JP3364820B2 - Synthetic voice output method and apparatus - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a synthetic speech output method and apparatus, and more particularly, in a synthetic speech reading process of reading a Japanese document containing kanji and kana by synthetic speech, a compound word including a time word can be highly accurately defined. The present invention relates to a synthetic speech output method and apparatus capable of dividing a sentence into accent phrases and outputting a speech with unnaturalness removed.

[0002]

2. Description of the Related Art When reading a computerized Japanese sentence containing kanji and kana such as newspaper articles and e-mails with synthetic voice, at least the kanji and kana mixture sentence is read in order to output a more natural voice. Then, it is necessary to divide the sentence into accent phrases, determine the accent type of the accent phrase, and set the pose between the accent phrases at appropriate intervals.

At present, such text-to-speech synthesis method and device have been put into practical use, but they are considerably unnatural compared to human speech. This unnaturalness is mainly a problem of setting accuracy of prosody information. The prosody information includes accent phrase information and pose information. Next, those features will be described.

An accent phrase is a unit that is uttered as one unit when naturally spoken, and is expressed on a pitch pattern (time pattern of fundamental frequency F ₀ ). In the Japanese Tokyo dialect, it is defined as a unit having at most one accent each (Mora at which F ₀ begins to fall sharply), and corresponds to almost a phrase. However, a plurality of clauses may be one accent phrase, and one clause may be a plurality of accent phrases.

When an accent phrase is composed of a plurality of words, the accent type of the word is lost or moved, so that it is necessary to set the accent type as the accent phrase. The accent phrase information has the range of each accent phrase and the accent type information. Poses are set between accent phrases with proper intervals and lengths due to syntactic / semantic factors and vocal restrictions (breathing). The pose information has information about the presence / absence of a pose between each accent phrase and, if there is a pose, information about the pose length.

Conventionally, a method of dividing a sentence into accent phrases using morphological analysis has been used. Basically, "1 independent word + consecutive (plural) adjuncts" is defined as 1 accent phrase (stem of stem + conjugation is regarded as 1 independent word). For example, the sentence “Today's weather in Tokyo will be fine.” Is divided into “Today / Tokyo / weather / is fine.” (/ Is the accent phrase boundary) and accent phrases.

However, in compound nouns, two consecutive
There are cases in which independent words of about 4 words are grouped together into one accent phrase, and cases in which the word is divided halfway into a plurality of accent phrases. Therefore, it is necessary to consider the accent phrase division method of compound nouns. Here, the compound noun is a word string of continuous body words (noun, affix). Strictly speaking, there is a bunsetsu boundary between consecutive word strings of nouns, which may not be one compound noun. For example, "every year abroad" in the sentence "every year abroad" is not a compound noun,
There is a bunsetsu boundary between "every year" and "overseas" (however,
When referring to "annual event", "annual" is 1
Be one compound noun). However, since morphological analysis does not recognize bunsetsu, words that appear to be compound nouns such as "every year overseas" (hereinafter, words with such expressions are called compound nouns) and "annual It is not possible to distinguish what is actually a compound noun like ". Therefore, in both cases, the accent phrase segmentation method using morphological analysis must be treated as a compound noun.

[0008] Two typical documents are shown as a conventional accent phrase segmentation method using morphological analysis. As the first document, "Nomura: Accent phrase segmentation and accent addition of compound words using word classification, IEICE Transactions Vo
l.J75-D-II, No.9, pp.1479-1488 (1992-9) ”.
In this method, the independent word is divided into four types from the viewpoint of accent division, and the accent phrase division is performed using the classification and the part of speech.

As a second document, "Miyazaki: Automatic extraction method of compound word accent phrase using semantic connection between words, IEICE Transactions Vol.J68-D, No11, pp.25-"
32 (1985-1) ". In this method, accent phrase segmentation is performed based on the dependency analysis within the part of speech and compound nouns.

