JPS6148032A - Speech input type japanese document processor - Google Patents

Abstract

PURPOSE:To reduce a buffer capacity for writing a transition matrix used in suitability discrimination relative to a candidate syllable train by carrying out a preliminary discrimination according to the transition matrix of (M-1) order having one smaller power of number without using the transition matrix of M order directly. CONSTITUTION:A speech information entering a monosyllable recognition section 1 is compared with a pattern stored in a standard pattern memory 2 and the standard pattern approximating to the input speech is selected as a candidate in accordance with an order of the approximation and a candidate train comprising these combination is stored in a candidate train buffer 4. Then, plurality candidate trains stored in this buffer 4 are discriminated about its suitability by the use of a primary transition matrix in a transition matrix buffer 5, and in the next place, by the use of a secondary transition matrix and the rest is stored in a candidate train creating section 6. In a clause analysis division 7, high-order processing such as a dictionary collation is done. Finally, the selected clause is stored in a recognition result memory 8 and issued properly.

Description

DETAILED DESCRIPTION OF THE INVENTION <Industrial Application Field> The present invention relates to a human voice recognition method in a voice input type Japanese document processing device.

<Prior art> As a method for recognizing one section of speech, such as a phrase, into syllable units, a matrix is used that describes the connections between the syllable units to be recognized (2+many relationships), so that no transitions between syllable units occur. It has actually been proposed to exclude candidate syllable strings (hereinafter simply referred to as candidate strings) that have possible combinations and to perform the following processing only on the remaining candidate strings (for example, (Refer to Publication No.-58493).

The purpose of this is to increase the degree of recognition of phrases, etc., and to reduce the amount of higher-order processing as a result, and extends not only to first-order transitions but also to second-order transitions and even M-order transitions. If it is possible, and the degree of M is increased, the limit of the candidate sequence becomes stronger, and the accuracy of recognition can be increased.

<Problem to be solved by the invention> Recognition pair '4 Mi's mouth/door ah's 8 incense J1 Sat 111 (Kara l), 5 Required to describe the transition relationship between each of these tones (' The capacity of the amplifier (1λ is the base of +12 (=111+1).
Create a candidate sequence (Koji-Seru N1 dog transition matrix) consisting of γ (this becomes (M+1)414 of 112 bits,
Even though the larger the order of I'l=1, the stronger the limitation 1 and the more severe the effect, the required capacity of the transition matrix 1: for input/<゛・nofa is huge. This comes with the problem of becoming.

The present invention solves these problems and improves the transition! IJ is designed to reduce the required capacity of the buffer for writing multi-iteration rows.
It was done as a target.

<Means for determining the problem h'4> 1. To achieve the above-mentioned purpose l), the present invention provides (lvl+1
) suitability for a candidate sequence consisting of j″1 planes $11
When cutting, first, apply the CM+1) next transition matrix to divide the candidate columns having the transition peak multi-choice 11 drive combination, and only light up the remaining 1) candidate columns to create the IV4th transition row. W.I.J.
is assumed to be JTIL=.

For example, if we consider a first-order transition matrix, that is, a transition rA relation representing the connection between consecutive 21 syllables, this transition matrix is a second-order transition matrix, that is, the connection between three consecutive syllables. It describes the transition relations that represent . In other words, if the first-order transition relationship does not hold, the second-order transition relationship also does not hold.The present invention focuses on this point, and the transition is closed by the (M-1)-order transition matrix. Only the specified combinations of syllables are extended to the Mth order to determine suitability.

<Effect> In this way, suddenly apply the M-order transition matrix...
By the transition matrix of one less power (M-1),
In other words, because a preliminary judgment is performed, the number of candidate columns to be subjected to the M-th transition matrix is reduced, and the required capacity of the transition matrix writing buffer is significantly reduced. Equivalent recognition accuracy can be obtained.

<Example> The present invention will be specifically described below with reference to an example shown in the drawings.

In Figure 1, was the monosyllabic recognition unit 1 manually inputted? j
' ? ! 'Qing (41< is compared with the collar i1 (stored in the pattern memory 2 (j, The nodes are also stored in buffer 3 in chronological order, and candidate sequences consisting of these combinations are stored in candidate sequence buffer l. The candidate column is the first-order transition in the transition matrix buffer 5, the transition column (hereinafter M7\'[RIX-
1) is used to determine suitability, and then 2
The next transition matrix (hereinafter referred to as MATRIX-2) is JIJ
The suitability of the candidates is determined by 1'l+, and the remaining ones are stored in the candidate sequence creation section 6. To the dog, in the phrase analysis part 7, the following words:
High-level processing such as +1 matching is performed, and the finally selected phrase is stored in the recognition result memory and output as appropriate.

9 is C1-' (+) which controls these operations.

