JPH0257320B2 - - Google Patents

Description

[Detailed Description of the Invention] (A) Technical Field of the Invention The present invention relates to a speech recognition device, particularly a speech recognition device equipped with a dictionary in which a plurality of standard feature patterns are registered for each item to be recognized. The present invention relates to a speech recognition device that automatically deletes standard feature patterns that have been registered based on voice information that has been registered, in a registration/practice mode or in a recognition mode, thereby improving the quality of a dictionary.

(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 utterance patterns, or, for example, "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 happens, it is not easy to detect the 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, this method is based on the premise that the registered standard feature pattern is correct to some extent, and if the standard feature pattern is incorrect, there is a problem that convergence is slow.

(C) Purpose and Structure of the Invention The present invention aims to solve the above-mentioned problems by detecting and automatically deleting unsuitable standard feature patterns when there is an invalid standard feature pattern in the registration or practice mode or in the recognition mode. The aim is to improve the quality of dictionaries and increase the recognition rate without placing as much burden on dictionary creators or users as possible. Therefore, the speech recognition device of the present invention includes a dictionary in which one or more standard feature patterns are stored corresponding to each item to be recognized, and input feature patterns obtained by acoustic analysis of unknown input speech and the above-mentioned input feature patterns are provided. In a speech recognition device that performs recognition by comparing standard feature patterns in a dictionary, there is an error detection unit that detects errors in the recognition result, and when the error detection unit detects a recognition error, it detects a recognition error that is most similar to the input feature pattern. Compare the similarity between the most similar standard feature pattern and other standard feature patterns of the item to which the most similar standard feature pattern belongs, and the similarity with the standard feature pattern of the correct item corresponding to the input voice of the most similar standard feature pattern. and a registration for deleting the most similar standard feature pattern from the dictionary if the difference or ratio of the degrees of similarity is greater than a predetermined reference value based on the determination result of the similarity determination unit. It is characterized by having a pattern deletion section. Embodiments will be described below with reference to the drawings.

(D) Embodiments of the invention FIG. 1 is a diagram for explaining the relationship between the distribution of voice patterns and standard feature patterns, FIG. 2 is a diagram for explaining the outline of the processing according to the present invention, and FIG. The structure of an embodiment of the invention, FIGS. 4 to 6 are diagrams illustrating processing of each embodiment of the similarity determining section, and FIG. 7 is a diagram illustrating processing of an embodiment of the pattern deletion section.

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 to B ₃ 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 between B ₁ , .

If the standard feature patterns registered in the dictionary include a defective standard feature pattern or an erroneous standard feature pattern that deviates from the distribution of speech patterns, the recognition rate will deteriorate. The present invention aims to improve the recognition rate by deleting such unsuitable standard feature patterns.

For example, as shown in FIG. 2, it is assumed that the distribution of the voice pattern of the word "Shibuya" is S in the diagram, and the distribution of the voice pattern of the word "Hibiya" is H in the diagram. When creating a dictionary, when registering multiple standard feature patterns, due to an operational error or a pronunciation error, the user mistakenly pronounced "hibiya" instead of "shibuya", resulting in the standard feature pattern being registered. Assume that the characteristic pattern S3 has been registered. Although the standard feature pattern S3 is actually a speech pattern for "Hibiya", it will be stored in the dictionary as belonging to the word "Shibuya".

Once registered as described above, for example, when recognizing the utterance of "Shibuya",
Only standard feature patterns S1 and S2 are matched, and pattern S3 is not matched. However, the fact that pattern S3 is incorrect is not detected, and it is simply treated as if no utterance corresponding to pattern S3 is made. On the other hand, it is assumed that the input feature pattern X for "Hibiya" is uttered as shown in FIG. 2, for example. Since the distance d ₁ between the input feature pattern X and the standard feature pattern S3 is smaller than the distance d ₂ between the input feature pattern X and the standard feature pattern H3, the pattern
will be recognized as the word "Shibuya". In this case, according to the conventional learning method, rather than determining that the standard feature pattern S3 is incorrect, it is determined that the standard feature patterns H ₁ , H ₂ , H ₃ of the word "Hibiya" are inappropriate, and " Standard feature patterns belonging to "Mebiya" were added and modified. Therefore, the error standard feature pattern S3 will be left as is in the dictionary.

