JPS60169900A - Word registration system - 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 (al) Technical Field of the Invention The present invention relates to a word registration method in a word speech recognition device that recognizes limited words by a specific speaker. The development of complex, sophisticated logic circuits for speech recognition algorithms, and means requiring high speed and large capacity have been supported by the realization of miniaturized and low cost highly integrated circuit elements (LSI) due to advances in semiconductor integration technology. Rare.

Speech recognition devices using various recognition methods have appeared on the market and are becoming widespread in various industrial fields.

In particular, speech recognition devices that recognize speech by specific speakers have improved speech recognition rates and are used in various fields, but they are limited to specific speakers! 1′! Regarding the word registration method for recognizing words, various registration methods are being considered to be simpler and to improve the recognition rate.

tc+ Prior Art and Problems A conventional speech recognition device of this type for a specific speaker will be explained. FIG. 1 shows a block diagram of a conventional speech recognition device for a specific speaker. When registering a word by a specific speaker, specify a certain word from the input terminal 1, manually input the notation of the word and the word pronunciation by the specific speaker, and then pass the input speech through multiple band filters in the speech analysis feature extraction unit 2. The spectral outputs applied to the group and tl are sampled with frame synchronization of several ms for each spectral output, extracted as word feature variation time series by linear time normalization, and the extracted feature parameter time series is It is temporarily stored in the feature parameter buffer 3. The feature parameter time series stored in this feature parameter buffer 3.

When the solid line side of the mode changeover switch 4 is turned on, the specified word is written into the word set memory 5a of the ff1 word SE/NODO file group 5 as the designated word standard pattern. The word memory of this word set file group 5 is provided with n term memory 5a to 5n corresponding to the limited number of words to be registered.
Then, other words are sequentially designated and the word sounds are inputted, and are written as the word standard pattern of the designated word into the word set memories 5a-5n of the word file group 5 through the same route as above. n (IIi 1 minute 5 words are registered, the registration is completed, and the word cent dictionary for the specific speaker is completed.

Next, when recognizing an unknown human-generated voice of a specific speaker, switch the mode selector switch 4 to the recognition mode side (dotted line side). The characteristic parameters are stored as a time series in the characteristic parameter folder 3 via the path, and are temporarily stored as an input pattern in the memory of the input pattern buffer 6 via the mode changeover switch 4. The memory of this input pattern nopanopher 6 is updated and rewritten each time a voice is input. The word standard pattern of n1tlil in word memory 5a-5n of word cent file group 5 is as follows.

Input cover turn of input cover turn buffer 6 and recognition comparison unit 7
The degree of similarity between the two is sequentially compared in step 9, and the requested degree of similarity is inputted into the recognition result selection section 8, and the word standard pattern having the maximum degree of similarity with the input keyword is selected, and the recognition result is Output from output terminal 9. However, in order to avoid misrecognition, the recognition result selection unit 8 has a monitoring function that monitors the degree of similarity, and outputs the recognition result only when the degree of similarity exceeds a certain dark value, and outputs it when the degree of similarity is below the dark value. No recognition results are output from terminal 9. That is, 1, for example, similarity "100"
is a complete match and the darkness value is set to "90", and if the recognition comparison unit 7 determines the similarity of a certain unknown input voice to be "85", the determination output of the recognition result selection unit 8 is rejected. Therefore, a check is made by saying the rejected word several times.

The above is the registration and recognition procedure of limited words by a specific speaker in a conventional speech recognition device, but in this conventional method, there is a possibility that words with extremely high similarity are registered among the registered words. Even if the learning check that is performed several times during recognition mode is performed, if erroneous recognition occurs frequently during actual use, confusion will occur in the recognition operation, and it will be necessary to register the specific word again. During this time, the recognition operation is interrupted, which has the disadvantage of significantly reducing the operating rate of the device.
Inoko.

(d+ Purpose of the Invention The present invention has the drawback that even if a word is registered, it cannot be checked whether it is likely to be mistaken until the word is uttered several times. When doing so, the purpose is to have the system check for words that are likely to be mistaken in advance without having to check each utterance in order to register many easy-to-recognize words such as "ce" and "nod" that will not be misrecognized except by the influence of noise.

