JPH09204434A - Device and method for synthesizing voice and recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To exactly read sentence information, in which plural kinds of reading are existent corresponding to one description, by providing a dictionary in hierarchical structure hierarchizing the information of descriptions and reading of addresses, etc., for which plural kinds of reading are existent corresponding to one description. SOLUTION: Corresponding to the input of a sentence input part 1, a hierarchical dictionary retrieval part 2 retrieves candidates for the reading of a word contained in inputted sentence information from a stored hierarchical dictionary 3 together with the information in hierarchical structure hierarchizing the information showing the descriptions and reading of words according to the connection order of words by words such as addresses composed of word groups for which the reading of a word to appear next is decided by a word to appear first. Based on the information in the hierarchical structure, a sentence parsing part 4 selects the reading of a word string coincident with the word string in the hierarchical dictionary 3 contained in the sentence information out of the candidates of reading and converts the selected word string to the reading of the dictionary 3. A voice waveform generating part 6 generates a voice waveform from the information of reading. Thus, the word string of the address or the like is read in correct reading.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice synthesizing apparatus, a voice synthesizing method, and a computer program recording the voice synthesizing method for reading out sentence information such as an address having a plurality of readings in one notation with correct reading. Regarding the medium.

[0002]

2. Description of the Related Art In a conventional speech synthesizer, when sentence information is input, it refers to a dictionary that stores information such as word reading and accents for synthesizing speech in association with word notation. Then, the text information is converted into reading, a voice waveform is generated from the reading information and the information such as accent, and the text information is read aloud by synthetic speech.

[0003]

However, when the text information includes an address or the like in which a plurality of readings are present in one notation, for example, the text information of "Honmachi" is changed to "Honcho" and "Ho". There is a possibility that it may not be possible to accurately determine which of a plurality of readings such as “n'machi” and “moto'machi” (“′” is an accent symbol) to be read out, and may be read out with an incorrect reading.

The present invention has been made in order to solve such a problem, and has a hierarchical structure in which notation such as an address having plural readings for one notation and reading information are hierarchized. By having a dictionary, a voice synthesizing device, a voice synthesizing method and a recording medium recording a computer program of this method for reading out sentence information such as an address in which a plurality of readings exist for one notation with correct reading. For the purpose of provision.

[0005]

FIG. 1 is a basic block diagram of a speech synthesizer according to the present invention. When the text input unit 1 for inputting text information directly from a recording medium such as a CD-ROM or a magneto-optical disk or via a public line or the like inputs text information, the hierarchical dictionary search unit 2 appears first. Information about the notation and reading of each word of a word string such as an address formed of a group of words that determines the reading of the word that appears next to the word,
From the hierarchical dictionary 3 that is stored together with the information of the hierarchical structure in which the words are connected in order, for example, in the case of an address, the order of prefectures, city counties, and towns and villages is stored. Searching for reading candidates, sentence analysis unit 4
Is selected from among the reading candidates based on the hierarchical structure information, the reading of the word string that is included in the sentence information and matches the word string in the hierarchical dictionary 3, and this word string is selected in the hierarchical dictionary 3. A sentence analysis unit 4 for converting the reading into a reading, a speech waveform generating unit 6 for generating a speech waveform from reading information, and a speaker 7 for outputting the speech waveform. As a result, a word string such as an address having plural readings in one notation is read aloud with the correct reading.

Further, the speech synthesizer of the present invention searches the hierarchy belonging to the designated area by inputting the information for identifying the area and designating the area based on the judgment from the place name included in the sentence information. To reduce the search time.

Further, in the speech synthesizing apparatus of the present invention, the hierarchical structure information is a hierarchy dictionary which is information for identifying a parent word in a higher hierarchy of each word which determines the reading, and a reading dictionary retrieved from the hierarchical dictionary. By referring to this information of the candidates, the connection relation of the candidates for reading the word string included in the sentence information is set.

Further, the speech synthesizer of the present invention searches a hierarchical dictionary on the basis of a word string in which notation of a word in any one of the word strings is omitted, and thereby a part of the word string in the document is searched. Read correctly even if is omitted.

The speech synthesizer of the present invention reads the hierarchical dictionary when the notation of a predetermined number or more of characters or words matches the notation of the characters or words included in the word string stored in the hierarchical dictionary. Is determined to be the reading of the character or word, except for a word string such as an address in which a plurality of readings exist,
A general word whose notation matches a part of this word string is not erroneously read out when reading an address or the like.

Further, the speech synthesizer of the present invention is provided with a suffix dictionary in which information of notation and reading of suffixes connected to the word string is stored, and the word string matching with the word string in the hierarchical dictionary is provided. By searching the suffix dictionary for information of the notation that matches the notation of the immediately following sentence information and selecting the notation of the notation of the suffix dictionary that matches the notation of immediately after as the reading of the notation immediately after, this word string Read the suffix with the correct reading, which becomes different from the reading of a general word when connected.

