JP3357796B2 - Speech synthesis apparatus and method for generating prosodic information in the apparatus - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION The present invention relates to a technique for obtaining a boundary strength between prosodic phrases composed of prosodic words in a sentence from a sentence structure when an arbitrary sentence is given, and using the boundary strength to obtain a prosody. The present invention relates to a voice synthesizing apparatus that synthesizes a highly natural voice by performing control, and a method for generating prosody information in the voice synthesizing apparatus.

[0002]

2. Description of the Related Art In general, it is unnatural to synthesize speech of a sentence that is a word chain using only rules in accordance with syllable-specific features in a word and accent-type features that are a word-specific feature. In speech synthesis of an arbitrary sentence, it is necessary to analyze a sentence syntactic structure in order to utter a synthesized speech with higher naturalness.
Sentence analysis results provide structural information for determining prosodic features, so for natural control of prosody (intonation), not only accurate syntactic structure but also semantic structure and contextual information It is said that it is necessary.

[0003] In a conventional text-to-speech conversion system, limited parsing is performed. These analyzes have been performed in various ways. A method of estimating prosodic features only from the features of two adjacent morpheme chains by classifying the parts of speech in detail has been proposed. There is also a method of performing prosodic control using the relation between three connected phrases.

[0004] When calculating the structure of the entire sentence, it is costly to eliminate the ambiguity of the sentence structure and uniquely determine the entire structure of the sentence. It is pointed out that the prosodic control is possible to some extent without obtaining the structure of the entire sentence.

[0005]

However, the pitch shape differs depending on whether the end of the sentence is "Is it?" In a sentence with a contrast structure, a pause is used to indicate the contrast part (△ part: “~ wa ~ (△) ~ wa ~”)
However, the pose of the sentence "~ wa ~ de ~" (△ part: "~
Is (△) ~~~)
If prosody control is performed using syntactic information with a limited sentence structure, the meaning will not be transmitted correctly, and it will be difficult to hear and unnatural monotonous utterances.

[0006] Also, when trying to generate a natural prosody of a sentence using the syntactic information of the entire sentence, for example, one prosody such as the sentence "All the reality has been twisted toward yourself." Syntactic structure information does not correspond to appropriate processing when the phrase becomes long.

SUMMARY OF THE INVENTION The present invention has been made in view of the above circumstances, and has as its object to provide a speech synthesis apparatus capable of realizing flexible prosody control for improving the naturalness of intonation of a whole sentence and a method for generating prosody information in the apparatus. Is to do.

[0008]

The speech synthesizer of the present invention performs a morphological analysis on input text information, and generates a prosodic word as a minimum unit for determining one accent type from the analysis result. A prosodic word forming means for forming a chain of the prosodic words, a prosodic word dependency strength determining means for obtaining a dependency strength between the prosodic words in the chain of the prosodic words, and a prosodic word between the prosodic words in the chain of the prosodic words. Prosodic inter-word dependency analysis means for obtaining dependency relations between prosodic words based on dependency strength, prosodic word count determining means for obtaining the number of prosodic words from the obtained dependency relations, and prosody word obtained Means for determining a connection strength between adjacent prosody words using at least one of a dependency strength between words, a dependency relationship, and the number of received words, Depending on the bonding strength Prosodic phrase boundary strength determination means for determining the boundary strength of prosodic phrase, in consideration of the prosodic phrase boundary strength is characterized in that a prosodic information generating means for generating a prosodic information of the text information.

In the speech synthesizing apparatus having such a configuration, an adjacent prosody word is obtained from a prosody word in a sentence by using at least one of information on the dependency strength between the prosody words, the dependency structure, and the number of receptions. By determining the strength of the connection between the prosodic phrases and determining the strength of the boundary between the prosodic phrases and using this strength of the prosody to generate prosodic information for prosodic control, natural prosody is achieved by simple processing using prosodic words as control units. Control becomes possible.

Here, in order to obtain the dependency strength between prosodic words, an interdependency table showing the interdependency between grammatical attributes of each prosodic word may be used. In addition, as a dependency relationship used for determining the strength of connection between adjacent prosodic words, the difference between the number of prosodic words in which the target prosodic word in the text information reaches the end of the sentence and the number of prosodic words in which the succeeding prosodic word reaches the end of the sentence is indicated. The depth of the boundary may be used.

