JP2008026452A - Speech synthesizer, method and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech synthesizer capable of performing high speed speech synthesizing processing by reducing a processing amount in the speech synthesizer, especially the processing amount in preliminary selection. <P>SOLUTION: Based on a phoneme sequence which is selected with a predetermined reference, under condition for using certain phoneme for certain synthesizing object information, a cost degradation value data base is provided for indicating a cost degradation value of the phoneme for the synthesizing object information, which indicates a degradation degree of a synthesized speech waveform in the speech phoneme sequence selected with the predetermined reference. One or more pieces of synthesizing object information included in the cost degradation value data base are selected for each input synthesizing object information. Phoneme candidates for the input synthesizing object information are selected from a plurality of phonemes corresponding to the selected synthesizing object information, based on the cost degradation value. A series of speech phonemes used for the output synthesized speech waveform is selected from a series of input synthesizing object information pieces and the speech phoneme candidates for each selected input synthesizing object information piece. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a speech synthesizer, a method, and a program for synthesizing speech by selecting an optimum speech unit from speech units stored in a database and connecting the selected speech unit.

  In unit-connected speech synthesis, which is one of speech synthesis technologies, a large number of speech units are stored in a database in advance, and are close to each parameter of synthesis target information specified at the time of synthesis, The speech unit having a good connection relation is selected from the unit database and synthesized. Each speech segment is assigned parameters such as phoneme labels, acoustic parameters, and appearance environment in the speech corpus.

  In unit-connected speech synthesis, a speech unit to be used is selected based on designated synthesis target information (hereinafter referred to as unit selection). This unit selection is a distortion measure called cost. That is, it is performed based on an index indicating the degree of deterioration of the speech waveform synthesized by the selected speech segment from the target synthesized speech waveform. The cost can usually be divided into a target cost indicating an error between the synthesis target information and the speech unit, and a connection cost indicating the degree of discontinuity between the speech units, and segment selection minimizes the overall cost. To be done.

  In order to obtain synthesized speech with higher naturalness, it is necessary to use a large unit database. However, as the unit database becomes larger, the number of combinations of speech units to be considered is huge. Then, the problem that the segment selection process becomes difficult occurs. In order to solve this problem, before selecting the final speech segment considering both the target cost and the connection cost, a preliminary segment that narrows down speech segments prepared in the segment database based on the target cost. A configuration for making a selection has been proposed.

  However, in recent years, the segment database has become increasingly large, and as a result, the amount of preliminary selection performed to reduce the amount of segment selection processing cannot be ignored. For this reason, Patent Document 1 describes a configuration in which the amount of preliminary selection processing is reduced by performing preliminary selection in multiple stages. According to Patent Document 1, the target cost used for the preliminary selection is divided into sub-costs such as fundamental frequency, duration, and MFCC (Mel-Frequency Cepstrum Coefficient), and the number of calculations is reduced by sequentially narrowing down by sub-cost. ing.

  Further, Patent Document 2 has a configuration in which the frequency of selection processing is reduced by checking the selection frequency of each speech unit by speech synthesis using a test sentence and deleting a speech unit with a low selection frequency from the unit database. Are listed.

JP 2005-265895 A JP 2004-37605 A

  The configuration described in Patent Document 1 is the same as the prior art in that all speech unit candidates must be evaluated at the time of speech synthesis, and there is a limit to the reduction in processing amount by this configuration. Further, since different selection criteria are used in the preliminary selection and the segment selection, there is a problem that the speech segment candidate selected in the preliminary selection is not necessarily the most preferable selection criterion in the segment selection. . In order to cope with this problem, it is necessary to increase the number of speech unit candidates to be selected in the preliminary selection as much as possible. However, if the number of selections in the preliminary selection is increased, there is a problem that the processing amount of the unit selection increases. appear.

  Further, in the configuration described in Patent Document 2, when there are few test sentences, a frequently used speech element that was not included in the test sentence may be deleted from the element database, There may be a problem in that a speech unit that is included and remains infrequently used remains in the segment database. In order to deal with this problem, it is necessary to prepare a large amount of test sentences as compared with the segment database, but it becomes difficult to prepare test sentences as the scale of the segment database increases.

