JPS6126095A - Automatic calculation of word-to-word distance - 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] [Field of Industrial Application] The present invention relates to a method for automatically calculating the distance between words, and is effective when used to evaluate in advance the suitability of a set of words to be recognized by a speech recognition device. .

[Conventional technology]

Some speech recognition devices have a predetermined group of words that can be input by voice (known as recognition target words).
In order to improve the recognition rate of input words of a speech recognition device, it is important to make sure that the word set does not contain any (IJ) pronunciation (pronunciation).The QX words themselves included in the word set are It is determined by the purpose of use of the device and is expected to be difficult to change, but changing the pronunciation of a word will not cause any particular problem, so the pronunciation of each word in the word set should be selected so as not to be confusing or confusing. Should I leave it?

There are some things that are easy to tell whether they are confusing or not.

For example, the number ``7-j'' can be read as ``shichi'' or ``nana'', but if you pronounce it as ``1-shichi'', it will be pronounced as the number ``1'', that is, ``
It is easy for the two to be mistaken for each other. Therefore"
If ``7'' is pronounced as ``nana'', it will be clearly distinguished from ``ichi J'', and this is known from experience. However, as the number of words to be recognized increases, it is not easy to tell which words are confusing (and when you find confusing words and substitute one pronunciation for the other, it becomes easier to understand which ones are confusing). There is also the problem that the word set for speech recognition is a word that has 2 or 3 different pronunciations, and 12 are selected to fix the pronunciation of the word set. We use a cut-and-try method to test whether or not there is a misrecognition by performing voice recognition, and if there is a misrecognition, we change the pronunciation of the word to a different reading and test again. However, this method requires an extremely large amount of time and effort.

Therefore, rather than actually conducting a speech recognition test, it is effective to check the pronunciation of each word in the word set in advance, while it is still at the character stage. Considering that the problem of confusion occurs between any two words in the word set, we can increase the degree of confusion with all remaining words for every word in the word set, and It is effective to reject a word set if there is a word pair (pronunciation) that poses a risk of recognition. Word distance is a numerical expression of the degree of similarity and dissimilarity between words.

DP matching method @ (Velichko et al
, Int.

J, Man-Machine S, tudie
s, vol, 2+ p223+ 19
In step 70), this distance between words is determined as the distance between character strings. Simply put, the local distance between each syllable of two words is accumulated, and there are various possible combinations of these syllables, and different combinations result in different cumulative values, but the minimum value among them is taken as the inter-word distance. It is.

[Problem that the invention seeks to solve]

However, this DP method does not reflect the number of syllables in each word. 1; Word with 2 syllables becomes 3 syllables
Suppose we get the result that the distance between word B and word B is equal to the distance between word C, which has two syllables, and word R, which has two syllables.
It is said that the misrecognition rate for the set of words A and B and the set of word C1D is the same, but empirically speaking, it is said that the set of words A and B has the same number of syllables.
The recognition rate of B should be higher than that of sets C and D, which have the same number of characters. This point is not reflected in distance calculation using the conventional DP matching method.

The present invention aims to compensate for the insufficiency of the DP method described above and to make the preliminary evaluation of recognition target words more practical.

[Means for Solving Problem c] The present invention provides a method for calculating the distance between each word in a word set to be speech recognized. A step of calculating the distance between a word and a word with the number of syllables N using the DP matching method, and using the obtained distance as follows: ■ If two words have the same number of syllables, the distance is smaller than if the two words do not have the same number of syllables. have the same word length, the distance to the end point calculated by the DP method divided by the word length is the normalized distance. ■The distance is always symmetrical about the two distances. This method is characterized by the step of calculating the distance between words by multiplying it by a normalization constant that makes corrections that satisfy three conditions. Next, the structure and operation will be explained in detail with reference to embodiments.

〔Example〕

FIG. 1 is a schematic diagram of a system in which two words A and B written in kana are decomposed into syllables in processing block 1.2, and the distance between the words is determined by the DP method in processing block 3. 4
Kana and syllable correspondence table used when breaking down into syllables, 5
is the phoneme distance matrix used when calculating the distance.

The kana notation for words A and B consists of 50 syllables (46 syllables >+' s, nasal rasp, semi-voiced, consonant, rasp, sul, loanwords such as ``sui'' and ``ti'', and any of these long syllables. The syllables decomposed by processing block 1.2 consist of consonants + vowels, and the vowels a,
i, u, etc., the consonants are s, ni, t, etc.
Consists of...etc. A part of the romaji notation for kana is shown below.

Table 1 Processing block 3 calculates the distance between two words A and B using a distance matrix between phonemes. A part of the distance matrix is shown in the table below.

