JPS5991499A - Voice recognition 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] The present invention relates to a voice recognition system using monophonic string frequency information, and in particular converts speech into kana code and kanji code, and inputs symbol code and special symbol code for typesetting. The present invention relates to a voice input device for printing.

Conventionally, when inputting a printing manuscript (which the operator refers to during phototypesetting), the operator operates a keyboard or the like with his or her hands or fingers to input the manuscript in rows 1. C.

Therefore, there are disadvantages in that inputting the manuscript requires a great deal of effort, requires technical skill, and is accompanied by physical fatigue.

A voice input device for computerized phototypesetting has been proposed as a solution to these shortcomings, but conventional voice recognition devices have not been able to achieve a sufficiently satisfactory recognition rate. However, there was a drawback that a large amount of 9Y power was required.Therefore, the purpose of the present invention is to provide a speech recognition system that does not have the above drawbacks. The optimal candidate is selected from monosyllable candidates using the frequency information obtained from the frequency information of the monosyllable string, and when correcting misrecognition, the above evaluation value and frequency information of the monosyllable string can be used to easily correct the misrecognition. I'm trying to get it.

This invention will be explained below.

This invention relates to a voice recognition system, and as shown in FIG.
For example, a parameter extraction unit 11 that extracts %S1''
,′. ', 5'', etc. symbol code, Ant Lupi Hajime', '
Compare the keyboard 2 for inputting special symbols such as ``2-minute wata'' and ``4-minute kuwata'' with the characteristic parameter A sequence of corresponding monosyllable candidates C(c,,c2.・-
・cl), and also output the similarity value V(
v,,v2,...vl), and this feature parameter determining device 1.
Recognition that corrects the sequence C of monosyllable candidates using the sequence C of monosyllable candidates from 3 and its similarity value ■, and the extra-monophone frequency information MT stored in advance in the monosyllable sequence information storage device 14. Candidate correction device] 5 inputs the monosyllabic code ZT from the recognition candidate correction device 15 and the key code KC sent from the keyboard 2 via the key manual device 16 and outputs code information SC, and also outputs code information SC. The voice recognition device 10 is comprised of a control device 17 that controls the speech recognition device 10. and,
The coder n report 8C output from the voice recognition device IO is processed by the kanji processing device, and then phototyped by the computer phototypesetting system 3.
It is now ready to print. That is, the kanji processing device hole has a word processor 21 that converts code information SC from kana to kanji, and also converts kanji code, kana code,
It has a storage device 22 that stores printing code PCs corresponding to symbol codes and special symbol codes for typesetting, and prints the printing code PCs from the WTh storage device 22 to the computer phototypesetting system 3 in response to the input of code information SC. I am trying to output it.

In addition, the key human power device outputs a key code 1<C1 corresponding to the symbol keys arranged on the keyboard 2, special symbol keys for typesetting, and kanji/kana/alphabetic keys, and also outputs a code corresponding to the misrecognition correction key. It is set to output KC2. In addition, the control device 17 inputs the monosyllabic code ZT and the key code kc1 to the control symbols S1 and S.
2, and when the code KC2 is input, the next most suitable monosyllable candidate is output. Further, the extra-monophone frequency information MT stored in the monophonic string information storage device 14 is, for example, "Psychological Review" published in April 1960, pages 85-100,
"Relative frequency of Japanese city grams and their characteristics", 1973
National Institute for Japanese Language and Linguistics Report No. 49, pp. 94-12
The possibility of the appearance of monosyllable candidates is considered by referring to the sound sequence information table obtained from page 0. For example, the probability of a single syllable in the A stage following the single syllable "TA'" is as shown in Table 1 below.

Table 1 Furthermore, the probability of the monosyllable following the monosyllable 'SA', for example, in the A stage, is as shown in Table 2 below.

Table 2 The operation of such a configuration will be explained with reference to the flowchart in FIG.

The audio signal VS input from the microphone 1 is input to the parameter extraction device 11, and is subjected to audio preprocessing 2 frequency analysis,
From the feature extraction process, the feature parameter X is extracted (step s't'i). For example, the duration length of the input audio signal vS is divided into equal time units, and the frequency spectrum of the audio Bakun in all the corresponding divisions is calculated.
Find X2 + X3 +...Xn and calculate 1, X2.・
...Xk...Xn) is obtained.

Next, the percentage characteristic parameter judgment device 13 determines the characteristic parameters Pi=(p+ +
Between p2 + Ps r...I'n)i, a'r == 1 ma fan 12-j, (xJk-p,
k) 2・・・・・・・・・・・・(1) Calculate the degree of similarity and find out that the value of sociality G is small! A sequence of monosyllabic candidates C-(c, , c2
,...C1) and the sequence of similarity values ■-
Obtain (v+l v21 ”' Vl) (step S
'l"2゜S'l"3). Note that ■≦1≦1.

Then, it outputs a sequence of 11th monosyllable candidates from the power 1 similarity high (・), and also outputs the 1st degree of similarity.

The recognition candidate correction device 15 inputs each recognition candidate into 1. Frequency information MT is calculated for each monosyllable candidate using the appearance frequency information of the sound sequence stored in advance in the monosyllable information storage device 14.
Evaluation value “”” (Yl r Y2 +...
.yl) can be calculated (step ST4). Here, the evaluation value y1 is yi=γi ” vi ”'″−゛−−
--Find it using (2). Then, as shown in Table 2, the evaluation coefficient γ is set to 110.0'' when the frequency probability is 0φ, and all cases where the frequency probability is other than 0.0 are ``1.0''. // You can close it. thus,
A single note string? The sequence C of monosyllabic candidates is modified by the evaluation value Y obtained from the evaluation coefficient γ in consideration of the frequency information MT, and the optimal 7f monosyllabic candidate Crff is selected and output as a monosyllabic code ZT. Note that the sequence of optimal monosyllable candidates ()l/−(C; ,
c'2. ,,,ci) is Y-(yI,y2゜...
・yl), for example, becomes a roar when the corresponding monosyllables are arranged in descending order of value.

