JP2014115577A - Read-aloud sentence generation device for voice synthesis, and program of the same - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a read-aloud sentence generation device for voice synthesis capable of preventing duplicated generation of read-aloud sentences and reducing the number of generated read-aloud sentences.SOLUTION: A read-aloud sentence generation device 1 for voice synthesis includes: graph branch deleting means 11 for generating a branch deleted graph; integration graph generation means 12 for integrating an initial graph and the branch deleted graph with an integration graph; graph comparison means 14 for generating list exclusion information by comparing read-out text data and the integration graph; list exclusion means 15 for excluding words and phrases from the list based on the list exclusion information; conditional expression generation means 16 for generating a first conditional expression and a second conditional expression; minimum number of times of passing calculation means 17 for calculating the minimum number of times of passing by solving the first conditional expression and the second conditional expression; and read-out sentence generation means 18 for generating the same number of read-out sentences as the minimum number of times of passing.

Description

  The present invention relates to a speech synthesis read-out sentence generation device that generates a read-out sentence necessary for constructing a speech synthesis database and a program thereof.

  Conventionally, an invention for generating a text-to-speech for speech synthesis has been proposed (see Patent Document 1). The invention described in Patent Document 1 generates a conditional expression for generating a read-out sentence for speech synthesis from the minimum number of times a node passes in a directed graph, and solves the generated conditional expression by a simplex method. Is generated.

JP 2010-33462 A

  In the invention described in Patent Document 1, even when there is a sentence that has already been registered and read out in the speech synthesis database (hereinafter referred to as “read-out sentence”), this read-out sentence cannot be reflected. There is a problem that the same read-out sentence is generated twice and the number of read-out sentences increases.

  Therefore, an object of the present invention is to provide a speech synthesis read-out sentence generation device and a program therefor that can prevent duplicate generation of read-out sentences and reduce the number of read-out sentences generated.

  In view of the problems described above, the speech synthesis read-out sentence generation device according to the first invention of the present application is composed of a plurality of nodes indicating words included in a sentence and edges indicating connection relations between the nodes, and can be replaced with nodes. A speech synthesis speech generation device for speech synthesis that generates a speech for a speech synthesis database required for speech synthesis using a directed graph in which a sentence is represented by a directed graph to which a simple word or phrase is assigned. An integration unit, a text data input unit, a graph comparison unit, a list exclusion unit, a first conditional expression generation unit, a second conditional expression generation unit, a minimum passage number calculation unit, and a reading sentence generation unit are provided. It is characterized by that.

  According to such a configuration, the speech synthesizing speech generation apparatus receives a directed graph and a list that stores one or more words that correspond to the nodes of the directed graph and that are replaced by the nodes by the graph input unit. In addition, the speech synthesis read-out sentence generation device integrates (clusters) the directed graphs input to the graph input unit by the graph integration unit. Then, the text-to-speech read-out sentence generating apparatus receives read-out text data indicating the utterance content of the sentence already read out by the text data input means.

  In addition, the speech synthesizing device for speech synthesis includes, by the graph comparison unit, the read-out text data input to the text data input unit and the integrated directed graph, thereby including the read-out text data from the list. Ask for a phrase Then, the speech synthesis read-out sentence generation device excludes the phrase obtained by the graph comparison unit from the list by the list exclusion unit.

  Further, the speech synthesizing speech generation apparatus uses the first conditional expression generation unit to set the number of passages of the node to be equal to or more than the number of words stored in the list corresponding to the node and excluded by the list exclusion unit. A first conditional expression is generated. Then, the speech synthesizing speech generation device uses the second conditional expression generation means as the second conditional expression, a conditional expression in which the sum of the number of times of passing of the edges input to the node is equal to the number of times of passing of the node, A conditional expression is generated in which the sum of the number of passes of the output edge is equal to the number of passes of the node.

Here, the number of times the node has passed means that when connecting each node by replacing the content indicated by the node in the path connecting the start point to the single end point while maintaining the node and edge structure in the directed graph, Refers to the number of times a node passes.
In addition, the number of times the edge passes is determined by replacing the contents indicated by the node in the path connecting the start point to the single end point while maintaining the structure of the node and edge in the directed graph. Refers to the number of passes.

  In addition, the speech synthesizing speech generation apparatus for speech synthesis calculates the minimum number of passages that minimizes the number of passages at the head of the sentence so as to satisfy the first conditional expression and the second conditional expression by the minimum passage number calculation means. Here, since the first conditional expression and the second conditional expression are linear equations or linear inequalities, the minimum passage number calculating means can solve the first conditional expression and the second conditional expression using the simplex method. .

  In the speech synthesis speech generation device for speech synthesis, the speech generation unit changes the number of passages equal to the minimum number of passages calculated by the minimum passage number calculation unit and the combination of words stored in the list deleted by the list exclusion unit. Generate a spoken sentence.

In the speech synthesis device for speech synthesis according to the second invention of the present application, the graph integration unit determines whether the edge of the directed graph input to the graph input unit after the second branch is branched, and the edge branches. A graph branch deletion that generates a branch deleted graph in which the branch of the edge is deleted separately from the directed graph in which the edge branches from the directed graph in which the edge branches Means and DP matching method (Dynamic Programming Matching) to determine whether the comparison target graph pair consisting of one of the directed graph or the integrated graph first input to the graph input means and the branch deleted graph is similar, If the comparison target graph pairs are similar, combine the comparison target graph pairs into one integrated graph, and if the comparison target graph pairs are not similar, An integrated graph generation unit that sets each comparison target graph pair as a new integrated graph, and the graph comparison unit compares the integrated graph input from the integrated graph generation unit with the read-out text data. And
According to such a configuration, the speech synthesizing text generation device for speech synthesis can be integrated after deleting the branch of the edge of the input directed graph.

In the speech synthesizing speech generation device according to the third invention of the present application, when the first conditional expression generation unit divides the same node into a plurality of nodes in the branch deleted graph by the graph branch deletion unit, A first conditional expression is generated in which the sum of the number of passages of nodes corresponds to the same node and is equal to or greater than the number of words stored in the list excluded by the list excluding means.
According to this configuration, the speech synthesis read-out sentence generation device can generate the first conditional expression that reduces the number of read-out sentences to be finally generated.

Further, the speech synthesis read-out sentence generation device according to the fourth invention of the present application, when the integrated graph generation means includes a plurality of replaceable words / phrases in the node of the integrated graph, from separate lists storing a plurality of words / phrases, Generates a unified list that stores all combinations of words contained in separate lists, and applies a greedy algorithm to the unified list, thereby generating a minimal compact list for speech synthesis and supporting nodes and compact lists It is characterized by attaching.
According to such a configuration, the speech synthesizing text generation device for speech synthesis can reduce the number of words even if the node includes a plurality of words that can be replaced.

  Here, the speech synthesis read-out sentence generator according to the first invention of the present application is a speech synthesizer in which hardware resources such as a CPU (Central Processing Unit), a memory, and a hard disk provided in a computer are cooperatively operated as the respective means described above. It can also be realized by a read-out sentence generation program (the fifth invention of the present application). This program may be distributed through a communication line, or may be distributed by writing in a recording medium such as a CD-ROM or a flash memory.

