JP2001350489A - Voice synthesizer - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice synthesizer in which the number of errors is reading a same shape different sound word is made small, rhythm/voice quality that are appropriate for the contents of a mail are used for the reading and degree of understanding is made high. SOLUTION: The synthesizer is provided with means (115, 116 and 117) which conduct grouping for mail addresses, and store hierarchial data that are made by hierarchizing each group, means (118 and 106) which add node names of the hierarchial data to the words in the mail and register these words in a user's word dictionary and a means which generates list of the node names of the data from information containing the names of originators of the mails/ addresses/destinations of carbon copies. During the retrieval of the user's word dictionary (106), the members of the list above and the node names that are added to the words in the dictionary are collated and text analysis is conducted by using the matched words only.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizer for reading out an electronic mail.

[0002]

2. Description of the Related Art Conventionally, as a method of distinguishing between homonymous words in text-to-speech conversion, reference 1: (“Separation of homonymous words using a decision list”) Yoshiyuki Umemura, Tsukasa Shimizu, Toyota Central Research Laboratory Machine Recognition Room, The 4th Annual Meeting of the Language Processing Society of Japan, March 1998). Homomorphic words are words that have the same notation and have multiple readings. For example, "Market" is a homonym with two readings as common nouns, "Shijo" and "I ↓ Chiba"("↓" indicates a descending position of an accent).

[0003] Natural language analysis using a decision list is described in Document 2: ("Extraction and Evaluation of Japanese Dependent Dependency Preference Information from Corpus", Takehito Utsuro et al., Natural Language Processing Vol. 7, 1999 10
Month).

A method of morphological analysis for dividing a Japanese sentence into words and further determining part of speech and pronunciation is described in Reference 3:
("Japanese Morphological Analysis Using Probabilistic Decision Trees", Hideki Kashioka, et al., Proc. Of the 3rd Annual Meeting of the Language Processing Society of Japan, 19
(March 1997). In this method, learning is performed by preparing an attribute related to a word, classifying the word according to the value of the attribute, and forming a decision tree with probability.
At the time of analysis, word probabilities are obtained from the decision tree using a set of values of the word attribute, and word combinations with higher appearance probabilities are used as output word divisions to perform word division.

[0005] Accent, duration,
Reference 4 shows the differences in poses:
("Analysis of Prosodic Features of Japanese Conversational Voice in Roundtable and Rakugo", Shoichi Takeda, et al., Journal of the Acoustical Society of Japan, Vol. 54, No. 3
No. 1998). What is prosody?
It is not a segmental feature of individual vowels or consonants, but a feature given to a speech sequence consisting of a plurality of phonemes. Accent, intonation (intonat)
ion), stress, emphas
is), prominence, rhythm (rh)
ythm), tempo, pause
e) etc. ("Prosodic features" in [Acoustic Term Dictionary]
Description). In the speech synthesizer, selection of a prosody corresponds to selection of a prosody prediction table such as a duration prediction table, an accent prediction table, a pitch prediction table, and a pause length prediction table.

[0006] Further, when reading out a mail, the voice quality is changed for each mail sender to facilitate the identification of the mail sender, as disclosed in Japanese Patent Laid-Open No. 11-102198.
"Message processing device, message processing method and medium recording message processing program". Voice quality refers to the auditory characteristics of the whole voice, other than phonemes as linguistic information, perceived from speech waves, personality information of the speaker as to who is talking, and what kind of mental state is being spoken (Information of "voice quality" in the [acoustic dictionary]). In the speech synthesizer, selection of voice quality corresponds to selection of a segment set (a set of segments created from one speaker).

[0007]

In the case of synthesizing an e-mail by voice, the field of the e-mail is diversified. Therefore, in the conventional method of using prosody and voice quality, reading is far from the context of the e-mail. The present invention uses the information of the sender, destination, copy destination, etc. attached to the mail, and a database in the device to select a reading suitable for the context of the mail, thereby enabling the homomorphic word to be read. A first object is to provide a speech synthesizer with few reading errors.

[0008] Further, in the case of synthesizing an e-mail, the field of the e-mail is diversified. Therefore, the conventional method using the prosody and voice quality often has a prosody and voice quality that does not match the sender or contents of the mail. Was. According to the present invention, information such as a sender, a destination, and a copy destination added to a mail is provided.
To provide a speech synthesis device with high intelligibility that can read aloud with the prosody and voice quality suitable for the contents of the mail by selecting the reading, prosody and voice quality suitable for the mail using the database in the device. This is the second purpose.

[0009]

For this purpose, in the voice synthesizing apparatus according to the first aspect of the present invention, there are provided means for registering words used in text analysis of a received mail, means for storing a mail address database, and mail address. Means for adding / modifying / deleting the contents of the database, grouping the mail addresses of the database, storing each group as hierarchical data, and storing the words as words in the hierarchical data. Means for adding a node name to the user word dictionary, and means for creating a list of node names of the hierarchical data from information on the sender, destination, and copy destination of the mail. In the search, the members of the created list and the node names added to the words in the user word dictionary Collating, and performing text analysis using only words that the node name matches any of the members of the list.