Further, in the generation of the prosody information by the prosody information generating means, the prescribed number of mora, the number of mora of the prosody phrase, and the prosody phrase boundary strength, which are allowed in one expiration, are determined for the chain of prosody words. , A prosody control command to be set at the boundary between each prosody phrase and the following prosody phrase may be determined. For example, for a prosodic phrase having a mora number equal to or greater than the specified mora number, the prosodic words in the corresponding prosody section are separated based on the corresponding adjacent prosody word connection strength, and a mora number smaller than the specified mora number is determined. The prosodic phrase having the same prosody phrase may be combined with the following prosodic phrase based on the connection strength between adjacent prosodic words of the subsequent prosodic phrase to form one exhalation paragraph.

[0012]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a speech synthesizer according to one embodiment of the present invention.

The speech synthesizer shown in FIG.
It comprises a language processing unit 12, a phoneme processing unit 13, an acoustic parameter generation unit 14, a speech waveform generation unit 15, and an output unit 16.

Here, the prosodic words will be described before describing the above-mentioned components. A “prosodic word” is defined as a phrase phrase that is the smallest unit of meaning in a linguistic sound, in which a pitch change is recognized in the chain (referred to as an autonomous prosody), and in which a pitch change is not recognized (other law). A prosodic element that belongs to at least one autonomous prosodic element or base (a word that cannot be further divided), or at least one autonomous prosodic element or base And other ones that belong to one or more subsequent particles and affixes.

In general, a "morpheme" is a minimum unit of a meaningful linguistic form and is a component of a word (Kenji Morioka,
"Formation of vocabulary", Meiji Shoin 1987). Then, morphemes are classified into bases, affixes, and particles. A base is a form that indicates a concept or idea, and there are free form and bound form. When the morpheme is further decomposed, it becomes a syllable chain, which corresponds to a syllable chain. Therefore, morphemes and prosodic elements correspond to each other through a phrase chain. Autonomous and heterogeneous prosodic elements are used as basic units of tone processing in speech synthesis, and those composed of a chain thereof are used as prosodic words.

Next, each of the above components will be described. First, the language processing unit 12 is a part that generates a prosodic word from an arbitrary sentence (text information) input by the input unit 11, for example, a sentence mixed with a kanji kana, and determines a connection strength between adjacent prosodic words. The language processing unit 12 includes the prosodic word forming unit 1
21, a prosodic word dependency strength determining unit 122, a prosodic word dependency analyzing unit 123, a prosodic word receiving number determining unit 124, and an adjacent prosodic word coupling strength determining unit 125.

The prosodic word forming unit 121 receives the kanji kana mixed sentence sent from the input unit 11 and performs morphological analysis (registered in a morphological analysis dictionary (not shown)) on the kanji kana mixed sentence. Perform morphological analysis with reference to the information,
An autonomous / other rhythm element is determined, and based on a rhythm element chain, it is determined whether the element is an autonomous rhythm element or another rhythm element.
By other autonomous prosodic elements belonging to one autonomous prosodic element or base, or by other autonomous prosodic elements belonging to at least one autonomous prosodic element or base and one or more subsequent particles or affixes A prosodic word is constructed by combining prosodic and prosodic elements. Thus, the prosody word forming unit 121
Forms a chain of prosodic words.

The morphological analysis information includes grammatical attribute names, readings, morphological classifications (bases, particles, affixes), and prosodic classifications (autonomous prosody, other prosody) for each prosodic element. In addition, if it is an autonomous prosodic element, its basic accent type is included. By referring to the morphological analysis information, the prosodic word forming unit 121 sets and outputs attributes (reading, accent type, grammatical attribute, number of mora of the prosodic word) to the prosodic word.

A prosodic word dependency strength determining unit 122 receives the prosodic word chain formed by the prosodic word forming unit 121 as an input,
Based on the grammatical attribute information set for the prosodic words, a dependency between the two prosodic words is set with reference to an interdependence table (not shown) between the grammatical attributes. FIG. 2 shows an example of the contents of the interdependence table between the grammatical attributes (interdependency table).
Here, for each combination of the grammatical attributes of the two prosodic words,
The degree of dependency between the prosodic words (dependency) is 2
It is registered in advance in five stages of 0, 40, 60, 80, and 100.

The inter-prosodic word dependency strength determining unit 122 refers to the interdependence table between the grammatical attributes from the first prosodic word to the last prosodic word of the prosodic word chain formed by the prosodic word forming unit 121. By doing so, a dependency is set between all two prosodic words, and a dependency strength is set between the prosodic words.