  Accordingly, an object of the present invention is to provide a speech synthesizer, a method, and a program that perform speech synthesis processing at high speed by reducing the amount of processing in the speech synthesizer, in particular, the amount of processing in preliminary selection.

According to the speech synthesizer of the present invention,
Based on a set of input synthesis target information, a series of speech units is selected from a plurality of speech units prepared in the unit database, and speech waveforms corresponding to each of the selected series of speech units are connected and synthesized. A speech synthesizer for outputting a speech waveform, which is based on a series of test synthesis target information and based on a speech segment sequence selected from a segment database according to a predetermined standard, with respect to a certain speech synthesis target information. A cost deterioration value of the speech unit with respect to the test synthesis target information indicating the degree of degradation of the synthesized speech waveform in the speech unit sequence selected by the predetermined criterion under the condition for using the segment, and each test synthesis Cost deterioration value data for each of the input deterioration target information and cost deterioration value calculation means for generating a cost deterioration value database indicating the cost deterioration value of the corresponding speech segment for the target information Pre-selection means for selecting one or more test synthesis target information from the source and selecting a speech unit candidate for the input synthesis target information from the speech unit corresponding to the selected test synthesis target information based on the cost deterioration value And, based on a series of input synthesis target information, a unit selection for selecting a series of speech units to be used for an output synthesized speech waveform from among speech unit candidates for each input synthesis target information selected by the preliminary selection means Means.

According to another embodiment of the speech synthesizer of the present invention,
The preliminary selection means groups the test synthesis target information included in the cost deterioration value database, selects a plurality of speech units based on the cost deterioration value from the speech units corresponding to each of the grouped test synthesis target information, Preliminary selection result database generating means for recording the selected speech unit in association with the group, and determining a group to which each input synthesis target information belongs, and determining the speech unit corresponding to the determined group to the input synthesis target information It is also preferable to include selection means for selecting as speech segment candidates.

According to another embodiment of the speech synthesizer of the present invention,
Preliminary selection result database generation means uses all test synthesis target information belonging to the same unit speech as a root node, and groups test test target information included in a leaf node of a decision tree generated by sequentially dividing the root node into one group It is also preferable that the selection unit determines a group to which the input synthesis target information belongs according to the decision tree.

Furthermore, according to another embodiment of the speech synthesizer of the present invention,
Preliminary selection result database generation means generates a decision tree by repeating set partitioning that minimizes the evaluation value, and the evaluation value is a test included in each leaf node by using the evaluation value of each leaf node generated by the partitioning. The sum of values weighted based on the number of synthesis target information, and the evaluation value of the leaf node is a predetermined number selected based on the cost deterioration value from the speech segment corresponding to each test synthesis target information included in the leaf node It is also preferable that the average value of the cost deterioration values of the speech segments.

Furthermore, according to another embodiment of the speech synthesizer of the present invention,
It is also preferable that the unit selection unit selects a series of speech units in consideration of only the degree of deterioration of the synthesized speech waveform due to the connection of speech units.

Furthermore, according to another embodiment of the speech synthesizer of the present invention,
The cost deterioration value calculating means selects one or more speech segments corresponding to each series of test synthesis target information from the segment database, and arranges the selected speech segments in the same order as the series of test synthesis target information. , Each speech unit is associated with a node of the search graph, and cost calculation indicating the degree of degradation of the synthesized speech waveform is performed by dynamic programming from the beginning and the end, respectively. , Acquiring the cost of the speech unit sequence as a reference and the speech unit sequence under the condition of using the certain speech unit, and combining each test synthesis based on the acquired cost of each speech unit sequence It is also preferable to obtain the cost deterioration value of the speech unit selected from the unit database for the target information.

According to the speech synthesis method of the present invention,
Based on a series of test synthesis target information, based on a speech unit sequence selected by a predetermined standard from a segment database, the predetermined standard is used under the condition of using a speech unit for a certain test synthesis target information. Using the cost deterioration value database having the cost deterioration value of the speech unit with respect to the test synthesis target information, which indicates the degree of deterioration of the synthesized speech waveform in the speech unit sequence selected in step 1, and synthesized speech from a series of input synthesis target information A speech synthesis method for outputting a waveform, wherein, for each input synthesis target information, one or more test synthesis target information is selected from a cost deterioration value database, and a cost is calculated from a speech unit corresponding to the selected test synthesis target information. A step of selecting speech segment candidates for the input synthesis target information based on the deteriorated value and a set of input synthesis target information for each input synthesis target information. Selecting a series of speech units to be used for the synthesized speech waveform to be output from the speech unit candidates to be output, and connecting the speech waveforms corresponding to each of the selected series of speech units to generate a synthesized speech waveform And generating.