Further, in the speech synthesizer of the second invention, means for controlling the prosody or voice quality when reading the received mail, means for storing a mail address database, and adding / modifying the contents of the mail address database A voice synthesizing apparatus having means for deleting the mail, grouping the mail addresses of the database, storing each group as hierarchical data in a hierarchy, and adding information on the sender, destination, and copy destination of the mail; Means for selecting a node of the hierarchical data based on the prosody and voice quality associated with the selected node when reading out the mail.

Further, in the speech synthesizing apparatus according to the third aspect of the present invention, there is provided a means for reading out homomorphic words using a decision list, a means for storing a mail address database,
Means for adding / modifying / deleting the contents of a mail address database, wherein the mail addresses of the database are grouped,
Means for storing each group as hierarchical data, and means for selecting a node name of the hierarchical data from information on the sender, destination, and copy destination of the mail. The method is characterized in that words are distinguished by using a decision list including a rule in which a certain attribute is used as evidence.

[0012]

Embodiments of the present invention (hereinafter, embodiments) will be described below in detail with reference to the drawings. <First Embodiment><Configuration> FIG. 1 is a diagram showing a configuration of a first embodiment. Reference numeral 101 denotes a user interface for operating the speech synthesizer, 102 denotes a text-to-speech conversion unit that reads out the contents of the mail received by the mail transmission / reception unit 112 in the mail management unit 111, and 103 denotes an input from the received mail. Text input for reading text, passing sender / destination / copy / destination information on the mail to the text-to-speech conversion information setting unit, and passing two types of information returned from the processing unit to the subsequent processing unit Department. Reference numeral 104 denotes a text analysis unit that generates an intermediate language (reading, accent position, phrase start position, pause position) from the input text, and 105 denotes word information (notation, reading, part of speech, accent type,
System word dictionary that stores accent-binding type, etc., 1
Reference numeral 06 denotes a user word dictionary with a mail address that stores word information added by the user. A synthesis parameter generation unit 107 generates synthesis parameters (synthesis unit / duration / pitch / pause / amplitude) from the output of the text analysis, and 108 stores a prediction table used to determine synthesis parameters. This is a prediction table storage unit. The prediction table storage unit includes a duration prediction table, a pitch prediction table, a pause prediction table, and an amplitude prediction table. In order to cope with the selection of the prosody, a plurality of tables are stored for any type of table. Reference numeral 109 denotes a speech synthesis unit that generates a waveform from synthesis parameters, and reference numeral 110 denotes a speech unit dictionary that stores a speech unit set used in speech synthesis. A plurality of unit sets are stored in the speech unit dictionary in order to support selection of voice quality.

Reference numeral 111 denotes a mail management unit for sending / receiving and storing mails; 112, a mail sending / receiving unit for sending / receiving mails; and 113, a mail storage unit for storing received mails.

Reference numeral 114 denotes a database management unit for managing a database related to the mail address.
Is a mail address management unit that manages the mail address database 116 and the mail address hierarchy database 117. Reference numeral 116 denotes a mail address database storing information on mail addresses and groups, and 117 denotes a mail address hierarchy database storing mail addresses and hierarchical relationships between groups. A user word dictionary management unit 118 adds, deletes, and corrects words in the user word dictionary with user addresses.
Reference numeral 19 denotes a prosody and voice quality setting database management unit that adds, deletes, and corrects data in the prosody and voice quality setting database 120, and 120 denotes the mail address database 1
A prosody / voice quality setting database 121 for associating the mail address or group stored in 16 with the set of prosody attributes and the set of voice attributes, 121 is a set of prosody attributes used in the prosody / voice quality database 120. A prosody / prediction table correspondence database that associates the prediction table stored in the prediction table storage unit 108 with the prediction table. Reference numeral 122 denotes a voice quality / unit set correspondence database that associates a set of voice quality attributes used in the prosody / voice quality setting database 120 with a unit set of the voice unit dictionary 110.

Reference numeral 123 denotes a “user dictionary search mail address list” and a “prosodic / voice quality setting mail address” based on the sender, destination, and copy destination information attached to the mail.
This is a text-to-speech conversion information setting unit that determines the following two types of information. The "sender", "destination", and "copy destination" of an email are defined as follows. The sender is the e-mail address of the person who sent the e-mail. The destination is a mail address of a mail destination specified by the person who sent the mail, and a plurality of addresses can be specified. The copy destination is a mail address that is added to the destination and specified by the sender of the mail for the purpose of sending a copy, and a plurality of destinations can be specified. It is assumed that it is not possible to omit the destination and specify only the copy destination. The sender, destination, and copy destination of the received mail are the same as the settings at the time of transmission. Therefore, there is a case where a mail which passes through the mailing list does not have its own mail address at both the destination and the copy destination.