[0011]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 2 is a schematic diagram showing an example of a speech synthesizer of the present invention. In the voice synthesizer of this example, a disc drive of a general-purpose personal computer having a voice output function stores a dictionary such as an address dictionary and an address suffix dictionary, and a computer program of a voice synthesizing method as described below. Magneto-optical disk, CD-RO
A recording medium D such as M is loaded, a computer program is loaded, voice is synthesized from text information input from a disk of a text database or via a public line (not shown), and a speaker connected to the computer main body From or through a public line.

The voice synthesizer according to the present invention is not limited to the configuration in which the software is loaded on the general-purpose personal computer as described above, but also a voice for reading the character string information such as traffic information received from the FM multiplex broadcast by the synthesized voice. It may be a dedicated synthesizer, in which case the text information is input via the antenna.

[First Embodiment] FIG. 3 is a block diagram of a first embodiment of a speech synthesizer of the present invention. The text input unit 101 inputs text information from a CD-ROM, a magneto-optical disk, or the like, and the address dictionary search unit 102 writes a word string that matches the word string of the address included in the text information and displays the word string of the address and Hierarchical address dictionary 103 in which reading information is hierarchized and hierarchical information is stored together with the hierarchical information.
Search from. The sentence analysis unit 104 sets the reading of the word retrieved from the address dictionary 103 as the reading of the address dictionary 103, searches the basic dictionary 105 for words that match other general words, and converts the sentence information into the reading. The voice waveform generation unit 106 generates a voice waveform from the reading and accent information and outputs a synthetic voice from the speaker 107, which reads the text information.

FIG. 4 is a conceptual diagram of an example of the address dictionary 103. The address dictionary 103 stores notation and reading information of each word constituting the word string of the address in a hierarchical manner.
The top of the hierarchy is the 47 prefecture names, and as the hierarchy below it, the place name of the town or village, and the place names such as large letters and small letters are hierarchized in the order of connection.
For each word, in addition to such hierarchical information, pronunciation information that includes prosodic information such as Kanji notation, accent phrase boundary information, and accent type that serves as a search key, that is, reading information is included.

FIG. 5 shows the address dictionary 103 according to the first embodiment.
3 is a flowchart of a basic search algorithm of.
First, a text pointer is set at the beginning of the text buffer that stores the input sentence information (S101), and the address dictionary search pointer is set at the beginning of the hierarchical structure of the address dictionary 103 (S102). While shifting the text pointer one character at a time, the word starting at the position of the text pointer is compared with the search candidate word in the address dictionary 103 to determine whether they match (S103, 104), and the word in the address dictionary 103 becomes a text. Check if it exists in the buffer. When the word starting with the text pointer exists in the address dictionary 103, the address dictionary search pointer is set to the next layer (S105) and the text pointer is set to the next word position (S106). It is checked whether or not there is a word in the next layer at the next word position (S107, S1).
08).

When there is no word matching the word in the address dictionary 103, the section matching the word in the address dictionary 103 is regarded as the address section, and the reading string set in the matching word string is Pronunciation information is set as the reading of the address section (S109) and passed to the sentence analysis unit 104. In addition, since the sentence after the address section may include the address, the address dictionary search pointer is set at the beginning of the hierarchical structure (S110), and the same process is performed on the entire sentence until the end of the text information. (S111, S112).

[Second Embodiment] FIG. 6 is a block diagram of the second embodiment. The same parts as those in the first embodiment described above are designated by the same reference numerals and the description thereof is omitted. In the second embodiment, no word exists in the basic dictionary 105 used by the sentence analysis unit 104, but if the pronunciation information is known by the search of the address dictionary 103, the word is not registered and the pronunciation is input in the input sentence. An address pronunciation setting unit 108 that embeds information as a pronunciation designation character string in a text and a pronunciation designation analysis unit 10 that identifies and analyzes the pronunciation designation character string embedded in the text.
9 and are provided.

Here, when the format of the pronunciation designation character string is defined as "<pronunciation: Kanji notation: pronunciation information>",
"<:>" Is a symbol for giving a special meaning in a sentence, and the character string "pronunciation" is a keyword for identifying pronunciation designation. "Kanji notation" is merely a role of comment, and "pronunciation information" describes pronunciation information expressed by katakana and accent marks.

Next, the operation will be described. Text input section
For example, when the sentence "I live in Ota-ku, Tokyo" is input from 101, the address dictionary search unit 102 searches the address dictionary 103 as in the first embodiment, and then "Ota-ku, Tokyo" is entered. Is the address word string, and the pronunciation of that address is "Tokyo" Otaku '("'" is the accent symbol, " Is an accent ward boundary symbol). Further, in the address pronunciation setting unit 108, the section of the address of the input sentence is replaced with the pronunciation designation character string,
"Pronunciation: Ota-ku, Tokyo: Tokyo ' Ota '
I live in The pronunciation designation analysis part
Output to 109.

The pronunciation designation analysis unit 109 identifies the characters delimited by parentheses (<>) and analyzes them as pronunciation designations, and in that portion, the pronunciation is "Tokyo". It is assumed that there is a noun "Otaku", while the other parts of the sentence are passed to the sentence analysis unit 104 as they are. The sentence analysis unit 104 analyzes the pronunciation specifying information and other sentences, and sets correct reading information.

