JP3626398B2 - Text-to-speech synthesizer, text-to-speech synthesis method, and recording medium recording the method - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a text-to-speech synthesizer that generates speech from text, a text-to-speech synthesis method, and a recording medium that records the method.
[0002]
[Prior art]
This type of conventional apparatus is configured, for example, as shown in FIG. In FIG. 5, the text analyzer 102 inputs text information (for example, “left”) indicating a word or sentence mixed with Japanese kanji or kana from the input terminal 101, and reading information ( For example, the phoneme string “hidari”) is output to the prosody generator 103 and the speech element selector 104. Note that reading information indicating reading of a word, a sentence, or the like may be input to the input terminal 101 and supplied to the prosody generator 103 and the speech element selector 104. In this case, the text analyzer 102 is omitted.
[0003]
When the prosody generator 103 receives the reading information, the prosody generator 103 generates prosody information from the reading information and outputs the prosodic information to the speech synthesizer 106. This prosodic information indicates the height and size of speech, or the duration. The pitch of the voice is set by the pitch frequency of the voice waveform indicating the vowel phoneme. For example, since the phonemes of the vowels in the phoneme string “hidari” are “i”, “a”, and “i”, the pitch frequency of the speech waveform indicating the phonemes of these vowels is set. The magnitude and duration of the speech are set for each phoneme as the amplitude of the speech waveform indicating the phoneme and the duration of the speech waveform indicating the phoneme. For example, the amplitude and duration of the speech waveform are set for each phoneme “h”, “i”, “d”, “a”, “r”, and “i”.
[0004]
When the speech element selector 104 receives reading information, the speech element selector 104 extracts a speech unit composed of at least one phoneme from the phoneme string that is the reading information.
[0005]
As a voice unit, there are consonant + vowel (CV: Consonant, Vowel) or vowel + consonant + vowel (VCV). The former CV audio unit is, for example, “ka” or “gu”. The latter VCV speech unit retains the characteristics of the transitional part of the phoneme chain in order to improve speech quality, and is, for example, “aki” or “ito”.
[0006]
Here, VCV audio units are used. In this case, the speech segment selector 104 extracts each VCV speech unit “* hi”, “ida”, “ari”, and “i **” (* is a silence symbol) from the phoneme string “hidari”. To do. Then, the speech unit selector 104 searches the speech unit database 105 for each speech unit data corresponding to these speech units, and outputs these speech unit data to the speech synthesizer 106.
[0007]
In the speech unit database 105, a large number of speech unit data is registered in advance. These speech segment data are obtained by appropriately cutting out speech data indicating an announcer's speech for each speech unit, and further converting the speech data into a format suitable for speech synthesis. The number of VCV speech units used for general Japanese text-to-speech synthesis is about 800, and each speech unit data corresponding to these speech units is registered in the speech unit database 105 in advance.
[0008]
When the speech synthesizer 106 receives the prosodic information from the prosody generator 103 and the speech unit data from the speech unit selector 104, the speech synthesizer 106 receives the speech information indicated by the speech unit data based on the prosodic information. After adjusting the height, size, and duration, voice data is formed by smoothly connecting these voices in each vowel section, and this voice data is output from the output terminal 107.
[0009]
By the way, in such a device, one speech unit data is registered for one speech unit. If any word includes the same speech unit, the same speech unit data is stored. One piece of data is used. For this reason, even if the speech unit data is appropriate for a speech unit included in one word, it may be inappropriate for the same speech unit included in another word. That is, when voice data is converted into voice, the former word can be heard clearly, but the latter other words cannot be heard.
[0010]
For example, in the “speech synthesizer” described in Japanese Patent Laid-Open No. 2-205896, a proper noun such as a person name is distinguished from a normal sentence, a speech unit data group for proper nouns, A speech unit data group for text is registered, and two speech unit data groups are used properly. As a result, proper nouns are pronounced clearly, and normal sentences are pronounced smoothly.
[0011]
[Problems to be solved by the invention]
However, in the above conventional apparatus, the same speech segment data is used for any proper noun as long as the same speech unit is included. For example, since “yamaguchi” and “taguchi” contain the same speech unit “agu”, the same speech segment data corresponding to the speech unit “agu” is used for these proper nouns. . In this case, for example, the pronunciation of one “yamaguchi” becomes uncharacteristic, and it becomes difficult to distinguish between the pronunciation of “yamaguchi” and the pronunciation of another similar “yamaguchi”.