In these cases, accent phrase division can be successfully performed in the following cases (1) and (2) (/ indicates a correct accent phrase boundary, and "" indicates one word. The example uses this notation). (1) "Chemical""Factory" (2) "Small" / "Process""Device""

[0011]

However, in the above-mentioned conventional method, although there is no problem in the above-mentioned examples (1) and (2), verbs (nouns representing times and sometimes used as adverbs) There is a problem in the case of compound nouns such as (3) and (4) below that include. However, (3) is, for example, an apparent compound noun that corresponds to the "" part of the sentence "I made a profit similar to" Monthly salary "."

(3) "Monthly" / "Salary" (4) "Semi-annual""Course" In the above-mentioned first document, no special rule is applied to the toki. Therefore, in the case of a verb used in the same way as a general noun as in the example (4), accent phrase division can be performed well. However, as in example (3), since it treats apparent compound nouns in the same way, "monthly salary"
Becomes one accent phrase.

On the other hand, in the above-mentioned second document, "toki (in the document," adverbal noun ("this year",
"A noun used as an adverb to represent times such as" morning "and" spring ""
Has a high degree of independence, and is therefore an independent accent phrase (an accent phrase having a number of independent words of 1). ". Therefore, in the case of the example (3), accent phrase division can be performed well. However, there is a problem in that the accent phrase division is also performed when the word and the word immediately after that are strongly connected as in the example (4).

As described above, the conventional accent phrase segmentation method has a problem that the accent phrase segmentation cannot be properly performed when there is a compound noun including a punctuation as in Examples (3) and (4). . The present invention has been made in view of the above points, and provides a synthetic speech output method and device capable of more natural synthetic speech output by more accurately dividing an accent phrase in a compound noun including a verb. The purpose is to

[0015]

[Means for Solving the Problems] FIG. 1 is a diagram for explaining the principle of the present invention. The present invention is a synthetic speech output that reads out a Japanese document containing a mixture of Japanese kanji and kana with synthetic speech.
In the power method, each of the input Japanese kana mixed sentences
Word recognition processing (step 1) to certify words and word recognition
Morphological analysis is performed on words that have been certified by regular processing, and each
The word analysis process that obtains information for each word and the solution
Is it Japanese using each analyzed word and information for each word?
Kana grant process to grant the pseudonym reading the mixed statement use (step 2), the Japanese kana, when lyrics is the attributive manner
Information about whether or not you can enter, dependency information in compound nouns,
Part of speech of words forming compound noun and number of words in the compound word
Accent phrase division processing for dividing the accent phrase using the (Step 3), divided accessed accent phrase each
Accent type setting processing for setting the accent type (step 4)
And pose information is set between the divided accent phrases.
Pose setting process (step 5) and phonetic transcription process
Kana, accent type, and pose information given in
With prosody information corresponding to Japanese kana mixed sentences
It consists of a voice synthesis process (step 6) for generating a synthetic voice.
The accent phrase division process (step 3) is given.
When a compound word in a compound word is used as an adjunct
Is generated according to the type of part of speech of the word immediately preceding the time verb.
Change the word pattern identification Change the word pattern identification
Processing and word pattern identification change, word one word ahead
Depending on the processing, the sa verb type noun or the dependency relation
A word that is judged to be a general noun and is one word ahead of the time
When it is at the end of a compound word, or just after the verb
If the word is a suffix, adjunct, or symbol, then
The word word pattern is composed of the word and the word immediately preceding the word.
The trailing end judgment process is performed to judge that there is .