According to the present invention, the following excellent effects can be obtained.
According to the first and fifth inventions of the present application, a speech synthesizing speech generation device and its program are integrated in a directed graph, and are included in the read text data based on the comparison result between the integrated directed graph and the read text data. Exclude words from the list. Thus, according to the first and fifth inventions of the present application, the speech synthesizing text generation device and its program generate a text to be read only with words and phrases that are not included in the text that has been read out. It is possible to prevent and reduce the number of read-out sentences generated.

According to the second aspect of the present invention, the speech synthesis read-out sentence generation device can reduce the number of read-out sentences to be generated by integrating after deleting the branch of the edge of the input directed graph.
According to the third aspect of the present invention, the speech synthesizing speech generation apparatus for speech synthesis uses, as the first conditional expression, the sum of the number of passages of the nodes divided into a plurality corresponds to the same node before the division, and after the exclusion Since a primary inequality that is equal to or greater than the number of words stored in the list is used, the number of read-out sentences to be generated can be reduced.
According to the fourth aspect of the present invention, the speech synthesizing speech generating apparatus for speech synthesis can reduce the number of replaceable words / phrases to one even if the node includes a plurality of replaceable words / phrases. Regardless, it is possible to share the generation process of the read-out sentence and simplify the configuration.

It is a block diagram which shows the structure of the read-out sentence production | generation apparatus for speech synthesis which concerns on embodiment of this invention. It is a figure explaining the directed graph input into the graph input means of FIG. 1, (a) is a 1st example of a directed graph, (b) is a 2nd example of a directed graph. It is a figure explaining the list input into the graph input means of FIG. FIGS. 2A and 2B are diagrams for explaining a first example of generation of a branch deleted graph in the graph branch deletion unit of FIG. 1, in which FIG. It is a figure explaining the 2nd example of the production | generation of the branch deletion completed graph in the graph branch deletion means of FIG. It is a figure explaining the production | generation of the integrated graph by the integrated graph production | generation means of FIG. 1, (a) is a comparison object graph pair, (b) is an integrated graph. It is a figure explaining the production | generation of the integrated graph by the integrated graph production | generation means of FIG. 1, (a) is a comparison object graph pair, (b) is an integrated graph. (A) is a figure explaining read-out text data, (b) is a figure which shows the directed graph which corresponds to the read-out text data of (a), (c) is a phrase included in the read-out text data. It is a figure which shows the list | wrist excluded. It is a flowchart which shows operation | movement of the speech synthesizing text generation apparatus for speech synthesis of FIG. It is a figure explaining the read-out sentence production | generation apparatus for speech synthesis which concerns on the modification 1 of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings as appropriate. In each embodiment, means having the same function are denoted by the same reference numerals and description thereof is omitted.

(Embodiment)
[Configuration of a text-to-speech generator for speech synthesis]
With reference to FIG. 1, the structure of the speech synthesizing text generating apparatus according to the embodiment of the present invention will be described.
As shown in FIG. 1, the speech synthesis read-out sentence generation device 1 generates a read-out sentence for a speech synthesis database necessary for speech synthesis using a directed graph representing a sentence. For this reason, the speech synthesis read-out sentence generation device 1 includes a graph input unit 10, a graph branch deletion unit 11, an integrated graph generation unit 12, a text data input unit 13, a graph comparison unit 14, and a list exclusion unit 15. A conditional expression generation unit 16, a minimum passage number calculation unit 17, and a reading sentence generation unit 18.
The graph branch deletion unit 11 and the integrated graph generation unit 12 correspond to the graph integration unit described in the claims.

  The graph input means 10 receives a directed graph G (FIG. 2) and a list L (FIG. 3). Then, the graph input unit 10 outputs the input directed graph G and the list L to the graph branch deletion unit 11.

<Specific examples of directed graphs and lists>
A specific example of the directed graph and the list will be described with reference to FIGS. 2 and 3 (see FIG. 1 as appropriate).
In FIG. 2, the head of the sentence in the directed graph G is illustrated as “START”, and the end of the sentence in the directed graph G is illustrated as “END”.

  As shown in FIG. 2, the directed graph G is a directed graph used in graph theory and expresses a sentence (for example, a read-out original of a weather forecast program). The directed graph G includes a node N and an edge E (an arrow connecting the nodes N).

The node N represents a part of a sentence and can replace a phrase included in the sentence.
The edge E indicates the connection relationship of the node N such as connection, branching and joining of the node N. Here, the connection of the node N means that the node N is connected one-to-one via the edge E. The branching of the node N is to branch the edge E from one node N and connect to a plurality of nodes N. Further, the merging of the nodes N means that the edges E from the plurality of nodes N are merged and connected to one node N.
The list L is associated with one or more nodes N in advance, and stores one or more words / phrases to be replaced with the associated node N.

In the example of FIG. 2A, the directed graph G ₁ includes a node N _1S indicating the head of a sentence, a node N ₁₁ indicating “near the center”, and a node N ₁₂ indicating “within [time] in the future” , A node N ₁₃ indicating “of [direction]”, a node N ₁₄ indicating “a strong wind is expected to blow”, and a node N 1 _E indicating the end of the sentence, for a total of six nodes N. In the directed graph G ₁ , the edge E connects the node N _1S to the node N _1E in order without branching and joining the node N.

Further, the node N indicates that the word / phrase included in the “[]” can be replaced with the word / phrase stored in the list L. For example, node N ₁₂ can replace the phrase [time] with the phrase stored in list L ₁ of FIG.
On the other hand, if node “N” is not included, node N indicates that the word cannot be replaced. For example, since the node N ₁₁ does not include “[]”, the phrase cannot be replaced, and always indicates “near the center”.

In the example of FIG. 2B, the directed graph G ₂ includes a node N _{2 S} indicating the head of the sentence, a node N ₂₁ indicating “near the center”, and a node N ₂₂ indicating “from [numerical value]”. , Node N ₂₃ indicating “within [time] in the future”, node N ₂₄ indicating “[wind speed] meters”, node N ₂₅ indicating “expected to blow strong wind”, and end of sentence The node N _2E is composed of a total of seven nodes N.

In the directed graph G ₂ , the node N ₂₁ branches to nodes N ₂₂ and N ₂₃ via the edge E. Further, the nodes N ₂₂ and N ₂₃ merge with the node N ₂₄ via the edge E. In other words, the directed graph _{G 2} is illustrates a sentence one node _{N 22} or the node _{N 23} is included in the selectively.

As shown in FIG. 3, the list _{L 1} is associated with the node _{N 12, N 23} in FIG.
The list L ₁ stores two words “12 hours” and “24 hours” corresponding to [time] of the nodes N ₁₂ and N ₂₃ . That is, the nodes N ₁₂ and N ₂₃ are replaced with the words and phrases stored in the list L ₁ to represent two types of “within the next 12 hours” and “within the next 24 hours”.
In FIG. 3, the word / phrase replaced at the node N is illustrated at the top of the list L in order to make the correspondence between the list L and the word / phrase replaced at the node N easier to understand.

