JPH0331274B2 - - Google Patents

Description

[Detailed Description of the Invention] (A) Technical Field of the Invention The present invention relates to a speech recognition device, in particular, when modifying or adding a standard feature pattern, based on the similarity to the standard feature pattern of other dictionary items, The present invention relates to a speech recognition device that can improve the quality of a dictionary by determining the validity of a standard feature pattern or a newly registered standard feature pattern.

(B) Prior art and problems In general, in speech recognition, in order to improve the recognition rate, it is important to determine what kind of feature parameters are extracted from speech information and used for matching. It is also important to determine how best to prepare the standard feature parameters representative of each item in the dictionary through extraction. No matter how good the method of feature extraction or matching is, the standard feature patterns registered in the dictionary may contain defective standard feature patterns such as noise-added patterns, unclear speech patterns, etc., such as "a". If there are many erroneous standard feature patterns due to utterance errors such as uttering ``i'' when ``i'' should be registered, the recognition rate will not improve.

Standard feature patterns are stored as digital information in the dictionary, and there are many of them, and they are not visible like mechanical parts, and not all standard feature patterns are uniformly used, so once registered. If this occurs, it is not easy to detect the above-mentioned defective standard feature patterns, erroneous standard feature patterns, etc.

Conventionally, all standard feature patterns once registered are treated as correct, and if a recognition error occurs,
Generally, it was assumed that the input voice to be recognized was bad or that recognition was at its limit, and that it was unavoidable. In addition, a learning method has been proposed that improves the quality of the dictionary by averaging the input feature pattern extracted from the input voice that caused the misrecognition with the already registered standard feature pattern. However, there was a problem in that if the input speech for learning was poor, the quality of the dictionary would deteriorate.

(C) Object and Structure of the Invention The present invention aims to solve the above-mentioned problems, prevents the registration of defective feature patterns due to pronunciation errors, noise, etc. in the registration mode or practice mode, improves the quality of the dictionary, and improves the quality of the dictionary. The purpose is to be able to improve the speech recognition rate. Therefore, the speech recognition device of the present invention performs speech recognition by comparing an input feature pattern obtained by acoustically analyzing unknown input speech with a standard feature pattern stored in advance corresponding to each item in a dictionary. In the speech recognition device that performs standard feature pattern addition/modification, the degree of similarity between the input feature pattern extracted from the input voice and the most similar standard feature pattern among dictionary items different from the input feature pattern is determined as the input feature. a pattern addition determination unit that determines whether the similarity between a standard feature pattern belonging to the same dictionary item as the pattern and the most similar standard feature pattern is greater than or equal to a predetermined value; a pattern correction addition unit that uses the input feature pattern to add/modify a standard feature pattern when the difference in similarity is greater than a predetermined value; It is characterized by having a re-registration section for re-registering the pattern.

The speech recognition device, which is another aspect of the present invention, recognizes speech by comparing input feature patterns obtained by acoustic analysis of unknown input speech with standard feature patterns stored in advance corresponding to each item in a dictionary. In a speech recognition device that performs recognition, there is an error detection unit that temporarily recognizes the input speech when adding/modifying standard feature patterns and detects recognition errors, and an error detection unit that detects recognition errors and input feature patterns that cause recognition errors. a pattern addition determination unit that determines whether the degree of similarity with the dictionary item is greater than a predetermined value; and a pattern addition determination unit that determines whether the degree of similarity with the dictionary item is greater than a predetermined value; The present invention is characterized by comprising a pattern modification and addition section that modifies or adds standard feature patterns related to the pattern. This will be explained below with reference to the drawings.

(D) Embodiments of the Invention FIG. 1 shows a diagram for explaining the relationship between the distribution of voice patterns and standard feature patterns.

In Figure 1, the parts surrounded by solid lines A, B, and C indicate the distribution of actual speech patterns in the pattern space, and A ₁ and A ₂ are for word A (including monosyllables; the same applies hereinafter). The registered standard feature patterns B 1 and B ₂ represent standard feature patterns for word B _, and C ₁ represents a standard feature pattern for word C. In some cases, one word item is covered by one standard feature pattern, as shown in illustration C, but usually, multiple standard feature patterns are prepared and recognized for one item, as shown in illustrations A and B. It usually covers the distribution range of power speech patterns. For example, when an input feature pattern X of unknown input speech is extracted, the input feature pattern X and each standard feature pattern A ₁ , A ₂ ,
The matching distance with B ₁ is calculated, and the item to which the standard feature pattern with the short distance belongs is taken as the recognition result.