[0016]

FIG. 2 is a block diagram showing the principle of the present invention. The present invention synthesizes a Japanese document of Japanese kanji and kana mixed sentences into a synthesized voice.
Synthesis an audio output device, certify each word from the input Japanese kana performs morphological analysis on the words that have been approved, the word analysis unit 100 to obtain the information for each word read aloud, Using the words analyzed by the word analysis means 100 and the information for each word, the reading kana addition means 210 for giving the reading kana to the Japanese kana mixed sentence and the Japanese kana mixed sentence, the toki are connected Information about whether or not it is used in the compound noun, dependency information in the compound noun, the part of speech of the words forming the compound noun, and the number of words in the compound word. the accent type setting means 230 for setting the accent phrase for each accent type, a pause setting means 240 for setting a pause information between divided accent phrase, kana-type application means 210 Granted kana, accent type, and using the pose information, voice synthesis section 3 for generating prosody information with synthesized speech corresponding to Japanese kana
00 and have a, accent phrase division means 220, given et al.
When a verb in a compound word is used as an adjunct
In this case, depending on the part-of-speech type of the word immediately preceding the time verb,
Change word pattern identification change
And the word immediately preceding the verb is the word pattern identification variable.
Depending on the changing means, the sa noun type noun or the dependency relation
A word that is judged to be a common noun and is one word ahead of the verb
Is at the end of a compound word, or two times after the verb
If the word is a suffix, an adjunct or a symbol,
A time word word pattern is composed of a time word and the word immediately preceding the time word.
It has a tail judging means for judging that it is performing .

[0018]

[0019]

According to the present invention, information on whether a toki is used as an adjunct, information on dependency analysis in a compound noun used in the second document, information on a part of speech of a word forming a compound noun, and By using the number of constituent words that have not been divided into accent phrases in compound nouns, accent phrases of compound nouns including time words are divided. In other words, when there is a strong semantic connection between a verb and the word that immediately follows it in a compound noun, paying attention to the fact that the verb and the word that immediately follows become one accent phrase, and the verb is used as a continuum. The accent phrase is determined by using the information such as "ka", the part of speech of the word immediately after the verb, the number of nouns following the verb, and the dependency information in the compound word from the verb to the word immediately after the verb. As a result, it becomes possible to accurately perform accent phrase division of a compound noun including a verb.

Therefore, by performing such accent phrase division, the synthesized voice output device can output a voice close to a human voice.

[0021]

Embodiments of the present invention will now be described in detail with reference to the drawings. In the following description, the (partial) compound noun is W1W.
2W3 ... Wa (W1: first word of compound noun, W2: second word of compound noun, W3: third word of compound noun, Wa: a-th word (end word) of compound noun). The verb is Wt (t is an arbitrary value from 1 to a-1), the word immediately after that is Wt + 1, and the word immediately after Wt + 1 is Wt + 2.

Here, the partial compound noun is each divided word string obtained by dividing the compound noun according to the part of speech and the characteristics of the constituent words. First, a word that is often used as an adjunct (to) in compound nouns is likely to be associated with the immediately following word. Therefore, the verb Wt may be used as an adjunct (to) in a compound noun as a necessary condition for connecting the verb Wt and the word Wt + 1 immediately after the verb.

Information indicating whether or not the compound noun is used as an adjunct (to) is described in the word dictionary information. Next, pay attention to the part of speech of Wt + 1. Wt + 1
Is a transposed noun (a noun turned from a noun, eg, rest, born,
In the case of (upward), the relationship between Wt and Wt + 1 is syntactically modified, but is often converted to a noun. For example, "New Year holidays" is a continuation of the "holiday holidays" modifier combined with "New Year holidays." This connection is numerous because the word that represents time generally modifies the word that serves as the term. Therefore, in this case, the word Wt and the immediately following word Wt + 1 are always the same accent phrase.

In addition, even when the immediately following word Wt + 1 is a sa noun verb type noun, similarly, in many cases, the syntactically modified word is a noun. For example, "Forecast for this term" is
"Forecast for this term" is a continuous modification of "Forecast for this term"
Is a continuum. However, the syllable-type noun is a word that has a strong connection with the word immediately after it. For example, in the compound noun "predicted profit for the current period", the connection between "predicted" and "profit" is stronger than the connection between "current period" and "forecast", and the accent phrase is "current period / expected profit". Therefore, when Wt + 2 is a noun, the connection between “Wt + 1 and Wt + 2” is greater than the connection between “Wt and Wt + 1”, and Wt + 2 is not a noun.
Or, only if it does not exist, set “Wt and Wt + 1” to 1
One accent phrase.

