JP2000047680A - Voice information processing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To generate such a pitch pattern hardly causing unnatural synthesized sound by generating information on the modulation of the pitch of a unit for every unit when a mora or syllable is regarded as the unit and generating the pitch pattern by using the information on the modulation of pitches of those units. SOLUTION: A text analysis part 2 takes a text analysis of an inputted text such as a morpheme analysis to divide the text into analytic units such as morphemes and gives various information on characteristic accent types, parts of speech, etc., by readings and its analytic units. A unit pitch information generation part 3 generates unit text information on units from various information on analytic units when moras of a synthesized sound are regarded as the units and generates information on the modulation of pitches of the units for every unit. A pitch pattern generation part 4 generates a pitch pattern based on the information on the modulation of unit pitches and various information of the analysis units.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech information processing apparatus for generating a pitch pattern for synthesizing a desired speech from a text, and more particularly to a pitch pattern in which a synthesized sound accent is less likely to be unnatural. The present invention relates to a voice information processing apparatus and method capable of performing voice processing.

[0002]

2. Description of the Related Art In order to generate a synthesized speech that is naturally and correctly understood from an arbitrary text, it is necessary to add a correct accent to the speech to be generated.
However, in the current technology, there is an error in the addition of the accent, and the error has a large effect on a synthesized speech to be generated, so that an unnatural and incomprehensible synthesized speech may be generated. .

[0003] In addition, with the improvement in the quality of speech synthesis, it is thought that reading out by dialect or generation will be required, but the common word accent which is the reference and the way of giving the accent are different. In order to realize such reading, a technique that can easily reconstruct rules for adding accents is required.

In general, accents are given to synthesized speech in the following procedure. First, the input text is subjected to morphological analysis to divide the text into morpheme units, read each morpheme, and add information such as part of speech and accent type. Here, the accent type is information indicating where in the reading of the word there is an accent, and is defined as the number of mora (beats) from the beginning of the word to the accent nucleus (the position where the fundamental frequency starts to drop sharply).

In actual reading, even if the accent type of a morpheme is determined, a natural accent cannot be given. This is because the combination of connected morphemes deforms or disappears the morpheme-specific accent type, and a new accent type is generated. Therefore, next, several morphemes are put together to create an accent phrase having at most one accent nucleus, and an accent type is given to each accent phrase.

[0006] By generating a pitch pattern corresponding to the accent type thus given and generating a speech waveform according to the pitch pattern, an accent is given to the synthesized speech.

[0007] In the case of giving an accent in the above-described procedure, performance in a part for giving an accent type to an accent phrase is particularly important. In general, an accent type is assigned to an accent phrase by an expert in linguistic analysis by creating a rule for determining the accent type based on the properties of the morphemes constituting the accent phrase, and using the created rule. (Sakazaka and Sato: "Accent Rules for Japanese Word Chain" Transactions of the Institute of Electrical Information Engineers, Vol. J66-
D, No7, pp. 849-856)

[0008]

In the above-described method of setting an accent type using a rule created by a skilled person, if a word that is not taken into consideration by the rule appears, it is impossible to set an accent. The type may be set. In addition, there is a problem in that it is not easy to realize reading aloud by dialect / generation because a great deal of labor is required to construct rules.

Further, when determining the accent type of the accent phrase, it is necessary to consider not only the morphological properties of the accent phrase but also the movement of the accent nucleus due to the phonological properties. For example, "Sendai City"
Is composed of the morphemes "Sendai" and "City", and the most likely accent core position is "i", considering the nature of those morphemes, but in fact, the previous mora "da"
Is uttered with an accent nucleus. This is thought to be due to the fact that the accent nucleus hardly comes to the double vowel "i" and the accent nucleus moves to the previous mora.

In general accent rules described in an accent dictionary or the like, a rule for determining the position of an accent nucleus based on the properties of words constituting an accent phrase and a rule for moving an accent nucleus according to the type of phoneme at the position of an accent nucleus Are separately described. Even in the conventional method, the position of the accent nucleus is determined based on the morpheme property of the accent phrase, and then the accent nucleus is moved based on the phonological property. However, in this method, in order to create a rule that determines the position of the accent nucleus from the morphological properties, it is necessary to prepare data from which the accent type has been extracted without the effect of the movement of the accent nucleus by the type of phoneme ( In the case of "Sendai City", the data that the accent nucleus comes to "i"), but the accent type in the actual utterance is
It is not easy to distinguish whether or not the result is a movement of the accent nucleus depending on the type of phoneme.