  According to the program of the present invention, a computer is caused to function as the speech synthesizer.

  A cost deterioration value database indicating the cost deterioration value of each speech unit for the test synthesis target information is created in advance, and preliminary selection is performed based on the cost deterioration value. Here, the cost deterioration value of the speech unit with respect to the test synthesis target information is based on the synthesized speech waveform in the speech unit sequence selected by the predetermined standard, and the speech unit is used for the test synthesis target information. Under these conditions, this is a value indicating the degree of deterioration of the synthesized speech waveform in the speech segment sequence selected based on the same predetermined criterion, and the connection of speech segments is also taken into consideration. That is, the selection criterion for preliminary selection is the same as the selection criterion for segment selection, and the accuracy of preliminary selection is increased. Therefore, it is possible to perform speech synthesis processing efficiently and at high speed without increasing the amount of processing for preliminary selection and segment selection.

  Furthermore, since the accuracy of preliminary selection is increased, high quality speech can be synthesized even if the number of candidates in preliminary selection is reduced. Moreover, since the evaluation value about the availability of the speech unit can be obtained by using the cost deterioration value, unlike the conventional technology using the frequency information of the unit selection result, the cost deterioration value database is created. High-quality speech synthesis can be performed without using a large amount of test sentences.

  By preliminarily grouping test synthesis target information included in the cost deterioration value database and selecting speech segments for each group, the processing of the preliminary selection means at the time of speech synthesis is performed for the group to which the input synthesis target information belongs. Only the determination is made, and the processing amount of the preliminary selection can be reduced. A decision tree can be used for group creation and judgment, and even if input synthesis target information different from the test synthesis target information is given, an appropriate result reflecting the result for similar test synthesis target information is quickly returned. be able to.

  In addition, target cost calculation that requires acoustic feature parameters can be omitted by selecting a series of speech segments considering only the degree of degradation of the synthesized speech waveform caused by the connection of speech segments. It is possible to reduce the amount of processing. Furthermore, the cost deterioration value is calculated by calculating the cost by the dynamic programming from the beginning and the dynamic programming from the end after associating each speech unit with the node of the search graph. Thus, it is possible to reduce the burden of creating the cost deterioration value database.

  BEST MODE FOR CARRYING OUT THE INVENTION The best mode for carrying out the present invention will be described below in detail with reference to the drawings. Here, “unit speech” is a synthesis processing unit of the speech synthesizer according to the present invention, and a plurality of speech units (hereinafter simply referred to as “units”) correspond to one unit speech. . Specific examples of unit speech include phonemes, syllables, and words. Each segment may correspond to only one unit sound or may correspond to a plurality of unit sounds. Furthermore, the synthesis target information is an index for selecting a segment to be used for a speech waveform to be synthesized, and has a correspondence relationship with a unit speech in the same manner as a segment. That is, when the synthesis target information is given, a plurality of segments corresponding to the given synthesis target information are obtained through unit speech, and the speech synthesizer considers the target cost and the connection cost, Speech synthesis is performed by selecting the optimum segment from the segments.

  FIG. 1 is a block diagram of a speech synthesizer according to the present invention. According to FIG. 1, the speech synthesizer includes a preprocessing unit 1, a cost deterioration value calculation unit 2, a preliminary selection unit 3, a segment selection unit 4, a speech waveform connection unit 5, and a cost deterioration value database 6. The unit information database 7 and the speech waveform database 8 are provided.

  The segment information database 7 is a database for a segment ID for identifying a segment and each parameter of the segment, for example, fundamental frequency, MFCC, and duration, and the speech waveform database 8 is a segment ID. This is a database for the actual waveform of the specified segment, and the segment information database 7 and the speech waveform database 8 constitute a segment database.

