JPH1078795A - Speech synthesizing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain the rule capable of accurately determining the phoneme duration time length of mora phonemes, such as syllabic nasal and geminative consonants, and phoneme systems, such as vowel unvoiced phoneme systems and eventually to easily obtain more natural synthesized speeches. SOLUTION: A block and segment part 40 determines the phoneme blocks by segmenting the phoneme symbol strings in the uttered vowels. A section 42 for setting the length of the time between the energy centroids of the vowel parts determines the length DG of the time between the energy centroids of the vowel parts to the phoneme section with the kinds and utterance speeds of the phoneme system held between the vowel parts at both ends of the phoneme section as parameters by using the rhythm rule (A) 44. A section 46 for setting the duration time length of the respective phonemes segments the phoneme blocks to respective phoneme parts and determines the phoneme duration time length of the respective phoneme parts with DG and the kinds of the phoneme systems as parameters by using the rhythm rule (B) 48. The kinds of the phoneme systems are discriminated by the kinds of the consonants following, for example, the syllabic nasal, the kinds of the consonants constituting the geminative consonants and the vowel unvoiced phoneme systems.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizer for synthesizing speech using a phoneme duration.

[0002]

2. Description of the Related Art Conventionally, there is a speech synthesizer which performs speech synthesis by controlling a phoneme duration time in units of mora. Here, the mora is a unit of prosody corresponding to a single syllable in Japanese speech.
Corresponds to Mora. Further, the phoneme duration is a time length assigned to each phoneme when synthesizing a voice.

In a conventional speech synthesizer, as a rule for controlling the duration of a phoneme, two adjacent mora units are used as units, and a phoneme sequence based on a VCV (vowel / consonant / vowel) type contained in the mora is used. The duration was set. Specifically, the center of gravity (CEG) of the time integral of the energy of the speech waveform for the vowel part (V) of each mora
V: The Center of Energy Gravity of Vowels), the time length between these two CEGVs,
According to the type of the consonant part (C) existing between two vowel parts (V), the time length DG between the vowel part energy centroids is obtained, and the phoneme duration of each vowel and consonant is determined (Japanese Patent Laid-Open No. 5-281993, JP-A-6-266391
And JP-A-6-274195).

[0004]

The Japanese language has the same qualifications as other mora in terms of phonological duration, but cannot create monosyllables by itself and always stays in the adjunct component of other syllables A special type of mora called "Mora phonemes" (Miyaji: Japanese and Japanese education. Volume 1 Japanese Studies (1989) Meiji Shoin), such as "N", "G", and "-" Exists.

In the conventional speech synthesizer, when the above-mentioned mora phoneme is included between vowel parts, the mora phoneme is converted to a general vowel when calculating the time length DG between the vowel part energy centroids and the phoneme duration time. Or it was treated as a consonant. As a result, aspects of the features of the mora phoneme were discarded, and the two time lengths could not sufficiently represent the mora phoneme.

[0006] In Japanese, when a human generates vowels, the vocal tract shape is formed by the mouth and throat and the vocal cords are not vibrated or even the vocal tract shape is not formed. There is a phoneme sequence in which the voiced property is lost, that is, a phoneme sequence that causes vowel devoicing. The phoneme sequence that causes the vowel to be unvoiced is referred to herein as a vowel unvoiced phoneme sequence. In the conventional speech synthesizer, no special consideration is given to the phoneme sequence including the vowel to be devoiced, that is, the pattern of the VCV-type phoneme sequence is the same as in the case of a general voiced vowel. Is applied, and the vowel part energy centroid CEGV is forcibly set to the vowel that should be unvoiced. For this reason, the vowel part energy center-of-gravity point time length DG and the phoneme duration time obtained by this treatment cannot sufficiently represent the vowel devoicing.

For this reason, there has been a problem that when speech synthesis is performed on a text including a mora phoneme or a vowel unvoiced phoneme sequence, an unnatural speech is synthesized.

In order to solve the above-mentioned problems, the present invention provides a highly accurate phonological sequence for a phonological sequence having mora phonemes between vowel parts, such as a vowel sound, a consonant sound, and a prolonged vowel, and a phonological sequence causing vowel devoicing. It is an object of the present invention to provide a speech synthesizer that gives a duration and generates natural Japanese synthesized speech.

[0009]

According to a first aspect of the present invention, there is provided a speech synthesizer for generating a phoneme symbol sequence including a vowel portion, a consonant portion, and a mora phoneme from a text;
Partitioning means for partitioning the phoneme symbol string into phonemic sections by the vowel parts, and energy of each of the vowel parts located at both ends of the phoneme section according to the type of the consonant part and the mora phoneme in the phoneme section Vowel part energy barycenter time length generating means for obtaining a vowel part energy barycenter time length that is an interval between barycentric points, and a vowel part energy barycenter time length between the phonological sections, a consonant part in the phoneme section, Phoneme duration generation means for determining the phoneme duration of each phoneme in the phoneme section according to the type of the mora phoneme, and performing speech synthesis using the phoneme duration. .

[0010] According to the present invention, the present inventor has newly discovered "when a vowel transitioning from a vowel having the vowel part energy centroid CEGV to the next vowel, other than the vowel parts which must be realized. A phonological duration control rule is constructed based on the principle that the ease and difficulty of utterance of a phonemic sequence determines the time length between vowel energy center-of-gravity points. The principle used in the present invention removes the limitation that the principle used in the prior art is "a phoneme existing between vowels is a consonant." In other words, in the present invention, it is recognized in principle that not only consonants but also mora phonemes such as repellent, consonant, and prolonged sounds exist between vowel parts. Therefore, rules that take into account not only the types of consonants but also the types of mora phonemes are recognized. Is generated. That is, first, the phoneme symbol string generation means generates a phoneme symbol string which is classified into vowel parts, consonant parts, and mora phonemes from the text. Then, the section dividing means determines one vowel section to one vowel section as one phoneme section. Thus, the beginning and end of each phoneme section are vowel parts, and a consonant part and a mora phoneme may be interposed between them. The vowel part energy centroid point length generation means generates a vowel part energy centroid point length in consideration of not only consonants present between vowel parts but also types of mora phonemes. The phoneme duration generation means determines the phoneme duration of each phoneme in the phoneme section in consideration of not only the consonants present between the vowels but also the types of mora phonemes. Here, the type of the mora phoneme is based on, for example, a classification based on a combination of the mora phoneme and its subsequent consonant.

