JPS60179799A - Voice recognition equipment - 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] Industrial applications The present invention relates to a speech recognition device.

Configuration of conventional example and its problems Conventionally, a speech recognition device uses an n-dimensional feature vector series (&1.
P standard pattern vector sequences ("1 t "2 t """ b)) """ that are registered in advance in the device as a dictionary for a 2 , ... aI)
(b*+'2+...bU) The closest one or the one with the greatest similarity is used as the recognition result.In this case, the input vector series (al.

a 2 ,...-, al) and one of the standard pattern vector series, e.g. (b', bL..., b, 4
1 (however, 1 2' 7 = 1 ~ P), (a 1. a 2 , ・
....

The one-element vector ai of a and (bl, bl
、 、；、・、。

Calculate the city distance or Euclidean distance of the one-element vector b in 2 bM), use this as a measure of certainty, and use Guinaminok programming or linear expansion/contraction to calculate the total distance of the two vector series based on this. Most of the calculations are performed using methods such as

However, the city distance and Euclidean distance are given by the following formula.

ai = (ai, 1, ai, 2.--, ai
, N'1bt=(bt , bt ..., b cup, N
When l, m m'1 m2+ "l, m-Σ l al, r"! , rl (city distance)
−1 7t, m−Σ (a1′ r ” wealth, r) 2 (Yu 9 r par r = 1 distance) Below, we will explain the case where the above distance scale is used as a measure of the certainty value as an example. .

Based on the above principle, one of the speech recognition devices is to input words into a Japanese word processor by dividing them into syllables instead of inputting kana keys, such as ``a'', ``power'', etc. ","stomach".

Single syllable recognition, which recognizes "ha" and "su", and methods that divide continuous speech into syllable units and perform recognition on a syllable basis are used.

By the way, when using the above-mentioned syllable type recognition device as input to a Japanese word processor (hereinafter abbreviated as word processor), the word processor has a word dictionary 1 that cannot be used for kana-kanji conversion.
There is a method that corrects errors in syllable recognition using a combination of a sentence dictionary or a word dictionary and a language processing function.

That is, when the syllable recognition unit compares the input parameter vector sequence with the P marker patterns registered in advance as patterns for each syllable, instead of selecting one standard pattern that provides the minimum distance, it selects the one with the smaller distance. Select one (■ is a positive integer), for example, if the input audio is
(where t is a positive integer) syllables, the distance between IxJ syllable candidate matrices and the standard pattern for each of them is d□7. (However, 1+] is 1≦1≦I.

1≦j≦■), calculate the cumulative distance Sr(,) for each of the 11 possible sentences. r(1) +=1 (where r(1) is a function of i and 1≦j≦ in each stage of 1
■ Corresponds to one of the following. ),
Select N sentences from those with the smallest cumulative distance, search a pre-memorized sentence dictionary (the one used for general kana-kanji conversion can be used), and use the matching sentences as sentence recognition results. There is something. In this case, the number of additions required for cumulative distance calculation is 11. The operation during this time will be further explained using a specific example.

Figure 1 shows an example of I-6 and T-4, where four syllable candidates are output for one syllable input and the number of syllables in the sentence is five. ”.

For the first syllable (i=1), it is shown that (A7゛hata) (the lower row represents a distance of 13 62 63 71 distance) was obtained as a syllable candidate. The sentence that can be generated when I-5°T=4 is ``Adachii''.

"Atada Chibi"... "Tasasaishi" ■1-45
-1024 types exist.

For error correction, the aforementioned N sentences (N is a positive integer) with the smallest cumulative distance are used from among the 1024 sentences and compared with a pre-stored sentence dictionary.

In the example in Figure 1, the first candidate (adatashii) has a cumulative distance of 822 (adadashii) 8θ 3 (atarachii) 87 4 (new) 〃91, and the only sentence in the dictionary is ``new''. If it exists, the error can be corrected. In fact, in the text, the first
It is clear that there are no examples like the three candidates, so error correction is possible.

By the way, this method has the disadvantage that it is necessary to perform this calculation for each combination of steps when calculating the cumulative distance, and the amount of division is enormous and the processing takes time.

