JPS63229496A - Pattern updating system for voice recognition - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech recognition pattern update method, and more particularly to a standard pattern update method in a speech recognition device that recognizes word speech of an unspecified speaker.

Conventionally, in word speech recognition devices targeting unspecified speakers, one method of speaker adaptation in order to improve the recognition rate is to use the pattern of one speaker in advance (2 recorded). There is a method of updating a standard pattern, but if the updated pattern is defective, it may actually reduce the recognition rate.

It is cumbersome to update the pattern every time an erroneous recognition occurs, and there are many problems from the viewpoint of providing the user with an easy-to-use recognition device.

The present invention was made in view of the above-mentioned circumstances.
A speech recognition device, especially a speech recognition device for words uttered by an unspecified speaker, uses the utterance pattern of the speaker using the recognition device to pre-store standard patterns of recognition target words that are registered and stored in advance. This was done with the aim of adapting the performance of the recognition device to the user and improving the recognition rate by adding, updating, etc. without modifying the standard pattern for non-specific speakers. be.

Structure In order to achieve the above object, the present invention provides a voice L! that recognizes word sounds of unspecified speakers. ! In the cognitive device.

If a word uttered by a certain speaker is misrecognized more than a certain number of times against pre-registered word standard patterns for unspecified speakers, the registered standard pattern corresponding to the misrecognized word is used. Remove the speech pattern from the recognition target without changing it and add a new speaker's utterance pattern to the recognition target, or list similar words in the recognition target word set in advance before the speaker enters recognition mode. The feature is that all similar last words are uttered and added to the standard pattern for unspecified speakers.

Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a block diagram for explaining one embodiment of the present invention, in which 1 is a voice input section, 2 is a feature extraction section, 3 is an entry/recognition mode switching section, and 4 is a standard pattern 5 is a storage unit, 5 is a word speech recognition unit, and 6 is a recognition result output unit.The word speech input from the speech input unit 1 is analyzed by a group of bandpass filters and LPG analysis in a feature extraction unit 2, and then extracted with feature parameters. Converted to time series. Mode switching section 3
Now, when the switch SW is in the registration mode 31, a word standard pattern is registered and stored in the standard pattern registration section 4. When the switch SW is in the recognition mode 32, the standard pattern is quoted from the standard pattern storage section 4, and the word recognition section 5 performs pattern matching with the unknown input voice to select one of the most similar patterns. 1M! ! The P recognition result output section 6 outputs the result.

FIG. 2 is a block diagram for explaining the standard pattern updating method according to the present invention. In the figure, 7 is an erroneous recognition counting section, 8 is an update mode switching judgment section, 41 is an update pattern storage section, and C is an update section. Mode switching control (No. 1) Other parts having the same functions as those in the embodiment shown in FIG. 1 are given the same reference numerals as in FIG. 1. In FIG.

When the switch SW of the mode switching unit 3 is in the recognition mode 32, the unknown human voice input from the voice input unit 1 passes through the feature analysis unit 2 and is recognized by the word recognition unit 5. When erroneous recognition occurs, there are two possible cases: a case where the utterance pattern is defective and a case where the standard pattern 4 for unspecified speakers is defective. However, if erroneous recognition still occurs even after uttering the same word multiple times, it is more reasonable to consider that the standard pattern itself is significantly different from the speaker's utterance pattern than to consider that the utterance pattern is defective. From this point of view, if the same word is incorrect more than a certain number of times, the control signal C to the mode switching unit 3 is
and change the switch SW to update mode 31 to update the pattern.In this case, instead of deleting the existing standard pattern for unspecified speakers, a new pattern for a new speaker is added and registered. . Then, when the switch SW is switched to the recognition mode 32 again, the update pattern 41
, and take measures not to access the standard pattern for unspecified speakers that corresponds to the update pattern. This makes it possible to update the standard pattern applied to the speaker without modifying the non-specific pattern.