According to a second aspect of the present invention, there is provided a speech synthesizer for generating a phoneme symbol sequence including a vowel portion, a consonant portion, and a vowel unvoiced phoneme sequence from a text, and converting the phoneme symbol sequence into the vowel sound. Segmentation means for segmenting into phonological segments by segments, and intervals between the energy centroid points of the vowel portions located at both ends of the phonological segment according to the type of the consonant portion and the vowel unvoiced phoneme sequence in the phonological segment. Vowel part energy barycenter time length generating means for obtaining the vowel part energy barycenter time length, and the vowel part energy barycenter time length of the phoneme section, the consonant part in the phoneme section and the vowel devoicing Phoneme duration generation means for determining the phoneme duration of each phoneme in the phoneme segment according to the type of phoneme series, and performing speech synthesis using the phoneme duration. To.

According to the present invention, a vowel to be unvoiced is distinguished from a vowel part as a vowel unvoiced phoneme sequence. Based on the above-described principle of the present invention, when a vowel unvoiced phoneme sequence exists between voiced vowel parts, not only the type of consonant existing between vowel parts, the type is also considered. That is, in the present invention, it is recognized in principle that not only consonants but also vowel unvoiced phoneme sequences exist between vowel parts, and thus, not only the types of consonants but also the types of vowel unvoiced phoneme sequences are considered. A duration control rule is generated. That is, first, the phoneme symbol string generation unit generates a phoneme symbol string in which vowel parts, consonant parts, and vowel unvoiced phoneme sequences are distinguished from the text. Then, the section dividing means determines a phonological section in which the beginning and the end of the phoneme section are vowel parts, respectively, and a consonant part and a vowel unvoiced phoneme sequence can be interposed therebetween. The vowel unvoiced phoneme sequence is distinguished from a general voiced vowel part by the phonological symbol string generation means, and the partitioning means is used for the unvoiced vowels in the vowel unvoiced phonological sequence, the vowel energy at the beginning or end of the phonological section. No center of gravity is set. In other words, the partitioning means is one for a mora having a vowel uttered voiced.
Vowel energy center of gravity is set. The vowel part energy center-of-gravity point time length generating means generates the vowel part energy center-of-gravity point time length in consideration of not only the consonants existing between the vowel parts but also the type of the vowel unvoiced phoneme sequence. The phoneme duration generation means determines the phoneme duration of each phoneme in the phoneme section in consideration of not only the consonants existing between the vowel parts but also the type of the vowel unvoiced phoneme sequence.

According to a third aspect of the present invention, there is provided a speech synthesizer for generating a phoneme symbol sequence including a vowel portion, a consonant portion, a mora phoneme, and a vowel unvoiced phoneme sequence from a text, and the phoneme symbol sequence. Partitioning means for partitioning the vowel part into phonemic sections, the consonant part in the phoneme section,
A vowel part energy centroid for calculating a time length between vowel part energy centroids which is an interval between energy centroids of the vowel parts located at both ends of the vowel segment according to the type of the mora phoneme and the vowel unvoiced phoneme sequence. Inter-time length generation means, the time length between the vowel part energy centroids of the phoneme section, the consonant part in the phoneme section, the mora phoneme, and the type of the vowel unvoiced phoneme sequence in each of the phoneme sections. Phoneme duration generating means for determining the phoneme duration of the phoneme, and synthesizes speech using the phoneme duration.

According to the present invention, based on the above-described principle of the present invention, when not only the types of consonants existing between vowels but also mora phonemes and vowel unvoiced phoneme sequences exist between vowel parts, Types are also considered. That is, in the present invention, it is recognized in principle that not only consonants but also mora phonemes and vowel unvoiced phoneme sequences exist between vowel parts.
Therefore, a phoneme duration control rule is generated that takes into account not only the types of consonants but also the types of mora phonemes and the types of vowel unvoiced phoneme sequences. That is, first, the phoneme symbol string generation means generates a phoneme symbol string in which vowel parts, consonant parts, mora phonemes, and vowel unvoiced phoneme sequences are distinguished from the text.
Then, the section dividing means determines a phoneme section in which the beginning and the end of the phoneme section are vowel parts, respectively, and a consonant part, a mora phoneme, and a vowel unvoiced phoneme sequence can be interposed therebetween. Since the vowel unvoiced phoneme sequence is distinguished from the vowel part by the phonological symbol string generation means, the partitioning means adds the vowel part energy center point at the beginning or end of the vowel section to the unvoiced vowel in the vowel unvoiced phoneme sequence. Never set. The vowel part energy centroid time length generation means generates the vowel part energy centroid point time length in consideration of not only consonants present between vowels but also types of mora phonemes and types of vowel unvoiced phoneme sequences. Further, the phoneme duration generation means determines the phoneme duration of each phoneme in the phoneme section in consideration of not only consonants existing between vowels but also types of mora phonemes and vowel unvoiced phoneme sequences.

According to a fourth aspect of the present invention, in the speech synthesizer according to the first aspect, the means for generating the time length between the vowel part energy centroids is a linear function of a reciprocal of the speech speed. And the coefficient and constant term of the linear function are determined according to the type of the consonant part and the mora phoneme. In the speech synthesis apparatus according to a fifth aspect, in the second aspect, the vowel part energy centroid point time length generation means generates the vowel part energy centroid point time length by a linear function of a reciprocal of a speech rate. The coefficient and the constant term of the linear function are determined according to the type of the consonant part and the vowel unvoiced phoneme sequence. Further, in the speech synthesis apparatus according to a sixth aspect of the present invention, in the third aspect, the vowel part energy centroid point time length generating means calculates the vowel part energy centroid point time length by a linear function of a reciprocal of a speech rate. The coefficient and constant term of the linear function are determined according to the type of the consonant part, the mora phoneme, and the vowel unvoiced phoneme sequence.