The "user dictionary search e-mail address list" contains, among words in the user dictionary, the sender of the e-mail and
This is information used to extract only words commonly associated with the destination / copy / destination, and is a list of mail addresses or group names. The members of the list are given priorities as described below, and the members of the list are sorted in descending order of priority. The priority value is 0 for the mail address, -1 for the terminal group, and the minimum number of links when the non-terminal group follows the tree structure downward and reaches one of the terminal groups in the list. Minus. If there is only a common group in the list, the priority is not obtained, but 0 is set as the value of the priority.

User word dictionary with mail address 106
At the time of search, only when the email address or group name assigned to a word matches a member of this list, it is used for analysis (for words in the system dictionary,
There is no such restriction). The “prosody / voice quality setting mail address” is information for selecting the prosody and voice quality when reading out the mail, and is a mail address or a group name. For example, in an email addressed to a user,
It reads aloud according to the prosody and voice quality set for the sender of the email,
Since the user who is the recipient is the third party in the mail sent by copy sending, the mail is read out according to the prosody and voice quality set in the group to which the sender, copy destination, and destination belong in common. Also, e-mails sent via the mailing list are read out according to the prosody and voice quality set in the mailing list. The prosody / voice quality setting mail address is used by the synthesis parameter generation unit 107 for prosody selection, and is used by the speech synthesis unit 109 for voice quality selection.

[0018]

[Table 1] Table 1 shows a part of the contents of a user word dictionary with a mail address that stores word information added by the user. The user assigns a mail address or a group name in addition to information such as notation, part of speech, reading, and accent.

[0019]

[Table 2] Table 2 shows a part of the contents of the mail address database. The user stores information such as a mail address or group name, type, name, telephone number, affiliation, and the like. The name, telephone number, and affiliation are used by the user to set the destination when sending the mail and identify the destination.

[0020]

[Table 3] Table 3 shows a part of the contents of the mail address hierarchy database that stores the mail addresses and the hierarchical relationship between groups. If the mail address belongs to a plurality of groups, it is registered as a plurality of data. suz
uki @ oki. co. Since jp belongs to an in-house group and a voice group, there are two data.

[0021]

[Table 4] Table 4 shows a part of the contents of the prosody and voice quality setting database for setting the prosody and voice quality to the mail address or the group stored in the mail address database 116. For the mail address and the group, the prosody is designated by a set of attributes (including one attribute) prepared in advance, such as “speed” and “tone”. The voice quality is also designated by a set of attributes prepared in advance (including a case where there is only one attribute). An attribute for associating the prediction table / segment set with the mail address and the group is set. The user selects and sets only the attributes prepared in advance. Only the value of the attribute prepared in advance is used.

[0022]

[Table 5] Table 5 shows a part of the contents of a prosody / prediction table correspondence database that associates a set of prosody attributes used in the prosody / voice quality setting database with a prediction table stored in the prediction table storage unit 108. It was done. As the prediction table, there are four types of tables of duration, pitch, pause, and amplitude. "Duratio
n ”+ number,“ Pitch ”+ number,“ Pause ”+
The numeral, “Power” + the character string of the numeral is used as an identifier of the prediction table.

[0023]

[Table 6] Table 6 shows a part of the contents of the voice quality / unit set correspondence database that associates a set of voice quality attributes used in the prosody / voice quality setting database with the unit set of the voice unit dictionary 110. It is. "Male" + number,
The character string of “Female” + number is used as an identifier of a unit set. The segment set is made to correspond to the voice quality attribute setting pattern.

<Operation> In this embodiment, data called a mail address hierarchy database shown in FIG. 2 is used.
The “group”, “non-terminal group”, “terminal group”, and “common group” in this database satisfy the following conditions.

(1) The “group” is classified into a “non-terminal group” and a “terminal group”. (2) In the “non-terminal group”, the group can be at the lower level, but the mail address cannot be at the lower level. (3) In the “termination group”, the mail address can come at the lower position, but the group cannot come at the lower position. (4) It is assumed that there is a non-terminal group always called a “common group” at the root of the hierarchical structure. (5) Non-terminal groups other than the “common group” come under one non-terminal group. It does not allow loops that fall beneath yourself or a non-terminal group below yourself. (6) The mail address always comes under the terminal group. Allows email addresses to be subordinate to multiple termination groups.

In the following description, an operation is performed so as to maintain data consistency between databases even if not particularly described. The user can add / modify / delete the e-mail address or group. Therefore, always check and eliminate the processing that causes inconsistency. The e-mail address-added user word dictionary 106, the e-mail address database 11
6. Data consistency of the mail address hierarchy database 117 and the prosody and voice quality setting database 120 can be maintained.

The user adds, deletes, or changes the contents of the mail address database 116 and the mail address hierarchy database 117 by calling the mail address management unit 115 via the user interface 101. Also, by calling the user word dictionary management unit 118, words in the user word dictionary with mail address 106 are added / deleted / changed. Further, the contents of the prosody and voice quality setting database 120 are added, deleted, and changed by calling the prosody and voice quality setting database management unit 119.