  The preprocessing unit 1 divides the input text into unit speeches by morphological analysis, syntax analysis, word dictionary reference, etc., attaches prosodic information obtained by the analysis, and outputs it as text analysis information. Thereafter, acoustic feature parameters such as MFCC and fundamental frequency are added to the text analysis information and output as a series of synthesis target information.

  The cost deterioration value calculation unit 2 uses the segment information database 7 based on a series of synthesis target information generated from the input text input to the preprocessing unit 1, and uses the segment information database 7 for each synthesis target information. A cost deterioration value database 6 indicating the cost deterioration value is generated. The cost deterioration value database 6 is created before the start of speech synthesis processing using one or more test sentences as input text.

  Hereinafter, the creation of the cost deterioration value database 6 by the cost deterioration value calculation unit 2 will be described by taking as an example a case where the input test sentence is composed of four unit sounds. The preprocessing unit 1 divides the input text sentence into four unit sounds, and generates synthesis target information for each unit sound. The cost deterioration value calculation unit 2 acquires segment information of a plurality of segments corresponding to the four synthesis target information generated by the preprocessing unit 1 from the segment information database 7. In this example, it is assumed that four corresponding segments are prepared in the segment information database 7 for each synthesis target information.

  FIG. 4 is a diagram illustrating a search graph for calculating a cost deterioration value. Nodes n11 to n14 are four segments corresponding to the first synthesis target information, and nodes n21 to n24 are a second synthesis. The four elements corresponding to the target information, the nodes n31 to n34 are the four elements corresponding to the third synthesis target information, and the nodes n41 to n44 are the four elements corresponding to the fourth synthesis target information. Represents a piece. Here, the search graph is created by adding a node X and a node Y indicating silence to the beginning and the end of these.

  The line connecting the nodes corresponds to the connection cost, each node corresponds to the target cost, and the cost from one node to the other is obtained by sequentially adding each connection cost of the used route and the target cost of the passed node. . Note that the target cost corresponding to the node X and the node Y is zero. Here, among the plurality of routes from the node X to the node Y, the optimum route with the lowest cost is assumed to be the nodes n11, n22, n32, n43, and the node n23 is set to the second synthesis target information. It is assumed that the route having the minimum cost under the condition of use is the nodes n14, n23, n34, and n43. At this time, when the cost of the paths of the nodes n11, n22, n32, and n43 is used as a reference, the degree of deterioration of the cost of the paths of the nodes n14, n23, n34, and n43 is the node for the second synthesis target information. This is the cost deterioration value of the segment corresponding to n23. In other words, from a series of synthesis target information given in a test sentence, based on a condition that a segment is used for a certain synthesis target information, based on a sequence of segments selected as being optimal according to a predetermined criterion, The index indicating the degree of degradation of the synthesized speech waveform in the segment sequence selected based on the same predetermined criterion is the cost deterioration value of the segment relative to the synthesis target information. In the following explanation, the cost of the optimum path is The difference from the cost under the condition that a certain piece is used is used as the cost deterioration value.

  In this embodiment, the cost deterioration value of each segment is calculated by a simple method based on dynamic programming using the search graph of FIG. Specifically, for each node with t = 1, the sum of the connection cost with the node X and the target cost of the node is recorded as the minimum cost from the node X of the node. Next, for each node with t = 2, the connection cost of the route to the node immediately before four points, and the minimum cost recorded in the node immediately before point connected to the route are Add and record the smallest one plus the target cost for that node as the minimum cost from node X for that node. Thereafter, the above processing is repeated in order of time until the node Y is reached. By this operation, the minimum cost c (X → n) from the node X to any node n in the search graph is recorded.

Subsequently, the same processing is performed in the reverse direction of the time axis with the node Y as the head. However, c ′ (Y → n) = c (Y → n) −ct (n) is obtained as the value of the node n. Here, ct (n) is the target cost of the node n. For an arbitrary node n on the graph, the minimum cost of a plurality of nodes X → n → Y is given by c (X → n) + c (n → Y) −ct (n). Since (n → Y) = c (Y → n),
c (X → n) + c (n → Y) −ct (n) = c (X → n) + c ′ (Y → n)
And the cost deterioration value d (n) is
d (n) = c (X → n) + c ′ (Y → n) −c (X → Y)
Can be obtained as In this embodiment, by performing the process from the node X once and the process from the node Y once, it is possible to obtain the cost deterioration value of each unit for all the nodes, and individually from each node. There is no need to search for the minimum cost to node Y.