According to the present invention, the vowel part energy center-of-gravity point time length generating means is capable of generating the vowel part energy corresponding to the reciprocal of the speech speed, regardless of the phoneme sandwiched between the vowel parts of the phoneme section. Generate the time length between the centers of gravity. Moreover,
The time length between the vowel part energy centroid points is represented by a linear function of the reciprocal of the speech rate, and the proportional coefficient and the constant term of the reciprocal of the speech rate are determined for each type of phoneme sandwiched between the voiced vowel parts of the phoneme section. Can be The type of the phoneme includes, in addition to the type of the consonant part, the type of the mora phoneme in the speech synthesis device according to the fourth invention, the vowel unvoiced phoneme sequence in the speech synthesis device according to the fifth invention, and the sixth type. In the speech synthesizer according to the invention, the type of mora phoneme and the vowel unvoiced phoneme sequence are referred to. Thereby, the time length between the vowel part energy centroids is quantitatively and accurately determined, and a natural Japanese synthesized speech is realized.

[0017]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a system configuration of a Japanese speech synthesizer according to an embodiment of the present invention. In this Japanese speech synthesis system, a synthesized sentence to be synthesized is input from a keyboard or the like of the terminal 2. The text analysis unit 4 analyzes the structure of the synthesized sentence as a Japanese sentence, and generates a phoneme symbol string to which pronunciation information such as accent information, pause, and vowel devoicing necessary for speech synthesis processing is added. Output to the phoneme duration time generation unit 6. The phoneme duration generation unit 6 assigns a phoneme duration to a phoneme symbol string using a rhythm rule 8 described later.

The phoneme symbol string to which the phoneme duration is added is input to a sound source amplitude pattern generation unit 10, a pitch pattern generation unit 12, and a spectrum pattern generation unit 14. The sound source amplitude pattern generator 10 and the pitch pattern generator 1
2 and the spectral pattern generation unit 14 determine the intensity, the height of the fundamental frequency, and the timbre, which are the three elements of the voice, respectively. That is, the sound source amplitude pattern generation unit 10
While maintaining the vowel portion energy center-of-gravity point time length DG and each phoneme duration time length given by the rhythm rule 8 of the phoneme duration time generation unit 6, the power envelope of the speech is changed according to the power rule 1.
6 is determined. The pitch pattern generation unit 12
Determines the point pitch patterns for each accent phrase from the prosody control rules 18, interpolates them, interpolates them, and gives the vowel part energy centroid point time length DG and each phoneme continuation given by the rhythm rule 8 of the phoneme duration time generation unit 6. A continuous pitch pattern is generated according to the time length. Spectrum pattern generator 1
4 determines the type of phoneme such as a vowel or a consonant based on the phonological improvement rule 20 and the phoneme combination rule 22, and determines the vowel part energy barycenter time DG given by the rhythm rule 8 of the phoneme duration generation unit 6. And the envelope pattern of the frequency spectrum of each phoneme is combined in accordance with the duration of each phoneme to generate a spectrum pattern.

The sound source generator 24 generates sound source information based on the power pattern output from the sound source amplitude pattern generator 10 and the continuous pitch pattern output from the pitch pattern generator 12. The speech synthesizer 26 modulates the sound source information from the sound source generation unit 24 according to the spectrum pattern input from the spectrum pattern generation unit 14 and adds a tone to generate a synthesized speech.

FIG. 2 is a block diagram showing the structure of the phoneme duration generating unit 6 shown in FIG. In the phonological duration generation unit 6, first, the partitioning unit 40 divides the phoneme symbol string generated by the text analysis unit 4 into phonemic partition units of "/ vowel part-non-vowel part-vowel part /". Classify.
The vowel part energy barycenter time length setting unit 42 is a vowel part energy barycenter time length generation unit, and includes a type of a phoneme sequence other than a vowel part and a speech rate SR in each of the phoneme segments.
(Speech Rate) and the rhythm rule (A) 44 stored as a parameter, a vowel portion energy center-of-gravity point time length DG at a desired utterance speed is calculated and set in each section. Each phoneme duration setting unit 46 is a phoneme duration generation unit, and converts the DG obtained by the vowel part energy centroid time length setting unit 42 and the type of the phoneme sequence other than the vowel part of each phoneme segment into parameters. The duration of each phoneme included in each phoneme section is determined from the rhythm rule (B) 48 stored as the following formula, and the result is used as the sound source amplitude pattern generation unit 10, the pitch pattern generation unit 12, and the spectrum pattern generation. Output to the unit 14.

Next, the division section 40, the rhythm rule (A) 44, and the rhythm rule (B) 48 for generating each phoneme duration in the phoneme duration generator 6 which is a characteristic part of the present invention. The details will be described below.

Based on the phoneme symbol string output from the text analysis unit 4, the partitioning unit 40 determines a vowel part energy centroid C
EGVs are set one by one, and a section from an arbitrary vowel part energy centroid CEGV to the next vowel part energy centroid CEGV is determined as one phoneme section. The inside of the phoneme section thus divided is clearly distinguished into a part that is clearly a vowel and each phoneme part other than the vowel and is labeled. The classification of phonemes within this phoneme section will be described below.

The phonological succession time length control rule using the vowel part energy center-of-gravity point time length DG is based on the hypothesis that "physical constraints on the vocal organs determine the time length." According to conventional rules, this hypothesis is referred to as "vowel part energy centroid C
The ease and difficulty of uttering a consonant existing between EGVs determines the vowel part energy center-of-gravity point time length DG. And a rule was constructed based on this. In this case, however, as described above, there was a problem that the error in the phoneme duration of the mora phoneme and the vowel unvoiced phoneme sequence might be large. Therefore, in the present invention, the above hypothesis is described as “easiness of vocalization of phonological sequences other than vowel parts, which is required to be realized when transitioning between vowels having vowel part energy centroids CEGV, The time length DG between the local energy centroids is determined. " According to this embodiment, in realizing the phonemes existing between the vowel parts, there is room for taking in various factors in addition to the attribute of a mere consonant. In other words, the present invention is to review the stance point which conventionally determined that the phoneme existing between the vowel parts is a consonant one step at a time, return to one step from there, and aim at another more correct direction. .

For example, conventionally, it has been generally assumed that a VCV-type phoneme sequence is labeled as / vowel to consonant to vowel /, and phonemes other than the vowel portion consist of one type of consonant. But,
In the realization of the above hypothesis, it is assumed that various state transitions accompany the realization of one type of consonant. As a specific example, the realization of the plosive sound “/ k /” between vowels always involves a state of “crossing from a vowel + (silence + bursting) + crossing to the next vowel”. It is necessary. Similarly, “/ ky /” means “over vowel part +
It is necessary to go through a series of states of “(silent part + burst part + resonant sound) + over the next vowel part”.