The inter-prosodic dependency analysis unit 123 determines the dependency in the sentence from the dependency between all prosodic words from the first prosodic word to the last prosodic word in the prosodic word chain formed by the prosodic word forming unit 121. Analyze the receiving relationship.

In the analysis of the dependency relationship, a dependency relationship between prosodic words is determined using a principle shown in the example of FIG. 3 for reducing ambiguity in order to express a sentence structure. The principle is as follows.

(Principle 1) Dependency links do not cross (see FIG. 3 (a)) (Principle 2) Modification to the forward prosodic word is prohibited (FIG. 3)
(Refer to (b).) (Principle 3) A certain prosodic word does not involve two or more prosodic words at the same time (see FIG. 3 (c)). (Refer to FIG. 3 (d).) (Principle 5) Depend on one prosodic word and give priority to the closest one (see FIG. 3 (e)). In FIG. 4, the input kanji kana mixed sentence (input text information) In the sentence "I twisted all the reality toward myself."
3 shows an example of a dependency analysis result in Example 3. Here, “any” with consecutive numbers 1, 2, 3, 4, 5, 6
The words "reality,""all,""own,""toward," and "twisted" are prosodic words. From FIG. 4, it can be seen that, for example, the relation of “all” is “reality” with the number 2, and the dependency strength is 60.

The number-of-prosodic-words determining unit 124 determines the number of prosodic words received from the dependency relationship in the sentence analyzed by the inter-prosodic-word dependency analyzing unit 123. The number of receptions is the number of links relating to the prosodic words. FIG. 4 also shows an example of the number of receptions.

The adjacent prosody word bond strength determining unit 125 determines the dependency strength between the two prosodic words obtained by the prosodic word dependency strength determining unit 122 and the dependency relation obtained by the prosodic word dependency analyzing unit 123. The connection strength between adjacent prosodic words is determined based on the information on the depth of the boundary in the dependency structure obtained from the above and the number of receptions obtained by the prosodic word reception number determination unit 124. Here, the depth of the boundary means the number of prosodic words in the sentence (input text information) to be determined for the connection strength between adjacent prosodic words until the end of the sentence, and the number of prosodic words in the succeeding prosodic word until the end of the sentence. FIG. 4 also shows an example of the depth of the boundary.

The connection strength between adjacent prosodic words represents the strength of the dependency of the boundary between prosodic words, mainly based on linguistic information and taking into account syntactic information. a1) ~ (a5)
Are classified into five types.

(A1) Special Boundary The relationship with a non-language one (one that is not uttered as speech), such as a symbol, parenthesis, or quotation, corresponds to this.
In this case, the interdependence table corresponds to a prosody word dependency strength of 20. In addition to this, the relationship with the case where the number of subsequent prosodic words is 0 is also applicable.

(A2) Independence Boundary The relation between nouns that have particles such as "ha", "mo", and "about" and the verbs, and the relation between the conjunctive particles "kara", "ri", "shi", etc. In addition to relationships, this is the dependency between the title and the narrative. Here, the interdependency table has a prosodic inter-word dependency strength of 40. In addition,
This also applies to the relationship where the reading point and the depth of the boundary are 2 or more and the number of subsequent prosodic words is 1 or more.

(A3) Concatenation Boundary The relation between a situational complement and a verb, the connecting particle "NODE",
This includes the relationship between "so" and the verb, the relation between the conjunctive form of the verbal and the verbal, and the relationship between the adverb indicating intention and the verbal. Here, the prosodic word dependency strength on the interdependency relation table is equivalent to 60. In addition, the relationship where the depth of the boundary is 0 (the relationship between adjacent prosody words) also corresponds to this.

(A4) Connection boundary Relation between subject, object, complement and verb, relation between conjunctive particles "tsu", "tsutsu" and verb, relation between adverbs such as time and place and verb, adjunct This is the relationship between words and nouns. Here, the interdependency table has a prosodic inter-word dependency strength of 80. In addition, the boundary depth is 1
And the number of subsequent prosodic words received is one or more.

(A5) Combination Boundaries The prefixes, suffixes and nouns, relations between words, words and auxiliary verbs, chains of auxiliary words, and the like that constitute words are applicable to this. Here, the prosodic word dependency strength on the interdependence relation table is 100 (corresponding to one phrase when connected).