In the conventional method, one accent type is given to one accent phrase, and a pitch pattern is generated in accordance with the accent type. Therefore, even if the accuracy of the given accent type is low, the accuracy is high. The same speech as in the case is synthesized. Thus, if the given accent type is wrong, the synthesized speech will be very unnatural.

SUMMARY OF THE INVENTION The present invention has been made in consideration of the above circumstances, and has as its object to provide a speech information processing apparatus which rarely gives an impossible accent to synthesized speech.

Another object of the present invention is to provide a speech information processing apparatus capable of automatically constructing rules necessary for giving accents from a database.

Still another object of the present invention is to generate information necessary for giving an accent by simultaneously considering various types of phonemes and the properties of words constituting a text.
It is to provide a voice information processing device.

[0015]

According to a first aspect of the present invention, there is provided a speech information processing apparatus comprising: a text input unit for inputting a text; a text analysis unit for analyzing the text input to the text input unit; A unit pitch information generating unit configured to generate information on a pitch of the unit in units of mora or syllables according to the text analysis result obtained by the text analyzing unit for each unit; And a pitch pattern generation unit for generating the pitch pattern using information on a unit pitch obtained by the unit.

In the audio information processing apparatus according to a second aspect of the present invention, the subjective height of the unit pitch is called a unit subjective pitch height, and when the unit subjective pitch height has a plurality of categories, The information about the pitch of the pitch, the voice information processing apparatus according to a third aspect of the present invention, characterized in that the unit corresponds to any of the categories of the unit subjective pitch height, the pitch of the unit Is characterized by being the probability that the unit corresponds to each category of the unit subjective pitch height.

In the voice information processing apparatus according to a fourth aspect of the present invention, the subjective change in pitch of the unit pitch is called a unit subjective pitch change, and when the unit subjective pitch change has a plurality of categories, The information on the pitch of the unit is characterized by which of the categories of the unit subjective pitch change the unit belongs to.

In the voice information processing apparatus according to a fifth aspect of the present invention, the information on the pitch of the unit is a probability that the unit falls into each category of the unit subjective pitch change. .

[0019] In a speech information processing apparatus according to a sixth aspect of the present invention, the information on the pitch of the unit is the accuracy with which the subjective pitch of the unit changes from high to low.

In the voice information processing apparatus according to a seventh aspect of the present invention, in the unit pitch information generating means, information relating to the pitch of the unit is sequentially generated in the order in which the unit appears in the text. When generating the information on the pitch of the unit, the information on the pitch of the unit generated for the unit before the unit is used.

[0021] In the voice information processing method according to an eighth aspect of the present invention, the pitch pattern generating means uses the information on the pitch of the unit generated by the unit pitch information generating means, using the input information. The text is divided into accent phrases and an accent type is set for the accent phrases.

A speech information processing method according to a ninth aspect of the present invention divides the input text into accent phrases and generates information on the accent phrases in accordance with a text analysis result obtained by the text analysis means. The unit pitch information generating unit generates information on the pitch of the unit using information on the accent phrase generated by the accent phrase dividing unit, and generates the pitch pattern. In the means, an accent type for the accent phrase divided by the accent phrase dividing means is set by using information on the pitch of the unit generated by the unit pitch information generating means.

In the audio information processing method according to a tenth aspect of the present invention, the information on the pitch of the unit is a probability that the unit falls into each of the categories of the unit subjective pitch, and the pitch pattern generation In the means, a pitch pattern corresponding to the accuracy is generated.

In the audio information processing method according to an eleventh aspect of the present invention, the information on the pitch of the unit is a probability that the unit falls into each category of the unit subjective pitch change, and the pitch pattern The generating means generates a pitch pattern according to the accuracy.

In the voice information processing method according to a twelfth aspect of the present invention, the information on the pitch of the unit is a probability that the unit takes each of a plurality of types of values representing subjective pitch. In the pitch pattern generating means, the accuracy of the accent type set for the accent phrase is calculated using the accuracy, and a pitch pattern corresponding to the accuracy of the accent type is generated. I do.