FIGS. 2 and 3 are diagrams for explaining the outline of processing according to the present invention, and FIG. 4 shows the configuration of an embodiment of the present invention.

As can be seen from the explanation in Figure 1, if there is a defective standard feature pattern or an erroneous standard feature pattern that deviates from the distribution of speech patterns among the standard feature patterns registered in the dictionary, the recognition rate will deteriorate. becomes. The present invention prevents registration of such invalid standard feature patterns as follows.

For example, as shown in Figure 2, standard feature pattern _A1 is registered for each dictionary entry "Shibuya",
Assume that standard feature pattern B ₁ has already been registered for the dictionary item name "Hibiya." In this state,
Furthermore, suppose that in order to modify or add the standard feature pattern of "Shibuya", a utterance about the item name "Shibuya" is input, and the input feature pattern is _A2 . First, a standard feature pattern that is most similar to the input feature pattern _A2 is searched out among dictionary items other than the item name "Shibuya." For example, if this is the standard feature pattern B ₁ of the item name "Hibiya", then in order to check the validity of each feature pattern A ₁ , A ₂ , B ₁ , we will check the similarity of the standard feature patterns A ₁ and B ₁ . The difference between the degree of similarity and the degree of similarity between the standard feature pattern B ₁ and the input feature pattern A ₂ is calculated.
If the similarity between patterns B ₁ and A ₂ is greater than the similarity between patterns A ₁ and B ₁ by more than a predetermined threshold, then at least one of patterns A ₁ , B ₁ and A ₂ is , there is a high possibility that it deviates from the normal speech pattern distribution. Therefore, in such a case, a request is made to re-speak the item names "Shibuya" and "Hibiya", and the registration of the standard feature patterns _A1 and _B1 is re-registered. Through this re-registration, any registration mistakes made at the time of initial registration, such as utterance errors, will be corrected. Note that the threshold value used in the above similarity comparison may be uniformly determined in advance by the system, but it may be determined in advance by considering the commonality of phonemes, etc. for each two types of dictionary items to be compared. It is desirable to set it appropriately.

In particular, in the case of re-registration, and in order to prevent defective standard feature patterns from being registered, standard feature patterns to be re-registered may be determined, for example, as follows. Multiple utterances are input for registration of one standard feature pattern of one dictionary item. and,
For example, as explained in FIG. 3 A to C, average patterns are selected.

For example, assume that the input feature patterns extracted from four utterances are P ₁ , P ₂ , P ₃ , and P ₄ . In the case shown in Figure 3A, the centroids of patterns P ₁ , P ₂ , P ₃ , and P ₄ are roughly calculated in the pattern space, and pattern P ₃ closest to the centroids is selected as the standard feature pattern to be registered. There is. In the case shown in FIG. 3B, there are four characteristic patterns P ₁ , P ₂ , P ₃ ,
The average value of P ₄ is determined, and the average pattern P _n is registered as a standard feature pattern. Further, in the case shown in FIG. 3C, pattern P ₂ which is far away from other patterns is removed, and an average pattern P' _n of the remaining patterns P ₁ , P ₃ and P ₄ is determined and registered. By selecting or creating one standard feature pattern from a plurality of utterances in this way, the re-registered standard feature pattern becomes good.

FIG. 4 is a block diagram showing the configuration of an embodiment of the present invention, in which reference numeral 1 is a microphone, 2 is an acoustic analysis section, 3 is a pattern extraction section, 4 is a switching section,
5 is a pattern addition determination section, 6 is a pattern correction addition section, 7 is a dictionary section, 8 is a re-registration section, 9 is a display section, 1
0 represents a verification determination section.

The audio signal input from the microphone 1 is subjected to frequency analysis in the acoustic analysis section 2. The acoustic analysis unit 2 includes, for example, a group of band filters, a parameter extraction circuit, etc., and extracts features (parameters) of the input speech, such as a moment _M1 corresponding to the first formant frequency and a moment M1 corresponding to the second formant frequency. The moment M ₂ and further low-frequency power and high-frequency power are extracted, sample points related to these feature quantities are determined, and time-series information on the feature quantities is obtained.