Also, the list _{L 2} are associated with the node _{N 13.} The list L ₂ stores 16 words (16 directions) corresponding to the [direction] of the node N ₁₃ such as “east-northeast”, “east”, and “east-southeast”.
Also, the list _{L 3} is associated with the node _{N 22.} The list L ₃ stores one word / phrase “15” corresponding to [numerical value] of the node N ₂₂ .
The list L ₄ is associated with the node N ₂₄ . The list L ₄ stores two words “20” and “30” corresponding to [wind speed] of the node N ₂₄ .
Note that the node N and the list L need not correspond one-to-one. For example, the nodes N ₁₁ , N ₁₄ , and N ₂₅ do not correspond to any list L.

The following description is based on the assumption that the directed graphs G ₁ and G ₂ in FIG. ₂ and the lists L _{1 to} L _{4 in} FIG.
The directed graph G that is first input to the graph input unit 10 may be referred to as an “initial graph”, and the directed graph G that is input after the second may be referred to as a “branch deletion target graph”. In Figure 2, directed graph G ₁ is with the initial graph, a directed graph G ₂ is branched deleted graph.
In the present invention, it goes without saying that the directed graph G and the list L are not limited to the examples of FIGS.

Returning to FIG. 1, the description of the configuration of the speech synthesis read-out sentence generation apparatus 1 will be continued.
The graph branch deletion unit 11 determines whether the edge of the branch deletion target graph input from the graph input unit 10 is branched.
When the edge is not branched, the graph branch deletion unit 11 treats the branch deletion target graph as it is as a branch deleted graph.
On the other hand, when the edge is branched, the graph branch deletion unit 11 generates a branch deleted graph in which each node at the branch destination of the edge is separately included from the branch deletion target graph.
Thereafter, the graph branch deletion unit 11 outputs the list L, the initial graph, and the branch deleted graph input from the graph input unit 10 to the integrated graph generation unit 12.

<Generation of branch deleted graph: first example>
With reference to FIGS. 4 and 5, two specific examples of generation of the branch deleted graph will be described (see FIG. 1 as appropriate).
As illustrated in FIG. 4A, the graph branch deletion unit 11 determines that the edge E branches at the node N ₂₁ . In this case, as shown in FIG. 4B, the graph branch deletion unit 11 branches from the branch deletion target graph G ₂ so that the nodes N ₂₂ and N _{23 at} the branch destination of the edge E are separated. Deleted graphs G ₃ and G ₄ are generated.

Specifically, the graph branch deletion unit 11 directly deletes the nodes N _2S , N ₂₁ , N _{24 to} N _2E where the edge E is not branched in the branch deletion target graph G ₂ , and the branch deletion deleted graphs G ₃ , G _4. To copy. Further, the graph branch deletion unit 11 connects one branched node N ₂₂ to the preceding and following nodes N ₂₁ and N ₂₄ in the branch deleted graph G ₃ . Further, the graph branch deletion unit 11 connects the other branched node N ₂₃ to the preceding and following nodes N ₂₁ and N ₂₄ in the branch deleted graph G ₄ .

In other words, the branch Deleted graph _{G 3} in FIG. 4 (b), the node _N 2S branch deletion graph _{G 2,} and _{N 21,} and one node _{N 22} which is branched, is composed of a node _N 24 _{to N 2E} The The branch Deleted graph _{G 4} are composed of the node _N 2S branch deletion graph _{G 2,} and _{N 21,} the other nodes _{N 23} branched, the node _N 24 _{to N 2E.}
In this way, the speech synthesis read-out sentence generator 1 can integrate the directed graph G after deleting the branch of the edge E.

<Generation of branch deleted graph: second example>
Further, as the branch deletion graph G ₅ in FIG. 5, a case where the edge E is branched at a plurality of locations.
First, the graph branch deletion unit 11 starts from the START node N _5S in the branch deletion target graph G ₅ , selects an appropriate edge E when there is a branch of the edge E, and selects the END node N _5E. Search for a path to reach. In this pass, the graph branch deletion unit 11 generates one branch deleted graph (not shown). Next, the graph branch deletion unit 11 returns from the node N _{5E of} END to the previous node N and determines whether there is a branch of the edge E.

If there is no edge E branch or an edge E that has not been traced, the graph branch deletion unit 11 returns to the previous node N and repeats the same. If there is an edge E that has not been traced, the graph branch deletion unit 11 selects this edge E and searches for a path that reaches the node N of END.
Here, the graph branch deletion means 11 saves the path history at the location where the edge E branches when searching for a path. The stored history group matches the history of the current path. If there is nothing to do, it is determined that the edge E has not been traced.

In the example of FIG. 5, it is assumed that the path starts from the node N _{5 S of} START and passes through the edge E to the node N ₅₂ . In this case, the graph branch deletion unit 11 stores, in the node N ₅₁ , a history of passing through the edge E to the node N ₅₂ . Next, assume that the path has passed edge E to node N ₅₅ . Then, the graph branch deletion unit 11 stores in the node N ₅₄ a history of passing through the edge E from the node N ₅₁ to the node N ₅₂ and passing through the edge E to the node N ₅₅ . Since this path reaches the node N _5E of the END, the graph branch deletion unit 11 generates a branch deleted graph passing through the nodes N ₅₂ and N ₅₅ .

Next, the graph branch deletion unit 11 returns to the node N ₅₄ and examines the history. As a result, the graph branch deletion means 11, because there is no history of the edge E to the node _{N 56,} select the edge E to the node _{N 56,} as it enters the node _{N 5E} of END, the node _{N 52, N 56} Generate a branch deleted graph through.
Similarly, the graph branch deletion unit 11 generates a branch deleted graph passing through the nodes N ₅₂ and N ₅₇ .

Subsequently, the graph branch deletion unit 11 has already traced the edge E to the nodes N ₅₅ , N ₅₆ , and N ₅₇ and returns to the node N ₅₁ . Here, the graph branch deletion means 11, because there is no history of the edge E to the node N _53, selects the node N _53. The graph branch deletion unit 11 searches for a path to the node N _5E of END. At this time, the graph branch deletion unit 11 examines the history at the node N ₅₄ , passes through the edge E from the node N ₅₁ to the node N ₅₂ , and passes through the edge E from the node N ₅₅ to the node N _51. through the edge E to the node _{N 52,} and history through the edge E to the node _{N 56,} through the edge E to the node _{N 52} from the node _{N 51,} the history is found that through the edge E to the node _{N 57} but the nodes through edges E from _{N 51} to the node _{N 53,} since it is not found history through the edge E to the node _{N 55,} through the node _{N 55,} enters the node _{N 5E} of END, the node _{N 53} , N ₅₅ to generate a branch deleted graph.

Next, the graph branch deletion unit 11 returns to the node N ₅₄ and examines the history. As a result, the graph branch deletion unit 11 through the edge E to the node _{N 53} from the node _{N 51,} because there is no history of the edge E to the node _{N 56,} select the edge E to the node _{N 56,} as it is A node N _5E of END is entered, and a branch deleted graph passing through the nodes N ₅₂ and N ₅₆ is generated.
Similarly, the graph branch deletion unit 11 generates a branch deleted graph passing through the nodes N ₅₂ and N ₅₇ .