In the second embodiment, an address pronunciation setting unit 108 is connected to the address dictionary search unit 102, and an input path from the address pronunciation setting unit 108 is provided between the address pronunciation setting unit 108 and the sentence analysis unit 104. In addition, by connecting the pronunciation designation analysis unit 109 having the input path of the sentence information from the sentence input unit 101, when reading a sentence that does not include an address, the sentence to be converted into the pronunciation information by referring to the basic dictionary 105. Information analysis part 10
You can type directly in 4. That is, in the figure, the address reading part surrounded by a broken line can be configured as an independent device, and the input path of the text information can be selectively used.

Therefore, first, when the address reading unit is constructed as an independent device, parallel processing becomes possible. The number of words in the address dictionary 103 generally exceeds 100,000 words, and the number of words in the basic dictionary 105 of the language processing unit may exceed tens of thousands words to 100,000 words. With such a configuration, the processing of the address reading part and the language processing for speech synthesis of the other parts can be processed in parallel by the two CPUs, so that an increase in processing time can be prevented. Is. Secondly, when the address reading unit is composed of software, independent commands can be created as commands for the address reading unit and the language processing unit, so that software maintenance work and system changes can be handled. Is easy.

[Third Embodiment] FIG. 7A shows a third embodiment.
FIG. The same parts as those in the above-described embodiment are designated by the same reference numerals and the description thereof will be omitted. In this embodiment, a designated area input unit 110 that sets the area name each time text information is input, and a search start position storage buffer that holds the designated area as the search start point of the hierarchical address dictionary 103. If 111 and are specified and the area name is specified, the processing time required to perform the search will be significantly increased by starting the search for each hierarchy in the hierarchical structure belonging to the specified area. To reduce.

The address may be written from the 47 prefectures in all, but the prefecture name is obvious from the context, or the prefecture name may be omitted because it is a well-known place name. In many cases, the prefecture name is omitted. However, if the sentence contains an address that starts in the middle of the hierarchy, the hierarchical address dictionary 103 can search from the middle hierarchy, but the number of search target words increases as the hierarchy decreases, There are hundreds of thousands of words to search for. Therefore, in this embodiment, information indicating which point on the hierarchy is to be used as the starting point for the search is specified by region.

FIG. 8 is a conceptual diagram of the search start position storage buffer 111 when the area "Asahikawa, Hokkaido" is specified by the specified area input unit 110, for example. When the area "Hokkaido Asahikawa City" is specified by the specified area input unit 110, the specified area is searched by the address dictionary search unit 102,
The search start position storage buffer 111 stores the search pointers in each of the three hierarchical structures of “top of hierarchy”, “Hokkaido”, and “Asahikawa city”. Address dictionary search unit 102
Searches the address dictionary 103 starting from each search pointer in the search start position storage buffer 111. Note that the information on the search start position is not limited to one area, and information on a plurality of areas may be stored.

FIG. 7B is a block diagram of a modification of the third embodiment. This modification is different from the third embodiment in that a designated area input unit 110 is replaced with a designated area acquisition unit 112 that obtains area information from the input text information. The designated area acquisition unit 112 is the address dictionary search unit 102.
When the address dictionary 103 is searched with, and there is a word string candidate in the address dictionary 103 that matches a word in the sentence, the area of the place name in the sentence is acquired and stored in the search start position storage buffer 111, and the search is started. The position information is used as described above.

[Fourth Embodiment] FIG. 9 is a block diagram of a fourth embodiment of the speech synthesizer of the present invention. The same parts as those in the above-described embodiment are designated by the same reference numerals and the description thereof will be omitted. In the fourth embodiment, when a method of searching all the following layers from the middle of the address dictionary is used to search for a place name starting from the middle layer, the hierarchy that the search takes a long time is solved. Figure 10 of the structure address dictionary 103
In addition to the above configuration, a connection relation setting unit 113 for determining the connection relation between words is provided.

That is, in the address dictionary 103 of FIG. 10, all the words in the hierarchical structure are included in the notation and reading information as parent numbers such as parent numbers that specify the parent word in the upper hierarchy, where the reading is this reading. Information is given and expressed in a parent-child relationship (eg, parent word # 1, parent word # 2, ...). Furthermore, in order to facilitate the search by the notation, they are sorted in the code order of the notation. At this time, words having different readings in one notation have parent information for each of the plural readings, and words having different readings have the same reading, so that each reading has multiple parent information. It can also be compressed.

FIG. 11 is a flowchart of the algorithm of the connection relation setting unit 113, and the steps surrounded by broken lines show the processing in the connection relation setting unit 113. The search character position is set to the first character of the sentence information (S201), and the word starting at the start position is searched (S202). It is determined whether or not there is a word that starts at the start position (S203). If there is no word that starts at the start position, the search character position is set to the next character (S204), and it is determined whether it is the last character in the text. It is determined whether or not (S205). If it is not the last character in the text, the process proceeds to step S202, and the word starting at the start position is searched (S202). The address dictionary search unit 102 searches all layers of the address dictionary 103 based on all the words included in the sentence, and extracts all candidates that match the words in the sentence.