FIG. 5 is a diagram showing the cost deterioration value database 6, and for each of the synthesis target information A1, A2, A3, A4, A5... Belonging to a certain unit voice, the unit belonging to the same unit voice and its cost deterioration. The value is shown. For example, if a unit speech is a phoneme “A”, the synthesis target information A1, A2, A3, A4, A5 is synthesis target information belonging to the unit speech “A” obtained from the test sentence, and the segment a _1. , A ₂ , a ₃ ... Are the segments belonging to the unit voice “A” among the prepared segments. Actually, for each unit voice, for example, although not shown, the same prepared pieces b ₁ , b ₂ ... For the synthesis target information B ₁ , B _2. Each cost deterioration value is included in the cost deterioration value database 6.

  The preliminary selection unit 3 refers to the cost deterioration value database 6 for each series of synthesis target information generated from the input text by the preprocessing unit 1 in the actual speech synthesis process, and provides a segment for each synthesis target information. A candidate is selected, and the segment ID of the selected segment candidate is output.

  As an example of a method for selecting a segment candidate by the preliminary selection unit 3, when the same synthesis target information as the synthesis target information input from the pre-processing unit 1 exists in the cost deterioration value database 6, it corresponds to the synthesis target information. When a predetermined number of pieces are selected from those having a low cost deterioration value or a piece having a cost deterioration value smaller than the predetermined value and the same synthesis target information does not exist in the cost deterioration value database 6 There is a method of similarly selecting a segment candidate from the nearest synthesis target information existing in the cost deterioration value database 6.

  As another example of the method of selecting segment candidates by the preliminary selection unit 3, a predetermined number of synthesis target information close to the synthesis target information input from the preprocessing unit 1 is selected from the cost deterioration value database 6, and each selected synthesis is selected. From the segments corresponding to the target information, obtain an average value of cost deterioration values of the same segment, and select a predetermined number of segments or a segment whose average value is smaller than the predetermined value from the smaller average value. There is a way. Furthermore, when the synthesis target information identical to the synthesis target information input from the preprocessing unit 1 is in the cost deterioration value database 6, the same synthesis target information is not present in the cost deterioration value database 6 from the synthesis target information. In this case, a predetermined number of synthesis target information may be selected from the cost deterioration value database 6 and selected by a method of calculating an average of the cost deterioration values.

  The segment selection unit 4 acquires parameters from the segment information database 7 for the segment candidates selected by the preliminary selection unit 3, and from among the segment candidates based on a series of synthesis target information from the preprocessing unit 1. The segment that is the optimal combination is selected, and the segment ID of the selected segment is output.

  The speech waveform connection unit 5 acquires the waveform information of the corresponding segment from the speech waveform database 8 based on the segment ID output by the segment selection unit 4, and outputs a synthesized speech waveform by connecting the waveforms.

  In the speech synthesizer according to the present invention, the cost deterioration value of the segment belonging to the synthesis target information is obtained for each synthesis target information in advance based on the test sentence, and preliminary selection is performed based on the cost deterioration value. The cost deterioration value is an index including the degree of deterioration of the synthesized speech waveform when a segment is forcibly selected, that is, both the target cost and the connection cost. Therefore, the preliminary selection in the present invention is to narrow down the segments prepared in the segment database using an index that includes both the target cost and the connection cost, and the segment selection unit 4 does not perform the preliminary selection. When the segment selection is performed from all the segments, there is a high possibility that the segment selected is included in the segment candidate. That is, the accuracy of preliminary selection is increased. Since the segment selection unit 4 performs segment selection from such segment candidates, the speech selection process can be performed efficiently and at high speed without increasing the amount of processing in the preliminary selection unit 3 and the segment selection unit 4. Is possible.

  Furthermore, since the accuracy of preliminary selection is increased, high quality speech can be synthesized even if the number of candidates in preliminary selection is reduced. In addition, by using the cost deterioration value, evaluation values for availability can be obtained for all the unit candidates corresponding to the unit speech included in the test sentence. Unlike the prior art using information, high-quality speech synthesis can be performed without using a large amount of test sentences.

