JPS59157699A - Voice 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 Technical Field The present invention relates to an improvement in a voice registration method in a voice input device.

<Background technology> Generally speaking, when performing speech recognition in the dog language industry, the traditional method of uttering each word and registering the sounds would require all the enormous amounts of speech to be uttered in one go. It will also take a lot of effort. There is still a problem that as the number of words increases, errors between words become more noticeable, making it impractical.

Therefore, when considering speech recognition in the language industry, it is necessary to analyze speech in detail and extract as much of the characteristics of υ speech as possible.

Therefore, it has been proposed to divide speech into smaller units of time, such as phonemes and syllables, rather than words, and identify them in units of time.In addition, phonemes and syllables can range from tens to hundreds of types at most. Therefore, it is possible to perform speech recognition for a large number of languages by registering a small number of sounds, and it is considered to be effective.

However, each phoneme or syllable undergoes a wide variety of transformations due to variations in each utterance, differences in the way it is uttered, and the influence of preceding and following voices, so it is necessary to take these influences into consideration in advance when registering voices. .

Therefore, it becomes necessary to register a large number of sounds that have been transformed in advance for one phoneme or syllable. This results in an increase in the amount of processing and the number of voices to be registered, which eliminates the advantage of identifying each phoneme or syllable.

On the other hand, phonemes and syllables do not all appear with the same frequency, and their identification rates are not all the same.

For phonemes and syllables that occur frequently and have a low identification rate.

should be treated with more care, but it is not necessary to perform more processing for those that appear very infrequently.

By registering or re-registering phonemes and syllables in consideration of their frequency of occurrence and identification rate, it is possible to perform identification with less processing while maintaining overall performance. .

Purpose The present invention has been made from the above-mentioned viewpoint, and an object of the present invention is to provide a voice registration method that can maintain high recognition performance by registering a smaller number of voices. There it is.

<Implementation Example> Hereinafter, the present invention will be described in detail with reference to the drawings. FIG. 1 is a block configuration diagram showing an example of a voice input device implementing the voice registration method of the present invention. , 1 is a syllable sound identification unit, which recognizes the syllable parts extracted from single syllables and continuous sounds that are uttered separately, outputs the identification results of the syllables, and calculates the judgment results. Send it to the processing section 2. The syllable speech identification unit 1 has a memory 11 for storing standard patterns of syllables, and performs syllable identification based on pattern mapping between the standard pattern and an incoming cuff turn created from input and detected syllable sounds. .

The determination result processing unit 2 has a recognition result storage memory 21 that stores the recognition results sent from the syllable voice recognition unit 1, and once the contents stored in the memory 21 are stored,
Display section 3 displays syllable codes in units of syllables, words, phrases, sentences, etc.
send to The display section 3 displays these character strings in an appropriate format.

The operator looks at the identification results displayed on the display unit 3, finds out which syllable has the incorrect identification result, and performs a manual input operation using the manual switch, keyboard, light ben, etc. that constitutes the incorrect identification result indicating means 4. Then, a code or number specifying the incorrectly identified syllable is input to the judgment result processing section 2.

On the other hand, the count of the number of occurrences of each syllable identified by the syllable appearance number counting means 22 and the erroneous syllable number counting means 23 provided in the judgment result processing section 2 and stored in the memory 21 is incremented, and The count of the number of errors of the syllable specified by the misidentification result indicating means 4 is incremented, and the result is stored in the occurrence number storage memory 24 and the error number storage memory 25 for each syllable.

In addition, the judgment result processing unit 2 calculates the appearance rate and identification rate of each syllable from the appearance count value and error count count 1 for each syllable stored in the memories 24 and 25 by the conversion means 26 of the processing unit 2H. Then, the contents of the syllable appearance rate table storage memory 51 and the syllable identification rate table storage memory 52 are corrected. - Specifically, for example, after calculating the sum of the counts of the number of times each syllable appears stored in the memory 24, the value obtained by dividing the count of the number of times each syllable appears by this sum is calculated as the appearance rate al of each syllable. (i is a number representing the type of syllable) and the appearance rate bi of the syllable i already stored in the syllable appearance rate table storage memory 51, for example, (kJ+ai)/(k+D) (k
The conversion means 26 is configured to replace Lbi with an appropriate value (for example, a certain value between 1 and IO) and store it in the memory 51. Similarly, the same process is performed for the number of errors, and the error rate di for each syllable is calculated by dividing the count of the number of errors for each syllable by the total number of syllable appearances, and the syllable identification rate table has already been stored. For example, if the error rate ei of the syllable i stored in the memory 52 is (he4 +
The value obtained by the calculation di)/(h"I) (h is an appropriate value) is replaced with ei and stored in the memory 52.

