JPH11161651A - Phonetic symbol generator - Google Patents

Abstract

PROBLEM TO BE SOLVED: To minimize a cost for producing a converting rule and to give the most natural reading of Japanese to an unknown character string in foreign language by automatically generating a decision tree to be a converting rule from a character string and corresponding data being its phonetic symbol string. SOLUTION: The decision tree to be the converting rule is generated from the character string and corresponding data being its phonetic symbol string. At this device, a decision tree generating part 105 generates the decision tree 106 to be the converting rule from an English character string 101 to a KATAKANA (one of two kinds of KANA script used for Japanese syllabary) character string 103 from the corresponding relation of an alphabet symbol and a KATAKANA symbol at correspondence data 104. The generated tree 106 for conversion constitutes the group of the questions of an YES/NO form for deciding a KANA (Japanese syllabary) character symbol equivalent to the alphabet symbol from the kind of alphabet symbols positioned within n-number of characters before/after the alphabet symbol concerning the alphabet symbol at a position in alphabets.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for giving Japanese pronunciation to a foreign language character string such as an English character string in a Japanese speech synthesizer.

[0002]

2. Description of the Related Art A rule-based speech synthesis technique capable of reading an arbitrary sentence in real time has already been put to practical use. Regarding Japanese rule speech synthesis technology, products that read Japanese text on a personal computer in the form of speech synthesis reading software have been commercialized. There are various types of text to be read out, but in recent years the Internet and CD-RO, which have rapidly spread remarkably,
Large amounts of text are readily available through electronic media such as the M. In these texts, character strings in other languages such as English as well as Japanese often appear. It will become increasingly important in the future to successfully read foreign language character strings contained in such Japanese texts. Rather than reading a string in another language as a sequence of alphabets, it is of course to read it with the pronunciation of that string. There is a problem that the pronunciation of that part is very difficult to hear. In other words, in order to read aloud like Japanese, it is an issue to give Japanese pronunciation to a character string of another language.

[0003] In the conventional Japanese-language rule speech synthesizer, in order to give Japanese phonetic symbols to these foreign language character strings, a method of providing rules of how to read the foreign language character strings has been adopted. . For example, for English character strings, determine the pronunciation of English using an English word dictionary, and convert the English pronunciation string to a Japanese phonetic symbol string according to the English-Japanese pronunciation conversion table (Method 1).

In a field of application in a word processor or the like, when a katakana character string of a foreign word is input, a method of inputting a foreign word as spelled out and converting it into a katakana character string has been considered. Taking English as an example, split the English character string, apply the conversion rules to Katakana character strings to each of the divided parts, generate Katakana character string candidates, and then refer to the Katakana dictionary. Then, a correct candidate after conversion into Japanese is selected from the candidates (method 2).

[0005]

However, in the method 1,
After obtaining the English phonetic symbol string of the English character string, and applying the conversion rules to Japanese phonetic symbols for each phonetic symbol, the obtained Japanese phonetic symbol string There was a problem that it was difficult to obtain pronunciation.

According to the method 2, since an application such as a word processor is considered, a foreign Katakana dictionary is prepared in advance, and a Katakana character string candidate generated from an English character string is checked using the dictionary. We take the method of doing. For unknown foreign words that are not in the Katakana dictionary, there was a problem that it was not possible to determine the correct candidate for Katakana words generated from English character strings.

Further, in method 1, it is necessary to manually create a conversion rule from English pronunciation symbols to Japanese pronunciation symbols, and in method 2, it is necessary to manually create a conversion rule from English character strings to katakana character strings. However, there is a problem that the cost of manually creating the file cannot be ignored.

An object of the present invention is to solve these problems.

[0009]

In order to solve these problems, a phonetic symbol generation device of the present invention automatically generates a decision tree as a conversion rule from a pair of a character string and its phonetic symbols. A conversion rule generation unit, a conversion rule unit that holds the conversion rule generated from the first conversion rule generation unit, and a conversion unit that converts a character string input with reference to the first conversion rule unit to a phonetic symbol string. Things.