The user operates the mail transmitting / receiving unit 112 to
Send and receive mail. The received mail or a copy of the transmitted mail is stored in the mail storage unit 113.
When the user performs text-to-speech conversion of the mail, the mail storage unit 113 is transmitted via the user interface 101.
Is specified, and the text-to-speech converter 102 is called.

In the text-to-speech conversion unit 102, first, the text input unit 103 reads a mail from the mail storage unit 113. Next, the text-to-speech information setting unit 123
To obtain a user dictionary search mail address list and a prosody / voice quality setting mail address from the destination, sender, and copy destination of the read mail, and pass them to the subsequent processing unit. After the processing of the text-to-speech information setting unit 123, the text analysis unit 104 divides the text of the mail into words, determines the reading, accent position, intonation start position, and pause position, and generates an intermediate language.

The synthesizing parameter generation unit 107 calculates the duration of the phoneme for the intermediate language output by the text analysis unit.
Determine parameters for pitch, pause length, and amplitude. The prediction table associated with the prosody / voice quality setting mail address determined by the text-to-speech conversion information setting unit 123 can be obtained by referring to the prosody / voice quality setting database 120 and the prosody / prediction table correspondence database 121. it can. Then, a prosody is generated using the prediction table.

The speech synthesis unit 109 synthesizes speech from the intermediate language output from the text analysis unit and the parameters output from the synthesis parameter generation unit. The voice quality associated with the prosody / voice quality setting mail address determined by the text-to-speech conversion information setting unit 123 is the prosody / voice quality setting database 120 and the voice / speech unit correspondence database 12.
2 can be known. Then, speech is synthesized using the segment set in the speech segment dictionary 110.

FIGS. 3, 4, 5, 6, and 7 show a text-to-speech conversion information setting unit 123 in which a text-to-speech conversion mail address list and a prosodic / voice-quality setting mail address are set. It is a flow of information setting processing for conversion.

FIG. 3 shows a main routine of the text-to-speech conversion information setting process. In processing 301 and processing 302,
By referring to the sender, destination, and copy destination of the mail, the mail is classified into the mail addressed to the user, the mail sent by copy, and the mail received via the mailing list. Processing 303, processing 304, processing 30
The subroutine 5 is executed. In step 306, the group names of all the groups that can be reached by following the link upward are added to the user dictionary search mail address list determined in each subroutine. If the list is empty, unconditionally add a "common group". In the process 307, a subroutine for assigning priorities to members of the user word dictionary search mail address list and sorting the members in descending order of priority is executed.

FIG. 4 is a processing flow of information setting for a mail addressed to the user. Process 401, Process 402,
Process 403 is a process for determining a mail address list for user word dictionary search, and includes processes 404, 405,
Process 406 is a process for determining a prosody / voice quality setting mail address.

FIG. 5 shows a subroutine for a mail that has been copied. Process 501 is a process for determining a mail address list for user word dictionary search.
2 is a process for determining a prosody / voice quality setting mail address.

FIG. 6 is a processing flow for setting information on a mail that has arrived via a mailing list. In the case of mailing via a mailing list, the destination attached to the mail is not the user but the mailing list address.
Process 601, process 602, and process 603 are processes for determining a mail address list for user word dictionary search.
Processes 604, 605, and 606 are processes for determining a prosody / voice quality setting mail address.

FIG. 7 is a flowchart of a process for sorting members of the user dictionary search mail address list. Processing 7
In step 01, the priority of the mail address in the list is set to 0, the priority of the terminal group in the list is set to -1, and the priority of the non-terminal group is set to 1 indicating that it is not set. Process 702, Process 703, Process 704, Process 70
5. Process 706 is a process of setting, as the priority, a value obtained by changing the sign of the distance when the node is traced downward to the shortest mail address in the mail address hierarchy database for the members of the non-terminal group. In process 707, members of the user dictionary search mail address list are sorted in descending order of magnitude using the priority as a key. An existing sorting algorithm is used.

FIG. 8 shows the address (sender / destination / copy / destination) assigned to the mail address, and the text / speech conversion information (user dictionary search mail) determined by the text / speech conversion information setting unit. Address list,
And a prosody / voice quality setting mail address).

The flow of processing for determining text-to-speech conversion information from the addresses illustrated in FIG. 8 will be described below with reference to FIGS. 3, 4, 5, 6, and 7. However, it is assumed that the databases of Table 1, Table 2, Table 3, and Table 4 are used for the processing.

First, in the main routine of FIG. 3, it is determined in step 301 that the mail is addressed to the user, and in step 30
At 3 the subroutine 1 is called. In FIG. 4, since there is a copy destination in step 401, step 403 is executed. asahi @ ide. co. jp is not registered in the mail address hierarchy database (Table 3), so that noguchi @ north. c
o. jp, kurobe @ north. co. jp, s
hiomi @ south. co. The group to which the largest group belongs to jp is selected. According to the mail address hierarchy database, all three mail addresses belong to the climbing group, and none of the addresses belong to the other groups. Therefore, the mail address list for user dictionary search is {climbing group}.