In the audio information processing method according to a thirteenth aspect of the present invention, in the information on the pitch of the unit, each of the plurality of types of values representing a subjective change in the pitch of the unit is used. Calculating the accuracy of an accent type set for an accent phrase using the accuracy, and generating a pitch pattern according to the accuracy of the accent type. Features.

[0027]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing a configuration necessary for implementing one embodiment of the present invention. 1 is a text input unit, 2 is a text analysis unit, 3 is a unit pitch information generation unit,
Reference numeral 4 denotes a pitch pattern generation unit. Hereinafter, a process from input of a text to generation of a pitch pattern will be described.

First, a text to be subjected to speech synthesis (for example, a sentence mixed with Kanji and Kana) is input to the text input unit 1. The text input unit 1 can receive any text capable of speech synthesis. The text input to the text input unit 1 is sent to the text analysis unit 2.

The text analysis unit 2 performs text analysis such as morphological analysis on the input text to divide the text into analysis units such as morphemes, and reads and analyzes each analysis unit. Various information such as an accent type and a part of speech specific to the analysis unit is added. These various types of information are stored in the unit pitch information generation unit 3.
Sent to

The unit pitch information generation unit 3 derives from various kinds of information on the analysis unit sent from the text analysis unit 2,
When the mora (or syllable) of the voice to be synthesized is defined as a unit, unit text information is generated for the unit,
From the unit text information, information on the subjective height of the pitch of the unit is generated for each unit. here,
The subjective pitch of the pitch refers to the relative pitch of the sound at the time of uttering the word. For example, in a linguistic accent, the accent type is a type-1 word, "voice (onsei)
, The first mora is “high” and the second and subsequent mora are “low”. The unit for generating the information on the pitch level is one mora (or syllable) here, but may be a plurality of mora (or syllables) such as two adjacent mora, and an example of the plurality of mora will be described later. To generate information relating to the subjective height of the pitch of the unit, a deformation rule describing the correspondence relationship from the above-mentioned "unit text information" to "the subjective height of the pitch" is used. The information on the pitch level generated for each unit is one of two categories indicating whether the pitch of the unit is “high” or “low”, or the probability that the unit corresponds to each category. The above unit text information includes
The text attribute (part of speech, inflection, accent type, etc.) of the analysis unit to which the unit belongs and the analysis units before and after it, the relative position of the unit with respect to the accent kernel of the analysis unit,
The phoneme type of the unit and the unit before and after the same (such as whether or not the phoneme is a special phoneme such as a sound-repelling sound or a prompting sound) is included. The information on the pitch of each unit generated by the unit pitch information generation unit 3 is sent to the pitch pattern generation unit 4 together with various information of the analysis unit sent from the text analysis unit 2.

As information on the pitch of the unit, two or more categories or the above-mentioned units such as a category other than "high" and "low" are provided for a unit having a weak accent generally called a secondary accent. The unit may be the probability that each category falls.

Further, the information on the subjective height of the pitch of the unit may be such that the subjective pitch in the unit is “high”.
The accuracy may change from “to” to “low”, that is, the accuracy in which the unit becomes an accent nucleus.

As the unit text information, it is also possible to use information relating to the pitch height of the unit generated with respect to the previous unit. FIG. 2 shows an example in which the information on the pitch of the unit is the probability that the unit falls into each of the categories of “high” and “low”. First, the initial values P (A0 = H) and P (A0 = L) of the probabilities corresponding to the “high” and “low” categories are set to 0.5, and the unit number i is initialized (S1, S2). Next, unit text information on the i-th unit is generated (S3). In the unit text information generated in S3, the pitch of the previous unit is "high"
By adding the information (Ai-1 = H) and inputting this unit text information to the transformation rule (which describes the correspondence from "unit text information" to "subjective pitch height") , Ai-1 = H, unit pitch information P (Ai-1 = H * A
i = H) and P (Ai-1 = H * Ai = L) are generated (S4). Similarly, unit pitch information in the case of Ai-1 = L is generated by using unit text information to which information that the pitch of the previous unit is “low” is added (S5). Next, unit pitch information for the i-th unit is generated according to the following equation (S6).

