JP2003005776A - Voice synthesizing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To eliminate a need of two dictionaries and to avoid selection of words having expressions which are not normally used. SOLUTION: A dictionary 3 is provided where the expressions of words, readings, accents, part of speech information, and information on the appearance frequencies of the expressions of words in an input sentence are stored. A morpheme analysis means 1 receives the input sentence and uses the dictionary 3 to output word unit divided sentences obtained by dividing the input sentence in the unit of expressions of words. A word candidate selection means 2 uses frequency information in the dictionary 3 to select a sentence divided to a set of expressions of words having the highest appearance frequency when receiving a plurality of word unit-divided sentences. A voice waveform generation means uses the dictionary 3 for individual expressions of words in the selected sentence to generate pronunciation information to which readings and accents of words are imparted, rhythm information representing the intonation and the rhythm of the entire word unit-divided sentence is generated on the basis of pronunciation information, and a voice waveform is generated on the basis of the pronunciation information and rhythm information.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizer, and more particularly to a speech synthesizer for generating a speech waveform corresponding to a kanji / kana mixed sentence.

[0002]

2. Description of the Related Art Conventionally, this type of speech synthesizer has been used to read an input Kanji / Kana mixed sentence according to correct word reading and accent, and correct intonation and rhythm.

Referring to FIG. 6, which is a block diagram of this conventional speech synthesizer, the conventional speech synthesizer stores a word notation, word reading and accent, and word part-of-speech information for each word notation. And a kanji / kana mixed sentence (input sentence) is used to create a word unit divided sentence by dividing this kanji / kana mixed sentence into word notation units, and when there are a plurality of created word unit divided sentences, A morphological analysis unit that selects a word unit divided sentence having the smallest number of independent words and outputs the selected word unit divided sentence (There are various methods for morpheme analysis performed by this morpheme analyzing unit. For example, from the beginning of a sentence to a dictionary. There is a method called "left longest match method" that matches and checks the grammatical connection relations and sequentially divides it into word sequences, which is used for morphological analysis of many Japanese speech synthesizers. ) And a word unit divided sentence corresponding to a kanji / kana mixed sentence from the morpheme analysis means, and a dictionary is used for the notation of each word in the received word unit divided sentence to read and accent the word. A voice waveform is generated based on the pronunciation information of the assigned word unit divided sentence and the prosody information indicating the intonation and rhythm of the entire word unit divided sentence generated from the pronunciation information of the word unit divided sentence. And a voice waveform generating means for performing the same.

When a morpheme analysis means divides a kanji / kana mixed sentence into word notation units by using a dictionary to create a word unit divided sentence, a notated word not in the dictionary is processed as an unknown word and read. wear. Since the improvement of the performance of this unknown word reading process is effective in reducing the reading error rate of the speech synthesis / synthesis apparatus, various studies have been conducted on unknown word processing. Particularly, since the conventional synthetic speech device assumes a kanji-kana mixed sentence in the input text, the reading error rate of a character string consisting of only kana characters is very high. In order to avoid this, for example, in the technique described in Japanese Patent Laid-Open No. 2001-5479, even if the input is a character string in which Chinese characters are not mixed, a character string in which Chinese characters are not mixed is used so that accurate reading can be performed. By using a dictionary specialized for alphanumeric characters, we are trying to maintain the reading accuracy even for character strings that consist only of kana characters. However, this method cannot handle the case where the kanji are mixed even a little.

As another method for improving the reading accuracy of a character string containing a large number of kana character strings, there is a method of describing a lot of different notations (same notations) with the same sound and the same meaning for one word. For example, "sendin", "soshin", "soshin", "sendin", and "soshin" are registered as different expressions for "send", and kanji and kana mixed sentences, hiragana sentences, katakana sentences, The reading accuracy is maintained even when a spoken sentence is an input sentence. Some of these different notations include those that are not often seen in ordinary kanji / kana mixed sentences compared to other notations. When morphological analysis is performed using a dictionary in which many different notations are registered,
As a result of analysis, it may be divided into word strings that are not often used. For example, using a dictionary in which "scale" is registered as a different notation for "scale" (noun), the sentence "even if it collides with a truck scale ..." In some cases, "is not a tentative one," but a different notation, "balance," of "scales". In this case, "scale" means "provisionally" and "scale"
Whichever is interpreted, it is correct in morphological analysis, so it is not possible to judge which one is selected only by morphological analysis. In order to divide this into correct words, we need to consider the meaning of the sentence. However, in order to perform the semantic analysis, the number of man-hours required for maintaining a dictionary and constructing a system is large, and a method for realizing the system has not been established. Therefore, under the present circumstances, the more different the notation is added to the dictionary, the more the analysis accuracy of the normal kana-kanji sentence tends to decrease, and it is an important issue to find an optimal method for selecting a word-unit divided sentence.