If the result of determination in step S203 is that there is a word starting at the start position, words with different readings and words are divided (S206). The connection relation setting unit 113 determines whether or not the processing has been performed for all words (S207).
), If all words have not been processed, it is determined whether there is a word ending at the start position (S208).
), If it does not exist, information without parent word is set (S209). On the other hand, if the result of determination in step S208 is that there is a word ending at the start position, it is determined whether or not there is a word with a matching parent number (S210). When there is no word having the same parent number, the information without parent word is set (S209), and when there is the word having the same parent number, a pointer to the parent word is set (S211).

If the result of determination in step S207 is that all words have been processed, step S204 follows and the search character position is set to the next character, and steps S204 to S211 are set until the last character in the text is reached. repeat.

In this embodiment, when the parent word exists, the reading of that word is always selected. Address dictionary 10
By making 3 such a structure, the correct reading can be obtained even if the address starts from any hierarchy of the hierarchical structure. Further, as the address dictionary, a configuration may be adopted in which a hash index is given to the word representing each layer of the address and the reading is searched through the hash index. Even when the address dictionary has such a structure, the procedure for obtaining the connection relation after extracting the candidates matching the notation from the address dictionary by searching the address dictionary is the same as in the case of the configuration of FIG.

In addition, when a word string that matches a word string in a sentence is obtained from the address dictionary 103, a plurality of types of matching may occur. For example, Tokyo-Minato-Platinum
And using the address dictionary 103 that includes the place name “Tokyo-Minato-ku-Shirokanedai”, “Shirokanedai, Minato-ku, Tokyo is ... When synthesizing the sentence "," or "Yamagata Prefecture-Nanyo City-
Using the address dictionary 103 that includes the place names "Miyauchi-Shinmachi" and "Kumamoto Prefecture-Arao City-Miyauchi-Shinmachi", "Miyauchi Shinmachi in Arao City ... When the sentence “” is speech-synthesized, both place names are matched in the address dictionary 103. At this time, in the previous example, the total character length of << Tokyo-Minato-ku-Platinum >> is 7
It is a character, and the total character length of "Tokyo-Minato-ku-Shirokanedai" is 8
Since it is a character, the reading with the longer character length is selected. Also, in the example below, the place name in Yamagata Prefecture matches only 4 characters with "Miyauchi-Shinmachi", but the place name in Kumamoto Prefecture has "Arao City-Miyauchi-"
Since it matches 7 characters with "Shinmachi", select the Kumamoto reading with the longer character length.

The flow chart of the above algorithm is illustrated.
See Figure 12. A word candidate string is created from all the words (S301),
The word string having the longest number of characters is selected (S302). The selected word and the candidate string whose section overlaps with this word are deleted from the buffer (not shown) storing the word candidate string (S303). It is determined whether or not the buffer of the word candidate string is empty (S304), and step S302 is performed until the buffer becomes empty.
Repeat S303. In this algorithm, processing is performed in consideration of the case where a sentence contains a plurality of addresses.

[Fifth Embodiment] FIGS. 13 (a) and 13 (b) are block diagrams of the fifth embodiment of the speech synthesizer of the present invention. FIG. 13 (a) has a hierarchical structure as it is. FIG. 13 (b) is a block diagram when the address dictionary 103 having the structure of FIG. 4 is used, and FIG. 13 (b) is an address dictionary having the structure of FIG. 10 in which hierarchical information is expressed by parent-child relationships between words.
It is a block diagram when using 103. In the figure, the same parts as those in the above-described embodiment are designated by the same reference numerals and the description thereof will be omitted. In this embodiment, the layer omission information setting unit
The difference is that 114 is provided.

That is, as a part omitted when representing a place name, not only the top part starting from the prefecture name but also the middle level of the address may be omitted. For example, if it is exactly "Kamikichishokumura, Nishiyatsushiro-gun, Yamanashi Prefecture", but the name "Kamikichishokumura, Yamanashi Prefecture" is omitted, or more accurately, it is "Midori-ku, Yokohama-shi, Kanagawa Prefecture." Although it is Nagatsuda, the name of the ward may be omitted, such as "Nagatsuda, Yokohama-shi, Kanagawa". When there is an address notation in which one or several layers of the hierarchical structure of the address dictionary are omitted in the sentence, there is no matching candidate even if the address dictionary having the hierarchical structure is directly referred to for a search. I can't find it and can't read the address correctly.

Therefore, in the fifth embodiment, there is provided a rule that a combination skipped by one layer is possible when searching for a word or when setting a connection relation. With respect to FIG. 13 (a), description will be given based on a method of searching for words while shifting the pointer for each layer. In this case, when searching for words in a specific layer or all layers, not only the word in the layer indicated by the search pointer is searched, but also those words and one or several layers below those words. Search for all words. The layer omission information setting unit 114 is means for setting information on such layer omission in the search.