P (Ai = H) = P (Ai-1 = H * Ai = H) * A (Ai-1 = H) +
P (Ai-1 = L * Ai = H) * A (Ai-1 = L) P (Ai = L) = P (Ai-1 = H * Ai = L) * A (Ai-1 = H) + P (Ai-1 = L * Ai
= L) * A (Ai-1 = L) If i is the last unit number (N), the process is terminated. Otherwise, i is updated to i + 1 and the steps are repeated from S3.
By using the unit pitch information generated for the previous unit in this way, the accuracy of generating unit pitch information for a unit where the pitch of the previous unit is important can be improved.

The above-described transformation rule used to generate the information on the pitch level of each unit includes a “unit text information list” (FIG. 3) describing the text information of each unit for various texts, and a subjective description of those units. From the "unit pitch list" (Fig. 4), which describes the pitches of various pitches, by using a method such as a decision tree, a rule for automatically mapping "unit text information" to "unit pitch height" is automatically constructed. It is possible (Figure 5). When the decision tree method is used, information such as the probability that the unit corresponds to each category representing pitch height is obtained from the ratio of correctly set values among the number of data that meet the conditions at the end (leaf) of the decision tree be able to. However, not limited to the above, using any statistics obtained from the data,
Information such as the probability can be obtained. Further, by performing statistical processing on the above two lists, a plurality of units (here, 2
May be constructed as a deformation rule by constructing a statistical model for estimating the probability of taking each of the two values.

The pitch pattern generation unit 4 generates a pitch pattern according to the information on the pitch of the unit and the various information of the analysis unit sent from the unit pitch information generation unit 3. An example of a specific configuration of the pitch pattern generation unit 4 for generating a pitch pattern from information transmitted from the unit pitch information generation unit 3 will be described with reference to FIG.

FIG. 2 is a block diagram showing one form of the pitch pattern generation unit, which comprises an accent phrase division unit 41, an optimal accent type setting unit 42, and a pitch pattern generation processing unit 43. In the accent phrase division unit 41, the boundary between the analysis units in the text is the accent phrase boundary based on the information on the pitch level of each unit and the various information on the analysis unit sent from the unit pitch information generation unit 3. Then, the text is divided according to the thus determined accent phrase boundaries. At this time, whether or not the boundary of each analysis unit is a boundary of an accent phrase is determined based on the part of speech of the adjacent analysis unit and the height of the mora pitch immediately before and immediately after the boundary of the analysis unit. This can be easily realized by creating a rule that determines whether or not a phrase is a boundary. In addition, the accent phrase division unit 41 also gives the part of speech, the number of mora, and the like to each accent phrase according to the information on the morphemes constituting the accent phrase. Next, the optimal accent type setting unit 42 calculates the accuracy of realizing all possible accent types as shown in FIG. 7 by using information on the pitch level for each unit in the accent phrase, and among them, By selecting the accent type that has the highest accuracy, an optimal accent type is set for the accent phrase. Here, although the accuracy of realizing each accent type is calculated based on the probability that the pitch of each unit falls into each of the categories of “high” and “low”, it is sent from the unit pitch information generating unit 3. If the information about the pitch of each unit is in either the “high” or “low” category,
If the category is `` High '', the probability of the pitch being `` High '' or `` Low '' is 0.8, 0.2, etc. respectively, and if the category is `` Low '', 0.2, 0, respectively.
8 and so on. As described above, in the accent type in which the pitch in each mora is set based on the probability corresponding to each category of “high” and “low”, a mora that cannot be a low pitch has a low pitch or a high pitch has It is unlikely that a high pitch will be caused by a mora that cannot be sparse. The pitch pattern generation processing unit 43 generates an appropriate pitch pattern from information such as the accent type, part of speech, and the number of mora set by the optimal accent type setting unit 42. The pitch pattern generation here uses a well-known method using a point pitch model or a Fujisaki model,
Any method such as a method using HMM (Miyazaki etc. "Pitch pattern generation by HMM based on probability distribution in multiple spaces", Proceedings of the Acoustical Society of Japan, 1-7-18, Mar, 1998) Can also be achieved.