[0006]

In the conventional speech synthesizer described above, when a plurality of word unit divided sentences are created by the morphological analysis unit, for example, the word unit divided sentence having the smallest number of independent words is selected. Therefore, there is a problem that it is not clear whether or not the correct word unit division sentence is selected. In addition, by preparing a dictionary containing kanji and kana and an alphanumeric kana dictionary, using a kanji and kana dictionary for analysis of kanji and kana mixed sentences, and using an alphanumeric kana dictionary for sentences with only kana strings,
Although the accuracy of parsing only the kana string is maintained, it does not consider kana-kana mixed sentences. Therefore, if a large number of different notations are registered in a kanji-kana mixed dictionary, there is a problem that words of notations that are not often used may be selected.

The object of the present invention is to eliminate such drawbacks of the prior art by selecting a correct word-unit division sentence and without requiring two dictionaries having different characteristics, that is, a kanji-kana mixed dictionary and an alphanumeric dictionary. , It is to provide a speech synthesizer in which there is no fear that a word with a not-commonly used notation is selected.

[0008]

A first speech synthesizer according to the present invention receives a kanji / kana mixed sentence and divides the kanji / kana mixed sentence in word notation units and outputs the divided word-unit divided sentences. From the plurality of word unit divided sentences when there is a plurality of word unit divided sentences received by the morpheme analysis unit and the word unit divided sentence corresponding to the Kanji / Kana mixed sentence from the morpheme analysis unit, A word candidate selection unit that selects the word unit divided sentence divided into the set of the notation of the word having the highest appearance frequency, and each of the words in the word unit divided sentence selected by the word candidate selecting unit. The pronunciation and the accent of the word are given to the notation as the pronunciation information of the word unit divided sentence, and the word unit is divided based on the pronunciation information of the word unit divided sentence. Generates prosody information indicating the intonation and rhythm of the whole division statement is configured to include a speech waveform generation means for generating a speech waveform, the based on the sound information and the prosody information.

The second speech synthesizing device of the present invention includes a morpheme analysis means for receiving a kanji / kana mixed sentence and dividing the kanji / kana mixed sentence in word notation units and outputting the divided word unit divided sentences. The word unit divided sentence corresponding to the Kanji / Kana mixed sentence is received from the morpheme analysis unit, and when there are a plurality of word unit divided sentences received, the word unit divided sentence is selected from among the plurality of word unit divided sentences. The word unit divided sentence having the highest evaluation in the entire word unit divided sentence of the respective predetermined frequency information indicating the frequency of appearance in the kanji-kana mixed sentence corresponding to the notation of each of the words is selected. The word candidate selecting means and the reading and assigning of the word with respect to the notation of each word in the word unit divided sentence selected by the word candidate selecting means. Cents are given as the pronunciation information of the word-unit divided sentence, and prosodic information indicating the intonation and rhythm of the whole word-unit divided sentence is generated based on the pronunciation information of the word-unit divided sentence. And a voice waveform generating means for generating a voice waveform based on the prosody information.

The third speech synthesizing device of the present invention provides word notation, reading and accent of the word, part-of-speech information of the word, and frequency information indicating the frequency of occurrence of the notation of the word in a kanji-kana mixed sentence. A morphological analysis unit that outputs a word unit divided sentence obtained by dividing the kanji / kana mixed sentence by the dictionary stored for each notation of the word and the kanji / kana mixed sentence using the dictionary. When,
The word unit divided sentence corresponding to the Kanji / Kana mixed sentence is received from the morpheme analysis unit, and when there are a plurality of received word unit divided sentences, the frequency in the dictionary is selected from among the plurality of word unit divided sentences. In the word unit divided sentence selected by the word candidate divided unit for selecting the word unit divided sentence divided into the set of the notation of the word having the highest appearance frequency using information, in the word unit divided sentence selected by the word candidate selecting unit Based on the pronunciation information of the word-unit divided sentence, the pronunciation and the accent of the word are added to the notation of each of the words to give pronunciation information of the word-unit divided sentence. Speech waveform generation for generating prosodic information indicating intonation and rhythm of the entire word unit divided sentence, and generating a speech waveform based on the pronunciation information and the prosodic information It is configured by including a stage, a.