[0012]
In addition, since a dedicated speech segment data group is registered for proper nouns, the amount of data to be registered is approximately doubled, which is enormous.
[0013]
Therefore, the present invention has been made in view of the above-described conventional problems, and clearly pronounces any word while minimizing the increase in the amount of pre-registered speech segment data. An object of the present invention is to provide a text-to-speech synthesizer, a text-to-speech synthesis method, and a recording medium on which the method is recorded.
[0014]
[Means for Solving the Problems]
In order to solve the above problems, the present invention includes a speech unit database in which each speech unit data corresponding to each speech unit composed of at least one phoneme is registered, and a phoneme string indicating a reading of a word or a sentence. Based on the speech unit database, and at least one speech unit data is retrieved, and based on the retrieved speech unit data and the prosodic information of the word or sentence, the text speech synthesizer for synthesizing the speech of the word or sentence, When synthesizing a speech of a specific word and an exceptional speech segment database in which exceptional speech segment data corresponding to a specific speech unit consisting of at least one phoneme included in a phoneme sequence indicating a specific word reading is registered For a specific speech unit, instead of searching the speech unit database for speech unit data corresponding to this speech unit, And a speech data retrieving means for retrieving exception speech unit data from the exception speech unit database.
[0015]
According to the present invention having such a configuration, the speech unit data corresponding to the speech unit is normally retrieved from the speech unit database and used. Further, when synthesizing the speech of a specific word, the exceptional speech segment data is searched from the exceptional speech segment database and used for a specific speech unit. Thereby, the voice of this specific word becomes clear.
[0016]
In the apparatus of the present invention, a vocabulary list in which a plurality of words are registered, and each word similar in reading from the vocabulary list are selected, and at least from a phoneme string indicating at least one reading of each selected word, Similar word extraction means for extracting a specific speech unit composed of one phoneme and creating an exceptional speech segment table for retrieving exceptional speech segment data corresponding to the extracted specific speech unit from the exceptional speech segment database And.
[0017]
Here, a specific speech unit is extracted from at least one of the words whose readings are similar, and the exceptional speech for retrieving exceptional speech segment data corresponding to the specific speech unit from the exceptional speech segment database Create a fragment table. As a result, the exceptional speech unit table can be automatically created. When synthesizing the speech of a word from which a specific speech unit has been extracted, the exceptional speech unit data corresponding to the specific speech unit is obtained by searching the exceptional speech unit table, and the exceptional speech unit database is retrieved from the exceptional speech unit database. Search for piece data. By synthesizing the speech of a word using this exceptional segment data, it is possible to clearly distinguish this word from other similar words.
[0018]
Furthermore, in the apparatus of the present invention, the exceptional speech segment data in the exceptional speech segment database indicates speech in which consonants are emphasized.
[0019]
As a result, the consonant of a specific word is emphasized, and the specific word is not mistaken for another similar word.
[0020]
On the other hand, the present invention uses a speech unit database in which each speech unit data corresponding to each speech unit consisting of at least one phoneme is registered, and based on a phoneme string indicating a word or sentence reading, In a text-to-speech synthesis method for retrieving at least one speech unit data from a unit database and synthesizing speech of a word or sentence based on the retrieved speech unit data and prosodic information of the word or sentence, When using an exceptional speech segment database in which exceptional speech segment data corresponding to a specific speech unit consisting of at least one phoneme included in a phoneme sequence indicating reading is registered, and synthesizing speech of a specific word, For a specific speech unit, instead of retrieving speech unit data corresponding to this speech unit from the speech unit database, this speech unit is supported. We are searching for exception speech segment data from the exception speech unit database.
[0021]
According to the method of the present invention, a specific word is pronounced clearly as in the apparatus of the present invention.
[0022]
Further, in the method of the present invention, a vocabulary list in which a plurality of words are registered is used, each word similar in reading is selected from the vocabulary list, and a phoneme string indicating at least one reading of each selected word Then, a specific speech unit composed of at least one phoneme is extracted, and an exceptional speech unit table is created for searching the exceptional speech unit data corresponding to the extracted specific speech unit from the exceptional speech unit database. Yes.
[0023]
Here too, as in the apparatus of the present invention, the exceptional speech unit table can be automatically created.