Here, the connection strength between adjacent prosodic words classified into the above (a1) to (a5) is such that the special boundary of (a1) is the weakest and (a2)
It becomes stronger in the order of → (a3) → (a4), and the bond boundary of (a5) is the strongest. The above-mentioned inter-prosodic inter-word connection strength (division) is the dependency strength between two prosodic words obtained by the prosodic word dependency strength determining unit 122 and the dependency relation obtained by the prosodic word dependency analyzing unit 123. Is determined by the adjacent prosody word connection strength determining unit 125 based on the information on the depth of the boundary in the dependency structure obtained from the above and the number of receptions obtained by the prosody word receiving number determination unit 124. The information on the strength of connection between adjacent prosody words in the prosodic word chain determined by the strength of connection between adjacent prosody words (within the language processing unit 12) is passed to the phoneme processing unit 13.

The phoneme processing unit 13 determines the boundary strength in the prosodic phrase based on the information on the connection strength between adjacent prosodic words in the prosodic word chain, and generates a phrase start position and a phrase command for generating a pitch pattern. Prosody information (prosody control information) such as size, pause position, and pause length is obtained. The phoneme processing unit 13 also obtains
Generate a prosody symbol string and a phoneme symbol string. This phoneme processing unit 1
The prosodic symbol string and the phoneme symbol string generated in step 3 are passed to the acoustic parameter generator 14.

The acoustic parameter generation unit 14 generates a prosodic symbol string and a phoneme symbol string passed from the phoneme processing unit 13 based on the
An acoustic parameter sequence for driving a speech synthesizer (not shown) in the speech waveform generator 15 is generated. This sequence is sent to the audio waveform generator 15 and output via the output unit 16.

The details of the phoneme processing unit 13 will now be described.
First, the phonemic processing unit 13 includes a prosodic phrase boundary strength determining unit 131,
Prosody information generation unit 132 and phoneme symbol / prosody symbol generation unit 1
33.

The prosodic phrase boundary strength determining unit 131 forms a prosodic phrase from a prosodic word chain, and determines the strength at the boundary of the prosodic phrase. A prosodic phrase is composed of one or more prosodic word chains,
Which prosodic words are to be combined is determined by referring to the prosodic phrase boundary strength determining unit 131 based on the coupling strength between adjacent prosodic words passed from the language processing unit 12 to the phoneme processing unit 13. Here, two prosody words having a connection strength higher than a predetermined value 100 (connection strength with a value of 100) are fused to generate a prosody phrase. Here, the value specified in advance may be a value appropriately designated.

The prosodic phrase boundary strength determining unit 131 determines the following prosodic phrase boundary strength at the generated prosodic phrase boundary with reference to the corresponding adjacent prosodic word connection strength passed from the language processing unit 12. I do. This corresponds to information for generating a pitch pattern in prosody information.

The above-described prosodic phrase boundary strength indicates the strength of a prosodic boundary, and is classified into the following four types of strength categories (b1) to (b4). (b1) Independent Boundary This mainly corresponds to a boundary where a large pause is inserted or a large phrase is started.

(B2) Connected Boundary This mainly corresponds to a boundary where a small pause is inserted or a phrase is started.

(B3) Boundary boundary This corresponds mainly to a boundary where a phrase is added. (b4) None Do not add both poses and phrases (do not form a phrase boundary)
Corresponds to the border.

The prosodic phrase boundary strengths of (b1) to (b4) and
The correspondence between (a1) to (a5) and the connection strength between adjacent prosody words is as follows. (b1) Independent boundary ← → (a1) Special boundary, (a2) Independent boundary (b2) Connected boundary ← → (a3) Connected boundary (b3) Connected boundary ← → (a4) Connected boundary (b4) None ← → (a5 ) Joining Boundary The prosodic phrase boundary strength determining unit 131 performs accent processing of each prosodic word constituting the prosodic phrase when determining the strength at the above-described prosodic phrase boundary, and determines the prosody in the prosodic word chain forming the prosodic phrase. Determine the reading of the entire phrase, as well as the number of mora. The result of determination by the prosody phrase strength determining section 131 is passed to the prosody information generating section 132.

The prosody information generating unit 132 responds to the input of the prosody phrase chain by setting a predetermined number of mora (a specified number of mora) allowed for one exhalation and a number of mora (prosody) already set in the prosody. With reference to the phrase mora number), based on the prosodic phrase boundary strength, what prosody control command is given to the boundary between the prosodic phrase and the following prosodic phrase (whether the prosodic control operation is performed)
Is determined.