Next, the process 404 is executed, and a
sahi @ ide. co. Since jp is not registered in the mail address database (Table 2), the process 406
Is executed. asahi @ ide. co. Is not registered in the mail address hierarchy database, and noguchi @ north. co. jp,
kurobe @ north. co. jp, shiomi
＠South. co. jp belong to the climbing group,
Since none of the addresses belong to another group, the e-mail address for setting the prosody and voice quality is “climbing group”. At this point, the subroutine 1 ends, and the process returns to the main routine of FIG.

In the process 306 of FIG. 3, since the mail address list for user dictionary search is {Climbing Group}, the "Private Related Group" and the "Common Group" which are the upper groups of the climbing group are added. Then, the user dictionary search mail address list is {climbing group, privately related group, common group}.

Next, at step 307, subroutine 4 is called. The priority is “-1” for the terminal group “climbing group”, −2 for the non-terminal group “private related group”, and −3 for the “common group”. (-
1), privately related group (-2), common group (-
3) It becomes｝. However, the numerical value in parentheses after the group name is the priority value.

Finally, the mail address list for user dictionary search is {climbing group (-1), private related group (-2), common group (-3)}, and the prosodic / voice quality setting mail address is "climbing Group ". Through the above processing, the text-to-speech conversion information in FIG. 8 is determined.

FIG. 9 shows an example of the input sentence included in the mail and the generated intermediate language. The text analysis unit 104 performs a process of dividing an input sentence into words in the process of creating an intermediate language. At this time, candidates for the words to be divided are a system word dictionary 105 and a user word dictionary with mail address 106. Use words taken from. When searching the user word dictionary with a mail address, only when the mail address or the group name assigned to the word matches any member of the mail address list for user dictionary search, the word is extracted as a candidate for word division.

FIG. 10 is a processing flow from dictionary lookup to word division. FIG. 11 shows a graph structure generated in the process 101. Virtual nodes corresponding to the beginning and end of the sentence are added to the graph.

In the process 102, the appearance probability is determined by the method of Reference 3. In Reference 3, using the attributes of words,
The appearance probability is calculated. In the present embodiment, an attribute “priority” is added. As the value of the priority as the attribute of the word, the value of the priority of the mail address or the group name given to the word in the user dictionary search mail address list is used. Since the priority of the mail address or the group differs depending on the text, the “priority” of the attribute of the word has a different value depending on the text even if the word is the same. For words taken from the system dictionary,
The priority value of "common group" in the user dictionary search mail address list is used. Since the common group is always included in the list, the value can be determined. However, this value differs depending on the mail address information. As a decision tree used for word division, the above-mentioned priority attribute is determined for a large amount of mail data with mail address information, and a decision tree including the priority attribute is created.

In the process 103, the path having the maximum appearance probability can be selected by using the Viterbi algorithm.

In the above processing, a word having a high priority is not always selected, but if other attributes are the same,
There is a high possibility that a word having a higher priority attribute value will be selected.

FIG. 12 is a diagram for explaining the progress of the process of distinguishing between homonymous words by word search using the mail address information given to the mail and the user dictionary. For processing, each database of Tables 1, 2, and 3 is used. The input sentence is "Overriding is a problem."

From the address given to the mail, the mail address list for user dictionary search is changed to {Climbing group (-1),
Private related group (-2), common group (-3)}.

On the other hand, in the user dictionary with a mail address in Table 1, two words, "Norikoshi" and "Norikoshi", are registered as homonyms of "Norikoshi". The address of the mountain climbing group is assigned to “Choking (Nokkoshi)”, and the address of the common group is assigned to “Choking (Norikoshi)”.

The priority in the user dictionary search mail address list is -1 for the climbing group and -3 for the common group. The value of the priority as the attribute of the word is "-1" for "override" and "-3" for "override". Using this attribute and other attributes, the appearance probability of two words is calculated. Both are words in the user dictionary. In this sentence, the difference between ordinary nouns and sa-variable nouns does not affect the word selection. Therefore, considering the method of creating a decision tree, “Norikoshi” has a higher priority. Is more likely to be greater. Therefore,
In such a case, “override” is selected.

Further, through processing for determining an accent position, a phrase start position, and a pause position, an intermediate language,
"P Nokkoshiga, Mondida." Is generated. In the case of this sentence, there is no particular clue for reading before and after the “override”. Therefore, it is unlikely that reading is correctly performed with attributes other than the priority in the present embodiment. For example,
In the case where the attribute of the priority according to the present embodiment is not used and the frequency of the word is used as the attribute, it is highly possible that “Norikoshi”, which has a high frequency in a general sentence, is selected.