[0024]
Furthermore, in the method of the present invention, the exceptional speech segment data in the exceptional speech segment database indicates speech in which consonants are emphasized.
[0025]
In this case as well, as with the device of the present invention, it is not necessary to mistake a specific word for other similar words.
[0026]
Furthermore, the present invention also includes a recording medium that records the text-to-speech synthesis method as a program.
[0027]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.
[0028]
FIG. 1 is a block diagram showing a first embodiment of a text-to-speech synthesizer according to the present invention. As shown in FIG. 1, a text-to-speech synthesizer 10 according to this embodiment inputs text information through an input terminal 11 and generates a text information 12 based on the text information. 12, a prosody generator 13 that generates prosody information based on the reading information, a speech unit database 14 in which speech unit data groups are registered, and an exceptional speech unit in which exceptional speech unit data groups are registered. A database 15, a speech unit database switch 16 that selectively uses the speech unit database 14 and the exceptional speech unit database 15, an exceptional speech unit table 17 that is referred to by the speech unit database switch 16, and reading information Is input from the text analyzer 12, and at least one speech segment data is converted into speech segments based on the reading information. The speech unit selector 18 for searching from the database 14 or the exceptional speech unit database 15 and the prosody information are input from the prosody generator 13, and the speech unit data is input from the speech unit selector 18. A voice synthesizer 20 that generates voice data based on the voice element data and outputs the voice data from the output terminal 19 is provided.
[0029]
In the speech unit database 14, speech unit data corresponding to a large number of VCV speech units are registered in advance. Here, since general Japanese text-to-speech synthesis is performed, about 800 speech segment data are registered.
[0030]
In the exceptional speech unit database 15, exceptional speech unit data corresponding to a specific speech unit included in a phoneme string of a specific word is registered in advance. For example, exceptional speech unit data corresponding to the speech unit “agu” included in the phoneme sequence “yamaguchi” of a specific word is registered. The number of exceptional speech segment data is not particularly limited.
[0031]
Further, the exceptional speech unit table 17 is for searching for each exceptional speech unit data in the exceptional speech unit database 15. For example, as shown in FIG. 2, phoneme sequences “yamaguchi”, “kishino”, “taguchi”, and “sakaguchi” for each specific word, and specific speech units “agu”, “ * Ki "(* is a silent symbol)," agu "and" agu ", and identification numbers" 1 "," 2 "," 3 "and" 1 "for identifying specific voice units. Registered in association. If the same identification number is assigned to the same speech unit, the speech units of these words correspond to the same exceptional speech unit data in the exceptional speech unit database 15. If different identification numbers are assigned to the same speech unit, these speech units correspond to the different exceptional speech unit data in the exceptional speech unit database 15. For example, since the same identification number “1” is assigned to the same speech unit “agu” of the phoneme strings “yamaguchi” and “sakaguchi” of each word, this speech unit “agu” is used as the exceptional speech unit database. 15 corresponding to the same exceptional speech unit data in 15. Also, since different identification numbers “1” and “3” are assigned to the same speech units “agu” of the phoneme strings “yamaguchi” and “taguchi” of each word, the speech units of these words “Agu” corresponds to each different exceptional speech unit data in the exceptional speech unit database 15.
[0032]
In the exceptional speech unit table 17, the number of exceptional speech unit data corresponding to one speech unit is not limited. In addition, since each speech unit data corresponding to all speech units is registered in the speech unit database 14, each speech unit data corresponding to each speech unit "agu" and "* ki" is registered. Is also registered.
[0033]
For example, when “Yamaguchi-san” is input as text information to the text analyzer 12 through the input terminal 11, the text analyzer 12 reads the reading information of “Yamaguchi-san”, that is, the phoneme string “yamaguchiisaN”. And output to the speech unit selector 18.
[0034]
When the prosody generator 13 receives this phoneme string “yamaguchisaN”, it obtains the speech height, size, and duration for each phoneme of this phoneme sequence, and the speech height, size, and duration of each phoneme. Prosody information indicating time is generated, and this prosodic information is output to the speech synthesizer 20. FIG. 3 is a chart illustrating this prosodic information. In this chart, the pitch indicating the phoneme duration and the phoneme speech height corresponding to each phoneme “y”, “a”, “m”, “a”, “g”, “u”,. A frequency (Hz) and a loudness level (dB) are set. The pitch of the phoneme is set only when the phoneme is a vowel, and is not set when the phoneme is a consonant. Therefore, the pitch of the voice is determined by the pitch frequency of the vowel.