In this way, recent frequency information (occurrence, errors) is incorporated into the syllable appearance rate and syllable identification rate.

As described above, the appearance rate and error rate of each syllable are stored in the judgment result processing unit 2 in the form of frequency and frequency, and are identified as the appearance rate as appropriate (for example, when the number of input syllables reaches a certain number of times). The rate is converted, and the frequency and number of times in the processing unit 2 are reset.

Incidentally, the storage contents in the memos I751 and 52 may be corrected each time the processing of each syllable is completed.
The appearance rate table and the identification rate table stored in 1 and 52 are used to determine which standard patterns need to be changed among the standard patterns registered in the memo IJ l l, and the result is processed by the determination result processing unit. 2, the syllable to be re-registered is displayed on the display section 3, and the standard pattern can be rewritten for the desired syllable in the memo IJII.

The judgment operation of the registration judgment unit 6 is based on the error rate e for each syllable.
This is done by determining that i has exceeded a certain threshold E,
As a result, the characters of the syllable i are displayed on the display section 3.

Note that it is desirable that a plurality of threshold values E be set according to the value of the appearance rate bi, for example, the appearance rate B+, B2
．． When Ba (Bl>B2>J), bi) e 1>Et for the Bl syllable, and ei>E2 for the J>J>B2 (7) syllable. For syllables with B2>bl>B3, if e i >Es (Et <E2 <Ea), re-registration is instructed, and for syllables with a higher occurrence rate, if the error rate is lower. A re-registration instruction will also be issued.

The operator checks the syllables displayed on the display section 3 and decides whether to redo the registration or not.

In addition, at the time of initial registration, the number of standard patterns for each syllable is calculated based on an appearance rate table or a discrimination rate table obtained in advance as a standard by another speaker.

Also, even in continuous speech, at the beginning of a sentence or word, the tip of the syllable appears following the ridge of the silent section.

Therefore, standard patterns for the same syllable can vary greatly depending on the pronunciation situation, and standard patterns for syllables must include one extracted from within a speech interval and one extracted after a silent interval. Considering the effects of other articulatory combinations, a variety of standard patterns are required for one syllable.

At the time of re-registration, the registration determination section 6 sends syllable information to be re-registered to the determination result processing section 2, including information on which utterance condition of one syllable is required to be registered.

If the occurrence rate of a certain syllable is high and the identification rate is low, re-registration is urgent. In such a case, the determination result processing unit 2
may instruct the display unit 3 to display the characters of this syllable in a distinct manner. For example, on a display, the characters can be made to flash in one corner or displayed in a different color.

The embodiment described above is applied to a device that can specify identification errors in syllable units, but when recognizing words, phrases, etc. by syllable unit identification, it may not be possible to specify syllable errors.

For example, in the case of word recognition, it is easier to understand if the word recognition results are displayed in kanji rather than in hiragana or katakana character strings. Therefore, the overall processing efficiency may be higher when errors are specified on a word-by-word basis than on a syllable-by-syllable basis.

An example of a processing method in such a case is shown below.

Assume that the syllable candidates ``ka'', ``ba'', and ``ta'' are obtained for the syllable 1-da'' as shown in FIG. 2 1bl) as a recognition result of the voice input uttered ``tamagawa''.

Note that the numerical value shown below the syllabary represents the amount related to the reliability of that candidate.

This value is a quantity that represents reliability, likelihood, similarity, and distance.

A distance ratio to the first candidate, etc. can be used.

Here, the Euclidean distance between the input syllable pattern and the syllable standard pattern is shown as an example. The reliability of each syllable candidate decreases as the distance increases.