FIG. 4 shows a processing flow from inputting a word to obtaining a distance output. In Flock IA, input word A in hiragana is converted into a syllable string using the conversion table shown in Table 1 above. Fifth
Figure (al is a flowchart showing the details of this conversion process)
In the same figure, bl is a specific example. The two people in the block also perform the same processing on word B. The syllable string A obtained by these processings.

B is block 3A, and the distance is calculated using the 1 table for inter-phoneme distance shown in Table 2 above. These blocks IA, 2A
, 3A correspond to cock 1.2.3 in FIG.

In the DP method using the processing block 3, distances are calculated for all word combinations, so no windows are provided. And between each monosyllable! Find the local distance, accumulate the distances of ¥111 between each single syllable from the starting point, and find the distance to the ending point. To explain with Figure 2, let the distance (local distance) between the i-th monosyllable of word A and the first monosyllable of word BO be CI, J (not shown), and the cumulative distance j31i of the sesame seeds is d′1. Then, d'+ and i are expressed by the following formula.

dl, 4-min (dZ-+++ l d +-1,
h-+.

``l r J-+ ) '' C1r ] That is, j, distance to number j ``i,+, distance to three points in front of it dZ-
1,1+ d′l−1+ J l + ”+++ l
It is the sum of 10-cal distance 1 [c, , , , to the minimum of . If i and j are increased and M and N are reached, word A is obtained.

Find the distance between 8. Here, M is the number of syllables in word A, N
is the number of single 33B syllables.

The inter-word distance obtained by the DP method as described above does not take into account the number of syllables in each word, as described above. Therefore, in the present invention, once the absolute distance d' or iU to the final syllables M and N is determined, it is multiplied by the normalization constant K and used as the inter-word distance Ml.
t (+. In other words, α and β are weighting coefficient normalization constants. The above values were used because the following three points were taken into consideration. ■ If two words have the same number of syllables, then if they do not The distance is smaller than when the two words have the same length.The normalized distance is calculated by dividing the distance 11 to the end point obtained by the D and P method by the word length. ■The distance is always symmetrical about the two distances (the result is the same whether the order or the reverse).When M+N-k (constant), use the above equation as = (t-N)2~ The formula for 2~ is the above ■~■ satisfies the conditions.

A specific example will be explained below. FIG. 3 shows the flow of each process when calculating the distance between "se" and "tsu".

The distance between small words "se" and "tsu" by the DP method is obtained as the sum of the distances between each phoneme. In this case the consonants S and T
Suppose that the distance between S is d c (S, TS) −〇,'5, and the distance between vowels E and U is dv (E, u) = 500.0. The absolute distance is d'+
, +=dc (S, TS) +dv (E, u)
-500.5. The distance between consonants and vowels, such as 0.5.500, is determined by linguistically knowing the degree of similarity and dissimilarity between each consonant and vowel. The distance between the same consonants and vowels is O. Next, in the present invention, normalization is performed, and the normalization constant is M=1. Since N=1, α−
1, β-1, then the normalized distance d is d=l·500.5=500.5.

FIG. 6 is a diagram showing a hardware image of the above-mentioned word distance system. 10 is an input unit into which word pairs are input; 12 is a conversion unit that converts a character string into a syllable string using a conversion table, that is, a kana-syllable correspondence table 18; and 14 is a DP that performs distance calculations using a distance matrix 20. Arithmetic unit, 1
6 is an output section.

〔Effect of the invention〕

As described above, according to the present invention, differences in the number of syllables that cannot be reflected by the DP method alone can be incorporated into the distance between words.
It is possible to make the selection of readings even higher in recognition rate.

[Brief explanation of the drawing]

Figure 1 is a processing block diagram of distance calculation using the DP method, Figure 2
FIG. 3 is an explanatory diagram of the DP method, FIG. 3 is an explanatory diagram showing a specific example of the present invention, FIGS. 4 and 5 are flowcharts showing the processing procedure of the present invention, and FIG. 6 is a block diagram showing the system configuration. It is. In the drawing, 1 and 2 are means for decomposing into syllables, and 3 is a distance calculation means using the DP method.

Claims (1)

[Claims] A method for calculating the distance between each word of a word set to be speech recognized, comprising: decomposing each word into individual syllables; and distance between a word with M syllables and a word with N syllables. is calculated using the DP matching method, and the distance obtained includes (1) if the two words have the same number of syllables, the distance is smaller than if they do not, and (2) if the two words have the same word length. If it has, the normalized distance is the distance to the end point determined by the DP method divided by the word length, and (3) the distance is always symmetric about the two distances. 1. A method for automatically calculating a distance between words, the method comprising the step of calculating the distance between words by multiplying the distance by a normalization constant.