Further, even in a phoneme sequence other than the VCV type phoneme sequence, the phoneme sequence portion other than the vowel portion is an aggregate of various state transitions when transitioning from one vowel portion to the next vowel portion. You can think.

For example, a phoneme generally treated as a sound repellent is an aggregate of various sounds having a common feature that it is realized as a nasal having the same articulation point as the subsequent consonant. In other words, it is considered that the phoneme sequence including the repellent sound forms a series of states of “transition from vowel part + nasal state determined by succeeding consonant + next consonant + next vowel part”.
A phonological sequence including a consonant also has various changes in the following consonant, and is derived from a set of states “transition from vowel part + consonant state determined by subsequent consonant + next consonant + next vowel part”. It is considered to hold. Also, a vowel unvoiced phoneme sequence is considered to be composed of a set of states "over the vowel part + unvoiced state + next consonant + over the next vowel part". Further, also in the case of a phoneme sequence in which different vowels are continuous, it is considered that the transition portion is “transition from the vowel portion + transition to the next vowel portion”.

Based on the above considerations, examples of phonological sections output from the section dividing section 40 are shown in FIGS. FIG. 3 is a schematic diagram of a phoneme section of a general VCV phoneme sequence. The figure shows a phoneme sequence with the time axis taken horizontally rightward, and the elliptical shape schematically shows the power envelope. In this case, the phoneme section extends from the energy center of gravity CEGV of the vowel part of the i-th mora to the energy center of gravity CEGV of the vowel part of the (i + 1) -th mora, and the vowel part of the i-th and (i + 1) -th mora. Between (i +
1) There is a consonant part of the mora. FIG. 4 shows sound repellency,
FIG. 4 is a schematic diagram of a phonetic section of a mora phoneme such as a prompt sound and a long sound and a vowel unvoiced phoneme sequence, which is represented by the same expression method as in FIG. 3. In this case, each mora including mora phonemes and vowel devoicing constitutes one mora.
The phonemic section extends from the energy centroid CEGV of the vowel part of the i-th mora to the energy centroid CEGV of the vowel part of the (i + 2) -th mora. In other words, in the case of a phonological sequence in which devoicing, vowels, long sounds, and vowels have been devoiced,
Next, up to the mora determined to be a voiced vowel is 1
Are divided into two phoneme sections. And the ith, (i +
2) Between the vowel parts of the mora, the mora including the mora phonemes and vowels that are the (i + 1) th mora,
There will be a consonant part of the (i + 2) th mora and a transition part between them.

The vowel division section 40 sets the vowel part energy centroid point "sets one vowel part energy centroid CEGV for each mora in which a vowel clearly exists" regarding the handling of vowel parts. CEGV
Along with the restriction on the location of the vowel part energy centroid point CE, "The vowel part energy centroid CEGV is calculated using a part that can be clearly determined to be a vowel."
The vowel part used for calculating the GV is restricted. With this measure, in the case of a series of non-long vowels,
The transition from the vowel to the next vowel is excluded from the calculation of the vowel part energy centroid CEGV, and the vowel part energy centroid CEGV can be obtained more stably.
As a result, the error of the rules described below, which is derived using the analysis result of the vowel part energy centroid CEGV, is reduced. As described above, the partition section 40 has been described.

Next, the rhythm rule (A) 44 and the vowel part energy barycenter time length setting unit 42 will be described. The utterance speed SR is defined as the number of mora uttered per second. In the case of the VCV-type phonological sequence, the time length DG (second) between the vowel part energy centroids is the utterance speed SR (number of mora / second).
Can be expressed by the following approximate expression which is a linear function of a reciprocal of the following (Japanese Patent Laid-Open No. 6-266391). Where the coefficient A
The cons and the constant term Bcons are determined for each type of consonant part existing between the vowel parts.

[0030]

(Equation 1) The rhythm rule (A) will be sequentially described using the analysis result of the actual voice. First, in the analysis of the relationship between SR and DG of a general VCV-type phonological sequence, a sentence containing a 7-mora nonsense word, "It is" This is a OO Menkai "" was used. BX ₁ and CX ₂ are respectively entered in the o portions in the meaningless word portion. Here, X ₁ and X ₂ indicate arbitrary vowels, and C indicates an arbitrary consonant. In this way, the second and third mora in the meaningless word were changed into various phoneme sequences, and the vowel part energy centroid time length DG in each case was analyzed. Note that the difference in the utterance speed SR is approximately
They uttered 10 times each in three stages of normal speed / low speed. The speaker is a female speaker. FIG. 5 is a graph showing an example of an analysis result on the relationship between SR and DG of the VCV phoneme sequence. In this graph, the consonant C is represented by “t”.
y ”, three combinations of vowels X ₁ and X ₂ of bX ₁ and CX ₂ are shown, and the measurement points for each combination and the approximate curve according to the relational expression (1) for each combination are shown in FIG. From this figure, it can be seen that SR and D
The relationship with G hardly depends on the combination of the vowels at both ends of the VCV-type phonological sequence, and the coefficient Acons and the constant term Bcons determined for each type of consonant described above are the types of vowels sandwiching the consonant. It does not depend on. That is, it can be seen that the relationship between SR and DG can be approximated by the function expression (1) using the coefficient Acons and the constant term Bcons determined for each type of consonant part.

The above-mentioned framework is extended to each new section shown in FIG. That is, a phonological sequence in which vowels, consonants, long sounds, and vowels are unvoiced is interposed between vowel parts. The coefficient Acons and the constant term Bcons are the types of phonological sequences other than the vowel parts existing between the vowel parts. The coefficient is determined by

The relationship between the SR and DG of the phoneme sequence including the sound repellency includes an insignificant word of 8 mora.
"I'm sorry." BX ₁ and CX ₂ are respectively entered in the ○ portions of the meaningless word portion. Here, X ₁ and X ₂ indicate arbitrary vowels, and C indicates an arbitrary consonant. In this way, the second to fourth moras in the meaningless word were changed into various phoneme sequences, and the time length DG between the vowel energy centroids in each case was analyzed. The conditions regarding the speaker, the speed, and the number of times are the same as those in the case of the VCV-type phoneme series. FIG. 6 is a graph showing an example of an analysis result on the relationship between SR and DG of a phoneme sequence including a sound repellent. In this graph, the vowel X ₁ of the second mora “bX ₁ ” is “a”, and the third mora “CX ₂ ” is “ba”.
The measurement points for the meaningless word portion "Kobaban Menkai" and the approximate curve based on the relational expression (1) are shown in FIG. A measurement result was obtained that the coefficient Acons and the constant term Bcons of the equation (1) characterizing this approximated curve did not depend much on the types of the vowel parts at both ends defining DG. On the other hand, a measurement result was obtained in which, when the consonant C following the phoneme “N” indicating the sound repellency was changed, significant differences appeared in the coefficient Acons and the constant term Bcons. This phoneme “N”
The change in DG according to the type of C in C "is different from the change according to the type of C obtained for the VCV-type phonological sequence. Cannot be approximated by adding to the DG for each C of the VCV-type phoneme sequence for a certain period of time.