The third speech synthesizing device of the present invention is also characterized in that the dictionary storing a plurality of pieces of frequency information corresponding to the field of the kanji / kana mixed sentence and the kanji / kana mixed sentence. Using the frequency information in the dictionary corresponding to this field information, which receives field information indicating the field of
From the plurality of word unit divided sentences received from the morpheme analysis unit, the word candidate selection unit for selecting the word unit divided sentence divided into the notation set of the word having the highest appearance frequency, It is equipped with.

A fourth speech synthesizer of the present invention provides word notation, reading and accent of the word, part-of-speech information of the word, and frequency information indicating the frequency at which the notation of the word appears in a kanji-kana mixed sentence. A morphological analysis unit that outputs a word unit divided sentence obtained by dividing the kanji / kana mixed sentence by the dictionary stored for each notation of the word and the kanji / kana mixed sentence using the dictionary. When,
When the word unit divided sentence corresponding to the Kanji / Kana mixed sentence is received from the morpheme analysis unit, and when there are a plurality of word unit divided sentences received, in the word unit divided sentence of the plurality of word unit divided sentences A word candidate selecting unit that selects the word unit divided sentence having the highest evaluation in the entire word unit divided sentence of each of the frequency information in the dictionary, respectively corresponding to the notation of each of the words; The pronunciation and accent of the word are added to the notation of each word in the word unit divided sentence selected by the candidate selecting unit using the dictionary as pronunciation information of the word unit divided sentence. Generating prosodic information indicating the intonation and rhythm of the entire word unit divided sentence based on the pronunciation information of the word unit divided sentence, Is configured to include a speech waveform generation means for generating a speech waveform, the based on the information and the prosody information.

Further, a fourth voice synthesizing device of the present invention is the dictionary in which a plurality of the frequency information corresponding to the field of the kanji / kana mixed sentence is stored for each notation of the word, and the kanji / kana mixed sentence. Using the frequency information in the dictionary corresponding to this field information, which receives field information indicating the field of
Of the plurality of word unit divided sentences received from the morpheme analysis means, corresponding to the notation of each of the words in the word unit divided sentence, respectively, in the entire word unit divided sentence of each of the frequency information Word candidate selection means for selecting the word unit divided sentence with the highest evaluation,
It is configured with.

[0014]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of a speech synthesizer of the present invention.

In the present embodiment shown in FIG. 1, word notation,
Using the reading and accent of words, word part-of-speech information, and frequency information indicating the frequency of occurrence of word notation in a kanji-kana mixed sentence, and a dictionary 3 that stores kanji-kana mixed sentences, using the dictionary 3. The morpheme analysis unit 1 which outputs a word unit divided sentence obtained by dividing the Kanji / Kana mixed sentence into word writing units, and the word unit received by the morpheme analysis unit 1 corresponding to the Kanji / Kana mixed sentence When there are a plurality of divided sentences, the word unit divided sentence divided into a set of notations of the word having the highest appearance frequency is selected from the plurality of word unit divided sentences by using the frequency information in the dictionary 3. That is, of the plurality of word unit divided sentences, the frequency information in the dictionary 3 corresponding to the notation of each word in the word unit divided sentence is evaluated in the entire word unit divided sentence. The word candidate selecting means 2 for selecting the word unit divided sentence having the highest score and the dictionary 3 for the notation of each word in the word unit divided sentence selected by the word candidate selecting means 2 using the word reading and accent. Is given as the pronunciation information of the word unit divided sentence, and prosodic information indicating the intonation and rhythm of the entire word unit divided sentence is generated based on the pronunciation information of the word unit divided sentence, and the pronunciation information and the prosodic information are generated. And a voice waveform generating means for generating a voice waveform based on the voice waveform.