[0035]
Next, when the phoneme segment selector 18 inputs the phoneme string “yamaguchiisaN”, each voice unit “* ya”, “ama”, “agu”, “uchi”, “isa”, and “isa” of the VCV is input from this phoneme string. Extract “aN *”. Further, the speech element selector 18 gives the phoneme string database “yamaguchi” to the speech element database switcher 16.
[0036]
The speech segment database switcher 16 refers to the exceptional speech segment table 17 shown in FIG. 2 and searches for a phoneme string “yamaguchi” of a word, and a specific speech unit “agu” and an identification number corresponding to this word. Read “1”. When the speech unit selector 18 searches for a specific speech unit “agu”, the speech unit database switcher 16 gives the exceptional speech unit database 15 to the speech unit selector 18. The speech unit selector 18 includes exceptional speech unit data to which an identification number 1 is assigned from among a plurality of exceptional speech unit data corresponding to a specific speech unit “agu” in the exceptional speech unit database 15. Select. When the speech unit selector 18 searches for other speech units “* ya”, “ama”, “uchi”, “isa”, and “aN *”, the speech unit database switch 16 Gives the speech unit database 14 to the speech unit selector 18. The speech unit selector 18 searches each speech unit data corresponding to each speech unit “* ya”, “ama”, “uchi”, “isa”, and “aN *” in the speech unit database 14. To do.
[0037]
Thus, the exceptional speech unit data corresponding to the specific speech unit “agu” and the identification number 1, and the speech units “* ya”, “ama”, “uchi”, “isa”, and “aN *”, respectively. Speech unit data is retrieved, and the exceptional speech unit data and each speech unit data are supplied to the speech synthesizer 20.
[0038]
When the speech synthesizer 20 receives the prosodic information shown in FIG. 3 from the prosody generator 13 and the exceptional speech unit data and each speech unit data from the speech unit selector 18, the speech synthesizer 20 receives the exceptional speech based on the prosodic information. After adjusting the height, volume, and duration of each voice indicated by the segment data and each voice segment data, voice data is formed by smoothly connecting these voices in each vowel section. The audio data is output from the output terminal 19.
[0039]
As described above, in the present embodiment, each speech unit data corresponding to each basic speech unit is registered in advance in the speech unit database 14, and exceptional speech corresponding to a specific speech unit included in a specific word. When segment data is registered in advance in the exceptional speech unit database 15 and speech data of a specific word is synthesized, the exceptional speech unit data corresponding to a specific speech unit included in the specific word is converted into the exceptional speech. It is retrieved from the segment database 15 and used. For this reason, when the voice data indicating the specific word is converted into voice, the specific word can be clearly heard. For example, as is apparent from FIG. 2, the speech unit “agu” of “yamaguchi” and the speech unit “agu” of “taguchi” are distinguished from each other, and each speech segment data is used. Both become characteristic, and furthermore, it becomes possible to distinguish between the pronunciation of “yamaguchi” and the pronunciation of other similar “yamaguchi”.
[0040]
Further, here, proper nouns are illustrated as specific words, but any other kind of words can be treated as specific words. For this reason, if the contents of the exceptional speech unit table 17 and the content of the exceptional speech unit database 15 are appropriately set, it is possible to synthesize speech data in which similar nouns are pronounced differently.
[0041]
FIG. 4 is a block diagram showing a second embodiment of the text-to-speech synthesizer of the present invention. In FIG. 4, the same reference numerals are given to portions that perform the same functions as those of the apparatus 10 of FIG. 1, and the description is simplified.
[0042]
In the text-to-speech synthesizer 30 according to the present embodiment, an utterance vocabulary list 31 and a similar word extractor 32 are added to the apparatus 10 of FIG. Accordingly, as in the apparatus 10 of FIG. 1, text information is analyzed by the text analyzer 12, prosody information is generated by the prosody generator 13, and speech units in the speech unit database 14 are generated by the speech unit selector 18. The exceptional speech unit data in the data and exceptional speech unit database 15 is searched, and the speech synthesizer 20 synthesizes speech data.