The relationship between the SR and DG of the phoneme sequence including the prompting sound includes an insignificant word of 8 mora,
"I'm sorry." BX ₁ and CX ₂ are respectively entered in the ○ portions of the meaningless word portion. Here, X ₁ and X ₂ indicate arbitrary vowels, and C indicates an arbitrary consonant. In this way, the second to fourth moras in the meaningless word were changed into various phoneme sequences, and the time length DG between the vowel energy centroids in each case was analyzed. The conditions regarding the speaker, the speed, and the number of times are the same as those in the case of the VCV-type phoneme series. FIG. 7 is a graph showing an example of an analysis result on a relationship between SR and DG of a phoneme sequence including a prompt sound. In this graph, the vowel X ₁ of the second mora “bX ₁ ” is “a”, and the third mora “CX ₂ ” is “ka”.
The measurement points for the meaningless word portion “Kobakamenkai” and the approximate curve based on the relational expression (1) are shown in FIG. A measurement result was obtained that the coefficient Acons and the constant term Bcons of the equation (1) characterizing this approximated curve did not depend much on the types of the vowel parts at both ends defining DG. On the other hand, when the type of the consonant C of the phoneme "CC"("kk" in the example shown in the graph) indicating the prompting sound is changed,
The measurement results showed that significant differences appeared in the coefficient Acons and the constant term Bcons. The change in DG according to the type of C in the phoneme “CC” is the CV obtained for the VCV-type phoneme sequence.
Is different from the change due to the type of In other words, DG, which is the result of this measurement of the phoneme sequence including the prompting sound,
It cannot be approximated by the conventional method of adding to the DG for each C of the VCV-type phoneme sequence for a certain period of time.

The relationship between SR and DG of the vowel unvoiced phoneme sequence includes an insignificant word of 8 mora,
"I'm sorry." BX ₁ , C ₁ X _2, and C ₂ X ₃ are respectively entered in the ○ portion of the meaningless word portion.
Here, X ₁ and X ₃ are arbitrary vowels, X ₂ is “/ i /” or “/ u /”, and C ₁ and C ₂ are vowels X ₂ sandwiched therebetween.
Indicates an arbitrary set of consonants that causes devoicing. Thus, to create a variety of phoneme series such as occurs silent into vowels X ₂ of the third mora in meaningless words, analyzing the vowel portion energy centroid between duration DG between the second to fourth mora did. The conditions regarding the speaker, the speed, and the number of times are the same as those in the case of the VCV-type phoneme series. FIG. 8 is a graph showing an example of an analysis result on the relationship between SR and DG of a vowel unvoiced phoneme sequence. In this graph, the vowel X ₁ of the second mora “bX ₁ ” is set to “a”, and the third mora “C
₁ X ₂ "to" ki ", the fourth mora" C ₂ X ₃ "a" cha "
The measurement points for the meaningless word portion “Kobaki Chamenkai” and the approximate curve based on the relational expression (1) are shown in FIG. A measurement result was obtained that the coefficient Acons and the constant term Bcons of the equation (1) characterizing this approximated curve did not depend much on the types of the vowel parts at both ends defining DG. On the other hand, when the type of the vowel unvoiced phoneme sequence (in the example shown in the graph, “kich”) is changed, the coefficient Aco
The measurement result that significant difference appears in ns and constant term Bcons was obtained.

Although illustration of other phonological sequences is omitted, in the present rhythm rule (A), as described in the above phonological sequence, the coefficient Acons is equal to the number of types of phonological sequences other than vowel parts existing between vowels. , A set of constant terms Bcons is prepared.
In the present rhythm rule (A), the relational expression (1) is used as an approximate expression. However, the present invention is not limited to this relational expression, and any other approximate expression that can exhibit characteristics can be used. In the rhythm rule (A), the approximate expression to be used can be simplified by grouping the types of phonemic sequences other than vowel parts existing between vowels according to the purpose at the time of creating a synthesized speech. The rhythm rule (A) 44 and the vowel part energy center-of-gravity time length setting unit 42 have been described above.

Next, the rhythm rule (B) 48 and each phoneme duration setting unit 46 will be described. In the case of the VCV-type phonological sequence, as shown in FIG. 9, the vowel part energy centroid CEGV is divided into three parts: a preceding vowel part latter part, a consonant part, and a succeeding vowel part former part, and the duration time Dvp, D
c and Dvs are approximated by the following linear equation of DG (Japanese Patent Laid-Open No.
266391, JP-A-6-274195).

[0037]

(Equation 2) Here, Avp and Bvp are DG coefficients and constant terms relating to preceding vowels, Ac and Bc are DG coefficients and constant terms relating to consonants, and Avs and Bvs are DG coefficients and constant terms relating to succeeding vowels. These coefficients Avp, Ac, Avs, and the constant term Bv
p, Bc, and Bvs are determined for each type of consonant part existing between vowel parts.