If Wt + 2 is a noun, then W
Let t be one accent phrase. When Wt + 2 is a suffix, the ties of Wt, Wt + 1, and Wt + 2 are almost the same, and Wt, Wt + 1, and Wt + 2 are one accent phrase. If Wt + 1 is a general noun, W
No syntactic modification structure can be considered between t and Wt + 1. Therefore, we use dependency analysis in compound words that uses semantic information of words. It is determined that there is a connection between Wt and Wt + 1 when the general noun dependency (meaning that the semantic connection of words is strong) is established from Wt to Wt + 1.

However, when Wt + 2 is a noun,
In many cases, the connection between Wt + 1 and Wt + 2 is stronger than the connection between Wt and Wt + 1. Therefore, if Wt + 2 is not a noun or does not exist, Wt and Wt + 1
Is one accent phrase. When Wt + 2 is a noun, Wt is one accent phrase. When Wt + 2 is a suffix, the ties of Wt, Wt + 1, and Wt + 2 are approximately the same, and Wt, Wt + 1, and Wt + 2 are one accent phrase.

For example, in the "next term goal", the general noun dependency is established for the "next term" and the "goal". In this case, "next term goal" is one accent phrase. However, in the “target number for the next period”, the link between the “next period” and the “target”
The connection between "target" and "number" is stronger, and the accent phrase is "next term / target number".

Next, the structure of the synthetic speech output device of the present invention will be described. FIG. 3 shows a block diagram of a synthesized voice output device according to an embodiment of the present invention. The synthesized speech output device shown in the figure comprises a morphological analysis unit 100, a word dictionary 110, a reading / prosodic information setting unit 200, and a synthesized speech generation unit 300.

The morphological analysis unit 100 receives a kanji-kana mixed sentence to be output as a synthesized voice, and then receives the word dictionary 11
With reference to 0, the word-spacing word string b with word information is output. The separated word string b with word information is obtained by dividing a sentence into words and adding word information (part of speech, accent type, reading, etc.) for each word acquired from the word dictionary 110.

The reading / prosody information setting unit 200 comprises a reading adding unit 210, an accent phrase dividing unit 220, an accent type setting unit 230, and a pose setting unit 240.
The reading / prosodic information setting unit 200 receives the word-spacing word string b with word information as an input and outputs a kana string with prosody information.

The reading adding section 210 of the reading / prosody information setting section 200 adds reading to each word of the word-spacing word-spacing word string b. Reading information of word information is mainly used as reading, but rendaku processing, number processing (digit reading or,
(Determining stick reading) is also performed.

The accent phrase splitting unit 220 is inputted by utilizing the property that these two words are likely to become one accent phrase when the noun and the word immediately following it are strongly connected in the compound noun. The sentence is divided into accent phrases. Details of this processing will be described later.

The accent type setting unit 230 sets an accent type for each accent phrase when the accent phrase dividing unit 220 forms an accent phrase with a plurality of words. The pose setting unit 240 sets the presence / absence of a pose between the accent phrases. And
If there is a pose, set its length as well. From this pose setting unit 240, a kana string c with prosody information is output.

The prosodic information-added kana string c output from the pose setting section 240 represents reading information of a sentence to be output as synthesized speech, and has prosodic information such as accent type, pause information, male / female voice, and pitch level. . The synthetic speech generation unit 300 outputs a synthetic speech with the prosodic information-added kana sequence c as an input.

Next, the operation of the accent phrase dividing section 220 will be described in detail. FIG. 4 is a flow chart for explaining the processing of the accent phrase dividing unit according to the embodiment of the present invention. In the following explanation, accent phrase division is
Processing is performed in order from the beginning of the sentence to the end of the sentence.