That is, the prosody information generating unit 132 determines, for a prosodic phrase having a number of moras equal to or greater than the prescribed number of moras, the degree of connection between adjacent prosodic words between prosodic words in the prosodic word chain forming the prosodic phrase. By referencing, the prosodic words in the prosodic phrase can be separated, and conversely, for prosodic phrases with a small number of mora,
The subsequent prosodic phrases are merged into one exhalation paragraph by referring to the degree of connection between adjacent prosodic words, and an appropriate prosody control command is given to a phrase boundary.

As described above, the prosody information generating unit 132 continues the process of obtaining the prosody control commands of the phrase boundaries one after another while performing the recalculation until the processing of the final prosody phrase is completed. 5 and 6 show an algorithm for generating prosody information. Note that the value specified in advance may be a value specified as appropriate.

The prosody information indicates the approximate level of the prosody control command at each phrase boundary. The approximate level of the prosody control command includes the approximate level of the phrase start command, the approximate level of the pause (pause section length) insertion command, and the mora length command value (specified number of mora). There is an approximate level.

In the process of generating the prosody symbol string in the phoneme symbol / prosodic symbol generation unit 133, the position information (phrase start position, pause insertion position) of the prosody control command (phrase start command, pause insert command) is roughly calculated. The target level (the size of each command) is converted to a symbol. Similarly, a prescribed mora length command value such as a prescribed number of mora allowed for one exhalation is also indicated at an approximate level. This may be changed or designated as appropriate according to the utterance speed or the like.

In the present embodiment, the approximate levels of the magnitude of the phrase start command are (F1), (F2),
Three levels of (F3) are set, and the magnitude relation is (F1)>(F2)> (F3).

The approximate level of the magnitude of the pause insertion command is set at three levels (Po1), (Po2), and (Po3), and the magnitude relation is (Po1). )>
(Po2)> (Po3).

The approximate level of the mora length command value is set to three levels of S, T, and U.
The magnitude relation is U>T> S. Note that (F1) is inserted at the beginning of a sentence and (Po1) is inserted at the end of one sentence. It is assumed that (F1) and (Po1) are inserted at the reading points.

In the present embodiment, a chain of prosodic phrases in a sentence is given a number n (n = 1 to N, where N is the last prosodic phrase at the end) which is continuous from the beginning of the sentence. The generation of the prosody information at each prosodic phrase boundary by the prosody information generation unit 132 is 1
The processing starts from the ith prosody phrase and is performed according to the algorithm shown in FIG. 5 until the processing of the Nth prosody phrase ends. Hereinafter, the processing of the n-th prosodic phrase will be described according to the algorithm shown in FIG.

(Process 1) First, in the n-th prosodic phrase,
A process of acquiring (partition) the boundary strength at the boundary with the following prosody phrase from the result of the prosody phrase boundary strength determination unit 131 is performed (step S2). Note that the initial value of n is 1 (step S1).

(Process 2) Next, as a result of step S2, n
In the second prosodic phrase, if the boundary strength at the boundary (phrase boundary) with the following prosodic phrase is an independent boundary (Step S3), (F2) and (Po2) are inserted at the phrase boundary (Step S3). Steps S4 to S6).

However, if the distance from the immediately preceding phrase boundary (F1) or (F2) is equal to or less than S mora (step S
4) Insert (F3) and (Po2) at the phrase boundary (step S7).

If the distance to the end of the sentence immediately after is less than T mora, only (F3) is inserted at the phrase boundary (step S8). In this case, the distance from the immediately preceding independent boundary is U
If it is separated by more than a mora (step S9), (po3) is inserted at the connection boundary without pause (step S10).

Next, if n does not coincide with N, the n is incremented by 1 (steps S11 and S12), and the process returns to step S2 (of process 1) to continue the process. (Process 3) On the other hand, in the n-th prosodic phrase, if the (partition) of the boundary strength at the boundary with the following prosodic phrase is a connection boundary (not an independent boundary) (step S1)
3), (F3) is inserted at the phrase boundary (step S1)
4, S15).

However, if the distance from the immediately preceding phrase boundary (F1), (F2) or (F3) is equal to or less than S mora (step S14), nothing is inserted. This means that the subsequent prosodic phrase is merged into one exhalation paragraph. Next, n is incremented by 1 (step S12), and the process returns to step S2 (of process 1) to continue.