[0043]
In the utterance vocabulary list 32, for example, text information indicating a large number of names is registered in advance. The similar word extractor 32 compares these person names with each other and extracts two words having a high degree of similarity, that is, two sufficiently similar words.
[0044]
As a method of extracting two words having high similarity, for example, two words are converted into respective phoneme strings, and these phoneme strings are compared, and only one phoneme is a difference between each phoneme string. When only one phoneme is inserted into one phoneme string or when only one phoneme is dropped from one phoneme string, these words are regarded as having high similarity and are extracted. .
[0045]
As another method, each spectral parameter corresponding to all phonemes is registered in advance, and by referring to these spectral parameters, each spectral time series is generated from a phoneme sequence of two words, After these spectral time series are matched using the DP matching method widely used in the field of speech recognition, for example, the acoustic distance between words (more specifically, the spectral distance, etc.) is calculated. If this distance is close, it is considered that the similarity of these words is high, and these words are extracted.
[0046]
Next, the similar word extractor 32 obtains a characteristic speech unit of one of the two words having high similarity, that is, a difference between the words. For example, when one phoneme is inserted in the phoneme string of one word, the speech unit including the inserted phoneme is characteristic. Further, when the acoustic distance between the words is short, the sound unit of one word having the longest acoustic distance between the words is characteristic.
[0047]
For example, if each word of the phoneme sequence “yamaguchi” and the phoneme sequence “yamauchi” is extracted with a high degree of similarity, the phoneme “g” is inserted into one phoneme sequence “yamaguchi”. The speech unit “agu” including the phoneme “g” is obtained as a characteristic.
[0048]
After obtaining the characteristic speech unit of one of the two words having a high degree of similarity in this way, the similar word extractor 32 associates the word including the characteristic speech unit with the speech unit to generate an exceptional speech unit. Register in the single table 17.
[0049]
As described above, in this embodiment, two words having a high similarity are extracted from the utterance vocabulary list 32, and the exceptional speech unit table 17 is automatically associated with one word including a characteristic speech unit and the speech unit. Can be registered automatically. For this reason, the trouble for registering the contents of the exceptional speech unit table 17 can be saved. When synthesizing speech data of a word registered in the exceptional speech unit table 17, speech units included in the word are searched from the exceptional speech unit table 17, and exceptional speech unit data corresponding to the speech unit is obtained. It is read from the exceptional speech unit database 15 and the speech data of this word is synthesized using this exceptional speech unit data.
[0050]
Here, the speech unit data corresponding to the speech unit expected to be registered needs to be registered in the exceptional speech unit database 15 in advance. In addition, it is preferable that the exceptional speech segment data indicates a speech in which a consonant is emphasized. When speech data of a word is synthesized using speech segment data in which the consonant is emphasized, the consonant of the word is emphasized when the speech data is converted into speech. This makes it possible to clearly distinguish the voice of this word from the voice of other similar words.
[0051]
In this embodiment, a large number of names are registered in the utterance vocabulary list 32. However, a large number of place names and nouns or other types of words are registered in the utterance vocabulary list 32. You may extract each similar word from.
[0052]
Also, for other words that are not registered in the exceptional speech unit table 17, the speech data is generated using the basic speech unit data in the speech unit database 14, so that the pronunciation of the sentence can be reduced. There is no unnaturalness.
[0053]
In addition, this invention is not limited to the said embodiment, It can deform | transform variously. For example, although a VCV audio unit is illustrated as an audio unit, a CV audio unit or another type of audio unit may be applied.
[0054]
Alternatively, reading information indicating reading of a word or a sentence may be input to the input terminal 11 and supplied to the prosody generator 13 and the speech unit selector 18. In this case, the text analyzer 12 is omitted.
[0055]
Furthermore, the present invention also includes a recording medium that records the text-to-speech synthesis method as a program.
[0056]
【The invention's effect】
As described above, according to the present invention, normally, speech unit data corresponding to a speech unit is retrieved from a speech unit database and used. Further, when synthesizing the speech of a specific word, the exceptional speech segment data is searched from the exceptional speech segment database and used for a specific speech unit. This makes the voice of this particular word clear.