The voice waveform generating means is word candidate selecting means 2
Pronunciation information generation means 4 for generating pronunciation information of the word-unit divided sentence by using the dictionary 3 to give the pronunciation and accent of each word to the notation of each word in the word-unit divided sentence selected by From the pronunciation information generated by the pronunciation information generation means 4, prosody information generation means 5 for generating prosody information indicating the intonation and rhythm of the entire word unit divided sentence.
Based on the pronunciation information (eg, reading and accent of each word) and prosody information (eg, intonation and rhythm) of the word unit divided sentence, the voice is divided into fine units for voice synthesis. Voice waveform generation means 6 for generating a voice waveform using voice waveform data 7 that is data
It is composed of and.

The word candidate selecting means 2 selects a word unit divided sentence in which the sum of the frequency information, which is a predetermined numerical value indicating the frequency of appearance of a word notation in a kanji-kana mixed sentence, is the largest in the whole word unit divided sentence. I am trying to choose.

Next, the operation of the speech synthesizer according to this embodiment will be described in detail with reference to FIGS.

FIG. 2 is a diagram showing an example of a plurality of word unit divided sentences corresponding to a kanji / kana mixed sentence and these frequency information. A part of the inputted kanji / kana mixed sentence (believed so) and two words are shown. A part of the unit division sentence (belief / so, so / belief) and these frequency information (the frequency of "belief" is 2,
The frequency of "ji" is 1, the frequency of "yes" is 5, and the frequency of "belief" is 5), and this total is shown.

FIG. 3 is a diagram showing an example of the contents stored in the dictionary. The word notation, word reading and accent (there is an accent at the portion indicated by '), word part-of-speech information and word frequency information It shows a state of being stored for each notation. The frequency information (described as appearance frequency) is a numerical value indicating the frequency with which a notation appears in a sentence. For example, if the notation is rarely used, it is 1, and other notations are more common. If there is one but it is used more than 1, it is 2, if it is more than 2 but less than 4, it is 3,
Although it does not appear most often, it is given as 4 when it is higher than 3, and as 5 when it is used more obviously than other description.

In FIG. 1, a dictionary 3 is stored in advance for each word notation, word reading and accent, word part-of-speech information, and frequency information indicating the frequency at which the word notation appears in a kanji-kana mixed sentence. Prepare in advance. The morphological analysis unit 1 inputs a kanji / kana mixed sentence (for example, “Tanaka believes so, I did not doubt”), and inputs this kanji / kana mixed sentence (input sentence) using the dictionary 3 to write words. Create a word-based split sentence that is split into units. At this time, "Shinshin /
Same (“meaning when sending”) and “yes / believe” (meaning that “yes” and “believe” cease continuous use) (ie,
Two word-unit divided sentences) are obtained. The word candidate selection means 2 receives two word-unit divided sentences from the morpheme analysis means 1, and uses the dictionary 3 for these two candidates (“yes / believe” and “yes / believe”) to obtain individual words. The frequency of is calculated, and the one with the larger total is selected from these candidates. That is, based on the morphological analysis results (word division) of two candidates of “believe” and the frequency information and the total shown in FIG. 2, the total frequency of candidate 1 is 3, and the total frequency of candidate 2 is 10. Therefore, candidate 2 is Selected, "Yes / believe" in pronunciation information generation means 4
Only the word-unit division sentence including the word division result is sent. The pronunciation information generating means 4 uses the dictionary 3 to give the reading and accent of each word to the notation of each word in the word unit divided sentence selected by the word candidate selecting means 2 to generate a word unit divided sentence. Generate pronunciation information.

The prosody information generating means 5 generates prosody information indicating the intonation and rhythm of the entire word unit divided sentence from the pronunciation information generated by the pronunciation information generating means 4. That is, the prosody information generating means 5 generates a rhythm or intonation pattern when reading the input sentence based on the pronunciation information of the input sentence generated by the pronunciation information generating means 4. The prosody information generating means 5 has a duration control for determining rhythm and a pitch pattern control for determining intonation. As the duration control, for example, there is a model of duration control based on the control of the interval between vowel centers (time at which a vowel-like stable spectrum is shown) is kept constant. In this model, the time length between vowels in a synthetic speech is statistically estimated using information such as the speech rate, the type of vowel, the type of preceding phoneme, and the accent position. For pitch pattern control, for example, there is a method of estimating a normalized pitch pattern from the information of the part of speech of each accent phrase forming a sentence and the position information in the sentence, and then correcting the shape by the accent type. The normalized pitch pattern extracts the pitch frequency of the first vowel center and the peak frequency of each accent phrase in the sentence, and the pitch frequency of the last vowel center point of the accent phrase. It is the pattern created by dividing the value of.