In this way, the graph branch deletion unit 11 combines all of the branch destination nodes N ₅₂ , N ₅₃ , N _{55 to} N ₅₇ from the branch deletion target graph G ₅ (2 × 3 = 6), and deletes the branch. Generated graphs.

Returning to FIG. 1, the description of the configuration of the speech synthesis read-out sentence generation apparatus 1 will be continued.
The integrated graph generation means 12 obtains the node match, insertion error, missing error, and substitution error of the comparison target graph pair by the DP matching method, and determines whether or not the comparison target graph pair is similar. .

  This comparison target graph pair is a pair of directed graphs G to be compared by the integrated graph generation means 12. This comparison target graph pair is a pair of the initial graph G and the branch deleted graph G when the integrated graph T is not generated. Further, when the integrated graph T is generated, the comparison target graph pair is a pair of the integrated graph T and the branch deleted graph G.

When the comparison target graph pairs are similar, the integrated graph generation unit 12 integrates the comparison target graph pairs into one integrated graph T (FIG. 6) based on the result of the DP matching method. Furthermore, when there are a plurality of integrated graphs T, a plurality of comparison target graph pairs are also generated, so the integrated graph generation unit 12 integrates the integrated graph T with the highest degree of correctness by DP matching.
On the other hand, when all the comparison target graph pairs are not similar, the integrated graph generation unit 12 adds the directed graph G compared to the integrated graph T in the comparison target graph pair to the integrated graph T.
Thereafter, the integrated graph generation unit 12 outputs the list L and the integrated graph T input from the graph input unit 10 to the text data input unit 13.

<Generation of integrated graph>
The generation of the integrated graph will be described with reference to FIGS. 6 and 7 (see FIG. 1 as appropriate).
Here, the integrated graph generation means 12 will be described assuming that the initial graph G ₁ is input first, and the second and third branch deleted graphs G ₃ and G ₄ are input.

First, since the integrated graph T is not generated, the integrated graph generation unit 12 compares the initial graph G ₁ and the second branch deleted graph G ₃ with the comparison graph pair as shown in FIG. Let G ₁ and G ₃ . The integrated graph generation unit 12 then matches the node N match (H), insertion error (I), missing error (D), and substitution error (S) in the comparison target graph pair G ₁ and G ₃ . Is obtained by DP matching.

In this case, the match (H) is “2” between the nodes N ₁₁ and N ₂₁ indicating “near the center” and the nodes N ₁₄ and N ₂₅ indicating “expected to blow strong wind”. Moreover, alternative error (S), compared node _{N 12} indicating "within the next [Time]", and the node _{N 22} indicating the "from the Numeric]", the node _{N 13} indicating the "in the direction '' On the other hand, it becomes “2” with the node N ₂₄ indicating “[wind speed] meter”. Further, the insertion error (I) and the missing error (D) are “0”.

The integrated graph generation unit 12 calculates the correctness by substituting the DP matching result into the following equation (1). For example, when the DP matching result is substituted into equation (1), the accuracy is (2-0) / (2 + 2 + 0) = 0.5.
Accuracy = (HI) / (H + S + D) (1)

Further, the integrated graph generation unit 12 determines whether or not the correctness is equal to or higher than a preset threshold (for example, 0.3). In this example, since the accuracy = 0.5 is equal to or greater than the threshold = 0.3, the integrated graph generation unit 12 integrates the comparison target graph pair G ₁ and G ₃ into one integrated graph T.

In FIG. 6A, the nodes N ₂₂ and N ₂₄ become substitution errors (S) with respect to the nodes N ₁₂ and N _{13 of the first} initial graph G ₁ . Therefore, the integrated graph generation unit 12 adds the node N ₂₂ to be parallel to the node N ₁₂ and adds the node N ₂₄ to be parallel to the node N ₁₃ based on the _first initial graph G _1. To do. Here, when the substitution error (S) continues, the integrated graph generation means 12 does not merge the edge E from the node N ₂₂ to the node N ₁₃ , and does not merge the edge E connected from the node N ₂₂ to the node N _24. to add.

  In FIG. 6A, an example of the substitution error (S) is given. However, when it is determined as a missing error (D), the integrated graph generation unit 12 determines that the node N is determined as a missing error (D). Edge E is added so as to skip. Here, when a plurality of missing errors (D) continue, the integrated graph generation means 12 adds an edge E so as to skip these nodes N, as in the case of the substitution error (S).

  If it is determined that there is an insertion error (I), the integrated graph generation unit 12 adds the node N to the location determined as the insertion error (I), and also adds edges E to both ends of the added node N. Here, when a plurality of insertion errors (I) continues, the integrated graph generation means 12 inserts these nodes N continuously as in the case of the substitution error (S).

If the degree of correctness is less than the threshold (for example, when threshold = 0.6), the integrated graph generation unit 12 converts each of the comparison target graph pairs G ₁ and G ₃ into separate integrated graphs T. Treat as.

Then, the integrated graph generation unit 12, with respect to the third and subsequent branching deleted graph G _4, as a base integrated graph T, continue to add nodes N. In other words, integrated graph generation unit 12, as shown in FIG. 7 (a), an integrated graph T, 3-th branch Deleted graph G ₄ and comparative graphs pair T, and G _4. Then, the integrated graph generation unit 12 determines the node N match (H), insertion error (I), missing error (D), and substitution error (S) in the comparison target graph pair T, G ₄ . And obtained by DP matching.

In this case, the coincidence (H) indicates that the nodes N ₁₁ and N ₂₁ indicating “near the center”, the nodes N ₁₂ and N ₂₃ indicating “within [time]”, and “a strong wind is expected to blow”. It becomes '3' with nodes N ₁₄ and N ₂₅ indicating Moreover, alternative error (S), compared node N ₁₃ indicating the "in the direction]", becomes the node N ₂₄ indicating the "[wind] Meters"'1'. Further, the insertion error (I) and the missing error (D) are “0”.

When the result of this DP matching is substituted into the above-described equation (1), the accuracy is (3-0) / (3 + 1 + 0) = 0.75. In this example, since the accuracy = 0.75 is equal to or greater than the threshold value = 0.3, the integrated graph generation unit 12 integrates the comparison target graph pair T and G ₄ into one integrated graph T.

In FIG. 7A, the nodes N ₁₃ and N ₂₄ become substitution errors (S). Therefore, as shown in FIG. 7B, the integrated graph generation unit 12 branches the two edges E from the node N ₁₂ so that the nodes N ₁₃ and N ₂₄ are parallel to each other. Add to.
Note that the integrated graph generation unit 12 has already integrated the nodes N ₁₁ and N ₂₁ , the nodes N ₁₂ and N _23, and the nodes N ₁₄ and N ₂₅ that are matched (H) at the stage of FIG. 6B. Since it is added to the graph T, no processing is performed.

