JP2848604B2 - Speech synthesizer - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION <Industrial Application Field> The present invention relates to a speech synthesizer that generates rule-synthesized speech.

<Prior Art> In order to generate a natural synthesized voice, it is important to make a predetermined vowel unvoiced. Conventionally, rules for devoicing vowels include the following: "Shigeru Sakurai" What to Watch Out for Common Language Pronunciations "Japanese Pronunciation Accent Dictionary (revised new edition) NHK,
Commentary / Appendix P.128 1985 ”is shown as a general rule. It describes in detail the typical phonological environment in which vowel silencing occurs.

Examples of the devoicing rules that are actually used in a speech synthesizer include, for example, “Yamato Sato, Kazuo Hakoda,“ Speech Synthesis by Law ”, Research and Application Report, Vol. 27, No. 12, P.2562
(62) Dentsu Public Corporation, 1978. "

FIG. 3 is a flowchart of a devoicing determination routine using the conventional vowel devoicing rule. Hereinafter, the conventional voiceless judgment routine will be described with reference to FIG.

In step S _31, the vowel of interest are those vowels (/ i /, / u
/) Is determined. As a result, the process proceeds to step S ₃₂ if it is such a vowel, the process proceeds to step S ₃₇ it is determined that the voiced otherwise.

In step S _32, whether or not a vowel of interest is sandwiched unvoiced consonant is determined. As a result, the process proceeds to step S ₃₃ if it is pinched, the process proceeds to step S ₃₇ it is determined that the voiced otherwise.

In step S _33, the vowel of interest accents nuclear whether it has the (pitch of a sound from a relatively high position changes to low) is determined. As a result, the process proceeds to step S ₃₉ if it has an accent nucleus, step S otherwise
Proceed to ₃₄ .

In step S _34, whether or not a vowel of interest is the first mora is determined. As a result, if the first mora proceeds to step S _36, the process proceeds to step S ₃₅ otherwise.

In step S _35, whether or not the preceding vowel is already devoiced is determined. As a result, the process proceeds to step S ₃₉ as a quasi-unvoiced if it is silent and proceeds to step S ₃₆ otherwise.

In step S _36, whether or not a vowel of interest is sandwiched unvoiced fricative of the same type is determined. As a result, the process proceeds to step S ₃₉ as a quasi-unvoiced if sandwiched, the process proceeds to step S ₃₈ otherwise.

In step S _37, it is determined to voicing the vowel of interest.

In step S _38, the vowels of interest is determined to be unvoiced.

In step S _39, the quasi-unvoiced processing such as shortening the duration of the vowels of interest is performed.

<Problems to be Solved by the Invention> However, the above-mentioned conventional devoicing rules do not show a clear devoicing criterion between vowel devoicing and a phonological sequence and between vowel devoicing and an accent pattern. However, there is an ambiguous point as a criterion for devoicing. In addition, when the generation speed changes, the rate of voicelessness generally changes, but the relationship between the generation speed and the voiceless reference is not described at the time of quantification. For this reason, the rule of silence may not always match the actual voice. Therefore, there is a problem that the synthesized speech generated by the above-mentioned conventional rule of silence sounds unnatural.

Therefore, an object of the present invention is to make the target vowel voicing / voicing based on the degree of unvoiced occurrence for each phoneme sequence, the degree of unvoiced occurrence for each accent pattern, and the unvoiced threshold set according to the utterance speed. It is an object of the present invention to provide a speech synthesizer that can generate a natural synthesized speech close to a real speech by judging unvoiced voice.

<Means for Solving the Problems> In order to achieve the above object, an invention according to claim 1 is a character string analysis unit that determines a phoneme sequence / accent pattern from an input character string, and a phoneme sequence / accent pattern / accent pattern.
A speech synthesis device comprising: a speech synthesis parameter creation unit that creates speech synthesis parameters according to rules stored in a rule file; and a speech synthesis unit that performs speech synthesis based on the speech synthesis parameters.
First unvoiced occurrence degree setting means for setting a degree of unvoiced occurrence for each vowel based on the phoneme sequence and the degree of unvoiced occurrence for each phonological sequence stored in the rule file; Based on the degree of unvoiced occurrence for each accent pattern stored in the file,
Second unvoiced occurrence degree setting means for setting an accent pattern unvoiced occurrence degree for each vowel, and unvoicedness calculation for each vowel from the phonological sequence unvoiced occurrence degree and accent pattern unvoiced occurrence degree. Means for calculating a degree of voicing occurrence, setting means for setting a voicing determination threshold based on the specified vocalization rate and a voicing determination threshold for each vocalization rate stored in the rule file; Vowel voicing / voicing determining means for determining voicing / voicing for each vowel based on the voicing determining threshold, and generating voice synthesis parameters according to the determination result of the vowel voicing / voicing determining means. And

<Operation> An arbitrary character string is input to the character string analysis unit, and the phoneme sequence and the accent pattern of the character string are determined. When the phoneme sequence and the accent pattern from the character string analysis unit are input to the speech synthesis parameter creation unit, speech synthesis parameters are created according to the phoneme sequence and the accent pattern and rules stored in the rule file.