FIG. 8 is a block diagram for explaining another embodiment of the present invention. The description will focus on the differences from FIG. FIG. 8 is characterized in that an accent phrase division unit 5 is inserted before the unit pitch information generation unit 3. In this case, the accent phrase division unit 5 performs division into accent phrases using only various types of information of the analysis unit. On the other hand, the unit pitch information generation unit 3 can also use the information of the accent phrase obtained by the accent phrase division unit 5 (for example, the position of the mora in the accent phrase) as a determinant of the unit pitch information. In this configuration, the performance of accent phrase division may be inferior to the above method, but on the other hand, an improvement in the performance of generating information of unit pitch can be expected. In the pitch pattern generation unit 4, the optimal accent type selection unit 42 sets the optimal accent type for the accent phrase determined by the accent phrase division unit 5 in the same manner as in the first embodiment described above. In the pitch pattern generation processing unit 43, a pitch pattern is generated using information on accent type and morpheme. If the accent phrase division unit 5 can output a plurality of division candidates,
As shown in FIG. 9, the most appropriate division candidate can be selected from the division candidates.

Specifically, first, the score (S
core), the division candidate number (j) and the optimal candidate number (jbe
Set the initial value to (st) (S1, S2). Next, a division candidate j is selected (S3), and the pitch height of each mora and its accuracy are estimated using information on the accent phrase of the j-th division candidate.
(S4). Next, using the pitch height and accuracy of each mora estimated in S4, an optimal accent type is set for each accent phrase, and the accuracy of each set accent type is calculated (S5). Next, the accuracy (Pj) at which the optimal accent type is set for each accent phrase in the j-th division candidate is calculated by multiplying the accuracy of each set optimal accent type (S6). Pj and Score are compared (S7), and Pj
If Score is larger, update Score to Pj and jbest to j
(S8). If j is equal to the total number of division candidates (N), the process ends. Otherwise, j is updated to j + 1 and the steps are repeated from S3. When the step is completed, the jbest-th division candidate is selected as an optimal division candidate, and the optimal accent type set for each accent phrase in the jbest-th division candidate is used for generating a pitch pattern.

FIGS. 10 and 11 show a method of transforming a pitch pattern generated for an accent phrase using the set accuracy of the accent type.

FIG. 10 shows an example of a modification when the accuracy of the set accent type is small (for example, accuracy <0.25). For example, first, a constant α (0 <α <1) is determined according to the accuracy, and the deformation is performed by the following equation using a function f0base (t) serving as a reference of the pitch pattern deformation.

F'0 (t) = f0base (t) + (f0 (t) -f0base
(t)) * α (f '(t): pitch at time t after deformation, f (t):
(Pitch at time t before deformation) In FIG. 10, f0base (t) is a straight line represented by the following equation.

F0base (t) = f0 (t0) + (t-t0) * a (t0:
Reference point time, a: slope of straight line (all constants determined by set accent type)) Reference point time t0 is the center time of the first mora when the set accent type is 0, otherwise Is the center time of the last mora, and the slope a is also set to an appropriate value depending on the set accent type (generally 0 or a negative value)
. With this modification, when there is a high possibility that the set accent type is wrong, it is possible to perform an inconspicuous accent reading. Note that the method of calculating the accuracy of the set accent type used here is not limited to the method shown in the figure, but may be a method such as a method of correcting the calculated accuracy according to the number of mora of the accent phrase as shown in FIG. The formula for performing the deformation of the pitch pattern is not limited to the above formula, and various formulas can be used.

FIG. 11 shows an example of a modification in which the accuracy of the set accent type first and second candidates is close to each other. If the accuracy of the first candidate is P1 and the accuracy of the second candidate is P2, for example, P1 <P2 * 2, the pitch pattern generated for the first candidate and the pitch pattern generated for the second candidate Are weighted according to the accuracy and added, and this pitch pattern is used. For example, the addition is performed by the following equation.