In Figure 3, prior to recognition, similar words in the word set to be recognized are listed in advance, and the speaker utters the sounds of those similar words - column table to create a standard for unspecified speakers. This is a diagram for explaining the method of additionally registering a pattern. In the diagram, 9 is a similar m-word uniform table derived from the recognition target word column of the standard pattern storage unit 4. It is now possible to display. Before entering the recognition mode, the speaker should utter the similar words shown on displays 1 and 0 and add the standard pattern.
, and after switching the switch SW to recognition mode 32, recognition is performed. At this time, priority is given to accessing the additional patterns in the additional registration section 41, as in the embodiment shown in FIG.

FIG. 4 is a diagram for explaining an example of creating similar words in the similar words column table 9 in FIG. 3. In FIG. 1, data (DEETA), delta (DERUTA),
Correction (TEESEE), formation (KEESEIE), body (KOKU) and 6 (U in 110) are based on the number of different phonemes, and the distance between words is defined as 2.0.1.0.1.0, respectively. In this way, a method can be considered in which a list of word pairs in which the distance between words in the recognition target words is equal to or less than a certain value (for example, 2.0) is considered.

FIG. 5 is a diagram for explaining another example of creating the similar word column of the similar word column table 9 in FIG. 3, that is, a diagram showing an example of a confusion matrix between words;
In the figure, the numbers of the words to be recognized are 1, 2, etc. ..., N (
N is the number of words), and the value of the fusion between the nth and mth words is p<ntm). If there is no confusion between words, the matrix becomes an identity matrix (all diagonal elements are 1 and all other elements are O), but in general, for pairs of similar words, O<p(n, m) takes a value of <1. Therefore, a pair of words (π, depth) (E (
A method of selecting all (n, m)l p (n, m)≧Th) can also be considered.

In this way, the speaker's voice can be additionally registered with a simple procedure without modifying the standard pattern.
A standard pattern adapted to the speaker can be created, and the recognition rate can therefore be greatly improved. Also,
The pattern can be adapted and recognized by applying the same procedure to other speakers, without causing the problem of patterns being adapted only to a specific speaker, which previously existed when updating standard patterns. rate can be improved.

Effect-1 As is clear from the above explanation, according to the present invention, when using a word speech recognition device for unspecified speakers, the speech pattern of the speaker can be used without erasing the standard pattern.
By adding and updating the pattern, listing similar words from among the words to be recognized in advance, and additionally registering the speaker's voice, it is possible to speak without changing the original standard pattern for non-specific speakers. It can be applied to a person's vocal pattern. In other words, there is no need to modify the pre-registered standard pattern for unspecified speakers for use by a specific speaker, so the standard pattern can be applied only to a specific user and not to other users. Recognition can be improved by providing a standard pattern that can be applied to all users without causing adaptational harm.

[Brief explanation of drawings]

FIGS. 1 to 3 are block diagrams for explaining an embodiment of the pattern update method according to the present invention, and FIGS. 4 and 5 each illustrate a method for creating a list of similar words. It is a figure for explaining. DESCRIPTION OF SYMBOLS 1... Voice input section, 2... Feature extraction section, 3... Registration/recognition mode switching section, 4... Standard pattern storage section. 5... Word speech recognition unit, 6... Recognition result output unit, 7
. . . Erroneous recognition count section, 8. Update mode switching determination section, 9. Similar words column table, 10. Display.゛, 2-

Claims (4)

[Claims] (1) In a speech recognition device that recognizes word sounds from unspecified speakers, words uttered by a certain speaker are incorrectly recognized more than a certain number of times against pre-registered word standard patterns for unspecified speakers. 1. A pattern updating method for speech recognition, characterized in that, in the case of a misrecognized word, a registered standard pattern corresponding to an erroneously recognized word is removed from the recognition target without altering it, and a new speaker's utterance pattern is registered and added. (2) List similar words in the word set to be recognized in advance, and before the speaker enters recognition mode,
A speech recognition pattern update method characterized by uttering all similar words and additionally registering them in a standard pattern for unspecified speakers. (3) As a method for listing similar words in a set of words to be recognized, a method is selected in which the distance between kana sequences or the distance between phoneme sequences in a word is less than a certain threshold. A speech recognition pattern update method according to claim (2). (4) As a method for listing similar words in a set of words to be recognized, the values of the elements of the matrix are The speech recognition pattern updating method according to claim (2), characterized in that a word pair having a value equal to or greater than a certain threshold is selected.