Step 100) The segment division process is a process of dividing a sentence into segments. The word at the beginning of the segment is a noun (general noun, sa noun type noun, adjective verb type noun, metamorphic noun, toki, number, pronoun, proper noun)
Or if the beginning of the segment is a prefix and the second word is a noun, then (prefix) + noun + (consecutive nouns or suffixes) +
(Consecutive attached words) is one segment. This segment is a compound noun phrase segment.

In the cases other than the above, (prefix) + independent word other than noun + (suffix) + (consecutive adjunct) is defined as one segment. This segment is an independent segment. However, the above () means that the word is included in the segment only when it exists. The total number of segments is n.

Here, auxiliary words include adjectives and adjectives. The compound noun phrase segment is a combination of a compound noun and an adjunct word string. However, if the prefix is included in the compound noun, the segment is divided just before the prefix.

Step 200) Segment counter i
Substitute 1 for. Step 300) It is determined whether the i-th segment Sg (i) is a compound noun phrase segment. If the segment is a compound noun phrase segment, the process proceeds to step 400, and if not, the process proceeds to step 600.

Step 400) In the compound noun phrase accent phrase segmentation process, accent phrase segmentation of the compound noun phrase segment is performed. Step 410) The accent phrase is divided according to the word characteristics. The accent phrase division process based on word characteristics is a process of dividing accent phrases according to the part of speech of a word and the characteristic peculiar to the word. The characteristic peculiar to a word is based on the dictionary information of each word given at the time of morphological analysis. An example of this processing is shown below.・ Numbers are divided into accent phrases for each digit.

(Example) 123 (Hyaku Nijusan) → Hyakh / Nijou / San There are always words that independently form an accent phrase. (Example) "Moto" → "Tanaka / Former / Prime Minister", "Former / Representative", etc.-There is a word immediately before the accent phrase boundary.

(Example) “Strengthen (Zokyo)” → “Mission /
Augmentation ”・ There is a word immediately after the accent phrase boundary. (Example) “below (squid)” → “section chief or less / target” In this way, the accent phrase boundary is set for the segment according to the part of speech of the word or the characteristic peculiar to the word. Then, Sg (i) is decomposed into m sub-segments by the accent phrase boundary.

Step 420) Accent phrase division is performed according to the word pattern. Accent phrase division processing by word pattern is a process of extracting a word group having a strong connection among compound nouns in a partial segment as a word pattern and setting an accent phrase boundary based on the pattern.
Examples of word patterns include linguistic word patterns, proper noun word patterns, and numerical aid word patterns, which are objects of the present invention. Details of this processing will be described later with reference to FIG. In this step, the m partial segments are processed one by one. As a result, the number of segments Sg (i) is k (k ≧
m) is divided into sub-segments.

Step 430) The accent phrase is divided according to the number of remaining words. Accent phrase division processing based on the number of remaining words sets accent phrase boundaries for the partial segments for which no word pattern has been generated among the k partial segments, using the number of words of the partial segment and the word part of speech. It is a process to do. In this step, the accent phrase extraction based on the bond strength between words in the second document is used.

As described above, in step 104,
The segment Sg is set according to the set accent phrase boundary.
Divide (i) into accent phrases. Step 500) It is determined whether the segment counter i is n. If it is n, the process ends. If not n, the process proceeds to step 700.

Step 600) Accent phrase division processing of an independent segment is performed. In this step, Sg (i) is basically used as one accent phrase. After processing, the process proceeds to step 500. Step 700) 1 is added to the segment counter i. After the processing, the process proceeds to step 300.

Next, the accent phrase division process by the word pattern in the above step 420 will be described in detail. Figure 5
FIG. 6 is a flowchart of accent phrase division processing by word patterns according to an embodiment of the present invention.

Step 421) As initial setting processing,
Substitute 1 for the word counter j. Substitute the number of independent-word words of the corresponding partial segment into p. Step 422) It is determined whether the part of speech of the jth word Wj of the partial segment is a participle. If it is a tongue, the process proceeds to step 423, and if not, the process proceeds to step 425.