F'0 (t) = f0 * 1 (t) * P1 / (P1 + P2)
+ F0 * 2 (t) * P2 / (P1 + P2) (f '(t): pitch at time t after addition, f0 * 1 (t): pitch at time t for the accent type of the first candidate, f0 * 2 (t): pitch at time t with respect to the accent type of the second candidate) When actually speaking an accent phrase that may have the accent type of a plurality of candidates, the phrase is often read out with an accent that can be taken from either. By using the pitch pattern created by the addition as described above, by reading out the accent type that can be taken as both the first candidate and the second candidate accent type, it is possible to read out with an accent closer to the actual utterance It becomes possible. The expression used for adding the pitch pattern is not limited to the above expression, and various expressions can be used.

In the above-described embodiment, the unit for generating the information on the pitch in the unit pitch information generating unit 3 has been described as one mora (or one syllable).
This unit consists of multiple mora, such as two adjacent mora (
Or syllable). Information on the pitch level generated when two adjacent moras are used as a unit is as follows:
A plurality of values representing the change in pitch, such as “low → high”, “high → low”, and “no change”, or the accuracy of taking each value. In this way, in the method of considering the change in pitch, by providing a value of “no change”, the pitch is determined depending on the pitch of the previous word such as “ha” or “ga” of the particle, It is hoped that the rules for units for which it is not important to determine the pitch level itself will be simpler. In addition, when the information on the pitch level is the accuracy of taking each of a plurality of values representing the pitch level change, it is possible to obtain a clue on the degree of the pitch level change from the accuracy.
It is possible to generate a pitch pattern that changes in height according to accuracy.

In the above embodiment, an appropriate accent type is set for the information on the pitch of the unit obtained by the unit pitch information generating unit 3, and a pitch pattern is generated using the accent type. However, the pitch pattern generation unit 4 can directly generate the pitch pattern from the information on the pitch of the unit and various information of the analysis unit without setting the accent type. For example, a plurality of representative pitch patterns in units of 2 mora are prepared, and an appropriate pitch pattern is sequentially selected for information on the pitch height of two adjacent mora and various information of the analysis unit to which the two mora belong. By connecting them, a pitch pattern can be generated (FIG. 12). If the information generated by the unit pitch information generation unit 3 is an accuracy that takes each of a plurality of values representing the height of a pitch or a change thereof, a pitch pattern having a different degree of change in the pitch is selected according to the accuracy. You. Note that the unit of the prepared pitch pattern may be any unit, such as 3 moras, as long as it is not shorter than the unit for generating information in the unit pitch information generating unit 3. In addition, the method of generating the pitch pattern is not limited to the method of selecting and connecting the representative pitch patterns described above, and an appropriate pitch pattern may be used for the information on the unit pitch height and various information of the analysis unit. Any method can be used as long as the method can generate.

[0049]

As described above, according to the present invention,
In the case where mora or syllable is used as a unit, information on the pitch of the unit is generated for each unit, and a pitch pattern is generated using the information on the pitch of these units. It is less likely to set a natural pitch, and even if an incorrect accent is given, the synthesized voice is less likely to be unnatural.

Further, according to the present invention, the rule for generating information on the pitch of the unit can be automatically generated from the data, so that the rule can be easily constructed.

According to the present invention, information on the pitch of the unit is generated, and a pitch pattern is generated by using the information of the unit. Can be given at the same time.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a voice information processing apparatus according to an embodiment of the present invention.

FIG. 2 is a flowchart illustrating a processing method when information of a previous unit is used in a unit pitch information generation unit illustrated in FIG. 3;

FIG. 3 is a diagram for explaining a unit pitch list.

FIG. 4 is a diagram for explaining a unit text information list.

FIG. 5 is a diagram for explaining a method of creating a conversion rule from “unit text information” to “information regarding pitch height of a unit”.

FIG. 6 shows a configuration 1 of the pitch pattern generation unit 4 shown in FIG.
FIG. 4 is a block diagram showing an example.

FIG. 7 is a view for explaining an accent type selection method in the optimal accent type selection unit shown in FIG. 6;

FIG. 8 is a block diagram showing a voice information processing apparatus according to another embodiment of the present invention.

FIG. 9 is a flowchart illustrating a processing method in a case where a plurality of division candidates can be output by the accent phrase division unit 5 illustrated in FIG. 8;

FIG. 10 is a view for explaining a method of deforming a pitch pattern according to a set accent type.

FIG. 11 is a view for explaining a method of deforming a pitch pattern according to a set accent type.