(e) Structure of the Invention The above object is to convert a plurality of word input speech by a specific speaker into an input pattern and register it as a word standard pattern in a plurality of word center dictionaries. a selection means for selecting a pattern that is acoustically similar to the word standard patterns registered in the word standard patterns; and determining the similarity of the word standard patterns based on the number and time length of monosyllables of the word standard Bakuns selected by the selection means. and a notification means for notifying the word of the li word standard pattern determined by the similarity comparison means.

In the limited word speech recognition device that recognizes the unknown input speech of the specific speaker by comparing it with the word standard patterns registered in the plurality of word cent dictionaries, the plurality of words of the specific speaker are used as the word standard patterns. After registering the word in the word cent dictionary, the r'+ii selection means selects a standard word pattern that is acoustically similar.

This is achieved by the present invention, in which the selected word standard patterns are determined by the similarity determining means to determine the degree of similarity, and the notification means notifies the specific speaker of the words resulting from the determination.

That is, in a limited word speech recognition device that recognizes unknown input speech of a specific speaker by comparing it with word standard patterns registered in a plurality of word cent dictionaries, a plurality of word A selection means selects standard words that are acoustically close to each other.

Comparison of each word standard pattern is performed by configuring the system to determine the degree of similarity using the scale determination means and notify the specific speaker of the resulting word using the notification means.
If the similarity is within a certain level, it has a function that notifies you to change the pronunciation to a different reading as it has a high probability of misrecognition, and registers it in the word cent dictionary. The system can check in advance for words that are close to each other (which are likely to be misrecognized), and words that are likely to be misrecognized are rejected, and when actually used, many words that are not misrecognized except for the influence of noise etc. This provides a word registration method that improves recognition rate and operation rate because only easily recognized word sets are registered.

(“) Example of the Invention An example of the present invention will be described below.Specifically,
First, for the pronunciation of each word, we replaced it with a group name of close monosyllables created systematically, and statistically created a group name for each subsyllable only for words that matched the replaced Rahel. Find the similarity between words based on the similarity with each subsyllable.

If the value is smaller than a certain value, it is assumed that the word is a word that is likely to be misrecognized, and the word encyclopedia operator is notified after the completion of the word cent, thereby preventing the registration of words that are likely to be misrecognized in advance.Figure 2 1 shows a block diagram of the configuration of similar word selection of registered words according to the present invention, and the same objects are indicated by the same symbols as in FIG. 1. 11 is a corresponding statistics table.

Reference numeral 12 denotes a character string creation section, 13 a character string extraction section, 14 a similarity comparison unit, 15 a similarity comparison section, 16 a similarity determination section, and 17 a display device.

The correspondence statistics table 11 is a statistical table of the correspondence between the monosyllables of words and their group names. From the monosyllabic character strings of words, the correspondence between the characters in each character string is calculated, and the statistics are calculated for the correspondence between the monosyllabic character strings. to create. For example, in "Ichikawa" and "1 Ishikawa", the character strings "Ichikawa" and "Ishikawa" are in the same group. A character string creation unit 12 is a circuit unit that creates a character string by replacing the character string of each word with a group name for each word set.2 A character string extraction unit 13 is a circuit unit that extracts a character string that matches the group name character string. The similarity statistics table 14 shows the statistically created similarities between monosyllables,
and a continuation table of the time length between single syllables, the similarity comparison unit 15 is a counting circuit that calculates the similarity between terms for matching words, and the similarity determination unit 16 is a counting circuit that calculates the similarity between terms based on the number of middle syllables and the time length. A display 17 is a device for displaying a message and is a determining circuit that determines the degree of similarity and outputs words within a certain value.

In FIG. 2, the configuration block diagram and the procedure up to registering words in the word set memories 5a to 5n of the word cent file group 5 are the same as those described above, and will therefore be omitted. Word cent memory 5a for each word top node by registered operator
- When word registration is completed on 5n.