Next, LSTILIX-1 and MATRIX-2 will be described. A basic explanation of the transition matrix is described in detail in the above-mentioned patent publication, so it will be omitted here.

Figure 2 shows 84 ATRIX-1, Figure 3 shows MATRIX.
-2e, MATRIX-1ii, and MATRIX-2 block numbers are described in 2-byte units, and this point is similar to the transition matrix of the prior art described above, which describes the transition relationship between sound mJ units in 1-bit units. It is different from This M
The transition relationship of the candidates was checked using ATRIX-1, and if the transition was not possible, the block number <
0 (] (,1('l > 1-I E
A block number of ATRIX 2 is given.

MATRIX-2 has a block number for every 112 bits, and block 072 number O negates all first-order transition relationships, and is a block in which all 112 bits are 0'', for example, MATRIX-2. The transition relationship between certain syllable units in TRIX-1 is ゛< (l Of'10>HEX
If so, shamble to the block () of M]\TRlX-2. This link 0 has the ability to deny the second-order transition at the same time as the first-order transition does not hold. Also, if a transition is possible and a non-zero block number is indicated, jump forward to that number and check 112 hints from the beginning of the block. Here, * corresponds to each 1 bit or 1 syllable unit. In addition, the first bit of MAT RIX-2 is ``1''
If the remaining 111 bits are all "()", it means that it is a suffix.

As described above, when M=2, that is, three sounds @B'
The suitability of the transition relationship of the candidate sequence consisting of - digits is judged. According to the prior art mentioned above, there is a case where a second-order transition matrix is applied, and a case where a first-order transition matrix is applied according to the present invention. The number of bits when applying the transition matrix and then applying the second-order transition matrix is as follows.

(a) In the case of the prior art, 112 pieces = 1.404,928 [bit 10.・
(A) (b) In the case of the present invention 1=IAT old X-116X112”=200,704
[Hira) l... (B) i48 TRlX-21123X
1/2=702,464 [bin 1...(
C) (The occupancy rate of bit 1 in MATRlX-1 is calculated as 1/2 of the total. Therefore, the ratio of (11) to (a) is (B + C
)/A =903,168/”1,404,928=
0.6428, and in the case of the present invention, it is 65%1 compared to the prior art.
The same effect can be obtained with a buffer capacity of about 7.

Although the present invention is effective in reducing the buffer capacity when describing transition relationships between arbitrary livestock breeds, the buffer 7 capacity can be further reduced if used in conjunction with the following means. Incidentally, the following means can also be applied to the prior art, and even when used alone, they have some effect.

The first method is to use a transition matrix that describes only the connections between specific 41st syllables among the predetermined Japanese syllable units to be recognized. ” and a transition matrix representing the connections before and after it.

The second method is to group predetermined Japanese syllable units to be recognized into categories that have meaning in terms of speech recognition, and to use a transition matrix that describes the connection relationships between the groups. For example, 11) phonetically, (,) Pa line, Ri line, Ka line, (11) B line, G line, Ga line, (c) Su line, H line, (d) M line. A syllable group that includes different syllable units, such as , su line, and u line, is treated as one type of syllable unit, and the transition relationship that is the connection relationship between the remaining syllable units is described. It uses a te sequence.

<Effects of the Invention> As is clear from the explanation of the AJ4 example above, according to the present invention, it is possible to reduce the capacity of the buffer in which the transition matrix used for determining the suitability of candidate syllable strings is written.
This makes it possible to downsize the device, reduce costs, shorten processing time, etc.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention; FIG. 2 is a diagram showing an example of a second-order and first-order transition matrix;
The figure is a diagram showing an example of the quadratic transition matrix. 1... Single sound ff1i recognition unit 2... Standard pattern memory 3... Sound mi lattice buffer r 4... Candidate sequence buffer 5... Transition matrix buff) 7 6... Candidate sequence creation unit 7. ... Segment analysis unit 8 ... Recognition result memory 9 ... CPU

Claims (1)

[Claims] 1. When recognizing speech uttered in units such as phrases in syllable units and creating candidate syllable strings in the order of highly reliable combinations from multiple candidates recognized for each syllable unit, the predetermined Japanese language to be recognized is Using a transition matrix that describes transition relationships that represent connections between word syllable units, candidate syllable strings that have combinations in which transitions between syllable units are impossible are excluded, and the remaining candidate syllable strings are subjected to the following dictionary matching, etc. In a voice-input Japanese document processing device that outputs the recognition results of phrases, etc., when determining the suitability of a candidate syllable string consisting of (M+1) syllable units, first the (M-1) ) The following transition matrix is applied to exclude candidate syllable strings with combinations that cannot be transitioned, and the M-order transition matrix is applied only to the remaining candidate syllable strings to judge the suitability of the candidate syllable strings. A voice input type Japanese document processing device characterized by: