JP2005321520A - Voice synthesizer and its program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve naturalness of voice information in a voice synthesizer in which voice information of a synthesized sentence, that is made by replacing the variable portion of a reference sentence with replacement words and phrases, is generated by combining voice information obtained by rule-synthesizing the rhythm information of the the replacement words and phrases and the voice information of the reference sentence, and by reading the synthesized sentence using the voice information of the generated synthesized sentence for the synthesis of voice. <P>SOLUTION: The voice synthesizer is provided with a rule synthesizing means 4 in which when prescribed conditions are met, either one of the portion of the synthesized sentence preceding the replacement words and phrases or the portion of the succeeding synthesized sentence is made a rule synthesis expanding portion and the voice information of the rule synthesized expanding portion is rule-synthesized with the voice information of the replacement words and phrases. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a speech synthesizer that performs speech synthesis based on a sentence and a program thereof, and particularly, a synthesized sentence that is synthesized by combining a reference sentence having a replaceable variable part and a phrase that replaces the variable part. The present invention relates to a technology for voice synthesis.

  Conventionally, in speech synthesis technology for synthesizing a synthesized sentence generated by combining a fixed part of a reference sentence and a replacement phrase, the prosody of the fixed part is adjusted while adjusting the prosody of the variable part and referring to the prosody of the variable part. There was one that deformed (for example, Patent Document 1).

JP 11-38989 A

  Conventionally, according to the speech synthesis technology for such a synthesized sentence, the prosodic information of the replacement phrase is adjusted at the connection position of the prosody of the fixed form part and the replacement phrase, and the F0 (fundamental frequency) information of the fixed form part is changed. Smoothing (smoothing). However, in such a method, there is a problem in that connection is not smoothly performed depending on the contents of the variable part and the replacement phrase, and the naturalness of the prosody is deteriorated. The present invention has been made to solve such a problem, and an object of the present invention is to generate a natural prosody without causing deterioration of the prosody even if the variable part is diversified, and to make it easy to hear the synthesized speech. To do.

The speech synthesizer according to the present invention provides speech information of a synthesized sentence obtained by replacing a variable part of a reference sentence with a replacement phrase, speech information obtained by regular synthesis of prosodic information of the replacement phrase, and speech of the reference sentence The speech information generated by combining the information and smoothing the speech information obtained by the rule synthesis among the speech information of the generated synthesized sentence by increasing / decreasing by a predetermined smoothing amount and using the smoothed speech information In a speech synthesizer that performs speech synthesis of the synthesized sentence,
Extended smoothing means for further smoothing the voice information of the synthesized sentence by increasing / decreasing the voice information of the subsequent part of the synthesized sentence following the replacement phrase by the smoothing amount;
It is equipped with.

Such a speech synthesizer may be configured by combining a general-purpose computer and a speech synthesis program that causes the computer to perform speech synthesis processing. That is, the speech synthesis program according to the present invention includes speech information of a synthesized sentence obtained by replacing a variable part of a reference sentence with a replacement phrase, speech information obtained by regular synthesis of prosodic information of the replacement phrase, and the reference The speech information generated by combining the speech information of the sentence and smoothed by increasing / decreasing the speech information obtained by the rule synthesis among the speech information of the generated synthesized sentence by a predetermined smoothing amount. In a speech synthesis program that causes a computer to execute speech synthesis of the synthesized sentence using
Extended smoothing means for further smoothing the voice information of the synthesized sentence by increasing / decreasing the voice information of the subsequent part of the synthesized sentence following the replacement phrase by the smoothing amount;
Is executed by the computer.

  In this way, the speech synthesizer and the program thereof according to the present invention appropriately expands the rule synthesis range of the speech information in the synthesized sentence based on the relation between the reference sentence and the replacement phrase, and the synthesized voice information. Therefore, a natural prosody can be generated at the connection point between the fixed part and the rule synthesis range, and the synthesized speech can be easily heard.

Next, embodiments of the present invention will be described with reference to the drawings.
Embodiment 1 FIG.
FIG. 1 is a block diagram showing the configuration of a speech synthesis apparatus according to Embodiment 1 of the present invention. The speech synthesizer 1 shown in the figure is a speech synthesizer according to Embodiment 1 of the present invention, and includes a synthesized sentence generator 2 as a synthesized sentence generator and a rule synthesis range smoother 3 as a rule synthesis range smoother. A phrase similarity determination unit 3 as a phrase similarity determination unit, an extended smoothing unit 5 as an extended smoothing unit, a local smoothing unit 6 as a local smoothing unit, and a speech generation unit 6 as a speech generation unit. ing.

  The synthetic sentence generation unit 2 is a part that generates a synthetic sentence by replacing a part of a reference sentence with a replacement phrase, and further generates prosodic information of the generated synthetic sentence based on a rule. In this description and the following description, the term “part” is used to indicate a dedicated element or circuit configured to have such a function. However, a general-purpose computer central processing unit (CPU: Central processing Unit) is used. ) And a computer program that causes the CPU to execute equivalent processing may be combined.

  The reference sentence handled by the synthetic sentence generation unit 2 is a sentence composed of a combination of a fixed sentence and a variable part, for example, “Display a map near a nearby [station]”. In this example, the part [station] corresponds to the variable part. The variable part is a part that is assigned a temporary phrase, but is scheduled to be replaced with another phrase in generating a synthesized sentence later. On the other hand, “near” or “display a nearby map” is a part called a fixed part. By replacing the variable part of this reference sentence with a replacement phrase such as “Convenience store” or “Gas station”, “Display a map near a nearby convenience store” or “Display a map near a nearby gas station” Is generated.

  The purpose of the synthetic sentence generation unit 2 is to generate a synthetic sentence by performing character string manipulation processing, and regularly generate prosodic information of the generated synthetic sentence. Can be easily configured based on. Therefore, in the case of the compound sentence generation unit 2, it includes all the configurations that exhibit such an action, and is not limited to a specific configuration. However, in the following, for convenience of explanation, as an example of the configuration of the synthetic sentence generation unit 2, as shown in FIG. 2, a replacement phrase acquisition for acquiring a replacement phrase input from outside the speech synthesizer 1 is shown. It is assumed that the unit 21 includes a reference sentence storage unit 22 that stores a reference sentence, a character string operation unit 23, and a regular voice information generation unit 24.

  Here, the replacement phrase acquisition unit 21 is a part that acquires a replacement phrase applied to the variable part of the reference sentence through, for example, a keyboard. The reference sentence storage unit 22 is a part that stores information about the reference sentence using a storage element or circuit, or a storage medium such as a CD-ROM or a hard disk device. The character string operation unit 23 is a part that generates a character string in which the reference sentence text stored in the reference sentence storage unit 22 is analyzed by, for example, character string analysis, the variable part position is specified, and the replacement phrase is substituted into the variable part position. is there.

  FIG. 3 is a diagram illustrating an example of reference sentence data stored by the reference sentence storage unit 22 as information regarding a reference sentence. As shown in the figure, the reference sentence data 30 includes a plurality of records such as a record 31 and a record 32. Each of these records has an ID 301, a reference sentence 302, prosodic information 303, F0 information 304, and time length information 305 as fields (record items) as shown in FIG.

  The ID 301 is an identifier that uniquely identifies the record 31. The reference sentence text 302 is character string information represented by the record 31. In this example, a character string with the content “Display a map near a [station] nearby” is stored in the field of the reference text 302. Here, a portion surrounded by “[” and “]” indicates a variable portion, that is, a character string portion to be replaced with a replacement phrase. The other character string parts, that is, “near” and “display a map of the vicinity” are called fixed parts. The symbols “[” and “]” that specify the position of the variable portion may be other symbols or characters (including control characters). In addition, the reference text 302 is expressed as plain text (eg, “A map near the station is displayed”) without using the symbols indicating the start position and end position of the variable section, and the start of the variable section is displayed. The position (fourth character from the beginning) and the end position (fourth character from the beginning) may be held as separate fields of the record 31.

  The prosody information 303 is a field for holding prosody information when the reference text 302 is read out, and stores the number of mora and the accent type in this example. The F0 (read as F zero) information 304 is a field for storing an array of F0 information (fundamental frequency) when the character string indicated by the reference sentence text 302 is read out (however, a symbol indicating a variable part is not read out). . Here, the F0 information 304 stores the F0 information when the reference sentence 302 is read out without distinguishing between the fixed part and the variable part. In natural prosody, the speech information of the front and rear fixed parts varies depending on what words are actually present in the variable part. Therefore, the speech information (prosodic information, F0 information, time length information, etc.) for the reference sentence text 302 when the phrase “station” is assigned to the variable part is often different from the speech information when other words are assigned. . As a method of expressing the variable part of the reference sentence 302, an expression method (for example, “near * near * near” is assigned to an abstract variable field without assigning a specific word or phrase such as “station”. It is also conceivable to use “*” to indicate that it is a variable part and to be used as a symbol that does not read aloud, as in “Displays a map”. However, since humans do not read a sentence like "Display a map of nearby *" in a natural conversation or life, such an abstract expression is the source of speech synthesis. Natural voice information that cannot be obtained. For this reason, in Embodiment 1 of the present invention, a sentence that actually exists is designated as a reference sentence, and a part thereof is expressed as a variable part.

  The time length information 305 is a field in which time length information of each phoneme is digitized and stored. Also, the records after the record 32 are provided with fields corresponding to the ID 301, the reference text 302, the prosody information 303, the F0 information 304, and the time length information 305.

  Next, the configuration of the speech synthesizer 1 in FIG. 1 will be described. The phrase similarity determination unit 3 is a part that determines the similarity between the replacement phrase and the phrase originally in the variable part of the reference sentence. Here, as a similarity between the replacement phrase and the phrase originally in the variable part of the reference sentence, it is determined whether or not they are prosodic. The rule synthesizing unit 4 is a part that regularly synthesizes speech information for some words in the synthesized sentence. The part to be regularly synthesized is determined based on the determination result of the phrase similarity determination unit 3.

  The local smoothing unit 5 is a part that performs smoothing in order to smooth the connection between the speech information in the range that is regularly synthesized by the rule synthesis unit 4 and the speech information in the other part of the synthesized sentence. The voice generation unit 6 is a part that generates a voice signal based on the generated voice information and reproduces it as a voice that can be recognized through hearing.

  Next, the operation of the speech synthesizer 1 will be described. FIG. 4 is a flowchart showing the operation of the speech synthesizer 1. Here, it is assumed that “gas station” is specified as a replacement phrase from the outside, and information (for example, an ID value or a reference sentence itself) that identifies a record of reference sentence data to be combined with this replacement phrase.

  On the other hand, the replacement phrase acquisition unit 21 of the synthetic sentence generation unit 2 identifies information (here, ID number 001) that identifies the replacement phrase “gas station” and the record of reference sentence data to be combined with this replacement phrase. Is acquired (step S101). The character string operation unit 23 displays the contents of the reference sentence text 302 of the record 31 identified from the ID number 001 among the reference sentence data 30 stored in the reference sentence storage unit 22 “a map near the nearby [station]”. The variable part of this character string and the replacement phrase “gas station” acquired by the replacement phrase acquisition unit 21 are replaced with the compound sentence “Display a map near the nearby gas station.” Is generated (step S102).

  Next, the phrase similarity determination unit 3 determines prosodic similarity between the replacement phrase and the variable phrase (step S103). When prosodically similar, the speech information of the replacement phrase and the variable part phrase is similar and considered to be highly compatible. Therefore, in such a case, the connection between the voice information before and after the variable part and the voice information obtained by regularly synthesizing the replacement phrase is highly natural, and a voice without any sense of incongruity can be obtained even if connected as it is. On the other hand, if the replacement phrase and the variable part phrase are not prosodically similar, the effect of the replacement phrase on the prosody of the other part in the synthesized sentence is that the variable part phrase in the reference sentence is different from the other part. It is different from the effect. For this reason, even if the prosody of the variable part and the prosody of the replacement phrase are connected, naturalness cannot be obtained. Therefore, the phrase similarity determination unit 3 determines the prosodic similarity between the replacement phrase and the variable part phrase, and separates the subsequent processing depending on whether it is determined to be similar or dissimilar. It was decided.

  As an example of the prosodic similarity determination method, a method of comparing the number of mora and accent type of each of the replacement phrase and the variable part phrase can be considered. In this case, for example, when the accent type is the same and the number of mora is similar, the prosodic compatibility is increased, so that it can be determined that the prosodic similarity is high. Whether or not the number of mora is similar may be determined by obtaining a difference between the respective mora numbers and determining whether or not the difference is within a predetermined value. For example, if the difference in number of mora is 3 or less, it is determined as “similar”, and if it exceeds 3, it is determined as “dissimilar”. In the case of the part phrase “station”, the accent type is -4 and matches, but the number of mora is 8 and 2, respectively, so the difference is 6 and the condition of 3 or less is not satisfied. Accordingly, it is determined that the replacement phrase “gas station” and the variable phrase “station” are dissimilar in terms of prosody.

  Generally, it is known that the accent intensity increases as the number of mora increases. Therefore, when the difference in the number of mora between the replacement phrase and the variable part phrase increases, the error in the F0 information after replacement increases due to the difference in accent strength. The determination of similarity based on the above accent type and the difference in the number of mora is based on such knowledge.

  If it is determined in step S103 that the replacement word and the variable part are prosodically similar, the process proceeds to step S105. On the other hand, if it is determined that the replacement word and the variable part are dissimilar in terms of prosody, the process proceeds to step S105 after step S104. Therefore, in the following, step S104, which is executed when it is determined that the replacement phrase and the variable part are prosodically dissimilar, will be described first, and step S105 will be described later.

  Next, the rule synthesis unit 4 expands the range for rule synthesis (step S104). Since the reference sentence data 30 does not store the speech information of the replacement phrase part in the synthesized sentence, at least the speech information of the replacement phrase part is regularly synthesized. Therefore, the replacement phrase part (“gas station”) is the basic rule composition range. Further, here, in step S103, it is determined that the replacement phrase ("gas station") and the variable part ("station") are dissimilar in terms of prosody. Therefore, even in such a case, in order to make the speech information of the synthesized sentence highly natural, the range of rule synthesis is extended to a part other than the replacement phrase part. For example, the phrase immediately following the replacement phrase part (for example, “near”) is also included in the range for rule synthesis.

  In this way, not only the replacement phrase part but also the surrounding part is rule-synthesized, so even if the replacement phrase and the variable part are not prosodically similar, the speech information of the synthesized sentence Natural nature can be improved.

  As an example of a method for expanding the range of rule synthesis, there is a method in which the subsequent part “display a nearby map” is determined as a range for smoothing the first morpheme (in this case, “near”) by performing morphological analysis. is there. Further, the amount (number of morphemes) for extending the range of rule synthesis may be increased based on the similarity (degree of numerical value representing similarity) calculated by the phrase similarity determination unit 3. For example, in the phrase similarity determination unit 3, if the accent type of the replacement phrase is different from that of the variable part, and the number of mora also differs by a certain number or more, the degree of influence of the prosodic information in the subsequent part is greatly different. The morpheme (in this case “near”) or three morphemes (“near map”) may be determined as a range to be smoothed.

  In this way, since it is determined whether or not rule synthesis is performed for each morpheme following the substitution word, smoothing is performed in units of morphemes that often cause utterance breaks, and more natural speech information can be obtained. It is. Also, as the similarity between the replacement word and variable part becomes poor, the range of rule composition is expanded so that the length of the part that requires rule composition even when the influence of the word change is large. Can be determined dynamically, and natural speech can be obtained.

  It should be noted that the expansion of the range for rule synthesis is not limited to only the phrase that follows the replacement phrase. For example, depending on the vocabulary of the replacement phrase and how the variable part is configured (what the reference sentence 30 is designed as a variable part), the phrase preceding the replacement phrase may be included in the range of rule synthesis. Moreover, it is good also as both (both a preceding phrase and a succeeding phrase).

  Then, the rule synthesizing unit 4 generates speech information by rule synthesis for words in the range to be rule synthesized (step S105). As a rule synthesis method, a conventionally known speech synthesis method is used. For example, when the phrase to be subject to rule synthesis is “near a gas station”, the time length of each phoneme g, a, s, o, r, i, N, s, u, t,. The value is generated by a conventional speech synthesis method. One example of a rule-based voice generation method is a control model called point pitch. In this model, the pitch (F0) tends to decrease in the whole sentence. Therefore, the slope pattern is represented by a straight line, and the component that forms the accent added on it is represented by a trapezoid, and the center point of each mora The pitch (F0) is determined. The start and end points of the straight line, the height of the trapezoid, and the like are determined from the accent position, the number of mora, and the like of the synthesized sentence “Display a map near the nearby gas station” generated by the character string operation unit 23. The correspondence between the accent position, the number of mora, the start and end points of the straight line, and the height of the trapezoid can be described in advance in the table.

  Next, the local smoothing unit 5 first calculates a smoothing amount in order to smooth the speech information of the synthesized sentence that has been regularly synthesized (step S106), and then uses the smoothing amount for the speech information of the synthesized sentence. To smooth (step S107). The speech information that is rule-synthesized by the rule-synthesizing unit 4 does not take into account the connectivity between the fixed part of the reference sentence and the replacement phrase, and therefore often has low naturalness and is difficult to hear. Therefore, the local smoothing unit 5 smoothes the voice information of the synthesized sentence to enhance the naturalness.

式（１）において、Ｗ１及びＷ２は重み係数であって、Ｗ１＋Ｗ２＝１となる関係にある数値である。 Here, the calculation of the smoothing amount in step S106 is performed as follows. Now, as shown in FIG. 5, the F0 information of the last part of the fixed part immediately before the replacement phrase is B0, the F0 information at the beginning of the replacement phrase is H0, the F0 information at the end of the replacement phrase is T0, and the fixed part immediately after the replacement phrase is When the F0 information of the first part is A0, the smoothing amount Δ (delta) is calculated as a weighted average according to the equation (1).

In Expression (1), W1 and W2 are weighting factors, and are numerical values in a relationship of W1 + W2 = 1.

  In equation (1), W1 and W2 may be changed according to the variable part and the type of phoneme of the replacement phrase. By doing in this way, the content of smoothing processing can be switched individually and specifically based on the type of variable part or replacement phrase, and the naturalness of the synthesized speech can be increased and more easily audible speech information can be obtained. it can.

  As the F0 information B0 of the last part of the fixed part immediately before the replacement word, the F0 information of the last mora of the fixed part immediately before the replacement word can be considered as an example, but this is not restrictive. That is, since the F0 information of the mora adjacent to the variable part is often transformed by the variable part, it is not necessarily the last mora, not the F0 information of the last part of the fixed part immediately before the replacement phrase, The F0 information may be acquired from the vicinity of the last mora that is not easily deformed by the variable part. Similarly, the F0 information at the head of the replacement phrase may be used as the F0 information at the head of the replacement phrase, or F0 information near the head mora may be used. The F0 information T0 at the end of the replacement phrase and the F0 information A0 of the first part of the fixed part immediately after the replacement phrase may also be selected from the vicinity of the replacement phrase end mora or the first mora of the fixed part immediately after the replacement phrase.

  The speech information of the replacement phrase in step S107 is smoothed by uniformly increasing / decreasing (shifting) Δ calculated in step S106 to the F0 information of the replacement phrase in this section. In this way, when the first and last F0 information of the replacement phrase is far away from the F0 information of the fixed part connected to them, the difference between the F0 information is increased or decreased by increasing or decreasing the entire replacement phrase. The naturalness of the prosody is increased by reducing the size.

  After determining the range of the subsequent part to be smoothed in this way, the local smoothing unit 5 increases or decreases the audio information of the subsequent part to be smoothed by the same amount as the smoothing amount of the speech information of the replacement phrase part. The smoothing amount used here is, for example, Δ that has already been calculated by the equation (1). When each F0 information is increased by Δ as the speech information of the replacement phrase portion, each F0 information of the subsequent portion to be smoothed is also increased by Δ.

  In the above description, B0, H0, T0, and A0 of Equation (1) are calculated from the connection positions of the replacement phrase and the fixed phrases before and after it, and Δ is calculated from Equation (1), which is used as the subsequent part. The smoothing amount is used. However, the present invention is not limited to such a method. For example, the replacement word and the subsequent portion to be smoothed determined by the local smoothing unit 5 are regarded as an integrated range, and the integrated range and the front and rear fixed portions are Alternatively, B0, H0, T0, and A0 may be obtained from the connection position, Δ may be calculated first, and the entire F0 information in the integrated range may be increased or decreased using the calculated Δ. By using such a method, it becomes possible to reflect the connection state between the succeeding portion and other fixed sentences in the smoothing amount, so that a more natural connection is possible.

  Further, in step S107, when the boundary of the range to be rule-combined becomes a phrase delimiter, the local smoothing unit 5 includes the F0 information of the mora (for example, the last mora) near the end of the rule-synthesized range and its rule synthesis. The F0 information of the mora (for example, the second mora) in the vicinity of the beginning of the portion following the range to be connected is connected by a straight line, and the F0 information of the range to be ruled and the subsequent portion is transformed and further smoothed. May be. For example, if “near the gas station” is the range where the rule is synthesized, the boundary of the range where the rule is synthesized is a phrase delimiter. Therefore, the F0 information of the last mora in this range and the “map display” The F0 information of the second mora is connected by a straight line, and the range for rule synthesis and other ranges are transformed.

  In this way, the local smoothing unit 5 performs smoothing so that the speech information of the subsequent part of the replacement phrase and the subsequent synthesized sentence part is smoothly connected. Even if the natural smoothness cannot be obtained only by increasing or decreasing the simple smoothing amount for the portion, a sufficiently natural synthesized speech can be finally obtained as a whole synthesized sentence.

  It should be noted that the mora that is connected by a straight line is not limited to the second mora of the part that follows the last mora of the range to be regularly synthesized, but may be another mora, for example, the last mora instead of the last mora, A third mora may be used instead of the second mora in the subsequent portion. That is, it may be a nearby mora. Here, the term “neighboring” refers to a mora in a position close to the level expected to have a phonologically close property. This is because, in natural speech information, the speech information of mora that are close to each other often has a close value.

  Furthermore, not only the replacement phrase and the subsequent part of the replacement phrase, but also a mora near the beginning of the replacement phrase and a mora near the end of the preceding part of the replacement phrase may be connected with a straight line. Needless to say.

  Connecting with a straight line means, for example, interpolating the value of the F0 information of the mora existing in the middle using the F0 information of the mora at both ends serving as a reference. In addition, since the reason for performing such a process is to change the F0 information so that it smoothly changes, for example, a value on a continuous curve may be taken in addition to a straight line.

  The voice generation unit 6 generates voice from the voice information of the synthesized sentence obtained so far (step S109). Since this processing is the same as that of the conventional technology, the details are omitted.

  As described above, according to the speech synthesizer 1, not only the speech information of the replacement phrase but also the speech information of the subsequent part immediately after the replacement phrase is smoothed. It is possible to obtain voice information that generates a natural and easy-to-hear voice compared to the case of smoothing.

  In the above-described example, the description has been made mainly on the smoothing of the F0 information. However, the time length information may be modified similarly to the F0 information. In this case, the amount of deformation, that is, the smoothing amount, is the difference between the average time length of each mora in the fixed form part and the average time length of each mora of the replacement phrase or the phrase including the replacement phrase and the immediately following part. Calculate from Further, the time length of the mora of the fixed part adjacent to the variable part is changed by replacing the time length created by the rule corresponding to this mora with the weighted average of the time length of the mora of the fixed part. By doing in this way, the time length of the phoneme of the fixed part corresponding to the type of phoneme of the variable part can be obtained, and a natural combination of time length information can be obtained.

  Moreover, in the above-mentioned example, although the range which deform | transforms F0 information and time length information was made into the same range, you may determine these independently. That is, different boundary values are set for the difference in the number of mora for the F0 information and the time length information. For example, if the boundary value is set to 1 for F0 information and the boundary value is set to 3 for time length information, a map near “Station” will be displayed. When the variable part “station” of “is replaced by the replacement word“ convenience store ”, the number of mora of“ convenience store ”is 4, so the difference becomes 2 and exceeds the boundary value for F0 information, while the time length information The boundary value for is not exceeded. Accordingly, “station” and “convenience store” are treated as dissimilar for F0 information, and “station” and “convenience store” are treated as similar for time length information. Therefore, only the F0 information is subjected to smoothing of the subsequent portion by the extended smoothing unit 5.

  By handling F0 information and time length information independently in this way, the extended smoothing unit 5 increases or decreases only one of the F0 information and time length information of the subsequent portion by the smoothing amount. With such a configuration, smoothness is required for F0 information, but the time length information is sufficiently natural so that adjustment is necessary or vice versa. Is possible.

Embodiment 2. FIG.
In the speech synthesizer according to the first embodiment, whether or not to perform the extended smoothing process is determined based on the prosodic similarity between the replacement phrase and the variable part phrase. However, in addition to such a configuration, when smoothing (shifting) the speech information of the replacement phrase, extended smoothing processing is performed when the difference between the replacement phrase and the prosodic information at the connection point between the preceding and following fixed parts is large. You may make it perform. The speech synthesizer according to the second embodiment has such a feature.

  FIG. 6 is a block diagram showing a configuration of a speech synthesis apparatus according to Embodiment 2 of the present invention. In the figure, the components denoted by the same reference numerals as those in FIG. 1 are the same as those in the first embodiment, and the description thereof will be omitted. In the figure, a forward difference calculation unit 7 as a forward difference calculation means is a part that calculates a difference (forward difference) in prosodic information at a connection point between a replacement word and a fixed form part that precedes it. That is, the forward difference calculation unit 7 acquires the F0 information of the first mora of the replacement phrase from the speech information obtained by increasing or decreasing the smoothing amount of the speech information obtained by the rule synthesis, and the synthesized sentence preceding the replacement phrase Is obtained, and the difference between the F0 information of the head mora and the F0 information of the synthesized sentence is calculated.

  Next, the operation of the speech synthesizer 1 according to the second embodiment will be described. FIG. 7 is a flowchart showing the operation of the speech synthesizer 1. Since this flowchart is different from FIG. 4 which is the flowchart of the first embodiment only in steps S201 to S203 and S105-2, the following description will be made focusing on these processes. Also in the second embodiment of the present invention, similarly to the first embodiment, the synthesized sentence generating unit 2 generates a synthesized sentence until step S102. Subsequently, the rule synthesizing unit 4 synthesizes the speech information of the replacement phrase part with a rule (step S201). The rule synthesis method is the same as in the first embodiment.

  Thereafter, the forward difference calculation unit 7 calculates the difference in the speech information at the connection point between the replacement word part and the previous fixed part (step S202). It is the difference of F0 information that is used as this difference. For example, F0 information of the first mora of the replacement phrase part and F0 information of the last mora of the fixed part before the replacement phrase part are acquired, and the difference between them is obtained as a forward difference.

  Subsequently, the forward difference calculation unit 7 compares the calculated forward difference with a predetermined threshold value (step S203). If the forward difference exceeds the threshold value, the connection between the replacement word and other parts is unnatural, and the process proceeds to step S104. If the forward difference is less than or equal to the threshold value, the process proceeds to step S106. Since the process after step S106 is the same as that of Embodiment 1, description is abbreviate | omitted.

  In step S104, the rule synthesizing unit 4 extends the range for rule synthesis as in the first embodiment. The rule synthesizing unit 4 synthesizes the speech information in the expanded rule synthesis range (step S105-2). The rule composition method is the same as in step S201. Then, similarly to the case where the forward difference is equal to or smaller than the threshold value, the process proceeds to step S106 and subsequent steps.

  As a result of the smoothing process by the rule synthesizing unit 4 as described above, if the connection is not sufficiently smooth, natural speech information is obtained as the entire synthesized sentence by performing the smoothing process of the subsequent part. Easy voice can be generated.

  In this example, it is determined whether or not to expand the range of the words to be ruled based on the connection between the replacement word and the preceding word. However, instead of this, the range of words to be rule-combined may be expanded based on the connection between the replacement word and the subsequent word. That is, instead of the forward difference calculation unit 7, F0 information of the terminal mora of the replacement phrase obtained by rule synthesis is acquired, F0 information of the synthesized sentence part (rear part) after the replacement phrase is acquired, and the first mora is acquired. A backward difference calculating means for calculating a difference (backward difference) between the F0 information and the rear portion F0 information may be used.

Embodiment 3 FIG.
Whether or not the range to be smoothed is determined based on the similarity between the replacement word and the variable part as in the first embodiment, but whether the word following the replacement word and the replacement word have a certain relationship You may make it determine based on. For example, when the word “near” follows the replacement word “gas station”, the voice information of the word “near” is always smoothed. The speech synthesizer according to the third embodiment has such a feature.

  FIG. 8 is a block diagram showing a configuration of a speech synthesizer according to Embodiment 3 of the present invention. In the figure, the components denoted by the same reference numerals as those in FIG. 1 are the same as those in the first embodiment, and the description thereof will be omitted. In the speech synthesizer 1 in the figure, the related phrase determination unit 8 as a related phrase determination unit is a part that determines whether or not a replacement word and a word following the replacement word have a predetermined relationship.

  Next, the operation of the speech synthesizer 1 will be described. FIG. 9 is a flowchart showing the operation of the speech synthesizer 1. Since this flowchart is different from FIG. 4 which is the flowchart of the first embodiment only in step S301, the following description will be focused on this process. Up to step S102, the synthesized sentence generation unit 2 generates a synthesized sentence. In step S301, the related phrase determination unit 9 determines whether the phrase following the replacement phrase is a phrase having a predetermined relationship with the replacement phrase. For this determination process, the related word / phrase determination unit 9 stores the correspondence relationship between the replacement word / phrase and the subsequent word / phrase in a storage device (not shown) as a replacement word / phrase list. In step S301, when the related phrase determination unit 9 acquires the synthesized sentence, the related phrase determination unit 9 searches whether or not the replacement phrase and the subsequent phrase are stored in the replacement phrase corresponding phrase list. Judge words as related words. If the subsequent phrase is a related phrase, the rule synthesizing unit 4 proceeds to step S104 and expands the range for rule synthesis. If it is not a related phrase, the process proceeds directly to step S105 to synthesize the speech information within the range to be synthesized.

  As is clear from the above, according to the speech synthesizer 1 of Embodiment 3 of the present invention, when the replacement phrase and its subsequent phrase have a special relationship, the speech information including the subsequent phrase is regularly synthesized. As a result, it becomes possible to deal with the case where a special accent is generated by the word-to-word combination of the replacement phrase and the succeeding phrase, and it becomes possible to synthesize a voice that is easy to hear under natural prosody. .

  The speech synthesizer and the speech synthesis program according to the present invention are particularly applied to a speech synthesizer that synthesizes a synthesized sentence synthesized by combining a reference sentence having a replaceable variable part and a phrase that replaces the variable part. be able to.

It is a block diagram which shows the structure of the speech synthesizer by Embodiment 1 of this invention. It is a block diagram which shows the detailed structure of the speech synthesizer by Embodiment 1 of this invention. It is a figure which shows the example of the reference sentence data memorize | stored as information regarding a reference sentence. It is a flowchart which shows operation | movement of the speech synthesizer by Embodiment 1 of this invention. It is explanatory drawing for demonstrating the method of calculating the smoothing amount in Embodiment 1 of this invention. It is a block diagram which shows the structure of the speech synthesizer by Embodiment 2 of this invention. It is a flowchart which shows operation | movement of the speech synthesizer by Embodiment 2 of this invention. It is a block diagram which shows the structure of the speech synthesizer by Embodiment 3 of this invention. It is a flowchart which shows operation | movement of the speech synthesizer by Embodiment 3 of this invention.

Explanation of symbols

2 compound sentence generator,
3 Phrase similarity determination unit,
4 Rule composition part,
5 local smoothing part,
6 voice generator,
7 Forward difference calculation unit,
8 Related phrase determination unit.

Claims (12)

Generates speech information of a synthesized sentence obtained by replacing a variable part of a reference sentence with a replacement phrase by combining voice information obtained by regular synthesis of prosodic information of the replacement phrase and voice information of the reference sentence. In a speech synthesizer that synthesizes speech that reads out the synthesized sentence using the synthesized speech information,
When a predetermined condition is met, any part of the synthetic sentence part preceding the replacement phrase and the subsequent synthetic sentence part is defined as a rule synthesis extension part, and the speech information of the rule synthesis extension part is defined as the rule synthesis extension part. Rule synthesis means for synthesizing rules together with the speech information of the replacement phrase,
A speech synthesizer characterized by comprising: A phrase similarity determination means for determining whether or not the variable part phrase and the replacement phrase are prosodically similar;
The rule synthesizing means performs rule synthesis of the speech information of the rule synthesis extension part when the phrase similarity judgment means judges that the variable part phrase and the replacement phrase are not prosodically similar. The speech synthesizer according to claim 1. The phrase similarity determination means has a difference between the number of mora of the phrase of the variable part and the number of mora of the replacement phrase within a predetermined value, and the accent type of the phrase of the variable part and the accent type of the replacement phrase The speech synthesizer according to claim 2, wherein when the two match, it is determined that the phrase of the variable part and the replacement phrase are prosodically similar. 4. The speech synthesizer according to claim 3, wherein the rule synthesizer synthesizes morpheme speech information preceding or succeeding the replacement word as the rule synthesis extension part. The phrase similarity determination means calculates a numerical value indicating the similarity between the variable part phrase and the replacement phrase,
5. The speech synthesizer according to claim 4, wherein the rule synthesizing unit increases or decreases the number of morphemes included in the rule synthesis extension portion according to the numerical value. Local smoothing means for further smoothing speech information from a mora near the end of the replacement phrase to a mora near the beginning of the portion following the replacement phrase;
The speech synthesis apparatus according to claim 1, further comprising: Local smoothing means for further smoothing speech information from a mora near the beginning of the replacement phrase to a mora near the end of the portion preceding the replacement phrase;
The speech synthesis apparatus according to claim 1, further comprising: Forward difference calculating means for calculating a difference between the F0 information of the first mora of the speech information obtained by rule synthesis from the replacement phrase and the terminal F0 information of the synthesized sentence part preceding the replacement phrase;
2. The speech synthesizer according to claim 1, wherein the rule synthesizing unit synthesizes the speech information of the rule synthesis extended portion when the difference calculated by the forward difference calculation unit is equal to or greater than a predetermined value. . A backward difference calculating means for calculating a difference between the F0 information of the terminal mora of the speech information obtained by rule synthesis from the replacement phrase and the head F0 information of the synthesized sentence portion following the replacement phrase;
2. The speech synthesizer according to claim 1, wherein the rule synthesizing unit synthesizes the speech information of the rule synthesis extended portion when the difference calculated by the backward difference calculating unit is equal to or greater than a predetermined value. . Related phrase determining means for determining whether the preceding phrase or the succeeding phrase of the replacement phrase in the synthetic sentence is a predetermined phrase related to the replacement phrase;
The rule synthesizing means, when the subsequent phrase is a predetermined phrase related to the replacement phrase, synthesizes the speech information using the preceding phrase or the subsequent part as the rule synthesis extension part. The speech synthesizer according to claim 1. Generates speech information of a synthesized sentence obtained by replacing a variable part of a reference sentence with a replacement phrase by combining voice information obtained by regular synthesis of prosodic information of the replacement phrase and voice information of the reference sentence. In a speech synthesis program that causes a computer to execute a process of synthesizing speech that reads out the synthesized sentence using speech information of the synthesized sentence,
Rule synthesizing means for rule-synthesizing the speech information of the rule synthesis extension part together with the speech information of the replacement phrase, with any part of the part preceding and following the replacement phrase of the synthesis sentence as a rule synthesis extension part ,
A speech synthesizer characterized by causing the computer to execute. 12. The speech synthesis program according to claim 11, wherein the rule synthesis means determines whether to synthesize the speech information of the rule synthesis extension portion based on a predetermined condition.

Images