As described above, the first embodiment has the following advantages. (1) The range of use of words registered by the user can be limited, and reading errors due to unexpected side effects can be reduced. (2) Since the field of a sentence in which a word registered as a user word is used is limited, it is less necessary to consider an adverse effect when registering a word in a user dictionary, and the labor for word registration can be reduced. (3) Since the text-to-speech is read based on the prosody and voice quality in consideration of the sender, destination, and copy destination of the mail, the requirements of the mail can be efficiently heard.

<Second Embodiment><Structure> In this embodiment, “group”, “prosody
The definition of the "voice quality setting mail address" is the same as in the first embodiment. FIG. 13 is a diagram illustrating the configuration of the second embodiment. Reference numeral 201 denotes a user interface for operating the speech synthesizer, 202 denotes a text-to-speech conversion unit that reads out the contents of the mail received by the mail transmission / reception unit 214, and 203 denotes reading of input text from the received mail, The sender / destination / copy / destination information to the text-to-speech conversion information setting unit, and
This is a text input unit that transfers one piece of information returned from the processing unit to a subsequent processing unit.

Reference numeral 204 denotes an intermediate language (reading
Accent position, phrase start position, pause position)
Is a system word dictionary for storing word information (notation, reading, part of speech, accent type, accent combination type, etc.) used in the text analysis, and 206 is a word analysis unit for storing word information added by the user. It is a user word dictionary to be stored. Reference numeral 207 denotes a reading separation processing unit that performs reading separation for homomorphic words.
This is a reading determination list storage unit that stores a reading determination list used for reading homomorphic words. The decision list is a list of rules for determining a class D based on a certain evidence E in the order of priority, in the form of a list.
At the time of application, application is tried in order from the rule with the highest priority (see Document 2). In the present embodiment, reading is used as the class of the decision list, the ratio of appearance frequencies (likelihood ratio) is used as the priority, and the terms "reading" and "likelihood ratio" are also used below.
The reading determination list is created from a corpus and stored in advance for each homomorphic word.

A synthesis parameter generation unit 209 generates synthesis parameters (synthesis unit, duration, pitch, pause, amplitude) from the output of the text analysis.
Is a prediction table storage unit that stores a prediction table used for determining a synthesis parameter. Reference numeral 211 denotes a speech synthesis unit that generates a waveform from synthesis parameters.
A speech unit dictionary 12 stores a speech unit set used in speech synthesis. In the prediction table storage unit 210,
A duration prediction table, a pitch prediction table, a pause prediction table, and an amplitude prediction table are stored.

Reference numeral 213 denotes a mail management unit for sending / receiving and storing mails; 214, a mail sending / receiving unit for sending / receiving mails; and 215, a mail storage unit for storing received mails.

Reference numeral 216 denotes a database management unit for managing a database related to the mail address.
Is an e-mail address management unit that manages an e-mail address database and an e-mail address hierarchy database. 218 is an e-mail address database with attributes that stores information on e-mail addresses and groups. It is a mail address hierarchy database which stores a hierarchy relationship.
Reference numeral 220 denotes a text-to-speech conversion information setting unit that determines one piece of information to be used for text-to-speech conversion from the sender, destination, and copy / destination information attached to the mail. The text-to-speech conversion information to be determined is the same information as the “prosodic / voice-quality setting mail address” of the first embodiment. In Example 2, the purpose of the determined information is different.
The "voice quality setting mail address" is referred to as a "reading mail address".

[0061]

[Table 7] Table 7 shows a part of the contents of the user word dictionary. Unlike the first embodiment, the words in the user dictionary do not include information on the mail address or the group.

[0062]

[Table 8] Table 8 shows a part of the content of the decision list for reading the homomorphic word “market”. In the conventional device, evidence in the input sentence such as "notation of preceding and following words" (notation of independent words and some attached words within 10 words before and after) is used, but in the present embodiment, evidence other than the input sentence is used in the present embodiment. , Using the attribute of the mail address database. In Table 8, Rule 3 is a rule using “business type of the mail address database”. Further, in order to set a default value, a rule with the type of evidence “default” is added as a rule having the lowest likelihood. In Table 8, Rule 8 is a rule for setting a default value.

[0063]

[Table 9] Table 9 shows the contents of the attributed mail address database. Information such as mail address or group name / type, information such as name / telephone number / affiliation, and attributes used by a user in a decision list for reading distinction are added. Attributes that can be easily set by the user, such as “business type” and “field”, are used. The type of the attribute and the attribute value are prepared in advance, and the user selects from the attribute values displayed by the device.

<Operation> In the following description, an operation is performed so as to maintain data consistency between databases even if not particularly described. Since a user can add, modify, or delete a mail address or a group, the process of losing consistency is always checked and eliminated, and data consistency of the attributed mail address database 218 and the mail address hierarchy database 219 can be maintained. To do.

The user adds / deletes / changes the contents of the attribute-added mail address database 218 and the contents of the mail address hierarchy database 219 by calling the mail address management section 217 via the user interface 201. The user interface 20
The e-mail is transmitted / received by calling the e-mail transmission / reception unit 214 via the communication unit 1. A copy of the received mail or the transmitted mail is stored in the mail storage unit 215.