A character string creation unit 12 converts the pronunciation of each word into a monosyllable character string, and uses the recognition method to call and replace the corresponding group name of the group from the statistically created correspondence statistical chiffle 11. The character string of the group name replaced by the character string creation part 12 is input to the character string extraction part 13, and only one matching word from the group name character string input by the two character string extraction part 13 is extracted. Extract. The extracted words are manually input to the similarity comparison unit 15, and the similarity comparison unit 15 further uses the similarity table between each monosyllable statistically created in the similarity statistical table 14 to compare the words between the words. Calculate similarity. For example, in the above example, calculate the total similarity of "ii", "chi-shi", [(1] degree) of "ka-ka", and "wa-wa" similarity. Class A or higher determination unit 16
Enter. The similarity determination unit 16 performs determination based on the relationship between the number of monosyllables and the time length. This determination assumes that if the result obtained by the comparison unit 15 is a degree of similarity with a moderate value to some extent, there is a possibility that misrecognition will occur between the Clj words. Therefore.

The similarity determination unit 16 displays words whose similarity is within a certain value on the display 17. At this time, candidates for monosyllables that can be replaced with other monosyllables are listed together with their positions from the character string of the reading of the word.

Words that are likely to be misrecognized are displayed in a list on a display I7 and the registered operator is notified. Register the words displayed on the display 17, re-enter a meaningful pronunciation at P, -, 4, and re-check based on that.

Many difficult-to-recognize word sets that will not be misrecognized are registered before switching to the recognition mode.

As mentioned above, statistically create a monosyllable similarity table in advance, refer to the chiful from the pronunciation of each word,
Let's compare each word.

For those whose similarity is within a certain degree.

This is a word registration method that has a function that notifies you to change the pronunciation to a different pronunciation because the probability of misrecognition is high.

4gl As described in detail of the invention, in a limited word speech recognition device that recognizes unknown input speech of a specific speaker by comparing it with a word standard pattern registered in a plurality of word top dictionaries, After registering in the Nodo dictionary, the 5 selection means selects an acoustically similar word pattern, the similarity determination means determines the degree of similarity, and the notification means notifies the specific speaker of the resulting word. According to the present invention, standard patterns of each word are compared, and if the degree of similarity is within a certain level, the probability of misrecognition is high, and a notification is sent to change to another pronunciation. , you can check in advance for similar words (easily misrecognized) in the word cents that are registered in the word cent dictionary, and when actually using the word cents.

Since only many easy-to-recognize word cents that are not misrecognized except by the influence of noise etc. are registered, the recognition rate and the operating rate of the device are improved.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of a conventional speech recognition device for a specific speaker, and Fig. 2 is a similar flt of registered words according to the present invention.
Word selection configuration Block 1: Shows a J-Tsuku diagram. In the drawing, numeral 5 indicates a group of word nodes files, numeral 11 corresponds to a correspondence statistics table, numeral 12 a character string creation section, numeral 13 a character string extraction section, numeral 14 a similarity statistic table and a time length statistical table between monosyllables, and numeral 15 a correspondence statistics table. 16 is a similarity comparison section, 16 is a similarity determination section, and 17 is a display. Kaya 1 Fallen Qin 2 Mouths

Claims (1)

[Claims] A plurality of word input speeches by a specific speaker are converted into input patterns and registered as word standard patterns in a plurality of word set dictionaries, and the word standard patterns registered in the plurality of word center dictionaries are exchanged. a selection means for selecting a pattern that is acoustically similar to the word standard pattern;
a similarity determining means for determining the similarity of the (to) pattern; and a notifying means for notifying the words of the word standard pattern determined by the In((hereinafter) comparison means; In the limited word speech recognition device that recognizes words by comparing them with word standard patterns registered in the plurality of word cent dictionaries, after registering the words of the specific speaker as word standard patterns in the plural word cent dictionaries. , the selection means selects a word standard pattern that is acoustically similar; the similarity determination means determines the similarity of the selected word standard patterns; A word registration method characterized by being configured to notify.