The rhythm rule (B) will be described sequentially using the analysis result of the actual voice. Dv of general VCV phoneme series
For the analysis of the relationship between p, Dc, Dvs and DG, a sentence containing a 7-mora nonsense word, "It's" this is the name of xx "is used. BX ₁ and CX ₂ are respectively entered in the o portions in the meaningless word portion. Here, X ₁ and X ₂ indicate arbitrary vowels, and C indicates an arbitrary consonant. In this way, the second and third mora in the meaningless word are changed into various phoneme sequences,
The relationship between Dvp, Dc, Dvs, and DG in each case was analyzed. The speaker is a female speaker. FIG.
It is a graph which shows an example of the analysis result of the relationship between Dvp, Dc, Dvs and DG of a CV type phoneme series. In this figure, the horizontal axis is D
The G value and the vertical axis indicate the values of Dvp, Dc, and Dvs, and the measured values of Dvp, Dc, and Dvs are plotted for each DG value obtained by changing the speech rate. The straight line of the equation (2) obtained by the least squares method for each of the measured values of Dvp, Dc and Dvs is also shown in FIG. Thus, Dvp, Dc, Dv
It can be seen that the relationship between s and DG is well approximated by equation (2), which is a linear function.

The above-mentioned framework is extended to each phoneme section including a vowel sound, a vowel sound, a long sound, and a vowel unvoiced phoneme sequence among the phoneme sections output from the section dividing section 40. In the case of a phonological sequence in which a vowel sound, a consonant sound, a prolonged sound, and a vowel sound are unvoiced, a portion up to the next mora determined to be a voiced vowel portion is defined as one phoneme segment, and each portion corresponding to each phoneme included in each phoneme segment. Is approximated by an expression similar to the expression (2). When different vowels are continuous, the vowel part and the transition part are clearly divided and approximated by the equation (2).

FIG. 11 to FIG. 14 show examples of the division of each phoneme part in the case of a sound-repelling sound, a prompt sound, a long sound, and a vowel unvoiced phoneme sequence.
Shown in Note that the division of each phoneme portion is not limited to the division method shown in the example, and may be divided using another different criterion.

FIG. 11 is a schematic diagram showing an example of the division of a phoneme portion in a phoneme section of a phoneme series including a sound repellent.
As shown in FIG. 11, the phonological sections are sequentially arranged between the vowel part energy centroids CEGV, the latter part of the preceding vowel part, the sound repellent part,
The subsequent consonant part and the first half of the subsequent vowel part are divided into four parts, and the durations of these phoneme parts are respectively Dvp, DN, Dc, Dvs
And The analysis of the relationship between Dvp, DN, Dc, Dvs and DG includes the meaningless word of 8 mora,
"I'm sorry." BX ₁ and CX ₂ are respectively entered in the o portions in the meaningless word portion. Where X
_1, X ₂ is any vowel, C is represents any consonant. In this way, the second to fourth mora in the meaningless word are changed into various phoneme sequences, and Dvp, DN, Dc, D in each case are changed.
The relationship between vs and DG was analyzed. The speaker is a female speaker. FIG. 15 is a graph showing an example of an analysis result of the relationship between the durations Dvp, DN, Dc, Dvs and DG of each phoneme portion of the phoneme sequence including the sound repellency. In this example, the meaningless word is “Kobanba Menkai”. In this figure, the horizontal axis is DG
And the vertical axis indicates the values of Dvp, DN, Dc, and Dvs, and Dvp, DN, and Dv for each DG value obtained by changing the speech rate.
c, Dvs measurements are plotted. Dvp, DN, D
A straight line represented by a linear expression of DG obtained by the least square method for each of the measured values of c and Dvs is also shown in FIG. Thus, it can be seen that the relationship between Dvp, DN, Dc, Dvs and DG is well approximated by a linear equation similar to the above equation (2).

Here, Avp and Bvp are DG coefficients and constant terms in a linear equation relating to Dvp, AN and BN are DG coefficients and constant terms in a linear equation relating to DN, and Ac and Bc are DG coefficients in a linear equation relating to Dc. And constant terms, Avs, Bv
Assuming that s is a coefficient and a constant term of DG in a linear expression relating to Dvs, these Avp, AN, Ac, Avs, Bvp, BN, B
A set of c and Bvs is determined for each type of consonant "C" following the sound repellent "N" existing between the vowel parts.

FIG. 12 is a schematic diagram showing an example of the division of a phoneme portion in a phoneme section of a phoneme sequence including a prompt sound.
As shown in FIG. 12, this phonological section is divided into three parts between the vowel part energy centroid CEGV into a preceding vowel part latter part, a prompting part, and a succeeding vowel part former part, and the duration time of each of these phonological parts is Dvp, Dsok and Dvs. Dvp, Dsok,
In the analysis of the relationship between Dvs and DG, a sentence containing a meaningless word of 8 mora, "It is" ko * menkokai "" was used. In the ○ part in the meaningless word part, b
X ₁ and CX ₂ to enter. Here, X ₁ and X ₂ indicate arbitrary vowels, and C indicates an arbitrary consonant. In this way, the second to fourth moras in the meaningless word were changed into various phoneme sequences, and the relationship between Dvp, Dsok, Dvs and DG in each case was analyzed. The speaker is a female speaker. FIG. 16 is a graph showing an example of an analysis result of the relationship between the durations Dvp, Dsok, Dvs, and DG of each phoneme portion of the phoneme sequence including the prompt sound. In this example, the meaningless word is “Kobakamenkai”. In this figure, the horizontal axis indicates the DG value, and the vertical axis indicates the values of Dvp, Dsok, and Dvs, and the measured values of Dvp, Dsok, and Dvs are plotted for each DG value obtained by changing the speech speed.
A straight line represented by a linear expression of DG obtained by the least square method for each of the measured values of Dvp, Dsok, and Dvs is also shown in FIG. Thus, it can be seen that the relationship between Dvp, Dsok, Dvs and DG is well approximated by a linear expression similar to the above expression (2).

Here, Avp and Bvp are DG coefficients and constant terms in a linear equation relating to Dvp, and Asok and Bsok are DG coefficients and constant terms in a linear equation relating to Dsok, Avs and Bv.
Assuming that s is a coefficient and a constant term of DG in a linear expression relating to Dvs, these Avp, Asok, Avs, Bvp, Bsok, Bvs
Are determined for each type of consonant C constituting the prompting sound “CC” existing between the vowel parts.