(Process 4) In the case where the distance from the immediately preceding connection boundary is a connection boundary of T mora or more (step S1)
6) At the connection boundary between the connection boundaries, (F3) is inserted at the boundary having the weakest prosodic word connection strength (steps S17 and S18). However, if the adjacent prosodic word binding strengths at each connection boundary are all equal, (F3) is inserted evenly into the connection boundary so that all prosodic phrase chains between the connection boundaries are equal to or less than T mora (step S19).

In the prosody information generating unit 132, the above series of processing is repeatedly executed N times from n = 1 to n = N (step S11). The processing result of the prosody information generation unit 132 is passed to the phoneme symbol / prosodic symbol generation unit 133. The phoneme symbol / prosodic symbol generation unit 133 includes a prosody information generation unit 132
A phonological symbol string is generated based on the pronunciation of each vocabulary composing text information obtained from the processing result in step 1, and the tonal change in the surface layer of the vocal sound corresponding to the reading, and based on the tonal properties of the prosodic phrase boundaries and the prosodic information. To generate a prosodic symbol string. Here, a syllable symbol, a symbol indicating an accent, and a symbol specifying a pause interval length corresponding to a speech sound are generated.

The phoneme symbol sequence and the prosody symbol sequence are passed to the acoustic parameter generation unit 14. Sound parameter generator 1
Reference numeral 4 generates an acoustic parameter sequence for driving the speech synthesizer from the phoneme symbol sequence and the prosody symbol sequence. This sequence is sent to the audio waveform generator 15 and output via the output unit 16.

The basic function of each component in the speech synthesizer having the configuration shown in FIG. 1 has been described above. Next, a specific example of the operation in the speech synthesizer will be described with reference to FIG. 7 and FIG.
The prosody processing for the same kanji kana sentence (a) shown in FIG.

First, the kanji kana sentence sentence sent from the input unit 11 is that all the realities have been twisted toward ourselves. The language processing unit 12 (the prosodic word forming unit 12 in the
In the morphological analysis process in 1), based on the morphological analysis results, at least one autonomous prosodic element or another autonomous prosodic element belonging to the base, or at least one autonomous prosodic element in a chain of autonomous / otherrhythmic prosody. By combining other prosodic elements belonging to a prosodic element or a base and one or more subsequent particles or affixes belonging to affixes,
Prosody words are composed.

The chain of prosodic words formed from the result of the morphological analysis of the kanji kana mixed sentence as described above is as follows. [Every] [reality |] [all] [your | of] [to | | [bent | |] Here, those separated by "|" Those enclosed by "and"] "indicate prosodic words.

Next, the language processing unit 12 (the prosodic word inter-dependence strength determining unit 122 therein) generates the interdependency table of FIG. 2 for the chain of the prosodic words (formed by the prosodic word forming unit 121). By referencing, the degree of dependence between prosodic words is determined in units of prosodic words. Here, for each prosodic word (excluding the last prosodic word), the dependency (dependency) in the combination between all succeeding prosodic words is shown in FIG.
(B) is obtained. The dependence strength is indicated by numerical values in five stages as described above. A value other than 0 indicates that there is a dependency, and a larger value indicates a stronger dependency.

Next, the language processing unit 12 (the inter-prosodic word dependency analysis unit 123 therein) obtains the dependency relationship based on the dependency strength between prosodic words as shown in FIG. 7B. Here, by applying a principle for reducing ambiguity in order to approximate the sentence structure, a dependency structure as shown in FIG. 7C is obtained.

Next, the language processing section 12 (in which the number of prosodic words received is determined by the section 124) receives the respective prosodic words as shown in FIG. 7D from the dependency structure shown in FIG. 7C. A number is required. This number of prosodic words indicates how much the prosodic word is related from the preceding prosodic word in the structure of the dependency.

Next, the language processing unit 12 (in which the inter-prosodic inter-prosodic word connection strength determining unit 125) has already performed (the prosodic inter-word dependency strength determining unit 122, the prosodic inter-word dependency analyzing unit 123, and the Using the information of the dependency strength between the prosodic words, the dependency structure (depth of the boundary determined from it), and the number of receptions obtained by the determining unit 124), the adjacent words as shown in FIG. The connection strength between the prosodic words is determined.

The phoneme processing unit 13 generates prosody information from the prosodic word connection strength (division) in the prosodic word chain obtained by the language processing unit 12 as described below.
First, the phonological processing unit 13 (the prosodic phrase boundary strength determining unit 13
In 1), the boundary strength (division) of the prosodic phrase is obtained based on the connection strength between adjacent prosodic words in the prosodic word chain shown in FIG. 8A. In this example, each prosodic word in the prosodic word chain is a prosodic phrase, and the boundary strength (division) of each prosodic phrase is as shown in FIG. 8B. Here, the reading of the entire prosodic phrase in the prosodic word chain constituting the prosodic phrase and the number of mora are also determined. See FIG. 4B for the number of moras.