[0010]

Embodiments of the present invention will be described below. In this embodiment, the character string to be converted will be described as an English character string, and the converted phonetic symbol string will be described as a Japanese katakana character string.

FIG. 1 is a diagram showing the configuration of the present embodiment. 101 is an input English character string. 102 is
A conversion unit that converts the English character string 101 into a katakana character string. Reference numeral 103 denotes a katakana character string converted from the English character string 101 by the conversion unit 102.

Reference numeral 104 denotes correspondence data comprising an example of correspondence between English character strings and katakana character strings. FIG. 2 shows the corresponding data 1
04 shows an example of the content. The left side separated by '/' is an English character string, and the right side is a katakana character string. In the English character string, each alphabetic character is a character string symbol. In the katakana character string, the unit delimited by the space character '' is a single kana character symbol. Note that '-' indicates that there is no corresponding symbol. The top '/ take /
In take- / ', t = te, a =-, k = q, e
=-Represents a corresponding relationship.

Reference numeral 105 denotes a decision tree generation unit which generates a decision tree as a conversion rule from an English character string to a katakana character string from the correspondence between the English character symbol and the katakana character symbol in the correspondence data 104.

Reference numeral 106 denotes a conversion decision tree generated by the decision tree generation unit 105. Conversion decision tree 10 in the present embodiment
6 designates, for an English character symbol located at a certain position in the English character, a kana character symbol corresponding to the English character symbol by the type of the English character symbol located within n characters before and after the English character symbol (hereinafter referred to as English character context). ) To form a series of questions in YES / NO format to be determined from Therefore, the conversion decision tree 106 is a collection of 26 decision trees generated one by one for the English character symbols a to z. Since the procedure for generating the decision tree is the same for each English character symbol, the generation procedure will be described for only one English character symbol.

Hereinafter, a procedure in which the decision tree generation unit 105 generates a decision tree of “a” of the conversion decision tree 106 from the correspondence data 104 will be described.

Step 1: Data containing "a" is taken out of the data contained in the corresponding data 104, and 3 in FIG.
01 is created in the form of an English character context. In this English character context, the position where 'a' appears in the English character string of the corresponding data is set to 0, and the first three characters and the last three characters (if less than three characters, '#' is assigned) are arranged. It is a thing. -N indicates the position of the first n characters, and + n indicates the position of the last n characters. On the left side of the English character context 301 is the katakana character string 30
2. This is a kana character symbol or a kana character string to which 'a' at the position of 0 of each English character context corresponds in the corresponding data.

Step 2: Next, those having the same kana character string 302 are considered as one group. In FIG. 3, group 1 in which the kana character string is “/ − /” and “/ −”
Group 2 which is tsu / 'and Group 3 which is' / a /'
Divided into

Step 3: A group having a large number of English character contexts included in the group is selected. Examine the type of alphabetic character at each position in the alphabetic context that distinguishes that group from the others (for example, to distinguish group 1 from group 2 or group 3, use +
Note the 'e' at position 2). Such a position i
And a letter c is found, a branch node of the decision tree is generated, the branch condition at the branch node is set to i = c, and for a branch (Yes) satisfying the condition, the katakana character string of the group is assign. If there are two or more remaining groups for the branch (No) that does not satisfy the condition, the branch node generated by performing the procedure of step 3 is assigned to the remaining groups. If there is only one group left, the katakana character string of that group is assigned and the process ends. If the character position and the character type for distinguishing the groups are not found, these groups are ended as a result of step 3 (the previous step 3
Is assigned to the branch (No) that does not satisfy the condition (1).

According to the above procedure, the decision tree obtained from the English character context 301 and the kana character string 302 in FIG. 3 is as shown in FIG. Reference numeral 401 denotes a branch node that divides group 1, group 2, and group 3. 402
Is a branch node that divides group 2 and group 3.