FIG. 13 is a schematic diagram showing an example of the division of a phoneme portion in a phoneme section of a vowel unvoiced phoneme sequence. As shown in FIG. 13, this phonological section is divided into four parts between the vowel part energy centroids CEGV into a former half part of the preceding vowel part, a mora part having vowel devoicing, a succeeding consonant part, and a succeeding half part of the vowel part. The duration of each phoneme part is Dvp,
Dmusei, Dc, and Dvs. Dvp, Dmusei, Dc, Dvs
To analyze the relationship between DG and DG, we used a sentence that included an insignificant word of 8 mora, "It's" This is XXX. " In the meaningless word part, each of the ○ parts is bX ₁
And C ₁ X ₂ and C ₂ X ₃ are entered. Here, X ₁ and X ₃ are arbitrary vowels, X ₂ is “/ i /” or “/ u /”, and C ₁ and C ₂
Indicates an arbitrary set of consonants that causes devoicing. In this way, the second to fourth mora in the meaningless word are changed into various phoneme sequences, and Dvp, Dmusei, Dc, D in each case are changed.
The relationship between vs and DG was analyzed. The speaker is a female speaker. FIG. 17 is a graph showing an example of an analysis result of the relationship between the durations Dvp, Dmusei, Dc, Dvs and DG of each phoneme portion of the vowel unvoiced phoneme sequence. In this example, the meaningless word is “Kobaki Chamenkai”. In this figure, the horizontal axis indicates the DG value, and the vertical axis indicates the values of Dvp, Dmusei, Dc, and Dvs. For each DG value obtained by changing the speech speed, Dv
The measured values of p, Dmusei, Dc, and Dvs are plotted. A straight line represented by a linear expression of DG obtained by the least square method for each of the measured values of Dvp, Dmusei, Dc, and Dvs is also shown in FIG. Thus, Dvp, Dmusei,
It can be seen that the relationship between Dc, Dvs and DG is well approximated by a linear expression similar to the above expression (2).

Here, Avp and Bvp are represented by a coefficient and a constant term of DG in a linear expression relating to Dvp, and Amusei and Bmusei are represented by Dm.
A coefficient and constant term of DG in a linear expression relating to usei, A
Assuming that c and Bc are DG coefficients and constant terms in a linear equation relating to Dc, and Avs and Bvs are DG coefficients and constant terms in a linear equation relating to Dvs, these Avp, Amusei, Ac, and A
A set of vs, Bvp, Bmusei, Bc, Bvs is determined for each type of vowel unvoiced phoneme sequence existing between vowel parts.

FIG. 14 is a schematic diagram showing an example of the division of a phoneme part in a phoneme section of a phoneme sequence including a series of different vowels. As shown in FIG. 14, this phonological section is divided into three parts between the vowel part energy centroid CEGV into a preceding vowel part latter part, a transition part, and a succeeding vowel part former part, and the duration time of each of these phonological parts is Dvp, DseNi and Dvs. D
For the analysis of the relationship between vp, DseNi, Dvs and DG, a sentence containing a 7-mora nonsense word, "It's" this is the name of this girl "" was used. BX ₁ and X ₂ are respectively entered in the o portions in the meaningless word portion. Here, X ₁ and X ₂ indicate different vowels different from each other. In this way, the second and third mora in the meaningless word were changed into various phoneme sequences, and the relationship between Dvp, DseNi, Dvs and DG in each case was analyzed. The speaker is a female speaker. FIG. 18 is a graph showing an example of an analysis result of the relationship between the durations Dvp, DseNi, Dvs, and DG of each phoneme portion of a phoneme sequence including a series of different vowels. In this example, the meaningless word is “Kobai Menkai”. In this figure, the horizontal axis is the DG value, and the vertical axis is Dv
The values of p, DseNi, and Dvs are shown, and the measured values of Dvp, DseNi, and Dvs are plotted for each DG value obtained by changing the speech rate. A straight line represented by a linear expression of DG obtained by the least square method for each of the measured values of Dvp, DseNi, and Dvs is also shown in FIG. Thus, Dvp,
It can be seen that the relationship between DseNi, Dvs and DG is well approximated by a linear expression similar to the above expression (2).

Here, Avp and Bvp are the coefficients and constant terms of DG in the linear expression relating to Dvp, and AseNi and BseNi are DseN.
The coefficient and constant term of DG in the linear expression for i, Avs,
When Bvs is a coefficient and a constant term of DG in a linear expression relating to Dvs, these Avp, AseNi, Avs, Bvp, BseN
The set of i, Bvs is determined for each different set of vowels.

Although illustration of other phonological sequences is omitted, in the present rhythm rule (B), as described in the description of the above phonological sequences, the number of types of phonological sequences other than vowel parts existing between vowels is described. Only a set of a coefficient of a linear expression and a constant term corresponding to the expression (2) is prepared. Although the rhythm rule (B) uses a linear expression such as the relational expression (2) as an approximate expression, the present invention is not limited to this relational expression. Can be. Further, in the rhythm rule (B), the approximate expression to be used can be simplified by grouping the types of phoneme sequences according to the purpose at the time of generating the synthesized speech. Furthermore, by using that DG is the sum of the time lengths of the respective parts of each phoneme sequence, the number of prepared DG coefficients and constant terms is each (the number of phoneme parts included in each phoneme section minus one). Can be done.

As described above, the present invention has been realized as an apparatus in which the degree of freedom is increased and a new rule is found therein to synthesize a Japanese speech that matches the natural speech more. Note that, depending on the purpose of the synthesized speech device, the above-described precise rules can be simplified and implemented as necessary.

[0051]

According to the speech synthesizing apparatus of the present invention, "when a transition from a vowel having the vowel part energy centroid CEGV to the next vowel occurs, a phonological sequence other than the vowel part must be realized. And the difficulty of uttering determines the time length between vowel energy centroids. "Based on the principle, the location of the vowel energy centroid CEGV that further determines the phonological segment is determined as follows. A single vowel energy centroid point is set for an existing mora. ", The phonemes of a phonetic sequence such as a mora phoneme such as a sound-repelling sound or a prompting sound or a vowel unvoiced phoneme sequence which could not be expressed by the conventional method. A rule that can determine the duration time accurately can be obtained, and as a result, it is possible to obtain a more natural synthesized speech easily. That.

[Brief description of the drawings]

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

FIG. 2 is a block diagram illustrating a configuration of a phoneme duration length generation unit.

FIG. 3 is a schematic diagram of a phoneme section of a general VCV phoneme sequence.

FIG. 4 is a schematic diagram of a phoneme section of a phoneme sequence including a mora phoneme such as a vowel sound, a prompt sound, and a long sound and a mora in which a vowel has been devoiced.

FIG. 5 is a graph showing an example of an analysis result of a relationship between a speech speed of a VCV-type phoneme sequence and a time length between vowel part energy centroids.