Hereinafter, in order to simplify the description, in the integrated graph T of FIG. 7B, the code of the node N _1S is N _S , the code of the node N ₁₁ is N ₁ , the code of the node N ₁₂ is N ₂ , and the node N ₁₃ the code _{N 3,} the node codes _{N 4} nodes _{N 24} following the _{N 2,} node codes _{N 5} of _{N 22,} reference numerals the _{N 6} of the node _{N 24} following the node _{N 5,} the sign of the node _{N 14} N _7, called the sign of the node _{N 1E} and _{N E.}

Further, in the integrated graph T in FIG. 7 (b), the node _{N S} and node _{N 1 E S1} the sign of the edge connecting the, _{E 12} the sign of the edge that branches from the node _{N 1} to the node _{N 2,} node N E ₁₅ is an edge code that branches from ₁ to node N ₅ , E ₂₃ is an edge code that branches from node N ₂ to node N ₃ , E ₂₄ is an edge code that branches from node N ₂ to node N _4, and The sign of the edge connecting N ₅ and the node N ₆ is E ₅₆ , the sign of the edge joining from the node N ₃ to the node N ₇ is E ₃₇ , and the sign of the edge joining the node N ₄ to the node N ₇ is E _47. , the node _{N E 67} the sign of the edge that joins the node _{N 7} to _6, referred to as codes _{E 7E} of edges connecting the nodes _{N 7} and node _{N E.}

Returning to FIG. 1, the description of the configuration of the speech synthesis read-out sentence generation apparatus 1 will be continued.
The text data input means 13 is input with the read-out text data and outputs the input read-out text data to the graph comparison means 14.

  The read-out text data is text data indicating the utterance content of the read-out sentence. In other words, the read-out text data is, for example, text data of a read-out sentence in which the speech read out by the announcer is already registered in the speech synthesis database and does not need to be newly generated.

  The graph comparison unit 14 obtains words / phrases included in the text data from the list L by comparing the integrated graph T input from the integrated graph generation unit 12 with the read-out text data.

<Comparison of integrated graph and read-out text data>
A comparison between the integrated graph and the read-out text data will be described with reference to FIGS. 7 and 8 (see FIG. 1 as appropriate).
Here, the graph comparison unit 14 receives the lists L _{1 to} L _{4 in} FIG. 3 from the integrated graph generation unit 12, and the integrated graph T in FIG. 7B and the read-out text data in FIG. Will be described as a comparison.
In FIG. 8B, “[numerical value] =” and “[wind speed] =” are illustrated for easy understanding of the description.

As shown in FIG. 7B, the integrated graph T has a plurality of paths due to the branching of the edge E. In the integrated graph T of FIG. 7, a path passing through the nodes N _{1 to} N ₃ and N ₇ is called “path 1”, and a path passing through the nodes N ₁ , N ₂ , N ₄ and N ₇ is “path 2”. A path passing through the nodes N ₁ and N _{5 to} N ₇ is referred to as “path 3”.

When the integrated graph T has a plurality of paths, the graph comparison unit 14 determines which path matches the read-out text data. First, a graph comparing means 14, the path 1 of the integrated graph T, is compared with the reading preformatted text data, "in the center vicinity" of the node N ₁ up to but consistent, it will not match in the seventh character. For this reason, the graph comparison means 14 determines that the pass 1 does not match the read text data at that time, and ends the comparison between the pass 1 and the read text data. Next, when comparing the pass 2 and the read-out text data, the graph comparison unit 14 does not match at the seventh character as in the pass 1. For this reason, the graph comparison unit 14 determines that the pass 2 does not match the read text data at that time, and ends the comparison between the pass 2 and the read text data.

Finally, the graph comparison unit 14 compares the pass 3 with the read-out text data. Here, it is assumed that the path 3 of the integrated graph T includes [numerical value] = 15 and [wind speed] = 20. In this case, “near the center” of the node N ₁ matches up to the ending, and ““ [numerical value] = ”15” of the node N ₅ matches up to the ending, and ““ [wind speed] = ”20 of the node N ₆ meters of "do not match up to the ending," is expected to strong wind blows "of the node N ₇ is matched to endings, leading to the node N _E of END. For this reason, the input sentence addition graph comparison unit 40 determines that the path 3 and the read-out text data match.

The graph comparing means 14, the comparison result of the above, generated the phrase list L ₃ "15", and the phrase list L ₄ "20", but the list exclusion information indicating that contained in the speech already text data To do. Thereafter, the graph comparison unit 14 outputs the list L and the integrated graph T input from the integrated graph generation unit 12 and the generated list exclusion information to the list exclusion unit 15.
This list exclusion information is information for excluding words / phrases from the list L, which is used by the list exclusion means 15 described later.

  When [Numeric Value] = 15 and [Wind Speed] = 20 are included, if the path 3 of the integrated graph T does not match the read-out text data, it is assumed that another word / phrase stored in the list L is included. Repeat the above comparison.

Returning to FIG. 1, the description of the configuration of the speech synthesis read-out sentence generation apparatus 1 will be continued.
The list excluding unit 15 excludes words included in the read-out text data from the list L based on the list excluding information input from the graph comparing unit 14.

In this example, the list exclusion information, the phrase word "15" and a list L ₄ of the list L ₃ "20" is included in the reading preformatted text data. Accordingly, as shown in FIG. 8C, the list excluding unit 15 excludes the phrase “15” stored in the list L ₃ and excludes the phrase “20” stored in the list L ₄ . Then, the list excluding unit 15 outputs the lists L ₁ and L ₂ and the integrated graph T input from the graph comparing unit 14 and the lists L ₃ and L ₄ excluding the words to the conditional expression generating unit 16.

The conditional expression generation unit 16 generates a first conditional expression and a second conditional expression that are necessary for the reading sentence, and includes a first conditional expression generation unit 16a and a second conditional expression generation unit 16b.
The conditional expression generation unit 16 generates the lists L _{1 to} L ₄ and the integrated graph T input from the graph comparison unit 14, the first conditional expression generated by the first conditional expression generation unit 16 a, and the second conditional expression generation. The second conditional expression generated by the means 16 b is output to the minimum passage number calculating means 17.

<Generation of the first conditional expression>
Hereinafter, the generation of the first conditional expression will be specifically described.
For each node N of the integrated graph T, the first conditional expression generation unit 16a generates a first conditional expression that sets the number of times of passing F of the node N to be equal to or greater than the number of words stored in the list L corresponding to the node N. Is.
Here, the conditional expression generation unit 16 will be described on the assumption that the integrated graph T in FIG. 7B and the lists L _{1 to} L _{4 in} FIG. 8C have been input.