The voice waveform generating means 6 divides the voice into fine units for voice synthesis based on the pronunciation information generated by the pronunciation information generating means 4 and the prosody information generated by the prosody information generating means 5. A voice waveform is generated using the voice waveform data 7. That is, the voice waveform generating means 6
Edits the voice stored in advance in the voice waveform data 7 according to the rhythm and intonation pattern generated by the prosody information generating means 5 to generate a voice signal for the input sentence. In the voice waveform data 7, a large number of words and sentences that are uttered are divided and stored in fine units for synthesis. The unit to be synthesized is, for example, the first half (CV) of a consonant and a vowel, and the second half of a vowel and a consonant (V
The unit of C) can be used. Speech waveform generation means 6
Selects a unit voice from the voice waveform data 7 so that the phoneme environment before and after the unit voice matches the phoneme environment of the voice to be synthesized as much as possible. The selected unit voice is edited in accordance with the rhythm and intonation pattern generated by the prosody information generation means 5 while adjusting the duration time and changing the pitch frequency. The method of connecting the unit voices and the connecting position are determined in advance according to the characteristics of each phoneme. At the connection points, interpolation processing is performed so that the unit voices are smoothly connected. As a method for generating an actual voice waveform based on the voice data edited in this way, there is a waveform editing method, for example.

In the above description, one frequency information is stored in the dictionary 3 for each notation of a word, and the morpheme analysis means 1 stores the input kanji / kana mixed sentence in this dictionary 3.
Is output as a word unit divided sentence divided into units of word notation, and the word candidate selecting unit 2 receives the word unit divided sentence corresponding to the Kanji / Kana mixed sentence from the morphological analysis unit 1 and the received word unit When there are a plurality of divided sentences, a word-unit divided sentence divided into a set of word notation having the highest appearance frequency is selected from the plurality of word-unit divided sentences using this frequency information in the dictionary 3. However, it is equipped with frequency information for each field that stores a plurality of frequency information corresponding to fields of kanji / kana mixed sentences (for example, e-mail, newspapers, technical papers, address books, etc.) for each word notation. As shown in FIG. 5, which is a block diagram showing one embodiment of a voice synthesizing device using the field-specific frequency information, including the dictionary 9 shown in FIG. 4 showing an example of the stored contents of the dictionary, By the word candidate selection means 8 , Receives field information indicating the field of a kanji / kana mixed sentence (This information may be received from the outside of this device according to a user's instruction, and the input kanji / kana mixed sentence is checked to determine the field of this input sentence. The determined field information may be received.) Frequency information corresponding to this field information is obtained from this dictionary 9, and a plurality of word unit divisions received from the morphological analysis means 1 using this obtained frequency information. You may make it select the word unit division | segmentation sentence divided | segmented into the set of the notation of the word with the highest appearance frequency from among sentences.

[0026]

As described above, according to the speech synthesizer of the present invention, word notation, word reading and accent,
It is equipped with a dictionary that stores the word part-of-speech information and the frequency information indicating the frequency in which the word notation appears in a kanji-kana mixed sentence, and uses a dictionary that receives the kanji kana-mixed sentence by the morphological analysis means and uses this kanji character. A kana-mixed sentence is divided into word notation units, a word-unit divided sentence is output, and the word candidate selection means receives the word-unit divided sentence corresponding to the kanji kana-mixed sentence from the morphological analysis means. When there are multiple word-based split sentences, the frequency information in the dictionary is used to select the word-based split sentence that has been split into the set of words with the highest frequency of occurrence. By the means, a word unit divided sentence is added by using a dictionary to the notation of each word in the word unit divided sentence selected by the word candidate selection unit, using a dictionary. As pronunciation information, prosodic information indicating the intonation and rhythm of the entire word unit divided sentence is generated based on the pronunciation information of the word unit divided sentence, and a voice waveform is generated based on the pronunciation information and the prosody information. By the candidate selection means, the word unit divided sentence divided into the set of the notation of the word having the highest appearance frequency is selected using the frequency information in the dictionary, so that the correct word unit divided sentence can be definitely selected, In addition, because we have a dictionary that stores the frequency information that shows the frequency of word notation in a kanji-kana mixed sentence for each word notation, there are two dictionaries with different characteristics: a kanji-kana mixed dictionary and an alphanumeric kana dictionary. Further, it is not necessary, and further, a frequency information indicating the frequency of occurrence of a word notation in a kanji-kana mixed sentence is stored for each word notation, and the word candidate selecting means allows the frequency information in the dictionary to be stored. Since selecting the most occurrences it is divided into a set of frequent representation of word-word unit dividing statements using, usually there is no possibility of selected word notation not used very often.