According to this procedure, a decision tree is generated for other English character strings, and the conversion decision tree 106 is constructed.

Next, a procedure in which the conversion unit 102 converts an English character string into a katakana character string with reference to the conversion decision tree 106 will be described.

FIG. 5 shows a decision tree for English letters 't', 'p', and 'e'. Now, it is assumed that these decision trees and the decision tree for “a” in FIG. 4 are included in the conversion decision tree 106. 'Tape' is given as the input character string 101.
The conversion unit 102 adds three '#' characters before and after the input English character string.
Add. The corresponding decision tree of the conversion decision tree 106 is applied in order from the leading English character symbol excluding '#'.

With respect to the first 't', the letter immediately after the 't' is 'a', so that the condition of the branch node 501 is not satisfied. Therefore, branching to No, the branch node 502
Proceed to. Since the condition of the branch node 502 is not satisfied, N
o, and the katakana character corresponding to 't' is 'te'
Becomes

For the second “a”, the branch node 4
Since the condition of 01 is satisfied, the flow branches to Yes, and the corresponding katakana character string becomes "-".

For the third 'p', branch node 5
Since the condition of 03 is not satisfied, the process branches to No, and the corresponding katakana character 'P' is obtained.

For the fourth "e", only the correspondence with "-" appears in the correspondence data in the present embodiment.
It becomes a dummy katakana character '-' uniquely.

As described above, the correspondence data 1 of the present embodiment
Conversion is performed from the English character string “tape” that did not appear in 04 to “tape”.

In the present embodiment, the width of the English character context is set to three characters before and after the English character context. Further, in the present embodiment, the phonetic symbol string is a katakana character string, but may be an alphabet string in Roman notation.

[0029]

As is apparent from the above description, according to the present invention, a conversion tree is automatically generated from a character string and its corresponding phonetic symbol string as a conversion rule, whereby the conversion rule is obtained. Production costs can be greatly reduced. Further, according to the present invention, it is possible to obtain a phonetic symbol string by referring to the decision tree even for an English character string not included in the corresponding data.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a phonetic symbol generation device according to an embodiment of the present invention.

FIG. 2 is a diagram showing an example of correspondence data between English character strings and katakana character strings in one embodiment of the present invention.

FIG. 3 is a diagram showing a set of katakana character strings corresponding to an English character string context extracted from the corresponding data for 'a' and grouping of the katakana character strings according to the embodiment of the present invention;

FIG. 4 is a diagram showing an example of a decision tree obtained from an English character string context and a katakana character string for 'a' in one embodiment of the present invention.

FIG. 5 illustrates 't', 'p',
Diagram showing an example of a decision tree for 'e'

[Explanation of symbols]

 Reference Signs List 101 English character string 102 Conversion unit 103 Katakana character string 104 Corresponding data 105 Decision tree generation unit 106 Decision tree for conversion

Claims (3)

[Claims] A conversion unit configured to convert a character string in a certain language into a phonetic symbol string in another language; a conversion rule unit that determines a method of converting the character string into a phonetic symbol string; A conversion rule generation unit that generates a conversion symbol from conversion example data. 2. A conversion rule generation unit sets learning data based on correspondence data between a character string in a certain language and a phonetic symbol string in another language, and for a certain character string symbol, the character string symbol corresponds to any phonetic symbol. A decision tree representing a sequence of conditional branches conditioned on the type of a character string symbol located at a few characters before and after the character string symbol, which is a rule for determining whether to perform the decision, is automatically generated. The phonetic symbol generation device according to 1. 3. The method according to claim 1, wherein the conversion unit assigns phonetic symbols to each character of the input character string in accordance with a conditional branch of a decision tree stored in the conversion rule holding unit for types of characters located before and after. Item 1. The phonetic symbol generation device according to item 1.