FIG. 6 is a graph showing an example of an analysis result on a relationship between the utterance speed of a phoneme sequence including a sound repellent and the time length between vowel part energy centroids.

FIG. 7 is a graph showing an example of an analysis result on a relationship between a speech speed of a phoneme sequence including a prompting sound and a time length between vowel part energy centroids.

FIG. 8 is a graph showing an example of an analysis result on the relationship between the utterance speed of a phonological sequence including a mora that has caused vowel devoicing and the time length between vowel part energy centroids.

FIG. 9 is a schematic diagram showing sections in a phoneme section of a VCV type phoneme series.

FIG. 10 shows Dvp, Dc, and Dv in the analysis result of the meaningless word “Kobakamenkai”, which is an example of the VCV phoneme sequence.
It is a graph which shows the relationship between s and DG.

FIG. 11 is a schematic diagram showing the division of a phoneme portion in a phoneme section of a phoneme series including a sound repellent.

FIG. 12 is a schematic diagram showing a division of a phoneme portion in a phoneme section of a phoneme sequence including a prompting sound.

FIG. 13 is a schematic diagram showing the division of a phoneme portion in a phoneme section of a phoneme sequence including a mora in which vowels have been devoiced.

FIG. 14 is a schematic diagram showing the division of a phoneme portion in a phoneme section of a phoneme sequence including a series of different vowels.

FIG. 15 shows Dvp, D in the analysis result of the meaningless word “Kobanba Menkai”, which is an example of a phoneme sequence including a sound repellent.
It is a graph which shows the relationship between N, Dc, Dvs and DG.

FIG. 16 shows Dvp and Dso in the analysis result of the meaningless word “Kobakamenkai”, which is an example of a phoneme sequence including a gong.
It is a graph which shows the relationship between k, Dvs, and DG.

FIG. 17 is a graph showing the relationship between Dvp, Dmusei, Dc, Dvs, and DG in the analysis result of the meaningless word “Kobaki Chamenkai”, which is an example of a phonological sequence including mora that has caused vowel devoicing. It is.

FIG. 18 shows Dvp and Dse in the analysis result of “Kobaimenkai”, which is an example of a phoneme sequence including a sequence of different vowels.
4 is a graph showing the relationship between Ni, Dvs and DG.

[Explanation of symbols]

4 text analyzer, 6 phoneme duration generator, 8
Rhythm rule, 10 sound source amplitude pattern generator, 12 pitch pattern generator, 14 spectrum pattern generator, 24
Sound source generation unit, 26 speech synthesizer, 40 division unit,
42 vowel part energy centroid time length setting part, 44
Rhythm rule (A), 46 phoneme duration setting part, 4
8 Rhythm rules (B).

Claims (6)

[Claims] 1. A speech synthesizer for regularly synthesizing speech from a text, comprising: a phoneme symbol string generating means for generating a phoneme symbol string including a vowel part, a consonant part and a mora phoneme from a text; A section dividing unit that divides the vowel section into phonological sections according to types of the consonant section and the mora phoneme in the phonological section, and an interval between energy centroid points of the vowel sections located at both ends of the phonological section. Vowel part energy barycenter time length generating means for obtaining a vowel part energy barycenter time length, the vowel part energy barycenter time length of the phoneme section,
Phoneme duration generation means for determining a phoneme duration of each phoneme in the phoneme section according to the type of the consonant part and the mora phoneme in the phoneme section, and using the phoneme duration. A speech synthesizer characterized by performing speech synthesis. 2. A speech synthesis apparatus for rule-synthesizing a speech from a text, comprising: a phoneme symbol string generating means for generating a phoneme symbol string including a vowel portion, a consonant portion, and a vowel unvoiced phoneme sequence from a text; Partitioning means for partitioning the vowel parts into phonological sections, and the energy of each vowel part located at both ends of the phonological section according to the type of the consonant part and the vowel unvoiced phoneme sequence in the phonological section. Vowel part energy barycenter time length generating means for obtaining a vowel part energy barycenter time length that is an interval between barycenter points, and the vowel part energy barycenter time length of the phoneme section,
Phoneme duration generation means for determining the phoneme duration of each phoneme in the phoneme section according to the consonant part in the phoneme section and the type of the vowel unvoiced phoneme sequence. A speech synthesizer characterized by performing speech synthesis using a length. 3. A speech synthesizer for regularly synthesizing a speech from a text, comprising: a phoneme symbol sequence generating means for generating a phoneme symbol sequence including a vowel portion, a consonant portion, a mora phoneme, and a vowel unvoiced phoneme sequence from the text; Partitioning means for partitioning a phoneme symbol string into phonemic sections by the vowel parts, and located at both ends of the phoneme section according to the type of the consonant part, the mora phoneme and the vowel unvoiced phoneme sequence in the phoneme section. A vowel part energy centroid time length generating means for obtaining a vowel part energy centroid point time length which is an interval between the energy centroid points of each of the vowel parts, and the vowel part energy centroid point time length of the phoneme section,
A consonant part in the phoneme section, the mora phoneme and the vowel unvoiced phoneme sequence, depending on the type of the phoneme duration, including phoneme duration generation means for determining the phoneme duration of each phoneme in the phoneme section, A speech synthesizer characterized in that speech synthesis is performed using the phoneme duration. 4. The vowel part energy barycenter time length generation means determines the vowel part energy barycenter time length by a linear function of a reciprocal of a speech rate, wherein the coefficient and the constant term of the linear function are consonant parts. 2. The speech synthesis device according to claim 1, wherein the speech synthesis device is determined according to a type of the mora phoneme. 5. The vowel part energy barycenter time length generation means determines the vowel part energy barycenter time length by a linear function of a reciprocal of a speech rate, and the coefficient and the constant term of the linear function are consonant parts. 3. The speech synthesis apparatus according to claim 2, wherein the speech synthesis apparatus is determined according to a type of the vowel unvoiced phoneme sequence. 6. The vowel part energy barycenter time length generation means determines the vowel part energy barycenter time length by a linear function of a reciprocal of a speech rate, and the coefficient and the constant term of the linear function are consonant parts. 4. The speech synthesizer according to claim 3, wherein the speech synthesis unit is determined according to a type of the mora phoneme and the vowel unvoiced phoneme sequence.