The distance to a word candidate is expressed as the sum of the distances of each syllable candidate, and when they are arranged in descending order of reliability, the result is as shown in FIG. 2, 1bl. Limit the input words to place names and enter "
If there were no ``Gamagawa'' or ``Bamagawa,'' candidates for place name words such as those shown in Figure 2 (c) would remain.

When recognizing words by identifying each syllable, the determination result processing unit 2 must be equipped with functional means such as a syllable candidate generation unit 27, a word dictionary 28, and a dictionary collation unit 29 as shown in FIG. There is. Here, from the syllable sound identification unit l, as shown in Fig. 2 (a
) regarding reliability.

At the same time, syllable candidates are output. The syllable string candidate creation section 27 creates syllable string candidates in order of reliability as shown in FIG. 2(b) and sends them to the dictionary matching section 29. The dictionary collation unit 29 detects whether or not the syllable string candidate exists in the word dictionary 28, excludes it if it does not exist, and sends the word candidate found in the word dictionary to the display unit 3.

Recognition of sentences and phrases also requires complex processing in addition to dictionaries and collations, but in any case, one sentence candidate from syllable candidates is output to the display unit 3.

As shown in FIG. 2(c), if the first word candidate displayed is "Kanagawa" even though "Tamagawa" is uttered, then in this case, the second word candidate is "Kanagawa". Display up to the third candidate,
If you manually send information to the judgment result processing unit 2 to specify that the correct word is ``Doyou'', syllable candidates (see Figure 2 (
a)) It is determined that ``ta'' has been incorrectly changed to ``ga'' from 75. The number of errors counting means 23 counts whether the identification is correct or incorrect.

However, from a practical standpoint, if the first word candidate is incorrect,
There are some cases where it is better to say the words again as soon as possible. In this case, we know that "Kanagawa" is incorrect, but "ka"
I don't know which syllable of "na", "ga", or "wa" was incorrect. Although it may be appropriate to exclude the frequency of each syllable or the degree of correctness from counting, this is a problem when counting is required.

In such cases, the wrong syllable string (in this case ``gamagawa'')
), and if you re-pronounce it and compare it with the correct syllable string (in this case, ``tamagawa''), you will be able to see that you have mistaken ``taJ'' as ``ka'', and calculate the frequency and degree of correctness. can be added to.

It is thought that the identification rate of syllables uttered separately is generally higher than the method of detecting syllable parts from continuous speech and identifying each syllable. Therefore, by detecting syllable parts from continuous speech and identifying them for each sound excitation, the syllables that were incorrectly identified can be moved to the corresponding position in the kana character string by moving the kernel on the display. It is also conceivable that the correction could be made by uttering and identifying only that syllable.

At this time, since the sound is a syllable sound of segmented utterances, identification is relatively easy. In this case, if the same syllable is mistaken for the same syllable again, it is automatically corrected to another syllable candidate, thereby avoiding the need for many utterances to correct one kana character.

An example of the configuration of a device for performing these operations is shown in FIG.

The configuration and operation of the device shown in FIG. 4 will be explained using the example shown in FIG. on the keyboard 41, and select "
"Kanazawa" is displayed again, and after specifying an error, "Doyo" is displayed again. The dictionary of word items at this time is stored in the memory 28. The character string “ka” “ma” “is the result of syllable identification
"ga" and "wa" are stored in the memory 71.

The memory 72 stores the number of occurrences and the number of errors for each syllable. In the case of the example shown in FIG. 2, after finding out that the correct answer is "Doyo" (it can be determined that the answer is correct at the same time as the next utterance), based on the matching result of the syllable matching section 20, The count of the number of occurrences of the syllables "ma", "ga", and "wa" in the memory 72 is incremented, and the count of the number of errors of the syllable "ka" is incremented.

In this way, the memory 72 is updated each time each word is recognized. If the speaker specifies from the keyboard at an appropriate time, the contents of the memory 72 are processed by the conversion means 26, and then the syllable appearance rate table and the syllable identification results stored in the memos IJ, 31 and 32 are converted using the processed contents. (This process is called table update.)