With reference to FIG. 13 (b), a description will be given based on a method of searching all the words in all the layers appearing in the sentence and then obtaining the connection relation between the words in the dictionary by the connection relation setting unit 113. To do. In this case, the connection relationship setting unit 113 needs to set the relationship that a parent-child relationship between words exists even if one layer or several layers are omitted, and the layer omission information setting unit 114 performs control therefor. . That is, by tracing the pointer of the parent information twice, it is possible to easily determine that there is a connection even if one layer is omitted.

As another method for solving the above-mentioned problems, words that may be omitted, such as a county name and a ward name, may be omitted at the time of word search or connection setting. A method in which the information that there is is used as an attribute of a word in the address dictionary 103 can be considered. Figure 14 (a) and Figure 14 (b)
FIG. 14 is a configuration diagram of a modification of the fifth embodiment, FIG. 14 (a) is a configuration diagram when the address dictionary 103 having the configuration of FIG. 4 having the hierarchical structure is used, and FIG. 14 (b) is hierarchy information. FIG. 11 is a configuration diagram in the case of using the address dictionary 103 having the configuration of FIG. 10 in which is expressed by a parent-child relationship between words.

With reference to FIG. 14 (a), description will be made on the basis of a method of searching for words while shifting the pointer for each layer.
In this case, only when the hierarchy omission information acquisition unit 115 examines the attribute of a word in the address dictionary 103 and includes information that the word may be omitted, similar to the above, those words and their Search for all words in one or several levels below the word. This method has an advantage that the amount of processing does not increase because only the part that may be omitted is searched when one or several layers are omitted.

With respect to FIG. 14 (b), based on the method of searching all the words in all the layers appearing in the sentence and then obtaining the connection relation between the words in the dictionary by the connection relation setting unit 113. explain. In this case, the layer omission information acquisition unit 115
When the attribute of a word in the address dictionary 103 is checked, and only if it contains information that may be omitted, by tracing the pointer of the parent information twice, it is described as one level omitted However, it can be determined that there is a connection relationship. If the parent word may be omitted, it has the parent word of the parent word as parent information.
It may be configured to have a plurality of pieces of parent information.

[Sixth Embodiment] FIG. 15 is a block diagram of a sixth embodiment of the speech synthesizer of the present invention. The same parts as those in the above-mentioned embodiment are designated by the same reference numerals and the description thereof will be omitted. To do. In this embodiment, after creating all possible word string candidates, the word string candidate selecting section 116 that selects the most probable candidate from among the word string candidates and the address dictionary 10 that matches in the sentence.
An address reading determination unit 117 that determines whether or not to select the address reading for the word string candidates in 3 is provided.

That is, if the address dictionary is always referred to when matching with the address dictionary, there is a problem that the following parts other than the address are replaced with the reading of the address.
For example, "fossil (bake'ishi)", "triangle (Mikado,
There may be a place name with a pronunciation different from a common noun, such as "Mi'sumi", "Yamadera", "lowercase (common)", "uppercase (da'imoji)", etc.
Further, even if the readings are the same, the accent type or accent combination attribute may be different, so that the accent to be read may be different. Therefore, in such a case, it is necessary not to read aloud the address. In order to avoid this inconvenience, in the sixth embodiment, the address reading is not selected when the number of words or the number of characters is smaller than the predetermined number of words or the number of characters.

For example, "Okayama Prefecture (Ken) -Kamibo-gun-Kayocho-"
Place names such as "Kita"-"Kado""," Oita Prefecture "-Higashi-Kunisaki-gun-Kunimicho-Middle-Shimo" Is registered in the address dictionary 103, "Wait at the north gate"
Even if a sentence such as "Upper middle lower" is entered, 2 of "North Gate"
If only letters or two letters of "middle lower" are matched, if the reading of the address is not selected and the basic dictionary 105 is referred to, the correct reading can be performed.

FIG. 16 is a flowchart of the algorithm of the sixth embodiment. The number of words or the number of characters in the candidate word string is obtained (S401), and it is determined whether the obtained number of words or the number of characters is smaller than a threshold value (S402). When the number of words or the number of characters of the candidate word string is equal to or larger than the threshold value, the reading of the address is selected (S403). On the other hand, if the calculated number of words or characters is smaller than the threshold value, the reading of the address is not selected and the process ends.

[Seventh Embodiment] FIG. 17 is a block diagram of a seventh embodiment of the speech synthesizer of the present invention. The same parts as those in the above-mentioned embodiment are designated by the same reference numerals and their description is omitted. To do. In the seventh embodiment, a prosody boundary setting unit 118 for setting prosody boundary symbols is provided in order to obtain a synthetic speech with a natural utterance.

That is, the address with a long total number of moras is read.
When setting, do not set exhalation paragraph boundary and phrase boundary
When the vocal pitch becomes too low or there is no breathing,
It sounds stuffy in your voice. For example, "Sapporo City, Hokkaido
Minami-ku Jozankei Jozankei Toyoha Mine Kurumizawa "
"Hocca 'Idoo Sapporo ’ Min ' Jo
User Joe Zankate Johanna ' Kuru
It's "Mi'sawa", but when you read it all at once, it sounds very unnatural.
I can. Also, it sounds unnatural even if it is separated too much. That
For prosodic boundary symbols such as phrase boundaries or expiratory paragraph boundaries
Need to be set to an appropriate position.