  Next, another embodiment of the speech synthesizer according to the present invention will be described. The block diagram of the entire speech synthesizer of this embodiment is the same as that of FIG. 1, but in this embodiment, the preliminary selection unit 3 includes a preliminary selection result database generation unit 31 and a selection unit 32 as shown in FIG. And.

The preliminary selection result database generation unit 31 creates a preliminary selection result database 9 from the cost deterioration value database 6 in advance. FIG. 7 is a diagram showing the preliminary selection result database 9. As shown in FIG. 7, the preliminary selection result database 9 is a group of synthesis target information belonging to the same unit voice in the cost deterioration value database 6 and grouped by a predetermined method as will be described later. As described later, a predetermined number of segments are selected in advance from a plurality of segments corresponding to. For example, in FIG. 7, the synthesis target information A1, A2, A3 is grouped, and k pieces of pieces a ₁ , a ₅ , a ₉ ... A ₃₈ are selected in this group. .

  The selection unit 32 determines to which group each synthesis target information from the preprocessing unit 1 belongs, and outputs the unit IDs of k units corresponding to the determined group as unit candidates. Hereinafter, creation of the preliminary selection result database 9 by the preliminary selection result database generation unit 31 will be described.

The preliminary selection result database generation unit 31 creates a decision tree for each unit voice for grouping the synthesis target information. FIG. 6 is a diagram showing a decision tree for creating the preliminary selection result database 9. First, let T be the set of all synthesis target information belonging to the same unit speech, and this be the root node of the decision tree. This set T is divided into a set T1 of composite target information that satisfies the question q0 by the question q0 and a set T2 of composite target information that does not satisfy the question q0. Further, the tree structure is constructed by recursively repeating the division of the set T1 by the question q1 and the set T2 by the question q2. Here, the cost deterioration value when the segment u is used for the synthesis target information t is d (u, t), t1 and n1 are the synthesis target information belonging to the set T1 and the number thereof, and t2 and n2 are the set T2. And the number of synthesis target information belonging to. Here, for all u, the average value of d (u, t1) with respect to all t1 is calculated, k1 is selected from the smaller one, and the average value is defined as m1. Similarly, for all u, the average value of d (u, t2) with respect to all t2 is calculated, k2 is selected from the smaller value, and the average value is m2. Note that k1 and k2 are appropriately determined integers. At this time, each question q is expressed by the following formula (1).
(N1 * m1 + n2 * m2) / (n1 + n2) (1)
And is determined by either Yes or No. That is, a decision tree is constructed by repeating set partitioning. The question determined by either Yes or No refers to “the central phoneme is a vowel” for the discrete amount included in the synthesis target information, and “the time of the segment” for the continuous amount included in the synthesis target information. Ask a question such as “The length is less than 50 ms”. Regarding these questions, a plurality of questions are prepared in advance by referring to information included in the synthesis target information.

In order to avoid an increase in the tree structure, for example, the following division stop condition is set.
(A) When the number of stages from the root node is a predetermined value (for example, 30 stages) or more.
(B) When the difference between the evaluation values after division is less than a certain threshold. However, the evaluation value before division is the average value when the average value of d (u, t) for all t is calculated for all u and k is selected from the smaller ones. m. The evaluation value after the division is a value given by the above formula (1).
(C) When the number of synthesis target information included in the node, that is, the number n of synthesis target information t in the group is equal to or less than a certain number.

  For leaf nodes that do not satisfy the division stop condition, the division processing is performed. When the division stop condition is satisfied, the division is stopped. Subsequently, when a set of synthesis target information included in each leaf node is Tm, each synthesis target information belonging to Tm is tm, and there are N synthesis target information belonging to Tm, an average value of cost deterioration values for a certain unit u

を総ての素片について計算し、平均値の小さい方からｋ´個を、この葉ノードに対する予備選択結果として予備選択結果データベース９を生成する。

Are calculated for all the segments, and the preliminary selection result database 9 is generated as the preliminary selection result for this leaf node with k ′ pieces having the smallest average value.