Step 423) A word pattern generation process of time words is performed. Details will be described in detail in FIG. Step 424) It is determined whether the word counter j is equal to the number of independent word p. If they are equal, the process proceeds to step 427, and if they are not equal, the process proceeds to step 426.

Step 425) Other word pattern generation processing is performed. This processing includes a proper noun word pattern generation processing and a numeral auxiliary number word pattern generation processing. After the processing, the process moves to step 424. (Step 426) The word counter j is incremented by 1, and the process proceeds to step 422.

Step 427) Affix renormalization processing to the word pattern is performed. This is a process of incorporating this prefix / suffix into the word pattern when the immediately preceding word in the same sub-segment of the generated word pattern is the prefix or the immediately following word is the suffix.

Step 428) A word pattern adjunct word renormalization process is performed. This is a process in which, when the word (column) immediately after the generated word pattern is an adjunct word (column), this adjunct word (column) is carried into the word pattern. next,
The word pattern generation process of the verb in the above step 423 will be described. FIG. 6 is a flowchart of the word pattern generation process according to the embodiment of the present invention.

Step 4231) It is determined whether or not the compound in-word adjunct flag of Wj is ON. If it is on, the process moves to step 4232. If it is off, the process moves to step 4239. Here, the in-compound word adnominalization flag is one of word information, and is obtained by referring to the word dictionary 110 in the process of the morphological analysis unit 100 in FIG. Then, it indicates whether the word is used as an adnominal (to) in a compound word. The words for which the compound word adnominalization flag is on and their compound words are:
For example, "large" → "large refrigerator" (large refrigerator) "modern" → "modern painting" (modern painting) "Christmas" → "Christmas holiday" (Christmas holiday).

Step 4232) It is judged whether the part of speech of the j + 1th word Wj + 1 of the partial segment is a metamorphic noun. If it is a transposition noun, the process proceeds to step 4237. Step 4233) It is determined whether the part of speech of Wj + 1 is a general noun. If it is a general noun, the process proceeds to step 4235. Otherwise, go to step 4234.

Step 4234) It is determined whether or not the part of speech of Wj + 1 is a sa-variative noun. If it is a sa-verb type noun, the process proceeds to step 4236. If not, the process proceeds to step 4239. (Step 4235) It is determined whether the general noun dependency is established from Wj to Wj + 1. When it is satisfied, the procedure proceeds to Step 4236, and when it is not satisfied, Step 42 is performed.
Move to 39.

The general noun dependency is a general noun /
It is a dependency within a compound word that is established from a toki / pronoun (as a front word) to a general noun or a suffix (as a rear word) behind it. This dependency is established when the backward-sentence semantic category of the forward word and the backward-word semantic category match, or when the forward-sentence semantic category of the backward word and the forward-word semantic category match.
A semantic category is a hierarchical classification of nouns according to their meaning.

Step 4236) 1 is set to the word counter j.
The added value is equal to p, or the part of speech of the j + 2nd word Wj + 2 of the partial segment is the suffix or
Determine if it is an adjunct or a symbol. If the condition is satisfied, the process proceeds to step 4237, and if it is not satisfied, the process proceeds to step 4239.

Step 4237) Wj and Wj + 1 are put together to form a word word pattern. (Step 4238) 1 is added to the word counter j. Step 4239) Let Wj be a word word pattern.
After the processing, the word pattern generation processing of the time verb ends.

Next, the accent phrase division processing of FIGS. 4, 5, and 6 will be described with reference to FIG. 8 using the example sentence of FIG. The inputted kanji / kana mixed sentence is “In New York State, a record big cold wave strikes every Christmas holiday” is input, and is analyzed by the morphological analysis unit 100 as shown in FIG. 8 (A).

Next, the above sentence is divided into five segments by the segment division processing of step 100 (FIG. 4), and Sg (1) to Sg (4) are compound noun phrases as shown in FIG. 8 (B). The segment, Sg (5), becomes an independent segment. Next, consider the case where i = 2.