As a first method of setting the inter-word boundary symbol, the number of mora of words matched with the address dictionary 103 is accumulated, and the phrase is kept so that the accumulated number of mora does not exceed the threshold value or every time it exceeds the threshold value. A method of setting a prosodic boundary symbol such as a boundary or an expiratory paragraph boundary can be considered.

FIG. 18 is a flowchart of this algorithm. Set the current word mora number to the cumulative mora number (S
501). The reading of the place name word is set in a buffer (not shown) (S502), and the place name word pointer is advanced to the next (S50).
3). It is determined whether or not the place name word candidate sequence ends (S504).
), If not finished, the number of mora of the place name word is added (S505). It is determined whether or not the cumulative number of moras exceeds the threshold value (S506), and step S50 is performed until the threshold value is exceeded.
Repeat steps 2 to S505. When the cumulative number of moras exceeds the threshold value, the breath paragraph mark is set in the buffer (S507), and the process returns to step S501 until the place name word candidate sequence ends.
Steps S501 to S507 are repeated.

In the above example of "Jyozankei, Jozankei, Jozankei, Toyoha Mine, Kurumizawa, Minami-ku, Sapporo, Hokkaido", the number of mora for each word is {6.
Since it is 5, 4, 6, 13, 5}, if the threshold value is set to 13 moras, it is divided into {(6, 5), (4, 6), 13, 5}. Here, when the expiratory paragraph mark is represented by "・", the reading is "Hocker'id Sapporo'Shi Minami '
K It becomes "Joza'nke Joe Sankate Yohako'zan walnut 'Sawa" and can read aloud naturally.

As a second method of setting the inter-word boundary symbol, the boundary symbol is included in the hierarchical data structure of the place name word in the address dictionary 103, and the prosody such as phrase boundary or exhalation paragraph boundary is referred to by referring to it. A method of setting a boundary symbol can be considered. FIG. 19 is a flowchart of this algorithm. The reading of the place name word is set in the buffer (S601).
The pointer of the place name word is advanced to the next (S602), and it is determined whether or not the place name word candidate sequence ends (S603). If the place name word candidate sequence is not finished, if there is a prosodic boundary symbol, it is acquired and set (S604), and the process returns to step S601 to repeat steps S601 to S604 until the place name word candidate sequence is finished.

In the second method, not only the number of mora but also the phrase boundary and the expiratory paragraph boundary symbol can be distinguished and inserted in advance, so that the utterance can be made more naturally. In the above example, it is more natural to say "Sapporo City" and "Minami Ward" together, so that information is stored in the address dictionary 103. Here, when the expiratory paragraph boundary symbol is represented by "." And the phrase boundary symbol is represented by "/", the above-mentioned example is "Hocca'ido / Sapporo 'system. It is possible to read aloud as "Minami'ku / Joza'nke Joe Sankate Yohako'zan / walnut'Sawa."

[Embodiment 8] FIG. 20 is a block diagram of an embodiment 8 of the speech synthesizer of the present invention. In the figure, the same parts as those in the above-mentioned embodiment are designated by the same reference numerals. The description is omitted. In the eighth embodiment, an address symbol conversion unit 119 for converting an address symbol represented by a minus sign, a long sound symbol, etc. into an address reading is provided. That is, when a sentence including an address written as "1-6-1 Marunouchi, Chiyoda-ku, Tokyo" is entered, the address "1-6-1" is read out as "Ichi-minus-lok-minus-ichi". If you do, it will give you an uncomfortable reading of the address. Therefore, when an address is written with consecutive numerical values as described above, it is read out as "1 of 6 of 1", that is, "Ichi'no-Roku'no-Ichi".

FIG. 21 is a flowchart of this algorithm. After recognizing the address determined from the word string in the address dictionary 103 that matches the sentence, the characters following the address section are searched. The next letter is a number (0-9, 〇-9,
It is determined whether it is ten, one hundred, ten, ten thousand) (S701), and if it is a numeral, the pointer is advanced by one character (S702), and step S70 is performed.
Return to 1. If the next character is not a number, it is determined whether or not the next word is a classifier (address, number, chome, number) (S703).
), In the case of classifiers, the pointer is advanced by the character of the word (S7
04), and returns to step S701.

If the next word is not a classifier, it is determined whether the next character is a delimiter (-,-) (S705). If it is a delimiter, the character is replaced with "no" (S706).
Here, the long-sound symbol "-" is included as a delimiter so that the minus sign "-" is often mistakenly written as a long-sound symbol so that even if it is written incorrectly, it can be read. If none of the above characters or words is found, the process ends. By performing these processes, the description of the address including the delimiter character between the numerical values can be correctly read.

[Ninth Embodiment] FIG. 22 is a block diagram of a ninth embodiment of the speech synthesizer of the present invention. The same parts as those in the above-described embodiment are designated by the same reference numerals and the description thereof will be omitted. For example, if the sentence “Toshima-ku Third Junior High School in Tokyo” is entered, if only the address section determined from the word string of the address dictionary 103 that matches the sentence is converted into pronunciation information, “Tokyo ′ Interpreted as "Toshima" + "Ritsu 3rd Junior High School", and as a result, "Ritsu" in the sentence analysis unit 104.
Is read as "Ta'chi".