In this way, at t7, the recognition candidate 1 is detected by the correcting device 15.
;jMT is taken into consideration, and if the monosyllabic code ZT is selected and the correction key is not operated on the keyboard 2,
The optimum monosyllable candidate C is outputted from the control device 1'l as code information SC (step ST6,
ST7). Note that this code information SC includes the key code KCI from the keyboard 2. Further, when the correction key of the keyboard 20 is operated, 6τ sequentially outputs the second candidate CG r the third candidate bow from the sequence C' of the most jf monosyllable candidates (step 5).
T (5, ST8). In this way, the code information SC output from the control device 17 in the M voice recognition device 10 is sequentially input to the word processor 21 in the kanji processing device 1.
The printing code PC stored in the storage device 22 in advance is read out and sent to the computer phototypesetting system 3, and the electronic phototypesetting system 3 performs phototypesetting and printing in accordance with the code information SCK.

Next, a concrete experimental example will be shown and explained.

For example, ゞA TA TA I (LANA...“
When inputting the sentence ``A '1' A
Suppose that you have finished inputting TA' and now say ``i<A''. First, calculate the feature parameter X for the utterance ``i(A'') The similarity with the characteristic parameter Pi is calculated Jj.At this time, the sequence C of monosyllable candidates and the similarity value V are as shown in 3rd -=1, and 'KA'' is the first candidate. If we consider the evaluation coefficient γ of each single sound hole shown in light 1 and calculate the evaluation value Y, we will get the evaluation value Y shown in Fig. 3, and the input audio The first candidate C1 for
is ``KA'', and the second candidate bow is %%TA//
, it can be seen that the third candidate C1 is % DA #. Also, when inputting the sentence 'SA KANA 'YA...'', if you have finished inputting % SA # and now say ``I (A''),
Similarly, the monosyllable tj:, complementary sequence C and its similarity value ■ are as shown in FIG. Then, when the evaluation value Y is determined by considering the evaluation coefficient γ of each monosyllable shown in Table 2, the evaluation value Y shown in FIG. 4 is obtained. Therefore, in this case, the optimal monosyllable candidate sequence C' is such that the first C1 is 'KA', the second C9 is 'DA/', the third c/ is %S G
It is clear that correct speech recognition can be achieved by considering the frequency information of the sound string, although sound @ information alone will give incorrect results.

As described above, according to the present invention, when selecting a recognition result, evaluation is performed not only based on the phonological information of the uttered voice (but also considering the frequency information of the sound sequence), which is unique to the Japanese language. By making a judgment that takes into account the characteristics of

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing one embodiment of this invention, FIG. 2 is a flowchart for explaining its operation, and FIGS. 3 and 4 are diagrams for explaining specific examples of this invention, respectively. . 1...Microphone, 2...Keyboard, 3...
Miyako 1 calculation coffin system, 10... Voice recognition cup 1 color, 11
. . . Parameter extraction U” device, 12 . . . Single note @ feature parameter storage device, 13 . . . Heavy machinery parameter judgment device, 14 . ...Control equipment, 20...Kanji processing device. Applicant's agent Yu Yasugata 3 proceedings
Amendment (Method) March 1, 1980 Kazuo Wakasugi, Commissioner of the Patent Office1, Indication of the case Patent Application No. 202358 of 19822, Name of the invention Voice recognition system3, Person making the amendment Case and Relationship: Patent Applicant Tatsu Ifukube (and one other person) 4. Agent February 2, 1982 (shipment date February 22, 1981) 8. Power of attorney subject to amendment, specification and drawings 7. , Subject of amendment (1) Amend the power of attorney as shown in the attached sheet. (2) Engraving of the specification (no change in content) 672-

Claims (1)

[Claims] 1. a) a parameter extraction device for extracting feature parameters of input monosyllabic speech; b) a keyboard for inputting required information; It compares the stored monosyllable *'ja parameters and sequentially outputs them as corresponding monosyllable candidates starting from the one with the highest degree of similarity.
a feature parameter determination device that outputs the similarity value, and d) determining the monosyllable candidate using the monosyllable candidate and its similarity value, and monophonic sequence frequency information stored in an internal storage device in advance. The present invention is characterized by comprising: a recognition candidate correction device that performs correction; and a control device that inputs information from the recognition candidate correction device and the keyboard and outputs code information, and controls each of the devices. Speech recognition device. 2.3) a parameter extraction device for extracting feature parameters of input monosyllabic speech; a keyboard for inputting required information; Compare the
a feature parameter determination device that sequentially outputs the monosyllable candidates with the highest degree of similarity and also outputs the value of the degree of similarity; the monosyllable candidates and their degree of similarity are stored in an internal storage device in advance; a recognition candidate correction device that corrects the monosyllable candidates using a single syllable string frequency record; inputs information from the recognition candidate correction device and the keyboard and outputs code information; a voice recognition device comprising: a control device for controlling the device; b) a word processor for converting the code information from kana to kanji; and a kanji code. It has a storage device "t" that stores printing codes corresponding to kana codes, symbol codes, and special symbol codes for typesetting, and outputs the printing codes from the storage device to a computer phototypesetting system in response to input of the code information. 1. A voice input device for stamp scraps, characterized by comprising a kanji processing device.