In the case of the present invention, when a recognition error is detected, it is checked whether the standard feature pattern S3 is valid as follows, and if it is not valid,
The standard feature pattern S3 is deleted from the dictionary. In addition, whether or not a recognition error occurred is determined by
When in registration mode or practice mode, the system knows what the input word is and can immediately detect it. Furthermore, even in the recognition mode, if there is a means for indicating recognition errors and correct answers, errors can be detected by the user's instructions.

When it is determined that _the speech recognition for _input feature pattern The similarity with standard feature patterns H ₁ , H ₂ , and H ₃ in the heterogeneous dictionary entries is examined. Then, based on these similarities, the validity of the standard feature pattern S3 is determined. As a criterion for similarity, for example, the average value of the illustrated distance D ₁ and distance D ₂ in Figure 2, and the distance D′ ₁ ,
Using the average value of distance D' ₂ and distance D' ₃ , the validity of pattern S3 is determined depending on whether the difference or ratio between these average values is larger than a predetermined reference value.

Also, for example, as a criterion for similarity, the distance D ₁ ,
The minimum value of distance D ₂ and the minimum value of distance D′ ₁ , distance D′ ₂ , and distance D′ ₃ may be used. In this way, when it is detected that pattern S3 is not valid, pattern S3 is deleted from the dictionary. At this time, for example, various other conditions such as confirmation from the user may be set as conditions for registration deletion.

FIG. 3 is a block diagram showing the configuration of one embodiment of the present invention. In the figure, 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 section, 6 is a dictionary, 7 is a matching judgment section, 8
9 represents a result display section, 9 a keyboard, 10 an error detection section, 11 a similarity determination section, and 12 a pattern deletion 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. Moment _M2 , low frequency power, high frequency power, etc. are extracted, sample points related to these feature quantities are determined, and time series information of 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 response to a mode switching instruction from the keyboard 9, for example. When a registration instruction is given, the pattern addition section 5 additionally registers the input feature pattern extracted by the pattern extraction section 3 in the dictionary 6 in association with its item name. The dictionary 6 is, for example, an external storage device such as a magnetic disk device, and stores and holds the item names to be recognized and information on standard feature patterns.

When performing recognition, the output of the pattern extraction section 3 is
The data is supplied to the comparison determination section 7. The comparison/judgment unit 7 sequentially reads out the contents of the dictionary 6, and compares the input feature pattern with the standard feature patterns registered in the dictionary using, for example, well-known dynamic programming (DP).
Verify by matching etc. The recognition results are displayed on a result display section 8 such as a display. The user can check whether the voice input by the user has been correctly recognized by viewing the displayed results.
If it is incorrect, the user must, for example, use the keyboard 9
, it is indicated that the recognition result is incorrect and what the correct answer is. This allows the error detection unit 10 to detect recognition errors. Note that in the registration/practice mode, when the word of the input voice is known in advance, the error detection unit 10 immediately detects an error from the determination result of the verification determination unit 7. You can also do it.

When the error detection unit 10 detects a recognition error, it notifies the similarity determination unit 11 to that effect. The similarity determination unit 11 executes the process shown in the flowchart of FIG. 4 or 5, for example, and checks the validity of the standard feature pattern that is most similar to the input feature pattern. Criteria for determining validity
T _Ld may be uniformly determined by the system, or may be determined in advance according to two types of dictionary items. The difference or ratio of the degree of similarity such as the average distance or minimum distance explained in Figure 2 is 0.
If it is larger than the above predetermined reference value T _Ld ,
Start the pattern deletion section 12. In addition, in order to reduce the risk of accidentally deleting the correct standard feature pattern, for example, as shown in FIG.
First, the number of times N that the most similar standard feature pattern associated with a recognition error has been selected at the time of a recognition error is counted, and the pattern is deleted only when this number of times N exceeds a predetermined specified value _NL . You can do it like this. Note that only when this recognition error occurs multiple times, the determination process for deleting the standard feature pattern may be performed not by the similarity determination unit 11 but by another processing unit.