Next, the phoneme processing unit 13 (the prosody information generating unit 132 in the phoneme processing unit) performs (the division of) the boundary strength in the prosody phrase.
And prosodic information is generated based on the information on the mora number. Here, according to the algorithm shown in FIGS. 5 and 6, FIG.
A prosody control command (phrase start command, pause insertion command) represented at an approximate level as shown in FIG. The number of mora indicated by S, T, and U is S = 4, T
= 15 and U = 35.

As described above, when the prosody information such as the phrase start position, the size of the phrase command, the pause position, and the pause length is obtained, the phoneme processing unit 13 (the phoneme symbol / prosody symbol generation unit 133 therein) Indicate the pronunciation of each vocabulary composing the text information, the tone change in the surface layer of the vocabulary to the pronunciation, the syllable symbol (a phonological symbol sequence consisting of), and the accent based on the tone properties and prosodic information of the prosodic phrase boundaries. A symbol (a prosody symbol string consisting of) that designates a symbol and a pause interval length is generated.

Then, in the acoustic parameter generation unit 14, the phoneme processing unit 13 (the phoneme symbol / prosodic symbol generation unit 133 therein)
Based on the prosody symbol sequence and the phoneme symbol sequence obtained in the above, an acoustic parameter sequence for driving the speech synthesizer is generated. The audio waveform generation unit 15 generates an audio waveform from the acoustic parameter sequence, and outputs the audio through the output unit 16.

As described above, in the present embodiment, from the prosodic words in the sentence, information on (at least one of) the dependency strength between the prosodic words, the dependency structure, and the number of receptions is used.
By determining the bond strength between adjacent prosodic words and determining the boundary strength between prosodic phrases, and performing prosodic control using this boundary strength, natural prosody control can be performed with simple processing using prosodic words as control units. It becomes possible.

The above-described speech synthesizer having the configuration shown in FIG. 1 includes a computer, for example, a personal computer 90 shown in FIG. 9 and an input unit 11, a language processing unit 12, a phoneme processing unit 13, an acoustic parameter generation unit 14, Waveform generator 1
5 and a recording medium on which a program for functioning as the output unit 16 is recorded, for example, a floppy disk (F
D) By using the 91, the floppy disk 91 is mounted on the personal computer 90, and the program recorded on the floppy disk 91 is read and executed by the personal computer 90.

[0073]

As described above, according to the present invention,
Analyzing text information with prosodic words as constituent units, setting the connection strength between adjacent prosodic words by the dependency relationship between prosodic words, dependency relationship, and the number of receptions, finds the connection strength of prosodic phrases based on it, By generating prosody information at phrase boundaries, natural prosody control of the entire sentence can be made possible.

[Brief description of the drawings]

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

FIG. 2 is a view showing an example of the contents of an interdependency table which is referred to when setting a dependency between prosodic words.

FIG. 3 is a diagram showing an example of principles applied to the analysis of dependency relations between prosodic words for reducing ambiguity in sentence structure.

FIG. 4 shows an example of input text information and the results of language processing performed by the language processing unit 12 on the input text information, such as the dependency strength between prosodic words, the dependency relation, the number of receptions, the number of mora of each prosodic word, and the boundary. The figure which contrasts and shows the example of the connection depth between adjacent prosody words.

FIG. 5 is a diagram showing a part of an algorithm for generating prosody information in a prosody information generating unit 132;

FIG. 6 is a diagram showing the rest of the algorithm for generating prosody information in the prosody information generation unit 132.

FIG. 7 is a diagram for explaining a specific example of the operation of the speech synthesizer having the configuration of FIG. 1;

FIG. 8 is a diagram for explaining a specific example of an operation in the speech synthesis device having the configuration of FIG. 1;

9 is a diagram showing the appearance of a personal computer on which a floppy disk storing a program for realizing the function of each unit of the speech synthesizer having the configuration shown in FIG. 1 is mounted.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 11 ... Input part, 12 ... Language processing part, 13 ... Phoneme processing part, 14 ... Sound parameter generation part, 15 ... Speech waveform generation part, 16 ... Output part, 121 ... Prosody word formation part, 122 ... Prosody word dependence intensity Determining unit 123: prosodic inter-word dependency analyzing unit 124: prosodic word receiving number determining unit 125: adjacent prosodic word connection strength determining unit 131: prosodic phrase boundary strength determining unit 132: prosodic information generating unit 133 ... Phonological / prosodic symbol generation unit.