As a method for this processing, for example, first find the total sum of the counts of the number of times each syllable appears, and then divide the count number of the number of times each syllable appears by this sum to calculate the appearance rate ai of each syllable.
(t is a number representing a syllable), and the value bi of syllable i already in the syllable appearance rate table and, for example, (kJ+al)/(k
+I) (k is an appropriate value, for example, l-10), and the value obtained by the operation is replaced with bi. by this,
The latest frequency information can be incorporated into the syllable appearance rate table.Similarly to 0, this type of processing can also be performed for the number of errors. Error rate for each syllable (syllable identification).

(stored in the rate table) is assumed to be ei.

When the registration determination unit 6 determines that this ei exceeds a certain threshold, the character of the syllable i is displayed on the display unit 3 via the determination result processing unit 2. Therefore, the speaker (user) can decide whether to redo the registration.

The appearance rate of each syllable does not necessarily have to be one; if there are five standard patterns per syllable, a counter or memory may be provided for each standard pattern.

Each syllable standard pattern is not necessarily obtained under similar vocalization conditions; for example, two of the five standard patterns are created from syllable sounds of segmented utterances, and the remaining three are created from word sounds. Sometimes it is made from the syllable part of. In this case, the syllable appearance rate table, syllable identification rate table, and memory 72 do not store values for each syllable, but for each standard pattern.

In the example in Figure 2, if we make a mistake because the first syllable ``ta'' has the highest degree of similarity to the third Q pattern of ``ka'', we increment the number of occurrences of ``ta'' to Increment the count of the number of errors corresponding to the third standard pattern of "?".

As a result of updating the table, if the value ei of the syllable identification rate table corresponding to the third standard pattern of "ka" exceeds a certain threshold, the eighth standard pattern of "ka" is the same as the created voice. It is necessary to re-register while speaking. For example, if the voice is created from the ``ka'' part of the voice ``squid'', the display section can display, for example, ``squid: re-registration required''. In this case, it is necessary to add a memory to the memory 72 for storing the creation status of each standard pattern in the syllable sound recognition unit.

As described above, syllables that require re-registration are displayed and output based on the syllable appearance frequency and error frequency.

Effect> As described above, according to the present invention, it is possible to instruct whether the syllable identification result obtained when performing voice input is correct, and to indicate the frequency of appearance of each syllable based on the syllable identification result and the correct or incorrect instruction. Since the syllables to be registered or re-registered are determined based on the frequency of occurrence and error frequency of each syllable, the overall character is maintained. , it becomes possible to efficiently find syllables that require re-registration and perform re-registration processing with a smaller amount of processing.

[Brief Description of the Drawings] Fig. 1 is a block diagram showing an example of a voice input device implementing the present invention, Fig. 2 is a diagram for explaining an example of speech recognition, and Fig. 3 is a block diagram showing an example of a voice input device implementing the present invention.
The figure is a block diagram showing an example of the determination result processing section 2, and FIG. 4 is a block diagram showing another example of the apparatus implementing the present invention. l...Syllable speech identification unit, 2...Determination result processing part 1,
22... Appearance number counting means, 23... Error number counting means, 24... Appearance number storage memory, 25... Error number storage memory, 3... Display section, 4... Erroneous identification result Instructing means, 51... Syllable appearance rate table storage memory, 52.
...Syllable identification rate table storage memory, 6..Registration judgment unit. Agent Patent attorney Aihiko Fukushi (and 2 others) Voice input
2-tywa (5,0) 2, Suffenakawa 1585
Nolbno 3. 1L"@f
t) (5, 4) 4, Power Nagiwa 15.5 no 5, Luma Biwa t 5.6 not Iku 1 Sen) 11 (c) 2. 4:w3, Gokawa

Claims (1)

[Scope of Claims] 1. Instructing whether the syllable identification result obtained when performing voice input is correct or incorrect, and determining the appearance frequency and error frequency of each syllable based on the syllable identification result and the correct or incorrect instruction; A speech registration method characterized in that a syllable to be registered or re-registered is determined based on the appearance frequency and error frequency of each of the waterlogged syllables. □