Specifically, the first condition expression generating unit 16a is a node for the list corresponding to the N _S L is absent, and the number of words stored in the list L '0'. Therefore, the first condition generation unit 16a, as in equation (1), pass count F _S node N _S generates a primary inequalities to be '0' or more. The first condition expression generating unit 16a, like the node N _S, as in equation (2), pass count F ₁ of the node N ₁ generates primary inequalities to be '0' or more.
F _S ≧ 0 (1)
F ₁ ≧ 0 Equation (2)

The first condition expression generating unit 16a, in order to correspond the list L ₁ to the node N _2, to count the number of words stored in the list L _1. The first condition expression generating unit 16a, as in equation (3), to generate a primary inequality pass count F ₂ node N ₂ is stored the number of terms '2' or higher in the list L _1. Furthermore, the first conditional expression generating unit 16a is similar to the node _{N 2,} as in Equation (4), the node _N number of words passing number _{F 3} are stored in a list _{L 2} of the ₃ '16' above Generate a linear inequality.
F ₂ ≧ 2 Formula (3)
F ₃ ≧ 16 Formula (4)

Here, the node _N 4, _{N 6,} after the node _{N 24} were identical branched deleted graph _{G 2} is, which is divided into separate nodes _{N 24} branched deleted graph _G 3, _{G 4,} integrated graph T (See FIG. 4). In this case, the first condition generation unit 16a, as in Equation (5), the sum of the number of passes _F 4, _{F 6} nodes _N 4, _{N 6} are in the list ₄ corresponding to the node _N 4, _{N 6} A linear inequality is generated that is greater than or equal to the number of stored words '1'.
F ₄ + F ₆ ≧ 1 Formula (5)

The first condition generation unit 16a, as in equation (6), the number of passes F ₅ node N _5, the node number of words stored in the list ₃ corresponding to N ₅ '0' or Generate a linear inequality. Since this list ₃ does not store words, the number of words is “0”.
F ₅ ≧ 0 Formula (6)

Similarly to the node N ₁ , the first conditional expression generation unit 16 a generates a linear inequality such that the number of passes F ₇ of the node N ₇ is equal to or greater than “0”, as in the expression (7). The first condition expression generating unit 16a is similar to the node N _1, as shown in equation (8), to produce a primary inequality pass count F _E becomes "0" or more nodes N _E.
F ₇ ≧ 0 Equation (7)
F _E ≧ 0 Equation (8)

<Generation of second conditional expression>
Hereinafter, the generation of the second conditional expression will be specifically described.
The second conditional expression generating unit 16b uses, as a second conditional expression, a conditional expression in which, for each node N of the integrated graph T, the sum of the number of times H of the edge E input to the node N is equal to the number of times F of the node N is passed. Is generated.
In addition, the second conditional expression generation unit 16b generates a conditional expression in which the sum of the number of passages H of the edge E output from the node N is equal to the number of passages F of the node N as the second conditional expression.
That is, the second conditional expression generation unit 16b generates a second conditional expression for one node N on both the input side and the output side.

Specifically, the second condition generating unit 16b, as the second condition of the output side of the node _{N S,} as in equation (9), number of passages _{H S1} of edge _{E S1} output from the node _{N S} is , it generates the equal expression to pass number _{F S} node _{N S.} Here, the node a second condition on the input side of the N _S, since the node N _S is not present input a top not generated.
F _S = H _S1 Formula (9)

The second condition generating unit 16b, as the second condition on the input side of the node _{N 1,} as shown in equation (10), number of passages _{H S1} of edge _{E S1} is inputted to the node _{N 1,} the node Generate an expression equal to the number of passes F ₁ of N ₁ .
F ₁ = H _S1 Formula (10)

The second condition generating unit 16b, as the second condition of the output side of the node _{N 1,} as shown in equation (11), number of passes _{H 12} and the edge _{E 15} of the edge _{E 12} output from the node _{N 1} the sum of the number of passages H ₁₅ of, generates the equal expression to pass number F ₁ of the node N _1.
F ₁ = H ₁₂ + H ₁₅ Formula (11)

Similarly to the node N ₁ , the second conditional expression generation unit 16 b generates the second conditional expression on both the input side and the output side of the node N ₂ as in the expressions (12) and (13). .
F ₂ = H ₁₂ Formula (12)
F ₂ = H ₂₃ + H ₂₄ Formula (13)

Further, the second conditional expression generation unit 16b uses the second condition expression on the input side of the node N ₃ as the expression (14), and the number of passes H ₂₃ of the edge E ₂₃ input to the node N ₃ generating a equal expression to pass number _{F 3} of N _3.
F ₃ = H ₂₃ Formula (14)

Further, the second conditional expression generation unit 16b sets the number of passes H ₃₇ of the edge E ₃₇ output from the node N ₃ as the second conditional expression on the output side of the node N ₃ as represented by the expression (15). _An expression equal to the number of passes F ₃ of ₃ is generated.
F ₃ = H ₃₇ Formula (15)

Similarly to the node N ₃ , the second conditional expression generation unit 16 b performs the first operation on both the input side and the output side of the nodes N ₄ , N ₅ , and N ₆ as in the expressions (16) to (21). Two conditional expressions are generated.
F ₄ = H ₂₄ Formula (16)
F ₄ = H ₄₇ Formula (17)
F ₅ = H ₁₅ Formula (18)
F ₅ = H ₅₆ Formula (19)
F ₆ = H ₅₆ Formula (20)
F ₆ = H ₆₇ Formula (21)

The second condition generating unit 16b, as the second condition on the input side of the node _{N 7,} as shown in equation (22), and number of passes _{H 37} edges _{E 37} is input to the node _{N 7,} the edge a number of passages _{H 47} of E _47, the sum of the number of passages _{H 67} edges _{E 67} generates a equal expression to pass number _{F 7} node _{N 7.
F 7 = H 37 + H 47} + H 67 ... formula (22)

Further, the second conditional expression generation unit 16b uses the second condition expression on the output side of the node N ₇ as the second conditional expression of the node N ₇ , as shown in Expression (23), the number of passes H _7E of the edge E _7E output from the node N ₇ is _An expression equal to the number of passes F ₇ of ₇ is generated.
F ₇ = H _7E Formula (23)

The second condition generating unit 16b, as the second condition on the input side of the node _{N E,} as in equation (24), number of passages _{H 7E} edge _{E 7E} is inputted to the node _{N E,} the node An expression is generated that is equal to the number of times N _E passes F _E. Here, the node a second condition of the output side of the N _E, since the node N _E is the last output are not present, are not generated.
F _E = H _7E Formula (24)

Minimum pass frequency calculation unit 17, the minimum that solves the first condition and a second condition that is input from the conditional expression generating unit 16 in the simplex method, the number of times of passage through F _S node N _S at the beginning of the sentence is minimized The number of passes is calculated.
At this time, the minimum passage number calculation means 17 also calculates F _{1 to} F _E and H _{S1 to} H _7E included in the first conditional expression and the second conditional expression by the simplex method.

Here, the results of solving the equations (1) to (24) are as follows. In other words, the minimum number of passes is a value equal '17' pass count F _S.
F _S = 17, F ₁ = 17, F ₂ = 16, F ₃ = 16, F ₄ = 0, F ₅ = 1, F ₆ = 1, F ₇ = 17, F _E = 17, H _S1 = 17, H ₁₂ = 16, H ₁₅ = 1, H ₂₃ = 16, H ₂₄ = 0, H ₃₇ = 16, H ₄₇ = 16, H ₅₆ = 1, H ₆₇ = 1, H _7E = 17

Thereafter, the minimum passage number calculation means 17 reads out the lists L _{1 to} L ₄ and the integrated graph T input from the conditional expression generation means 16 and the results of solving the first conditional expression and the second conditional expression. 18 is output.

  The spoken sentence generation means 18 uses the integrated graph T input from the minimum passage number calculation means 17 to generate a reading sentence by changing the number of phrases equal to the minimum passage number and the combination of words stored in the list L. is there.