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G10L 13/08

Claims (6)

(57) [Claims] 1. Prosody information and phoneme information are generated from input text information, sound parameters necessary for speech synthesis are generated from the prosody information and phoneme information, and a synthesizer is driven according to the sound parameters to generate a speech waveform. A morphological analysis is performed on the input text information, and a prosody word that is a minimum unit that determines one accent type is generated from the analysis result. Prosodic word forming means for forming a chain, prosodic word dependency strength determining means for obtaining the dependency strength between each prosodic word in the prosodic word chain formed by the prosodic word forming means, and said prosodic word dependency strength determining means Prosody word dependency analysis means for obtaining a dependency relationship between prosodic words based on the dependency strength between prosodic words in the chain of prosodic words obtained by A prosodic word receiving number determining means for determining the receiving number of each prosodic word in the chain of prosodic words from the dependency relation between prosodic words determined by the inter-prosodic dependency analyzing means, The prosody word dependency analysis unit and the prosody word reception number determination unit determine the prosody word dependency strength, the dependency relationship, and the number of receptions, respectively, using at least one of the prosody word. Means for determining the strength of connection between adjacent prosody words in the chain, and determining the strength of the boundary between prosodic phrases by the strength of connection between adjacent prosody words obtained by the means for determining strength of connection between adjacent prosody words. A prosodic phrase boundary strength determining means for determining, a prosody information generating means for generating prosodic information of the text information in consideration of the prosodic phrase boundary strength obtained by the prosodic phrase boundary strength determining means; A speech synthesizer comprising: 2. The method according to claim 1, further comprising an interdependence table indicating interdependence between grammatical attributes of each prosodic word, wherein said interprosodic word dependency strength determining means converts a prosodic word chain formed by said prosodic word forming means. 2. The speech synthesis apparatus according to claim 1, wherein the inter-prosodic word dependency strength is obtained by referring to the interdependence table based on grammatical attribute information of the prosodic word as an input. 3. The dependency relation used by the adjacent prosody word connection strength determining means for determining the strength of connection between adjacent prosody words is defined by the number of prosody words in which the target prosody word in the text information reaches the end of the sentence. 2. The speech synthesizer according to claim 1, wherein the following prosodic word is a depth of a boundary indicating a difference from the number of prosodic words up to the end of the sentence. 4. The prosodic information generating means calculates a prosodic phrase based on a predetermined prescribed number of mora, a number of prosodic mora, and a prosodic phrase boundary strength for a predetermined sequence of prosodic words. 2. The speech synthesizer according to claim 1, wherein a prosody control command to be set at a boundary between the prosody phrase and the following prosody phrase is determined. 5. The prosodic information generating means, for a prosodic phrase having a number of moras equal to or greater than the prescribed number of moras, determines a prosodic word in the prosodic area based on the corresponding connection strength between adjacent prosodic words. A prosodic phrase having a mora number smaller than the prescribed number of mora is merged with a subsequent prosodic phrase to form one exhalation paragraph based on the connection strength between adjacent prosodic words of the subsequent prosodic phrase. 5. The speech synthesizer according to claim 4, wherein: 6. Prosody information and phoneme information are generated from input text information, sound parameters necessary for speech synthesis are generated from the prosody information and phoneme information, and a synthesizer is driven according to the sound parameters to generate a speech waveform. A prosody information generation method applied to a speech synthesis device that generates and outputs a prosody word, wherein a morphological analysis is performed on the input text information, and a prosody word that is a minimum unit that determines one accent type from the analysis result. A first step of forming a chain of prosodic words, a second step of determining the dependency strength between each prosodic word in the chain of prosodic words, A third step of obtaining a dependency relationship between prosodic words based on the dependency strength; and the number of prosodic words in the chain of the prosodic words from the dependency relationship between the prosodic words. Determining a fourth step of determining a connection strength between adjacent prosody words in the chain of prosody words using at least one of a dependency strength between the prosodic words, a dependency relationship, and the number of receptions. A step of determining a boundary strength of a prosodic phrase based on a connection strength between the adjacent prosody words; and a seventh step of generating prosody information of the text information in consideration of the prosody phrase boundary strength. And a prosody information generating method.