In addition, a dictionary storing a plurality of frequency information corresponding to the fields of kanji and kana mixed sentences for each notation of a word is provided.
The word candidate selecting means receives the field information indicating the field of the kanji / kana mixed sentence, obtains the frequency information corresponding to this field information from this dictionary, and uses the obtained frequency information to obtain the plurality of words received from the morpheme analyzing means. From the word unit split sentences of
By selecting the word unit divided sentence divided into the set of words with the highest frequency of appearance, the word unit divided sentence is selected according to the frequency information according to the field of the kanji and kana mixed sentence, so more correct reading is done. It will be possible.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of a speech synthesizer of the present invention.

FIG. 2 is a diagram showing an example of a plurality of word-unit divided sentences corresponding to a kanji / kana mixed sentence and their frequency information.

FIG. 3 is a diagram showing an example of stored contents of a dictionary.

FIG. 4 is a diagram showing an example of stored contents of a dictionary having field-specific frequency information.

FIG. 5 is a block diagram showing an embodiment of a speech synthesizer using frequency information for each field.

FIG. 6 is a block diagram of a conventional speech synthesizer.

[Explanation of symbols]

1 Morphological analysis means 2 Word candidate selection means 3 dictionary 4 Pronunciation information generation means 5 Prosody information generation means 6 Voice waveform generation means 7 Voice waveform data 8 Word candidate selection means 9 dictionary

Claims (9)