In step 300 of FIG. 4, Sg
Since (2) is a compound noun phrase segment, step 4
The process shifts to accent phrase division processing by specifying 10 words. No processing is performed here because no corresponding word exists. Next, the process proceeds to the accent phrase division process based on the word pattern in step 420. Then, in step 422 of FIG. 5, since W1 = “every year” is the verb, in step 423, the process shifts to the word pattern generation process of the verb.

Next, in step 4231 of FIG. 6, W1 =
Since the compound word in the compound word for "every year" is off,
The process moves to step 4239. And W1 = "every year"
Is a time word pattern. Next, step 42 of FIG.
Returning to step 3, since 1 ≠ p (= 4), step 426.
Move to. Then, j = 2 is set and the process proceeds to step 422. Since W2 = “Christmas” is the verb, the process moves to the vocabulary word pattern generation process in step 423.

Next, in step 4231 of FIG. 6,
Since the compound word internalization flag of W2 = “Christmas” is ON, the process proceeds to step 4232. Where W
3 = “vacation” = general noun, so step 4233
To Step 4235. Then, since the general noun dependency is established from "Christmas" to "holiday", the process proceeds to step 4236. In step 4236, since the part of speech of W4 = “koro” is the suffix, the process proceeds to step 4237. Then, "Christmas holidays" is used as a word word pattern. Next, the process proceeds to step 4238 and j = 3.

Next, returning to step 424, 3 ≠ p
Since (= 4), the process proceeds to step 426. Then, j = 4 is set and the process proceeds to step 422. W4 =
Since "koro" is not a verb, the process proceeds to step 425.
No word pattern is generated in step 425, and the process proceeds to step 424. Since 4 = p (= 4), the process proceeds to step 427. In step 427,
The time word word pattern “Christmas holiday” is included in the term “about”, and the time shifts to step 428 as “Christmas holiday”. In step 428, "ni" is added to "around Christmas holidays" to be "around Christmas holidays".

Next, the process proceeds to the accent phrase division process based on the number of remaining words in step 430 of FIG. Here, since there is no partial segment for which a word pattern has not been generated, processing is not performed. The series of processes described above completes the process for the segment Sg (2).

Similarly, for other segments, accent phrase division can be performed by performing the processes shown in the flow charts of FIGS. 4, 5, and 6 described above. The present invention is not limited to the above-described embodiments, and various modifications and applications are possible within the scope of the claims.

[0067]

As described above, according to the synthetic speech output method and apparatus of the present invention, the accent phrase division of time words, which has been treated uniquely in the past, is provided with information as to whether the time words are used as adjuncts. , The word pattern generation process of time words is performed by using the information of the dependency analysis in the compound noun, and the number of words in the compound word which is not divided into the part of speech and the accent phrase of the words forming the compound noun. As a result, in a compound noun including a syllable, accent phrases can be divided with higher accuracy, and more natural synthesized speech output is possible.

[Brief description of drawings]

FIG. 1 is a diagram for explaining the principle of the present invention.

FIG. 2 is a principle configuration diagram of the present invention.

FIG. 3 is a block diagram of a synthesized voice output device according to an embodiment of the present invention.

FIG. 4 is a flowchart for explaining the processing of the accent phrase dividing unit according to the embodiment of the present invention.

FIG. 5 is a flowchart of accent phrase division processing using word patterns according to an embodiment of the present invention.

FIG. 6 is a flowchart of word pattern generation processing according to an embodiment of the present invention.

FIG. 7 is a diagram showing an example of synthetic speech processing according to an embodiment of the present invention.

FIG. 8 is a diagram for explaining segment division processing according to an embodiment of the present invention.

FIG. 9 is a diagram for explaining a compound noun phrase accent phrase segmentation process according to an embodiment of the present invention.