FIG. 12 is a view for explaining a method of generating a pitch pattern from information on the height of a unit pitch and various information on an analysis unit.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Text input part 2 Text analysis part 3 Unit pitch information generation part 4 Pitch pattern generation part 5 Accent phrase division part 41 Accent phrase division 42 Optimal accent type setting part 43 Pitch pattern generation processing part

 ──────────────────────────────────────────────────続 き Continued on the front page (72) Inventor Takehiko Kagoshima 4-2-1-12 Honcho, Chuo-ku, Osaka-shi, Osaka Toshiba Corporation Kansai Branch F-term (reference) 5D045 AA09

Claims (13)

[Claims] 1. Text input means for inputting text, text analysis means for analyzing text input to the text input means, and mora or syllable in units of mora or syllables according to a text analysis result obtained by the text analysis means. In this case, unit pitch information generating means for generating information on the pitch of the unit for each unit, and generating a pitch pattern using the information on the pitch of the unit obtained by the unit pitch information generating means An audio information processing apparatus, comprising: 2. The subjective height of the unit pitch is referred to as a unit subjective pitch height, and when the unit subjective pitch height has a plurality of categories, the information on the pitch of the unit is the unit of the unit pitch. 2. The method according to claim 1, wherein the category corresponds to a category having a high or low subjective pitch.
A speech information processing apparatus according to claim 1. 3. The audio information processing apparatus according to claim 1, wherein the information on the pitch of the unit is a probability that the unit falls into each category of the unit subjective pitch. 4. The change in the subjective height of the pitch of the unit is called a unit subjective pitch change. When the unit subjective pitch change has a plurality of categories, the information on the pitch of the unit is as follows. 2. The speech information processing apparatus according to claim 1, wherein the category corresponds to one of the categories of the unit subjective pitch change. 5. The speech information processing apparatus according to claim 1, wherein the information on the pitch of the unit is a probability that the unit falls into each category of the unit subjective pitch change. 6. The speech information processing apparatus according to claim 1, wherein the information on the pitch of the unit is a probability that a subjective pitch changes from high to low in the unit. 7. The unit pitch information generating means,
The information on the pitch of the unit is sequentially generated in the order in which the unit appears in the text, and when generating the information on the pitch of the unit, the information of the unit generated with respect to the unit before the unit is generated. 7. The information according to claim 1, wherein information on a pitch is used.
An audio information processing apparatus according to the item. 8. In the pitch pattern generating means,
The input text is divided into accent phrases, and an accent type is set for the accent phrases, using information on the pitch of the unit generated by the unit pitch information generating means. Item 8. The audio information processing device according to any one of Items 1 to 7. 9. An accent phrase dividing unit for dividing the input text into accent phrases and generating information on the accent phrases according to a text analysis result obtained by the text analyzing unit, wherein the unit pitch information The generating means generates information on the pitch of the unit using the information on the accent phrase generated by the accent phrase dividing means, and the pitch pattern generating means generates the information on the pitch of the unit by the unit pitch information generating means. The speech processing apparatus according to claim 1, wherein an accent type for the accent phrase divided by the accent phrase dividing means is set using information on the pitch of the unit of pitch. . 10. The information on the pitch of the unit is the probability that the unit falls into each category of the unit subjective pitch, and the pitch pattern generation means generates a pitch pattern according to the probability. The audio processing device according to claim 1, wherein the voice processing is performed. 11. The information on the pitch of the unit is a probability that the unit falls into each category of the unit subjective pitch change, and the pitch pattern generation means generates a pitch pattern according to the probability. The audio processing device according to claim 1, wherein the voice processing is performed. 12. The information on the pitch of the unit is the probability that the unit falls into each of the categories of the unit subjective pitch, and the pitch pattern generation means uses the probability to calculate the accent phrase using the probability. The speech processing device according to claim 8, wherein the accuracy of the accent type set for the accent type is calculated, and a pitch pattern corresponding to the accuracy of the accent type is generated. 13. The information on the pitch of the unit is the probability that the unit falls into each category of the unit subjective pitch change, and the pitch pattern generation means uses the probability to add to the accent phrase. The speech processing device according to claim 8, wherein the accuracy of the accent type set for the accent type is calculated, and a pitch pattern corresponding to the accuracy of the accent type is generated.