OBJECTS OF THE INVENTION In view of the above drawbacks, the present invention aims to reduce the number of calculations when correcting errors by searching a text dictionary using a plurality of phoneme or syllable recognition candidates.

Structure of the Invention In order to achieve this object, the present invention calculates, in each stage of an input syllable, the smallest cumulative distance N to the previous stage.
A cumulative distance storage means for selecting and storing the selected syllables and a transition storage means for storing the transition state thereof are provided, and the total distance between the N cumulative distances stored in the cumulative distance storage means and each candidate syllable in active use is The calculation is performed, N items with the smallest distances are selected, and these are newly stored in the cumulative distance storage means, and the ranks of the selected candidates and the cumulative distances at that time are stored as transition information in the transition N storage means. Then, after finishing voice input,
The structure is such that a plurality of sentence recognition candidates are obtained by retracing the transitions in the transition N storage means and are matched with a sentence dictionary.

DESCRIPTION OF EMBODIMENTS An embodiment of the present invention will be described from the outside with reference to the drawings.

FIG. 2 is a block diagram of a speech recognition device in one embodiment of the present invention. In the same figure, 1 is the input audio
The N-dimensional parameter vector sequence (a 1 ,
This is a parameter analysis section that sequentially converts a2, .

Reference numeral 2 is a standard pattern storage unit, which stores the voice whose parameters have been analyzed in advance into a syllable standard pattern (b11 bl,..., b))..., (b\,su,... , bi) and memorize it.

Reference numeral 3 denotes a comparison unit which compares the input parameter vector sequence (a 1
, a 2 , . . . , all and the standard pattern stored in the pattern storage section 3 are calculated.

Reference numeral 4 denotes a determining unit, which selects one of the distances obtained by the comparison unit 3 from the smallest distance (factor is a positive integer).

This is d09. shall be. However, 1≦1≦■ indicates the first appearing syllable, and 1≦j≦I indicates the first smallest distance.

6 is the cumulative distance storage unit, which is the cumulative distance 5i-In of the (i-1) syllables uttered prior to processing the i-th syllable (however, if 1≦n≦N, the final N sentences are ) is stored for dictionary matching as a recognition candidate.

Reference numeral 6 denotes a cumulative distance calculation unit which calculates a sum S1 for all T1N combinations of the cumulative distances 5i-1,n stored in the cumulative distance storage unit and the distances d□,j obtained by the determination unit 4. I put it on.

1, n i.e. S, f = S + d, ...... (1) 1
, n ” '+” 1+1 However, 1. n, ] is 1≦i≦I, 1≦n≦N.

1≦j≦1.

7 is a selection section, which is calculated by the cumulative distance section (N
10 J) cumulative distances @ S, i of the distances 1, n N in order from the smallest to the cumulative distance S in the syllable
1. It is stored in the cumulative distance storage unit 5 as n (1≦n≦N).

Reference numeral 8 denotes a transition storage unit, which stores, as transition information, a vector ``y, n-(+t, nl) whose elements are the combination of (i, n) in that case, for each n of the above-mentioned cumulative distances Si, n. .However, j,, n, ) is 1≦j,≦k.

One combination of 1≦nt≦N.

Reference numeral 9 denotes a transition tracing unit, which outputs sentence recognition candidates up to the Nth candidate by tracing the transition information r, n−(!t+J) stored in the transition storage unit 8 in reverse. . Reference numeral 10 is a text dictionary in which all texts to be used are stored.

Reference numeral 11 denotes a dictionary matching section which sequentially compares the N sentence candidates obtained by the transition tracing section 9 with the sentences in the sentence dictionary 10, and outputs the matched ones as the final recognition result λ.

Next, the operation of the apparatus configured as described above will be explained using a specific example shown in FIG.

Assume that the sentence ``Atarashi'' consisting of all five syllables is uttered. It is assumed that after parameter analysis in the parameter analysis section 1 and pattern comparison with a standard pattern in the comparison section 2, the matrices shown in FIG. 1 appear sequentially as the output of the determination section 4. Now, a case will be explained in which four (N=4) of sentences with small cumulative distances are selected from among these sentences.The distances given by the determination section are shown by numbers in parentheses in FIG.