<Generating aloud text>
With reference to FIG. 7 and FIG. 8, the generation of the reading sentence will be described (refer to FIG. 1 as appropriate).
Here, it is assumed that the read-out sentence generating unit 18 is input from the minimum passage number calculating unit 17 with the integrated graph T in FIG. 7B and the lists L _{1 to} L _{4 in} FIG.

As shown in FIG. 7 (b), in the beginning of the integration graph T, edge _{E S1} is present which is input from the node _{N S} to node _{N 1.} Therefore, the read-out sentence generation means 18 selects the edge _ES1 . Furthermore, since the list L does not correspond to the node N ₁ to which the edge E _S1 is input, the reading sentence generation unit 18 selects “in the vicinity of the center” indicated by the node N ₁ as the reading sentence.

Subsequently, edges E ₁₂ and E ₁₅ branch from the node N ₁ . Therefore, the read-out sentence generation means 18 determines whether or not the total number of times of selection of the edge E ₁₂ that is one branch destination exceeds the number of times of passage H ₁₂ of the edge E ₁₂ .

If total number of selection times of the edge E ₁₂ does not exceed the number of passes H _12, reading sentence generating means 18 selects the edge E _12, it increments the total number of times of selecting the edge E _12.
On the other hand, if the total number of selection times of the edge E ₁₂ is greater than the number of passes H _12, reading sentence generating unit 18 selects an edge E ₁₅ is the other branch destination, the total number of times of selecting the edge E ₁₅ Increment.

Here, reading sentence generating means 18 is to be selected the edge E _12. Further, the node N _{2 that} is the branch destination of the edge E ₁₂ corresponds to the list L ₁ . In this case, the reading sentence generation means 18 selects the first word “12 hours” in the list L ₁ and selects “within the next 12 hours” indicated by the node N ₂ as the reading sentence.

Subsequently, edges E ₂₃ and E ₂₄ branch from the node N ₂ . Therefore, the read-out sentence generation means 18 determines whether or not the total number of times of selection of the edge E ₂₃ which is one branch destination exceeds the number of passes H ₂₃ of the edge E ₂₃ .

If total number of selection times of the edge E ₂₃ does not exceed the number of passes H _23, reading sentence generating means 18 selects the edge E _23, it increments the total number of times of selecting the edge E _23.
On the other hand, when the total number of times of selection of the edge E ₂₃ exceeds the number of times of passage H ₂₃ , the reading sentence generation unit 18 selects the edge E ₂₄ which is the other branch destination, and sets the total number of times of selection of the edge E _24. Increment.

Here, reading sentence generating means 18 is to be selected the edge E _23. Further, the node N _{3 that} is the branch destination of the edge E ₂₃ corresponds to the list L ₂ . In this case, the reading sentence generation means 18 selects the first word “east-northeast” of the list L ₂ and selects “east-northeast” indicated by the node N ₃ as the reading sentence.

Subsequently, there is an edge E ₃₇ input from the node N ₃ to the node N ₇ . Therefore, the read-out sentence generation unit 18 selects the edge E ₃₇ . Further, the read-out sentence generation means 18 selects “a strong wind is expected to blow” indicated by the node N ₇ as the read-out sentence.

Finally, the edge _{E 7E} is present that is input to the node _{N E.} Therefore, the read-out sentence generation unit 18 generates the read-out sentence by sequentially connecting the words of the node N selected as the read-out sentence. In this example, the reading sentence generation means 18 generates a reading sentence that “a strong wind in the east and northeast is expected to blow within the next 12 hours near the center”.

Reading sentence generating unit 18, by following the edge E from node N _S to node N _E, and repeats a process of generating reading sentence by the minimum number of passes. Here, when the unselected word / phrase remains in the list L, the reading sentence generation unit 18 selects the word / phrase. Further, when the unselected word / phrase does not remain in the list L, the read-out sentence generating unit 18 repeatedly uses the selected word / phrase.
In this way, the reading sentence generation means 18 generates the same number of reading sentences as the minimum number of passages.

[Operation of the text-to-speech generator for speech synthesis]
With reference to FIG. 9, the operation of the speech synthesizing text generation device 1 of FIG. 1 will be described (see FIG. 1 as appropriate).
In the speech synthesis speech-to-speech generator 1, the directed graph (initial graph, branch deletion target graph) G and the list L are input by the graph input means 10 (step S1).

The speech synthesis read-out sentence generator 1 determines whether the edge E of the branch deletion target graph G is branched by the graph branch deletion unit 11.
When the edge E is not branched, the graph branch deletion unit 11 outputs the branch deletion target graph G as it is as the branch deletion deleted graph G.
On the other hand, when the edge E is branched, the graph branch deletion unit 11 generates a plurality of branch deleted graphs G from which the branched edge E is deleted from the branch deletion target graph G (step S2).

In the speech synthesis read-out sentence generation device 1, the integrated graph generation unit 12 determines whether the comparison target graph pairs are similar by the DP matching method.
When the comparison target graph pairs are similar, the integrated graph generation unit 12 integrates the comparison target graph pairs into the integrated graph T based on the result of the DP matching method.
On the other hand, if the comparison target graph pairs are not similar, the integrated graph generation unit 12 treats each of the comparison target graph pairs as a new integrated graph T (step S3).

In the text-to-speech generation device 1 for speech synthesis, text data that has been read out is input by the text data input means 13 (step S4).
The speech synthesis read-out sentence generation device 1 generates list exclusion information by comparing the integrated graph T with the read-out text data by the graph comparison unit 14 (step S5).

Based on the list exclusion information, the speech synthesizing speech generation apparatus 1 for speech synthesis excludes words / phrases included in the read text data from the list L based on the list exclusion information (step S6).
The speech synthesis read-out sentence generation apparatus 1 generates the first conditional expression by the first conditional expression generation unit 16a (step S7).
The speech synthesis read-out sentence generation apparatus 1 generates the second conditional expression on both the input side and the output side of the node N by the second conditional expression generation unit 16b (step S8).

The speech synthesizing speech generation apparatus 1 calculates the minimum number of passages by solving the first conditional expression and the second conditional expression by the simplex method by the minimum number of passages calculation means 17 (step S9).
In the speech synthesis device for speech synthesis 1, the speech generation unit 18 uses the integrated graph T to generate a speech by changing the number of words stored in the list L by the number equal to the minimum number of passages (step S1). S10).

  As described above, the speech synthesis read-out sentence generation device 1 excludes the words / phrases included in the read-out text data from the list L by the list exclusion unit 15. For this reason, the speech synthesis read-out sentence generation device 1 can generate a read-out sentence only with words and phrases that are not included in the read-out sentence, prevent duplicate generation of the read-out sentence, and reduce the number of read-out sentences to be generated. can do.

  In the above-described embodiment, the case where there is one integrated graph T has been described as an example, but the number of integrated graphs T is not particularly limited. For example, in the case where there are a plurality of integrated graphs T, the similar integrated graph T is further integrated (variable sharing), and the dissimilar integrated graphs T are treated as individual items so that the simplex method is used as a whole. Is possible.