When the user performs a text-to-speech conversion of an e-mail, the user specifies the e-mail stored in the e-mail storage unit via the user interface 201 and calls the text-to-speech conversion unit 202. In the text-to-speech conversion unit, first, the text input unit 203 is connected to the mail storage unit 215.
Read mail from. Next, the text-to-speech conversion information setting unit 220 is called, and a reading mail address is obtained from the destination, the sender, and the copy destination of the read mail, and passed to the subsequent processing unit. The method of determining the mail address for reading distinction is the same as the method of determining the prosody / voice quality setting mail address of the first embodiment.

After the processing of the text-to-speech information setting unit 220 is completed, the text analysis unit 204 extracts words that are candidates for the words to be divided from the system word dictionary 205 and the user word dictionary 206, and converts the text of the mail into words. To divide. At this point, after calling the reading division processing unit 207 and performing reading of homomorphic words, reading, accent position,
Determine the intonation launch position / pause position,
Generate an intermediate language.

FIG. 14 is a processing flow of the read classification processing unit. The reading process is called after text analysis divides an input sentence into words. The read separation processing unit performs processing 1
At 21, the mail address for reading is searched for in the mail address database with attributes, and the attribute is extracted. Processing 1
At 22, the words that the text analysis unit has divided into words are:
The scanning position is set to the first word. From step 123 to step 126, scanning is performed one word at a time toward the end of the sentence, and when the scanning point moves to the position following the last word of the sentence, the processing ends according to the determination in step 123. In step 124, a search is made to determine whether a reading list for the word being scanned is present in the reading list storage unit 208. If there is a decision list, the process 125 checks in the order of the likelihood of the rule whether the condition of the type of evidence and the value of the evidence in the decision list satisfies, and if satisfied, applies the rule. Since there are rules for default setting, reading is always determined. Processing 12
At 6, the scan point is moved to the next word and the process returns to step 123.

FIG. 15 shows a specific example of the homomorphic word reading process. It is assumed that there is a reading determination list in Table 8 and an attributed mail address database in Table 9. Suppose that the input sentence "Which market do you want to check?" The text analysis unit divides the words into words as shown in the figure. At this point, the reading has also been determined. In the text-to-speech conversion information setting unit 220, the read mail address is kato @ a
ozora-bank. co. jp. The reading division unit scans one word at a time from the beginning of the sentence, where n = 1 and n =
Regarding No. 2, since there is no decision list in process 124,
The scanning point moves. Consider the state in which N = 3 in process 126. The third word “market” satisfies the condition of the process 124 because there is the decision list in Table 8, and executes the process 125.

Referring to the attributed e-mail address database in Table 9, the "business type" is "finance". The words before and after include the words “where”, “no”, “wo”, “investigation”, “mas”, “ka”, and “?”. FIG. 15 shows a list of types and values of evidence for searching the decision list. It is checked whether the condition of the type of evidence and the value of evidence is satisfied in order from rule 1. As for rules 1 and 2, there are no words "stock" and "share" before and after "market", so rules 1 and 2 do not satisfy the condition,
Not applicable. As for rule 3, there is evidence that the type is "business type of the mail database with attributes" and the value is "finance", and the rule is applied to satisfy the condition of the rule. Therefore, the result of reading “market” is the same as that in Rule 3.
Will be given by the “Shijo”. Further, the processing 126, the processing 123, and the processing 124 are repeated, and when n = 9, there is no corresponding word, and thus the read-separation processing ends.

After the processing of the reading division unit 207 is completed, the text analysis unit 204 determines the reading, accent position, intonation start position, and pause position, and generates an intermediate language. The synthesis parameter generation unit 209 determines parameters for the duration, pitch, pause length, and amplitude of the phoneme for the intermediate language output by the text analysis unit. The speech synthesis unit 211 synthesizes speech from the intermediate language output by the text analysis unit and the parameters output by the synthesis parameter unit.

As described above, according to this embodiment, since the attribute obtained from the mail address is used for reading, even if there is no clue in the text, reading can be performed.

Note that the present invention is not limited to the above-described embodiment, and the first and second embodiments are used for reading out an electronic mail, but may be used for reading out a home page. In the case of a homepage, the address of the homepage, no copy destination, and the mail address of the user are used as the sender and destination, respectively.

[0074]

As described above in detail, in the speech synthesizer of the first invention, means for registering a word used for text analysis of a received mail, means for storing a mail address database, mail A voice synthesizing apparatus comprising: means for adding, modifying, and deleting the contents of an address database; a means for grouping mail addresses in the database and storing each group as hierarchical data; Means for adding a data node name to the user word dictionary and registering the data in a user word dictionary; and means for creating a list of node names of the hierarchical data from information on the sender, destination, and copy destination of the mail. At the time of searching, the members of the created list and the node names added to the words in the user word dictionary are Since the text analysis is performed using only the words whose node names match any of the members of the list, the use range of the words registered by the user can be limited, and unexpected Reading errors due to side effects can be reduced. In addition, since the field of a sentence in which the word registered as the user word is used is limited, it is less necessary to consider an adverse effect when registering the word in the user dictionary, and the labor for word registration can be reduced.