[Explanation of symbols]

100 Morphological analysis unit, word analysis means 110 word dictionary 200 Reading / prosodic information setting section 210 reading adding unit, reading reading adding means 220 Accent phrase division unit, accent phrase division means 230 Accent Type Setting Unit, Accent Type Setting Means 240 Pose setting part, Pose setting means 300 synthetic speech generation unit, speech synthesis means

Continuation of front page (56) Reference JP-A-2-93499 (JP, A) JP-A-2-5097 (JP, A) JP-A-3-37700 (JP, A) JP-A-3-58097 (JP , A) JP 4-36799 (JP, A) JP 5-134692 (JP, A) JP 6-118981 (JP, A) JP 6-149282 (JP, A) (58) Fields investigated (Int.Cl. ⁷ , DB name) G10L 13/00-13/08 G06F 17/21-17/28

Claims (2)

(57) [Claims] 1. A sentence containing Japanese kanji and kanaJapanese document of
In the synthetic speech output method that reads aloud with synthetic speech, Each word is recognized and recognized from the input Japanese kana mixed sentence.
Morphological analysis is performed on the specified words, and
Word analysis processing to obtain information of Each word analyzed in the word analysis process and the above for each word
Using the information, add a phonetic kana to the Japanese kana mixed sentence
Yomi kana giving process to give, Toki is used as a conjunctive sentence in the Japanese kana mixed sentence
Information of whether or not, dependency information in compound noun, the compound noun
Using the part-of-speech of the words that make up and the number of words in the compound
Accent phrase division processing to divide into accent phrases, Set the accent type for each divided accent phrase.
The xent type setting process, Set the pose information between the divided accent phrases.
Setting process ,The reading kana given in the reading kana addition process,
Using the cent type and the pose information, whether it is Japanese or not
Synthetic speech with prosodic information corresponding to various mixed sentences
Consists of voice synthesis processing, The accent phrase division processing is The verbs in a given compound word are
If it is, it depends on the type of part-of-speech of the word immediately preceding the time verb.
A word pattern that changes the word pattern generated by
Identification change process, The word preceding the verb is the word pattern identification change
Depending on the processing, the sa verb type noun or the dependency relation
It is judged to be a general noun, and it is a single noun that is one after the time.
When a word is at the end of the compound word, or
If the second word after is a suffix, adjunct, or symbol
In this case, the word and the word immediately preceding
To perform the trailing edge determination process that determines that the
A synthetic speech output method characterized by: 2.A Japanese document with a mixture of Japanese kanji and kana
A synthetic voice output device for reading out in a synthetic voice, Each word is recognized and recognized from the input Japanese kana mixed sentence.
For a given word Morphological analysis is performed for each word
Word analysis means for obtaining information of Each word analyzed by the word analysis means and the above for each word
Using the information, add a phonetic kana to the Japanese kana mixed sentence
Yomikana giving means to give, Toki is used as a conjunctive sentence in the Japanese kana mixed sentence
Information of whether or not, dependency information in compound noun, the compound noun
Using the part-of-speech of the words that make up and the number of words in the compound
Accent phrase dividing means for dividing into accent phrases, Set the accent type for each divided accent phrase.
Xent type setting means, Set the pose information between the divided accent phrases.
Setting means ,The reading kana given by the reading kana giving means,
Using the cent type and the pose information, whether it is Japanese or not
Synthetic speech with prosodic information corresponding to various mixed sentences
And a voice synthesizing means, The accent phrase dividing means, The verbs in a given compound word are
If it is, it depends on the type of part-of-speech of the word immediately preceding the time verb.
A word pattern that changes the word pattern generated by
Identification change means, The word preceding the verb is the word pattern identification change
Depending on the means, the sa verb noun or dependency
It is judged to be a general noun, and it is a single noun that is one after the time.
When a word is at the end of the compound word, or
If the second word after is a suffix, adjunct, or symbol
In this case, the word and the word immediately preceding
It must have a tail decision means to decide that the
And a synthesized voice output device.