  When the synthesis target information is given from the preprocessing unit 1, the selection unit 32 selects a unit candidate of the given synthesis target information based on a decision tree for unit speech corresponding to the given synthesis target information To do. Specifically, it is determined whether or not the given synthesis target information satisfies each question q of the decision tree, and the leaf nodes are sequentially traced from the root node. The corresponding k ′ element IDs are output as element candidates.

  According to the present embodiment, it is possible to quickly obtain segment candidates by preliminary selection from the preliminary selection result database 9 using the decision tree. Moreover, even if synthesis target information that was not included in the test sentence is given, an appropriate result reflecting the result for similar synthesis target information can be quickly returned, so without using a large amount of test sentences. High-speed selection processing can be performed. Note that it is not necessary to use a decision tree, such as evaluating all prepared questions and outputting a preliminary selection result for composite target information having the highest degree of matching of answers to the questions.

  Finally, the preprocessing unit 1 will be described. FIG. 3 is a block diagram of the preprocessing unit 1. According to FIG. 3, the preprocessing unit 1 includes a text processing unit 11 and a synthesis parameter generation unit 12. The text processing unit 11 divides the input text into unit speeches by morphological analysis, syntax analysis, word dictionary reference, and the like, attaches prosodic information obtained by the analysis, and outputs the text information. The synthesis parameter generation unit 12 attaches an acoustic feature parameter to the text information and outputs it.

  Here, the synthesis target information output from the preprocessing unit 1 can be output from the text processing unit 11 or output from the synthesis parameter generation unit 12. In the former case, since the acoustic feature parameter is not included in the synthesis target information, the segment selection unit 4 cannot calculate the target cost. Therefore, the target cost is not calculated for all segment candidates. Treat as 0. That is, the segment selection is performed in consideration of only the connection cost indicating the degradation of the synthesized speech waveform due to the segment connection. Since the error between each segment candidate segment output by the preliminary selection unit 3 and the synthesis target information is considered to be relatively small, in the subsequent selection process, all the target costs are considered to be 0. Even if one-sided selection is performed, the final cost can be considered to be relatively small. On the other hand, the calculation of the connection cost does not require the acoustic feature parameter included in the synthesis target information. Therefore, in the former configuration, the synthesis parameter generation unit 12 can be omitted, and the segment selection unit 4 can omit the target cost calculation that requires the acoustic feature parameter, and the processing amount of speech synthesis can be reduced. It becomes possible to reduce.

  In the above description, the cost deterioration value of a certain segment has been described as the difference between the cost of the optimal route and the optimal cost when the segment is forcibly selected. However, a method such as a ratio to the cost of the optimum route may be used. In addition, the segment to be calculated in the generation of the cost deterioration value database 6 by the cost deterioration value calculation unit 2 is narrowed down by the target cost even as all the pieces included in the segment information database 7, and the target cost value A small predetermined number of pieces may be used. When narrowing down by target cost, the generation processing burden of the cost deterioration value database 6 can be reduced, and when all the pieces are targeted, the more accurate cost deterioration value database 6 can be generated. In addition, when narrowing down by target cost, for the cost deterioration value of the segment that was not subject to calculation by narrowing down, the worst value or higher value among the segments calculated for the synthesis target information Is used.

  The speech synthesizer according to the present invention can be realized by a program that, when read into a computer, causes the computer to perform the operations of the functional blocks shown in FIGS.

1 is a block diagram of a speech synthesizer according to the present invention. It is a block diagram of a preliminary selection part. It is a block diagram of a pre-processing part. It is explanatory drawing of preparation of a cost deterioration value database. It is a figure which shows a cost deterioration value database. It is a figure which shows the decision tree for preparation of a preliminary selection result database. It is a figure which shows a preliminary selection result database.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Pre-processing part 11 Text processing part 12 Synthesis parameter generation part 2 Cost deterioration value calculation part 3 Preliminary selection part 31 Preliminary selection result database generation part 32 Selection part 4 Segment selection part 5 Speech waveform connection part 6 Cost deterioration value database 7 Element Single information database 8 Speech waveform database 9 Preliminary selection result database

Claims (8)