Further, in the speech synthesizer of the second invention, means for controlling the prosody or voice quality when reading the received mail, means for storing the mail address database, and adding / modifying the contents of the mail address database A voice synthesizing apparatus having means for deleting the mail, grouping the mail addresses of the database, storing each group as hierarchical data in a hierarchy, and adding information on the sender, destination, and copy destination of the mail; Means for selecting a node of the hierarchical data based on the
When reading out the mail, the mail is read out according to the prosody and voice quality associated with the selected node. Therefore, it is possible to efficiently listen to the requirements of the mail.

Further, in the speech synthesizing apparatus according to the third aspect of the present invention, means for distinguishing homonymous words using the decision list, means for storing a mail address database,
Means for adding / modifying / deleting the contents of a mail address database, wherein the mail addresses of the database are grouped,
Means for storing each group as hierarchical data, and means for selecting a node name of the hierarchical data from information on the sender, destination, and copy destination of the mail. Since the words are distinguished by using the decision list including the rule with the attribute as evidence, the distinction is possible even when there is no clue in the sentence.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a first embodiment.

FIG. 2 is a diagram showing an example of a mail address hierarchy database.

FIG. 3 is a main routine of a text-to-speech conversion information setting process.

FIG. 4 is a subroutine 1 of text-to-speech conversion information setting processing.

FIG. 5 is a subroutine 2 of a text-to-speech conversion information setting process.

FIG. 6 is a subroutine 3 of text-to-speech conversion information setting processing.

FIG. 7 is a subroutine 4 of text-to-speech information setting processing.

FIG. 8 is a diagram showing an example of an address and text-to-speech conversion information given to a mail address.

FIG. 9 is a diagram illustrating an example of an input sentence and an intermediate language.

FIG. 10 is a processing flow from dictionary lookup to word division.

FIG. 11 is a diagram showing a graph structure generated in processing 101 of FIG. 10;

FIG. 12 is an explanatory diagram showing the progress of a reading process using a mail address and a user word dictionary.

FIG. 13 is a configuration diagram of a second embodiment.

FIG. 14 is a process flow of a read-separation process.

FIG. 15 is an explanatory diagram of a reading process of homomorphic words “market”.

[Explanation of symbols]

 101, 201 User interface 102, 202 Text-to-speech conversion unit 103, 203 Text input unit 104, 204 Text analysis unit 105, 205 System word dictionary 106, 206 User word dictionary with mail address 107, 209 Synthesis parameter generation unit 108, 210 Prediction Table storage units 109, 211 Speech synthesis units 110, 212 Speech unit dictionaries 111, 213 E-mail management units 112, 214 E-mail transmission and reception units 113, 215 E-mail storage units 114, 216 Database management units 115, 217 E-mail address management units 116, 218 E-mail address database 117, 219 E-mail address hierarchy database 118 User word dictionary management unit 119 Prosody and voice quality setting database management unit 120 Prosody and voice quality setting database 121 prosody-prediction table corresponding database 122 voice-segment set corresponding database 123 prosody-voice quality setting database 207 diacritical 208 diacritical deciding list storage unit 220 text-to-speech conversion information setting unit

Claims (3)

[Claims] 1. A voice having means for registering a word used for text analysis of a received mail, means for storing a mail address database, and means for adding / modifying / deleting the contents of the mail address database. In the synthesizing device, a means for grouping the mail addresses of the database and storing each group as hierarchical data, a means for adding a node name of the hierarchical data to a word and registering the word in a user word dictionary, Means for creating a list of node names of the hierarchical data from information on the sender, destination, and copy destination of the
When searching the user word dictionary, the members of the created list are compared with the node names added to the words in the user word dictionary, and only the words whose node names match any of the members of the list are used. A speech synthesizer characterized in that text analysis is performed by using a speech synthesizer. 2. A speech synthesizer comprising: means for controlling the prosody or voice quality when reading a received mail; means for storing a mail address database; and means for adding, modifying, or deleting the contents of the mail address database. In the apparatus, means for grouping the mail addresses of the database and storing each group as hierarchical data in a hierarchy, based on information of a mail sender, a destination, and a copy destination
Means for selecting a node of the hierarchical data, wherein when reading out the mail, the mail is read out according to the prosody and voice quality associated with the selected node. 3. A means for reading homomorphic words using a decision list, a means for storing a mail address database, and the contents of a mail address database.
A voice synthesizing apparatus comprising: means for correcting / deleting; a means for grouping mail addresses of the database and storing each group as hierarchical data; and information on a sender, a destination, and a copy destination of the mail. Means for selecting a node name of the hierarchical data from the list, and performs the read-out of words using a decision list including a rule using an attribute set in the selected node as evidence. Synthesizer.