(Modification 1)
The present invention is not limited to the embodiments, and various modifications can be made without departing from the spirit of the present invention. Hereinafter, modifications of the present invention will be described.

In the above-described embodiment, it has been described that the node N includes at most one replaceable word / phrase, but the present invention is not limited to this.
The speech synthesis device for speech synthesis 1A according to the modified example of the present invention includes a point that two or more replaceable words are included in the node N of the integrated graph T, and an integrated graph generating unit instead of the integrated graph generating unit 12 The point provided with 12A is different from the above-described embodiment.

  When the node N of the integrated graph T includes a plurality of replaceable words / phrases, the integrated graph generation unit 12A calculates all combinations of words / phrases included in different lists L from different lists L in which the plurality of words / phrases are stored. A stored integrated list is generated, a greedy algorithm is applied to the integrated list to generate a compact list, and the node N and the compact list are associated with each other.

For example, as shown in FIG. 10A, a case is considered where the integrated graph T includes a node N indicating “when [monthly] [date] [division] [time]”. In addition, lists L _{10 to} L ₁₃ are associated with the node N.

List L _10, as shown in FIG. 10 (b), the phrase that corresponds to the monthly] of the node N is stored 12.
The list L ₁₁ stores 31 words / phrases corresponding to [date] of the node N.
The list L ₁₂ stores two words “AM” and “PM” corresponding to the [classification] of the node N.
The list L ₁₃ stores 12 words / phrases corresponding to [time] of the node N.

  In this case, as shown in FIG. 10C, the integrated graph generation unit 12A converts four words [monthly], [date], [division], and [time] that can be replaced by the node N into one Join the replaceable word [monthly-date-division-time].

The integrated graph generation unit 12A, as shown in FIG. 10 (d), to produce a consolidated list _{L 20} that all combinations of words in the list _L 10 _{~L 13} is stored. That, consolidated list _{L 20} is, 12 × 31 × 2 × 12 = 8928 pieces of words are stored.

For example, the first word / phrase stored in the integrated list L ₂₀ is “midnight on January 1,” which is a combination of the first word / phrases in the lists L _{10 to} L ₁₃ . Also, the second word stored in the integrated list L ₂₀ is “January 1 AM”, which is a combination of the first word in the lists L _{10 to} L ₁₂ and the second word in the list L _13. 1 o'clock ". Also, the third word stored in the integrated list L ₂₀ is “January 1 AM”, which is a combination of the first word in the lists L _{10 to} L ₁₂ and the third word in the list L _13. 2 o'clock ".

The integrated graph generation unit 12A, by applying the greedy algorithm to integrate the list L ₂₀ in FIG. 10 (d), to generate a minimum compact list (not shown) to the speech synthesis, and FIG. 10 (c) Correspond to the node N.

This greedy algorithm is described in, for example, the reference document “T. Colmen et al.,“ Algorithm Introduction ”, Volume 3, Modern Science, December 30, 1995, p313-317”, so that detailed explanations are given. Is omitted.
As the greedy algorithm, a sequential greedy algorithm described in Japanese Patent No. 4741208 can also be used.

  As described above, the speech synthesizing speech generation device 1A can reduce the number of words even if the node N includes a plurality of replaceable words. For this reason, the speech synthesizing speech generation apparatus 1A for speech synthesis can make the reading text generation process common regardless of the number of replaceable words included in the node N, and can simplify the configuration.

1,1A Speech Synthesizer for Speech Synthesis 10 Graph Input Unit 11 Graph Branch Deletion Unit (Graph Integration Unit)
12, 12A integrated graph generation means (graph integration means)
13 Text data input means 14 Graph comparison means 15 List exclusion means 16 Conditional expression generation means 17 Minimum pass count calculation means 18 Reading sentence generation means

Claims (5)

The sentence is expressed by a directed graph composed of a plurality of nodes indicating words included in the sentence and an edge indicating a connection relation between the nodes, and the replaceable words are assigned to the nodes. A speech synthesis generation device for speech synthesis that generates speech for a speech synthesis database necessary for synthesis,
A graph input means for inputting the directed graph and a list that stores one or more words corresponding to the nodes of the directed graph and replaced by the nodes;
Graph integration means for integrating the directed graph input to the graph input means;
Text data input means for inputting text data indicating the utterance content;
A graph comparing means for obtaining a phrase contained in the text data from the list by comparing the text data input to the text data input means and the integrated directed graph;
A list excluding unit for excluding the word obtained by the graph comparing unit from the list;
First conditional expression generation means for generating a first conditional expression that sets the number of passages of the node to be equal to or more than the number of words stored in the list that corresponds to the node and is excluded by the list exclusion means;
As a second conditional expression, a conditional expression in which the sum of the number of times of passage of edges input to the node is equal to the number of times of passage of the node, and a sum of the number of times of passage of edges output from the node are equal to the number of times of passage of the node Second conditional expression generation means for generating a conditional expression
Minimum passage number calculating means for calculating a minimum number of passages that minimizes the number of passages at the head of the sentence so as to satisfy the first conditional expression and the second conditional expression;
A reading sentence generation means for generating the reading sentence by changing a combination of words stored in the list deleted by the list exclusion means, a number equal to the minimum passage number calculated by the minimum passage number calculation means,
A speech synthesis apparatus for speech synthesis, comprising: The graph integration means includes:
It is determined whether or not the edge of the directed graph input to the graph input means for the second time and thereafter is branched, the directed graph in which the edge is not branched is defined as a branch deleted graph, and the directed graph from which the edge is branched , A graph branch deletion means for separately generating nodes at the branch destination of the edge and generating a branch deleted graph in which the branch of the edge is deleted;
By the DP matching method, it is determined whether the comparison target graph pair composed of one of the directed graph or the integrated graph first input to the graph input means and the branch deleted graph is similar, and the comparison target graph pair is When similar, the comparison target graph pair is integrated into one integrated graph, and when the comparison target graph pair is not similar, an integrated graph generating unit that sets each of the comparison target graph pairs as a new integrated graph; With
2. The speech synthesis read-out sentence generation apparatus according to claim 1, wherein the graph comparison unit compares the integrated graph input from the integrated graph generation unit with the text data.   When the same node is divided into a plurality of nodes in the branch-deleted graph by the graph branch deletion unit, the first conditional expression generation unit determines that the sum of the number of passes of the plurality of nodes is the same node. 3. The speech synthesis read-out sentence generation apparatus according to claim 2, wherein the first conditional expression corresponding to and greater than or equal to the number of words stored in the list excluded by the list exclusion unit is generated.   In the integrated graph generation means, when a plurality of replaceable words / phrases are included in the nodes of the integrated graph, all combinations of the words / phrases included in the separate lists are stored from different lists in which the plurality of words / phrases are stored. 3. An integrated list is generated, and a greedy algorithm is applied to the integrated list to generate a minimum compact list for the speech synthesis, and the node is associated with the compact list. Or the speech-speech production | generation apparatus for speech synthesis of Claim 3.   A speech synthesis read-out sentence generation program for causing a computer to function as the speech synthesis read-out sentence generation apparatus according to claim 1.

Images