Therefore, in the ninth embodiment, "standing, inside,
From the word strings in the address suffix dictionary 120 that stores the address suffixes and accent combining attributes and the like in the vocabulary such as “people, going, originating, arriving, ...” and the address dictionary 103 that matches the words in the sentence. The address suffix dictionary search unit 121 that searches the address suffix dictionary 120 for a word that matches the word that immediately follows the determined address section, and the address suffix dictionary search unit 121 searches for an address suffix after the address. If it exists, correct it so that the address section consisting of the word string in the address dictionary 103 that matches the word in the sentence includes the address suffix, and change the word string in the address dictionary 103 that matches the word in the sentence. An accent combination processing unit 122 is provided for combining the final word and the address suffix according to the accent combination attribute set in the address suffix, and setting the reading.

FIG. 23 is a flowchart of the algorithm of the ninth embodiment. It is determined whether or not the next word of the word string of the address is in the address suffix dictionary 120 (S801), and if it is in the address suffix dictionary 120, the accent combination processing unit 122 performs accent combination (S802). By performing such processing, the text information in the previous example can be converted to "Tokyo" The information "Toshimak'Ritsu" + "3rd junior high school" can be passed to the sentence analysis unit 105, so that it can be read correctly.

[Embodiment 10] FIG. 24 is a block diagram of a speech synthesizer according to Embodiment 10 of the present invention. The same parts as those in the above-described embodiment are designated by the same reference numerals and the description thereof will be omitted. For example, if an address appears repeatedly in a sentence and then the words contained in that address appear repeatedly, for example, "Kamiotadaka Community Center is in Kamiodanaka, Nakahara-ku, Kanagawa Prefecture." 』
In this sentence, the place name "Kamiotadanaka, Nakahara-ku, Kanagawa" can be read correctly, but the sentence "Kamiotadaka public hall is ..." that appears next, despite having read it earlier,
You may hear a wrong reading because another reading is registered with the same notation. Further, in the case of adopting the configuration as in the sixth embodiment, it may happen that the address reading cannot be performed because the number of characters (the number of words) is less than a predetermined threshold value.

Therefore, in the tenth embodiment, for all the constituent words of the address processed by the address dictionary search unit 102, a kanji notation and pronunciation information are paired to make the learning word buffer 12
4, an address word learning unit 123 to store in 4, a learning word buffer 124 to store the notation and reading of recently used addresses,
After searching the hierarchical address dictionary 103, the sentence for learning is searched for, and if there is a word that matches the learning word, the corresponding pronunciation information is embedded in the corresponding sentence. A learning word search unit 125 is provided. The words stored in the learning word buffer 124 are
By preferentially using the reading, it is possible to correctly read when a part of the address that appears even once in the sentence appears next time.

A learning word deleting unit 126 for initializing the contents in the learning word buffer 124 may be provided. In this case, the learning word buffer 124 is not initialized normally, but may be initialized according to the user's designation.
The configuration may be such that initialization is performed each time text information is input. Furthermore, the configuration may be such that initialization is performed when the number of sentences input after registration exceeds a predetermined number.

Note that the voice synthesizing method implemented in the voice synthesizing apparatus of the present invention is not limited to writing in the ROM of the voice synthesizing apparatus, but is also recorded in a recording medium D such as a compact disc as shown in FIG. Alternatively, the recording medium D may be loaded in a disc drive of a personal computer to synthesize voice.

In the above embodiment, the case where the word string is an address has been described, but the word string is not limited to an address and the same effect can be obtained as long as the word string has a hierarchical structure.

[0064]

As described above, the speech synthesizer of the present invention,
A voice synthesis method and a recording medium recording a computer program of this method have a hierarchical dictionary in which notation such as an address in which a plurality of readings exist for one notation and reading information are hierarchized. This has an excellent effect of reading text information such as an address in which plural readings exist for one notation with correct reading.

[Brief description of drawings]

FIG. 1 is a basic block diagram of a speech synthesizer of the present invention.

FIG. 2 is a schematic diagram of an example of a speech synthesizer of the present invention.

FIG. 3 is a configuration diagram of a first embodiment of a speech synthesizer of the present invention.

FIG. 4 is a conceptual diagram of an example of an address dictionary.

FIG. 5 is a flowchart of the algorithm of the first embodiment.

FIG. 6 is a configuration diagram of a second embodiment of a speech synthesizer of the present invention.

FIG. 7 is a configuration diagram of a third embodiment of the speech synthesis apparatus of the present invention and its modification.

FIG. 8 is a conceptual diagram of a search start position storage buffer.

FIG. 9 is a configuration diagram of a fourth embodiment of a speech synthesizer of the present invention.

FIG. 10 is a conceptual diagram of another example of an address dictionary.

FIG. 11 is a flowchart (part 1) of the algorithm of the fourth embodiment.