[0057]
Further, according to the present invention, a specific speech unit is extracted from at least one of words having similar readings, and exceptional speech unit data corresponding to the specific speech unit is searched from the exceptional speech unit database. An exception speech unit table is created to do this. As a result, the exceptional speech unit table can be automatically created. When synthesizing the speech of a word from which a specific speech unit has been extracted, the exceptional speech unit data corresponding to the specific speech unit is obtained by searching the exceptional speech unit table, and the exceptional speech unit database is retrieved from the exceptional speech unit database. Search for piece data. By synthesizing the speech of a word using this exceptional segment data, it is possible to clearly distinguish this word from other similar words.
[0058]
Further, according to the present invention, since the consonant of a specific word is emphasized, it is not necessary to misunderstand this word with other similar words.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a first embodiment of a text-to-speech synthesizer according to the present invention.
FIG. 2 is a diagram illustrating a configuration of an exceptional speech unit table in the apparatus of FIG. 1;
FIG. 3 is a chart showing the contents of prosodic information.
FIG. 4 is a block diagram showing a second embodiment of the text-to-speech synthesizer of the present invention.
FIG. 5 is a block diagram illustrating a conventional speech synthesizer.
[Explanation of symbols]
10,30 Text-to-speech synthesizer
11 Input terminal
12 Text analyzer
13 Prosody generator
14 Speech segment database
15 Exceptional speech unit database
16 Speech unit database switcher
17 Exceptional speech unit table
18 Speech segment selector
19 Output terminal
20 Speech synthesizer
31 Voice vocabulary list
32 Similar word extractor

Claims (7)

A speech unit database in which each speech unit data corresponding to each speech unit composed of at least one phoneme is registered, and at least one speech from the speech unit database is based on a phoneme sequence indicating reading of words and sentences. In a text-to-speech synthesizer that searches for segment data and synthesizes speech of words and sentences based on the searched speech unit data and prosodic information of words and sentences.
Exceptional speech segment data corresponding to a specific speech unit composed of at least one phoneme included in a phoneme string indicating the reading of a specific word, and the exceptional speech associated with the specific word and the specific speech unit An exceptional speech segment database registered to be identified by the segment data identification number ,
When synthesizing speech of a specific word, for a specific speech unit, instead of retrieving speech segment data from the speech segment database, an exceptional speech unit having an identification number corresponding to the specific word and the specific speech unit is used. A text-to-speech synthesizer comprising: speech data search means for searching for piece data from an exceptional speech unit database. A vocabulary list with multiple words registered,
Select each word with similar reading from the vocabulary list, extract a specific speech unit consisting of at least one phoneme from the phoneme string indicating the reading of at least one of the selected words, and extract the specific speech unit 2. The text-to-speech synthesizer according to claim 1, further comprising: a similar word extracting unit that creates an exceptional speech unit table for searching for exceptional speech unit data corresponding to 1 from the exceptional speech unit database. 3. The text-to-speech synthesizer according to claim 1, wherein the exceptional speech unit data in the exceptional speech unit database indicates speech in which consonants are emphasized. A speech unit database in which each speech unit data corresponding to each speech unit consisting of at least one phoneme is registered is used, and at least one from the speech unit database is based on a phoneme sequence indicating reading of words and sentences. In a text-to-speech synthesis method that searches for two speech unit data and synthesizes speech of a word or sentence based on the retrieved speech unit data and the prosodic information of the word or sentence.
Exceptional speech segment data corresponding to a specific speech unit composed of at least one phoneme included in a phoneme string indicating the reading of a specific word, and the exceptional speech associated with the specific word and the specific speech unit When an exceptional speech unit database registered so as to be identified by the identification number of the segment data is used. When synthesizing the speech of a specific word, the speech unit data is converted into a speech unit for a specific speech unit. A text-to-speech synthesis method, characterized in that, instead of searching from a database , exceptional speech unit data having an identification number corresponding to a specific word and a specific speech unit is searched from the exceptional speech unit database. Uses a vocabulary list in which a plurality of words are registered, selects each word that is similar in reading from the vocabulary list, and specifies at least one phoneme from a phoneme string indicating at least one reading of each selected word 5. An exceptional speech unit table for retrieving exceptional speech unit data corresponding to the extracted specific speech unit from the exceptional speech unit database is created. Text-to-speech synthesis method. 6. The text-to-speech synthesis method according to claim 4, wherein the exceptional speech unit data in the exceptional speech unit database indicates speech in which consonants are emphasized. A recording medium recording the text-to-speech synthesis method according to claim 4 as a program.

Images