[Claims] 1. A morpheme analysis unit that receives a Kanji / Kana mixed sentence and divides the Kanji / Kana mixed sentence in word notation units and outputs the divided word unit divided sentence; and the Kanji / Kana mixed sentence from the morpheme analysis unit. When there are a plurality of word unit divided sentences that have received the word unit divided sentence corresponding to the sentence, the word unit divided sentence is divided into the set of the notation of the word having the highest appearance frequency from among the plurality of word unit divided sentences. A word candidate selection unit that selects the word unit divided sentence, and the reading and accent of the word are given to the notation of each of the words in the word unit divided sentence selected by the word candidate selection unit. The pronunciation information of the word unit divided sentence, and based on the pronunciation information of the word unit divided sentence, the intonation and rhythm of the entire word unit divided sentence are indicated. Generates prosody information, speech synthesis apparatus characterized by comprising a speech waveform generation means for generating a speech waveform, the based on the sound information and the prosody information. 2. A morpheme analysis unit which receives a Kanji / Kana mixed sentence and divides the Kanji / Kana mixed sentence in word notation units and outputs the divided word-unit divided sentence; and the Kanji / Kana mixed sentence from the morpheme analysis unit. When there are a plurality of word unit divided sentences that have received the word unit divided sentence corresponding to the sentence, from among the plurality of word unit divided sentences, the notation of each of the words in the word unit divided sentence Word candidate selecting means for selecting the word unit divided sentence having the highest evaluation in the entire word unit divided sentence of the respective predetermined frequency information indicating the frequency of appearance in the corresponding kanji / kana mixed sentence, and the word candidate Before adding the pronunciation and the accent of the word to the notation of each of the words in the word unit divided sentence selected by the selection unit, As pronunciation information of the word unit divided sentence, based on the pronunciation information of the word unit divided sentence to generate prosody information indicating the intonation and rhythm of the entire word unit divided sentence, based on the pronunciation information and the prosody information And a voice waveform generating means for generating a voice waveform. 3. The voice waveform generating means adds the reading and accent of the word to the notation of each of the words in the word unit divided sentence selected by the word candidate selecting means, and Pronunciation information generating means for generating the pronunciation information of the word unit divided sentence, and prosody information for generating the prosody information indicating the intonation and rhythm of the whole word unit divided sentence from the pronunciation information generated by the pronunciation information generating means. A voice generating a voice waveform by using a generating means, and voice waveform data which is data divided into fine units for voice-synthesizing voice based on the pronunciation information and the prosody information of the word-unit divided sentence. The speech synthesizer according to claim 1 or 2, further comprising: waveform generating means. 4. The notation of the word, the reading and accent of the word, the part-of-speech information of the word, and the frequency information indicating the frequency of occurrence of the notation of the word in a kanji-kana mixed sentence are stored for each notation of the word. A dictionary, a morpheme analysis unit that receives the Kanji / Kana mixed sentence and outputs a word unit divided sentence obtained by dividing the Kanji / Kana mixed sentence into units of the notation of the word using the dictionary; When there are a plurality of word-unit divided sentences that have received the word-unit divided sentence corresponding to a kana-mixed sentence, the word-unit divided sentence most appears from the plurality of word-unit divided sentences using the frequency information in the dictionary. In the word unit divided sentence selected by the word candidate selected unit, which selects the word unit divided sentence divided into the notation set of the words having high frequency, Based on the pronunciation information of the word unit divided sentence, the pronunciation and the accent of the word are added to the notation of each of the words to give the pronunciation information of the word unit divided sentence. Voice waveform generating means for generating prosody information indicating the intonation and rhythm of the entire word unit divided sentence and generating a voice waveform based on the pronunciation information and the prosody information. Speech synthesizer. 5. The notation of the word, the pronunciation and accent of the word, the part-of-speech information of the word, and frequency information indicating the frequency of occurrence of the notation of the word in a kanji-kana mixed sentence are stored for each notation of the word. A dictionary, a morpheme analysis unit that receives the Kanji / Kana mixed sentence and outputs a word unit divided sentence obtained by dividing the Kanji / Kana mixed sentence into units of the notation of the word using the dictionary; When there are a plurality of received word unit divided sentences corresponding to the word unit divided sentence corresponding to the kana mixed sentence, among the plurality of word unit divided sentences, the word of each of the words in the word unit divided sentence Word candidate selection for selecting the word unit divided sentence with the highest evaluation in the entire word unit divided sentence of each of the frequency information in the dictionary, corresponding to each notation And the word unit division by adding the pronunciation and accent of the word using the dictionary to the notation of each word in the word unit division sentence selected by the word candidate selection unit. As pronunciation information of a sentence, prosodic information indicating the intonation and rhythm of the entire word unit divided sentence is generated based on the pronunciation information of the word unit divided sentence, and a speech waveform based on the pronunciation information and the prosody information. A speech synthesis apparatus comprising: a speech waveform generating means for generating 6. The dictionary storing the plurality of pieces of frequency information corresponding to the field of the kanji / kana mixed sentence for each notation of the word, and field information indicating the field of the kanji / kana mixed sentence Using the frequency information in the dictionary corresponding to, from among the plurality of word unit divided sentences received from the morpheme analysis means, the divided into the set of the notation of the word with the highest appearance frequency The speech synthesis apparatus according to claim 4, further comprising: the word candidate selection unit that selects a word unit divided sentence. 7. The dictionary, which stores a plurality of the frequency information corresponding to the field of the kanji / kana mixed sentence, for each notation of the word, and field information indicating the field of the kanji / kana mixed sentence. Using the frequency information in the dictionary corresponding to, from among the plurality of word unit divided sentences received from the morpheme analysis means, respectively corresponding to the notation of each word in the word unit divided sentence 6. The speech synthesis apparatus according to claim 5, further comprising: a word candidate selection unit that selects the word unit divided sentence having the highest evaluation in the entire word unit divided sentence of each of the frequency information. 8. The speech waveform generating means uses the dictionary for the notation of each of the words in the word unit divided sentence selected by the word candidate selecting means, and uses the pronunciation and accent of the word. Pronunciation information generating means for generating pronunciation information of the word unit divided sentence by giving each, and prosodic information indicating the intonation and rhythm of the whole word unit divided sentence from the pronunciation information generated by the pronunciation information generating means. A prosody information generating unit for generating a voice waveform using voice waveform data which is data divided into fine units for voice-synthesizing voice based on the pronunciation information and the prosody information of the word-unit division sentence. The speech synthesizer according to claim 4, 5, 6 or 7, further comprising: 9. The word candidate selecting means has the largest total of the frequency information, which is a predetermined numerical value indicating the frequency of appearance of the word in the kanji-kana mixed sentence, in the entire word unit divided sentence. 5. The word-unit division sentence that is is selected.
The speech synthesizer according to 5, 6, 7 or 8.