FIG. 12 is a flowchart (part 2) of the algorithm of the fourth embodiment.

FIG. 13 is a configuration diagram of a fifth embodiment of a speech synthesizer of the present invention.

FIG. 14 is a configuration diagram of a modification of the fifth embodiment of the speech synthesizer of the present invention.

FIG. 15 is a configuration diagram of a sixth embodiment of a speech synthesizer of the present invention.

FIG. 16 is a flowchart of an algorithm according to the sixth embodiment.

FIG. 17 is a configuration diagram of a seventh embodiment of a speech synthesizer of the present invention.

FIG. 18 is a flowchart (part 1) of the algorithm of the seventh embodiment.

FIG. 19 is a flowchart (part 2) of the algorithm of the seventh embodiment.

FIG. 20 is a configuration diagram of an eighth embodiment of a speech synthesizer of the present invention.

FIG. 21 is a flowchart of the algorithm of the eighth embodiment.

[Fig. 22] Fig. 22 is a configuration diagram of a ninth embodiment of a speech synthesis device of the present invention.

FIG. 23 is a flowchart of the algorithm according to the ninth embodiment.

[Fig. 24] Fig. 24 is a configuration diagram of a tenth embodiment of a speech synthesis device of the present invention.

FIG. 25 is a conceptual diagram of a recording state of the recording medium of the present invention.

[Explanation of symbols]

 1 Text Input Section 2 Hierarchical Dictionary Search Section 3 Hierarchical Dictionary 4 Text Analysis Section 5 Basic Dictionary 6 Speech Waveform Generation Section 7 Speaker

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Hidetoshi Tsujiuchi 1015 Kamiodanaka, Nakahara-ku, Kawasaki-shi, Kanagawa Fujitsu Limited (72) Inventor Shinta Kimura 1015, Kamitadanaka, Nakahara-ku, Kawasaki, Kanagawa Fujitsu Limited

Claims (8)

[Claims] 1. A voice synthesizer for synthesizing a voice from text information, each of a word string comprising an input means of text information and a word group in which a word appearing next to the word determines the reading of the word appearing next to the word appearing first. From the hierarchical dictionary, a hierarchical dictionary in which the notation of the word and the reading information are stored together with the information of the hierarchical structure in which the words are hierarchized according to the connection order, and the word reading candidates included in the input sentence information A sentence in which the reading of the word string included in the sentence information is selected based on the hierarchical structure information from the hierarchical dictionary search means for searching and the reading candidate, and the word string is converted into the reading. A voice synthesizing apparatus comprising: an analysis unit, a voice waveform generation unit that generates a voice waveform from reading information, and a voice waveform output unit. 2. The speech synthesizer according to claim 1, further comprising means for designating an area, wherein the hierarchical dictionary search means is means for starting the search from a hierarchy belonging to the designated area. 3. The information of the hierarchical structure is information for specifying a parent word in a higher hierarchy of each word whose reading should be determined as the reading, and refers to the reading candidate information searched from the hierarchical dictionary. 3. The speech synthesizer according to claim 1, further comprising a connection relation setting unit that sets a connection relation of candidates for reading the words of the word string included in the sentence information. 4. The hierarchical dictionary searching means includes means for searching a hierarchical dictionary based on a word string in which notation of a word in one of the word strings is omitted. The speech synthesizer according to. 5. The reading of the hierarchical dictionary is referred to as the reading of the character or word when the notation of the predetermined number of characters or words matches the notation of the character or word included in the word string stored in the hierarchical dictionary. The speech synthesis apparatus according to any one of claims 1 to 4, further comprising a reading determination unit for determining. 6. A suffix dictionary storing notation and reading information of suffixes connected to the word string and information of notation that matches the notation immediately after the word string are searched from the suffix dictionary. 6. The speech synthesis apparatus according to claim 1, further comprising: a suffix dictionary search unit that selects a reading of the suffix dictionary that matches the immediately following notation as the reading of the immediately following notation. 7. A notation of each word of a word string consisting of a word group in which the reading of the word appearing next to the word is determined by the word appearing first and the reading information is a hierarchy in which the words are hierarchized according to the connection order. A speech synthesis method for synthesizing speech from sentence information by referring to a hierarchical dictionary stored together with structure information, in which sentence information is input, and word reading candidates included in the inputted sentence information are searched. Searching from the hierarchical dictionary, selecting the reading of the word string included in the sentence information from the reading candidates based on the information of the hierarchical structure, converting the word string into the reading, and reading A voice synthesizing method comprising generating a voice waveform from information and outputting the voice waveform. 8. A hierarchy in which the notation of each word of a word string consisting of a word group in which the reading of the word that appears next to the word that appears first and the reading information are hierarchized according to the connection order. The step of inputting the hierarchical dictionary stored together with the structure information and the sentence information, and searching the hierarchical dictionary for a candidate for the reading of the word included in the input sentence information, the sentence from the reading candidates Selecting the reading of the word string contained in the information based on the information of the hierarchical structure, converting the word string into the reading, generating a speech waveform from the reading information, and outputting the speech waveform And a computer program including a step of performing recording.