be a little low. However, since the frequency of occurrence of such a pattern itself is low, it does not become a problem. Rather, the merit of being able to recognize such a pattern that is rarely uttered to some extent is more significant.

(4) Since the number of standard patterns used for matching does not increase due to training, the amount of calculation does not increase.

As a method of correcting the standard pattern, it is conceivable to use a (weighted) average of the old standard pattern and the training pattern instead of the above replacement, but in the following description, the correction is performed by replacement.

(Embodiment) FIG. 1 is a diagram showing an example of the configuration of a voice recognition device implementing a standard pattern registration method according to the present invention. The basic standard pattern storage unit 50 stores a preset standard pattern group. The word name will now be indicated by the number n. Each word n stores a plurality of standard patterns. These are designated by the number m. It is shown by the m-th standard pattern B (n, m) of the word n.

The reference standard pattern storage unit 50 stores a standard pattern B (n, m) as shown in FIG. As shown in the figure, the number of standard patterns may be different for each word.

Before the registration process, these standard patterns are moved to the standard pattern storage unit 40. There it is retained in the same format as in FIG. In the registration process, each word is uttered three times, for example, and the standard pattern of the corresponding word is corrected.
Below are the standard patterns B (n, 1), B (n, 2), B (n,
3) and B (n, 4) are modified.

The word n is uttered three times, but the following processing is performed for each utterance. The voice input from the microphone 10 is analyzed by the analysis unit 20 and input to the buffer 30 as a training pattern A. The standard patterns B (n, 1) to B (n, 4) are sequentially output from the standard pattern storage unit 40 and sent to the matching unit 60. There, the standard patterns are compared with the input pattern A, distances D (n, m), m = 1,2,3,4 are calculated corresponding to the respective standard patterns, and the decision unit 80 Sent to. There, the number m that minimizes D (n, m) is determined. This number m is sent to the standard pattern storage unit 40. As a result, the standard pattern B (n,
m) is replaced by the training pattern A in the buffer 30, and a new B (n, m) is generated. By repeating the same process for three training utterances, the registration for the word n is completed. The same applies to other words.

The recognition operation is performed using the standard pattern group created as described above. In that case, the input pattern is input to the buffer 30, and the matching unit 60 compares the input pattern with each standard pattern to calculate the distance D (n, m). The determination unit 70 determines the word name n that minimizes the distance D (n, m) and outputs it as a recognition result. With such a configuration, it is possible to accurately recognize a voice pattern having a type of variation that has not occurred in the operator's own standard pattern registration utterance. Further, since the number of standard patterns does not increase, it is possible to avoid increasing the size of the device.

However, in the above-described embodiment, one standard pattern is always corrected when one training pattern is input. Therefore, unlike the standard pattern group, when a pattern that shows a variation peculiar to the operator is actually input, the standard pattern in which the distance to any of the standard patterns is large and the distance is accidentally the smallest Will be replaced. As a result, it may happen that the standard pattern cannot cope with the type of variation.

An embodiment in which this point is also solved will be described below.

In this embodiment, a plurality of standard patterns are stored for the same category, and when a training pattern of the same category is input, a comparison is made with each of these standard patterns to calculate the distance, and the closest standard pattern is calculated. A pattern is defined, and when the distance is less than or equal to a predetermined value, the standard pattern is modified with this training standard pattern, and when the distance is greater than or equal to the predetermined value, this training pattern is additionally stored as a new standard pattern. To do.

Patterns for training The variation peculiar to the person can be detected by testing the distance from the standard pattern. When it is determined that the variation is peculiar to the person, the old standard pattern is not modified and is added as a new standard pattern. That is, the difference between this embodiment and the previous embodiment is as follows.

In FIG. 1, the determining unit 80 compares the minimum value of the distance D (n, m) with a predetermined value Θ, and if D (n, m) ≦ Θ, the corresponding standard pattern B (n, m) as described above. Are replaced by the training pattern A. If D (n, m)> Θ, specify a new number as the signal m (number m = 5 above B (n, 1) to B (n, 4) in the example of FIG. 2) for training. The pattern A and a new standard pattern B (n, 5) are additionally stored.

According to such a method, the standard pattern is not adversely affected by operator-specific variations.

The principle of the present invention has been described above based on the embodiments, but these descriptions do not limit the scope of rights of the present invention. In particular, the word has been described as a unit of voice, but it may be another unit such as a phoneme or a syllable. Further, the present invention is not limited to so-called discrete utterance recognition, and is also applied to continuous utterance recognition.

(Effects of the Invention) As described in detail above, according to the present invention, it is possible to reliably prepare variously varying standard patterns, have high adaptability to individuality, and have an economical memory and processing amount. Can realize a sufficient standard pattern registration method, and in particular, if the present invention is applied to voice recognition or the like, the practical advantage is extremely large.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, and FIG. 2 is an operation explanatory diagram thereof. In the figure, 10 ... Microphone, 20 ... Analysis section, 30 ... Buffer, 40 ... Standard pattern storage section, 50 ... Basic standard pattern storage section, 60 ... Matching section, 70 ... Judgment section, 80 ... The deciding part is shown.

Claims (1)

[Claims] 1. A plurality of standard patterns of the same category, which are uttered by a person who has specialized knowledge of voice with intentional variation, are stored in advance, and a training pattern of the same category is input. , Comparing with each of these standard patterns, calculating the distance, determining the closest standard pattern, if the distance is less than a predetermined value, correct this standard pattern with the training pattern, and if it is more than the predetermined value Is a standard pattern registration method characterized by additionally storing a training pattern as a new standard pattern.