FIG. 3 is a diagram illustrating the processing at each stage of input syllables according to the present invention. In the same figure, the arrangement of syllables in the entire matrix IJ Nox 35 is the same as that in FIG.

In the figure, the syllable "a" is first input at the time i=1. At this time, the output of the determination unit 4 includes four recognition candidates "A".
"Ka", "Ba", "Ta" and their distances (dl, 1.dl, 2.
dl, 3. dl, 4) = (13, 52, 6s, 7s) appears.

In the 1-1st stage, the initial value is the distance d12. (1≦1≦
4) are stored in the cumulative distance storage unit 6 (FIG. 2) as their "1." cumulative distances. That is, Sl
, n=d1'; , rl, n= (1, n), i.e. rl, 1-, as zero or transition information with (1≦j≦N)
(1,1)rl, 2=(1,2)rl, 3-(1,3)
rl, 4-(1, 4) is stored in the transition storage section 4. 21
゜23.25, 27.29 the contents of the cumulative distance storage section in stages 1-1 to 5, and 22, 24, 26, 28
30 shows the contents of the transition information in the transition storage section 8 (FIG. 2). i = When the syllable "ta" is uttered in the second row,
As the output of the determination unit 4 (FIG. 2):

ka, da, sa) is obtained, and the distance is (d2, 11d2,
2°d2,3. d2,4”” (19, 32, 53, 6
Suppose that 2) is obtained. At this time, cumulative distance S4. As n, (Sl, 11S1, 22S1, 31S1, 4) and (d2
, 1 "2.2 ld2, 31d2, 4) by the cumulative distance calculation unit 6 (Fig.
) is calculated by 0. This result is shown in Figure 3.
．． 34. The selection section 7 (Fig. 2) selects 34 out of 16 ways.
, select four with small distances from n and divide them into 1-
The cumulative distance storage unit 5 (second
・Store in Figure). In the example of Figure 3, (S2,1#S2,2・”2.3182,4)=(S
2,1・S2,1'S3. S') = (32, 45, 6
6,71) 2.1 2.2 (23 in Figure 3) is selected. At this time, transition information r2. n, the above 4
The subscripts (tl, J) of the two combinations are stored in the transition storage unit 8 (FIG. 2), i.e., r2,1=(1,1) r2,2=(2,1) r2,3−(3,1 ) r2,4=(1,2).

The same operation as performed for i = 2nd stage is also performed for i = 3.4°6. As a result, the transition storage section 8 (Fig. 2) contains the 3rd row.
As shown in FIGS. 21 to 3o, the values of the cumulative distance Si, n and the transition information ri, n are obtained.

The transition tracing unit 9 determines N=4 sentence candidates based on the transition information rinO value stored in the transition storage unit 8. In other words, if the transition information is rln-(+4.nt), the reverse trace of the n-th candidate in the first stage is
In the matrix 35 shown in the figure, the syllables corresponding to the matrices, , ps elements of (i, il) are taken as the syllable recognition results, and among the transition information in the first to first stages, r(1-IL"t is
-1) Transition information for the row. This first transitions ri,
n is rl 1. This is because it is determined from nl.

To explain using a specific example, for example, the third candidate N of the sentence,
= 3 by reverse tracing, the syllable in the i=5th row is determined by the matrix position (t', +()-(5 , 1), it can be determined that "Yes", and J
From -3, (i-1)-4th stage transition information is r4,3
- (1, 2), therefore, the syllable in the 4th row can be determined as "i" from the matrix position (i, 1l) = (4, 1), and trace r3,2 in the 3rd row from nt = 2. do.

The above operation is performed until the 1=1st row, and "Atarachii" is obtained as a sentence candidate.

If the same operation is performed for all of N-1 to N-4, the first
Candidate ``Adatachii'' 12 ＃ ``Adatashii'' 〃3〃 〃 ``Atarachii'' 14N ``New'' Get the result. The dictionary matching section 11 compares the four candidates with the standard sentences in the sentence dictionary 10, and takes the matches as the sentence recognition results. In the above example, since sentences such as the first to third candidates do not exist in general dictionaries, it is clear that the fourth candidate is selected.

Even if the above-described operation causes an error in selecting "Attachii" as the first candidate for syllable recognition, the error can be corrected by selecting a plurality of candidates.