The pattern deletion unit 12 deletes the most similar standard feature pattern notified from the similarity determination unit 11 from the dictionary 6. Then, if necessary, the result display section 8 displays the fact that it has been deleted. By the way, in order to prevent the pattern deletion function from working unnecessarily, for example when the input voice is unclear or recognition is performed in a noisy environment, it is necessary to allow pattern registration deletion. It would be convenient if, for example, the keyboard 9 could be used to instruct either a state in which registration or deletion of a pattern is prohibited. Therefore, it is preferable to provide a status storage unit (not shown) that stores whether the registration deletion is currently permitted or prohibited.
In this case, the pattern deletion section 12 refers to the state storage section, as shown in FIG. 7, for example, and after confirming that deletion is currently permitted, deletes the registration of the most similar standard feature pattern. Note that this state determination process may be executed by another processing unit.

The process for deleting the standard feature pattern may be performed only during the registration/practice mode, or may be performed during the normal recognition mode.

(E) Effects of the Invention As explained above, according to the present invention, it becomes possible to automatically delete defective standard feature patterns such as noise-added patterns and ambiguous patterns, as well as erroneous standard feature patterns caused by incorrect utterances during registration. , it becomes possible to improve the quality of the dictionary and improve the recognition rate.

[Brief explanation of the drawing]

FIG. 1 is a diagram for explaining the relationship between the distribution of voice patterns and standard feature patterns, FIG. 2 is a diagram for explaining the outline of processing according to the present invention, and FIG. 3 is a diagram for explaining the configuration of an embodiment of the present invention. 4 to 6 are explanatory diagrams of processing of one embodiment of the similarity determination section, and FIG. 7 is a diagram of illustration of processing of one embodiment of the pattern deletion section. In the figure, 2 is an acoustic analysis section, 3 is a pattern extraction section,
6 is a dictionary, 7 is a collation determination unit, 10 is an error detection unit,
11 represents a similarity determination section, and 12 represents a pattern deletion section.

Claims (1)

[Claims] 1. A dictionary in which one or more standard feature patterns are stored corresponding to each item to be recognized is provided, and input feature patterns obtained by acoustic analysis of unknown input speech and those in the dictionary are provided. A speech recognition device that performs recognition by comparing with a standard feature pattern includes an error detection unit that detects errors in the recognition result, and a most similar to the input feature pattern when the error detection unit detects a recognition error. Similarity that compares the degree of similarity between the standard feature pattern and other standard feature patterns of the item to which the most similar standard feature pattern belongs, and the degree of similarity between the most similar standard feature pattern and the correct item standard feature pattern corresponding to the input voice. a determination unit; and a registered pattern deletion unit that deletes the most similar standard feature pattern from the dictionary when the difference or ratio of the degrees of similarity is greater than a predetermined reference value based on the determination result of the similarity determination unit. A speech recognition device characterized by having the following features. 2. When comparing the degrees of similarity, the similarity determination unit calculates the average distance or minimum distance between the most similar standard feature pattern and the standard feature pattern in the same type of dictionary entry, and the standard feature in the different type of dictionary entry. The speech recognition device according to claim 1, characterized in that the difference or ratio of the average distance or the minimum distance to the pattern is used. 3. A patent claim characterized in that the number of recognition errors is stored for each of the standard feature patterns, and when the number of recognition errors exceeds a predetermined number, the corresponding most similar standard feature pattern is deleted. The speech recognition device according to item 1 of the scope of the invention. 4 The recognition error detection unit, the similarity determination unit, or the registered pattern deletion unit is provided with means for selecting either the registration deletion permission state or the registration deletion prohibition state, and only when the registration deletion permission state is set, the registration pattern deletion unit selects the registration deletion permission state. 2. The speech recognition device according to claim 1, wherein the speech recognition device is configured to delete similar standard feature patterns.