Based on a set of input synthesis target information, a series of speech units is selected from a plurality of speech units prepared in the unit database, and speech waveforms corresponding to each of the selected series of speech units are connected and synthesized. A speech synthesizer that outputs a speech waveform,
Based on a series of test synthesis target information, based on a speech unit sequence selected by a predetermined standard from a segment database, the predetermined standard is used under the condition of using a speech unit for a certain test synthesis target information. The cost degradation value of the speech unit for the test synthesis target information, which indicates the degree of degradation of the synthesized speech waveform in the speech unit sequence selected in Step 1, is obtained, and the cost degradation of the corresponding speech unit for each test synthesis target information A cost deterioration value calculating means for generating a cost deterioration value database indicating values;
For each of the input synthesis target information, one or more test synthesis target information is selected from the cost deterioration value database, and the voice corresponding to the input synthesis target information is selected from the speech unit corresponding to the selected test synthesis target information based on the cost deterioration value. A preliminary selection means for selecting a segment candidate;
A unit selection unit for selecting a series of speech units to be used for a synthesized speech waveform to be output from speech unit candidates for each input synthesis target information selected by the preliminary selection unit based on the series of input synthesis target information; ,
A speech synthesizer characterized by comprising: The preliminary selection means
The test synthesis target information included in the cost deterioration value database is grouped, and a plurality of speech units are selected based on the cost deterioration values from the speech units corresponding to the grouped test synthesis target information. A pre-selection result database generating means for recording in correspondence with a group;
Selecting means for determining a group to which each input synthesis target information belongs, and selecting a speech unit corresponding to the determined group as a speech unit candidate for the input synthesis target information;
The speech synthesizer according to claim 1, comprising: Preliminary selection result database generation means uses all test synthesis target information belonging to the same unit speech as a root node, and groups test test target information included in a leaf node of a decision tree generated by sequentially dividing the root node into one group Grouped as
The selection means determines a group to which the input synthesis target information belongs according to the decision tree,
The speech synthesizer according to claim 2. The preliminary selection result database generation means generates a decision tree by repeating set division to minimize the evaluation value,
The evaluation value is a sum of values obtained by weighting the evaluation value of each leaf node generated by the division based on the number of pieces of test synthesis target information included in each leaf node,
The evaluation value of the leaf node is an average value of the cost deterioration values of a predetermined number of speech units selected based on the cost deterioration value from the speech elements corresponding to each test synthesis target information included in the leaf node. ,
The speech synthesizer according to claim 3. The unit selection means selects a series of speech units considering only the degree of deterioration of the synthesized speech waveform caused by the connection of speech units,
The speech synthesizer according to any one of claims 1 to 4. Cost deterioration value calculation means
Select one or more speech segments corresponding to each series of test synthesis target information from the segment database,
Arrange the selected speech segments in the same order as the series of test synthesis target information,
Corresponding each speech unit to a node of the search graph, by performing cost calculation indicating the degree of degradation of the synthesized speech waveform by dynamic programming from the beginning and the end, the speech unit series as the reference, Obtain the cost of a speech segment sequence under the condition of using the speech segment,
Based on the cost of each acquired speech unit sequence, for each test synthesis target information, to determine the cost deterioration value of the speech unit selected from the unit database,
The speech synthesizer according to any one of claims 1 to 5. Based on a series of test synthesis target information, based on a speech unit sequence selected by a predetermined standard from a segment database, the predetermined standard Using the cost deterioration value database having the cost deterioration value of the speech unit with respect to the test synthesis target information, which indicates the degree of deterioration of the synthesized speech waveform in the speech unit sequence selected in step 1, and synthesized speech from a series of input synthesis target information A speech synthesis method for outputting a waveform,
For each of the input synthesis target information, one or more test synthesis target information is selected from the cost deterioration value database, and the voice corresponding to the input synthesis target information is selected based on the cost deterioration value from the speech unit corresponding to the selected test synthesis target information. Selecting a segment candidate;
Selecting a series of speech units to be used for a synthesized speech waveform to be output from speech unit candidates for each input synthesis target information based on the series of input synthesis target information;
Connecting a speech waveform corresponding to each of a selected series of speech units to generate a synthesized speech waveform;
A speech synthesis method characterized by comprising:   A program that causes a computer to function as the speech synthesizer according to any one of claims 1 to 6.

Images