The dictionary matching section 11 has methods using binary search, hash function, D, P, matching, etc., but since all of these methods are well known, their explanations will be omitted. In addition, a word dictionary may be used instead of the sentence dictionary 1o, and a word dictionary having grammatical attribute information, and a grammatical analysis processing means that uses part-of-speech information and conjugation information as the grammatical attribute information to determine connections before and after words. It is also possible to use As for the grammar analysis processing means, a known kana-kanji conversion technique can be used, so a description thereof will be omitted.

As is clear from the above explanation, the number of additions in the cumulative distance calculation in the above operation is Nx J
s, J=4, N=4, Nx T x I
=a.

■I=1024, which is a difference of more than 10 times.

As described above, according to the present embodiment, there is a cumulative distance calculation section 6 that calculates the cumulative distance of the first stage of appearing phonemes, a selection section 7 that selects N of the cumulative distances, and a selection section 7 that stores this result and uses the cumulative distance in the next stage. By providing a cumulative distance storage section 5 for calculating cumulative distance, a transition storage section 8 for storing its transition information, and a transition tracing section 9 for back tracing the transition information after voice input is completed, a given syllable matrix can be calculated. Efficient phoneme recognition error correction can be realized by determining sentence candidates from T + I by calculating the cumulative distance and matching with the sentence dictionary.

Note that in this embodiment, it is possible to combine this output with a kana-kanji conversion function and output the output result in a sentence with a mixture of kanji and kana. Further, this embodiment can be performed programmatically using a computer.

Further, in this embodiment, distance is used as a measure for phoneme candidate selection, but it can be similarly applied as a value (probability, etc.) expressing similarity or certainty.

Effects of the Invention As described above, the speech recognition device of the present invention selects a plurality of syllable recognition candidates with a small cumulative distance in each stage of the syllable based on the distance between the plurality of syllable recognition candidates and calculates the cumulative distance. It is stored and used in the next stage of cumulative distance calculation, and the transition information is also stored. After inputting the final syllable, multiple sentence candidates are determined by tracing the transition information in reverse, and matching with the sentence dictionary is performed. Compared to candidate selection by exhaustive syllable candidate selection, the amount of calculation can be significantly reduced and recognition performance can be improved, which has great industrial value.

[Brief explanation of the drawing]

FIG. 1 is a recognition diagram showing an example of the syllable recognition results of the present invention, FIG. 2 is a block diagram of a speech recognition device in an embodiment of the present invention, and FIG. 3 is a diagram for explaining the operation of the embodiment of the present invention. 0 1...Parameter analysis section, 2...Comparison section, 3...Comparison section, 4...Judgment section, 5
... Cumulative distance storage unit, 6... Cumulative distance calculation unit, 7... Selection unit, 8... Transition storage unit, 9...・Transition trace part, 10...
・Sentence dictionary, 11...Dictionary matching section.

Claims (2)

[Claims] (1) Phoneme identification means that identifies the phonemes of input speech and outputs a plurality of identification candidates and their likelihood values for each phoneme, and a phoneme identification means that identifies the top plurality of cumulative likelihood values of the phonemes that appeared earlier. a cumulative likelihood storage means for storing and calculating a cumulative likelihood value at the time of the next phoneme appearance; and a sentence candidate selection unit that stores transitions leading to selection of phoneme candidates having the plurality of top cumulative likelihood values. 1. A speech recognition device, comprising: transition storage means for storing data. (2) A phoneme identification means that identifies the phonemes of the input speech and outputs one identification candidate (factor is a positive integer) and its distance for each phoneme, and an accumulation of phonemes that appeared before the appearance of the corresponding phoneme. an output of the cumulative distance storage means for storing N items having small distances (N is a positive integer); a transition storage means for storing transitions leading to the selection of N items having small cumulative distances; and the output of the phoneme identification means. A cumulative distance calculation means calculates the sum by a combination of one distance given as . a selection means for selecting N items from among them, storing the distance in the cumulative distance storage means, and storing in the transition storage means the number and rank of the cumulative distance of which phoneme candidate was selected; tracing the transition information stored in the transition storage means;
A speech recognition device comprising: transition tracing means for outputting N discrimination result candidates.