The parameter time series information obtained by the acoustic analysis section 2 is input to the pattern extraction section 3. The pattern extraction unit 3 extracts an input feature pattern representing the characteristics of the input voice from this parameter time series information. The switching unit 4 switches registration or verification of pattern information in accordance with a switching instruction from a keyboard, for example.

The pattern addition determination unit 5 is activated, for example, when a standard feature pattern is further modified or added after registration of at least one pattern has been completed for each dictionary item. The pattern addition determining section 5 checks the validity of the input feature pattern for addition or modification extracted by the pattern extracting section 3, as described with reference to FIG. When the registered standard feature pattern is determined to be appropriate in relation to the degree of similarity with the input feature pattern, the input feature pattern is passed to the pattern modification/addition section 6 . The pattern modification/addition unit 6 performs a process of modifying the standard feature pattern or additionally registering it as a new standard feature pattern by performing a so-called averaging operation between the delivered input feature pattern and the already registered standard feature pattern of the same type of item. Execute.

If the pattern addition determining unit 5 determines that there is a possibility that the registered standard feature pattern is incorrect, the re-registration unit 8 is called. The re-registration unit 8 displays the suspected dictionary item name on a display unit 9 such as a CRT display, and instructs the user to re-input the registered voice. For example, from multiple utterances,
When each input feature pattern for re-registration is extracted by the pattern extraction unit 3, the re-registration unit 8 executes the process as explained in FIG. Decide on the pattern,
It is re-registered in the dictionary section 7 via the pattern correction/addition section 6. The dictionary section 7 is an external storage device such as a magnetic disk device, and stores and holds the item name to be recognized and the standard feature pattern in correspondence with each other.

When performing recognition on unknown input speech, the output of the pattern extraction section 3 is supplied to the matching determination section 10. The matching/judgment section 10 sequentially reads out the contents of the dictionary section 7, matches the input feature pattern and the standard feature pattern by, for example, well-known dynamic programming (DP), matching, etc., and outputs a recognition result.

FIG. 5 is a diagram for explaining the processing outline of the second invention, and FIG. 6 shows the configuration of an embodiment of the second invention.

For example, as shown in Figure 5, the dictionary entry "Shibuya"
Standard feature pattern A ₁ has been registered for
Standard feature pattern for dictionary entry “Hibiya”
Assume that B ₁ and B ₂ are registered. In registration mode or practice mode, for example, pattern A ₂
If the voice corresponding to "Shibuya" is input, it will be mistakenly recognized as the item "Hibiya". At this time, conventionally, it has been common practice to additionally register pattern _A2 in the item "Shibuya" or to perform a so-called averaging operation with the already registered standard feature pattern _A1 . However, it cannot be said with certainty that the voice "Shibuya", which is the basis of pattern _A2 , was uttered correctly. If there is a pronunciation error or the like, if pattern _A2 is reflected in the standard feature pattern, the quality of the dictionary will deteriorate. In the case of the present invention, the validity of pattern A _{2 is checked as follows by considering the similarity between pattern A 2 and} a registered standard feature pattern.

First, tentative recognition is performed using the input feature pattern _A2 . If a recognition error occurs, the degree of similarity between the input feature pattern _A2 and the dictionary item that resulted in the misrecognition, such as "Hibiya", is calculated. As the degree of similarity, for example, the average distance or minimum distance with standard feature patterns B ₁ and B ₂ included in the dictionary item "Hibiya" is used. If this degree of similarity is larger than a predetermined threshold for each dictionary item, that is, if the average distance or minimum distance is smaller than a predetermined value, there is a possibility that pattern _A2 is incorrect. Since it can be determined that it is large, the pattern
A ₂ is rejected without being used to add or modify the standard feature pattern. In this way, characteristic patterns of speech due to pronunciation errors and the like are prevented from being reflected in the dictionary.

FIG. 6 is a block diagram showing an embodiment of the second invention, in which reference numerals 1 to 7 and 10 correspond to those in FIG. 4, 11 represents an error detection section, and 12 represents a keyboard.

When audio is input from the microphone 1 in the registration mode, practice mode, etc., it is acoustically analyzed and a characteristic pattern is extracted by the pattern extraction unit 3. This input feature pattern is notified to the error detection section 11 via the switching section 4, and is also supplied to the matching determination section 10. The comparison determination unit 10 sequentially reads standard feature patterns from the dictionary unit 7 and performs provisional recognition in the same way as in normal recognition. This recognition result is displayed on a display (not shown) or the like. The error detection unit 11 detects this display by
For example, an incorrect instruction input from the keyboard 12 is detected. If there is no error instruction and the recognition result is correct, the same processing as before is continued.

If an erroneous instruction is detected, the pattern addition determination unit 5 is notified of the erroneously recognized item name and input feature pattern. The pattern addition determination unit 5
As explained with reference to FIG. 5, the degree of similarity between the input feature pattern and the dictionary item that was incorrectly recognized as a recognition result is calculated, and compared with a predetermined threshold, the validity of the input feature pattern is determined. Check.
If the similarity is determined to be greater than a predetermined threshold, there is a high possibility that the input feature pattern is defective, so the input feature pattern may be added to the dictionary section 7 or
Stop modifying the existing standard feature pattern by averaging it. Only when the above-mentioned similarity is smaller than a predetermined threshold value, the pattern modification/addition section 6 performs modification by additionally registering the input feature pattern in the dictionary section 7 or by averaging operation with an existing standard feature pattern. .

(E) Effects of the Invention As explained above, according to the present invention, invalid speech patterns, such as those caused by pronunciation errors, are prevented from entering the dictionary, and the quality of the dictionary is improved, so the recognition rate is improved. do.

[Brief explanation of the drawing]

FIG. 1 is a diagram for explaining the relationship between the distribution of voice patterns and standard feature patterns, FIGS. 2 and 3 are diagrams for explaining the outline of the process according to the present invention, and FIG. Embodiment configuration, FIG. 5 is a diagram for explaining the processing outline of the second invention, and FIG.
The figure shows the configuration of an embodiment of the second invention. In the figure, 1 is a microphone, 2 is an acoustic analysis section,
Reference numeral 3 represents a pattern extraction section, 4 a switching section, 5 a pattern addition/determination section, 6 a pattern correction/addition section, 7 a dictionary section, 8 a re-registration section, 10 a collation determination section, and 11 an error detection section.

Claims (1)

[Scope of Claims] 1. Speech recognition that performs speech recognition by comparing input feature patterns obtained by acoustic analysis of unknown input speech with standard feature patterns stored in advance corresponding to each item in a dictionary. In the device, when adding/modifying a standard feature pattern, the degree of similarity between the input feature pattern extracted from the input voice and the most similar standard feature pattern among dictionary items different from the input feature pattern is the same as the input feature pattern. a pattern addition determination unit that determines whether the similarity between a standard feature pattern belonging to a dictionary item and the most similar standard feature pattern is greater than a predetermined value by a predetermined value; a pattern modification/addition unit that uses the input feature pattern to add/modify a standard feature pattern when the difference in similarity is greater than a predetermined value; A voice recognition device comprising a re-registration unit for registering. 2. The re-registration unit inputs a plurality of utterances for each dictionary item, and registers the pattern closest to the center of gravity of the input feature patterns as the standard feature pattern. Speech recognition device described in section. 3. The re-registration unit inputs a plurality of utterances for each dictionary item, and registers an average pattern of the input feature patterns as a standard feature pattern. voice recognition device. 4 The above-mentioned re-registration unit inputs multiple utterances for each dictionary item, and registers the average pattern of those input feature patterns, excluding patterns that are significantly different from other patterns, as a standard feature pattern. Claim 1 characterized in that
Speech recognition device described in section. 5. In a speech recognition device that performs speech recognition by comparing an input feature pattern obtained by acoustic analysis of unknown input speech with a standard feature pattern stored in advance corresponding to each item in a dictionary, the standard feature pattern an error detection unit that temporarily recognizes input speech to detect recognition errors when adding/correcting the input voice; and an error detection unit that detects recognition errors by performing temporary recognition on the input speech; a pattern addition determination unit that determines whether the degree of similarity is greater than a predetermined value; and a pattern addition determination unit that determines whether or not the similarity is greater than a predetermined value; A speech recognition device comprising: a pattern correction/addition section for performing pattern correction/addition. 6. The speech recognition device according to claim 5, wherein an average distance or a minimum distance to a standard feature pattern in a dictionary item is used as the degree of similarity.