At this time, the first unvoiced occurrence degree setting means of the speech synthesis parameter creation unit sets the input character string based on the phoneme sequence and the unvoiced occurrence degree for each phoneme sequence stored in the rule file. A phonemic sequence unvoiced occurrence degree is set for each vowel. The second unvoiced occurrence degree setting means sets an accent pattern unvoiced occurrence for each vowel of the input character string based on the accent pattern and the degree of unvoiced occurrence for each accent pattern stored in the rule file. The degree is set.
Then, the unvoiced occurrence degree calculating means calculates the unvoiced occurrence degree for each vowel from the set phonological sequence unvoiced occurrence degree and the accent pattern unvoiced occurrence degree.

Further, the setting means sets a de-voice determination threshold corresponding to the specified utterance speed based on the de-voice determination threshold for each utterance speed stored in the rule file. Then, based on the calculated degree of occurrence of unvoicedness and the set unvoicedness determination threshold, voiced / unvoiced is determined for each vowel of the input character string by the vowel voiced and unvoiced determination unit. . Then, the speech synthesis parameters are generated according to the determination result by the vowel voiced / unvoiced determination means.

When the speech synthesis parameters are generated by the synthesis parameter generation unit as described above, speech synthesis is performed by the speech synthesis unit based on the generated speech synthesis parameters.

In this way, the vowel is devoiced in consideration of the phonological sequence and the accent pattern and the utterance speed, which greatly affect the vowel devoicing phenomenon, and a natural synthesized speech is obtained.

<Embodiment> The configuration and operation of the speech synthesizer according to the present invention will be described below with reference to the block diagram of FIG.

When an arbitrary character string is input to the character string analysis unit 1, the character string analysis unit 1 performs syntax interpretation of the input character string,
Determine the intonation pattern for the entire string. Further, a word included in the character string is searched with reference to the word dictionary 2, and the accent and phoneme sequence of each word in the character string are determined, thereby determining the phoneme sequence and accent pattern of the character string. In this way, the intonation pattern of the entire character string determined by the character string analysis unit 1 and the phoneme sequence and accent pattern of the character string are output to the rule control unit 3 as the speech synthesis parameter creation unit. You.

The feature parameter file 8 includes a target feature parameter file 6 and a time-series feature parameter file 7.
Outputs a target feature parameter representing a vowel feature to the rule control unit 3, and the time-series feature parameter file 7 outputs a time-series feature parameter representing a consonant feature to the rule control unit 3. On the other hand, rule file 4
Outputs to the rule control unit 3 a phonological control rule for connecting the target characteristic parameter and the time-series characteristic parameter output from the characteristic parameter file, and a prosody control rule for controlling each prosody. The prosody control rules include an unvoiced phoneme sequence described later, a vowel unvoiced occurrence degree for each phoneme sequence, and a vowel unvoiced occurrence degree for each accent pattern.

The rule control unit 3 stores the feature parameter file 8
With reference to the target feature parameter and the time-series feature parameter input from the above, the phoneme control rule for combining each phoneme input from the rule file 4 and the prosody control rule for controlling each prosody, The parameters necessary for speech synthesis are generated based on the intonation pattern of the entire character string, the phoneme sequence of the character string, the accent pattern, and the vowel voiced / unvoiced determination results described later, which are input from the character string analysis unit 1. Then, the generated parameters are output to the speech synthesizer 5.

The voice synthesizer 5 determines, based on the input parameters,
Speech synthesis is performed, and a rule synthesizer corresponding to the input character string is output.

FIG. 2 shows the voicing performed by the rule control unit 3.
It is a flowchart of a voiceless judgment routine. The second
The voiced / unvoiced determination routine will be described with reference to the drawings.

In step S ₁ , first, a vowel sequence is devoiced for each vowel from the phoneme sequence input from the character string analysis unit 1 and the unvoiced occurrence coefficient representing the degree of unvoiced occurrence for each phoneme sequence stored in the rule file 4. An occurrence coefficient is determined.

However, if either of the preceding and following consonants is a consonant, the phonological sequence unvoiced occurrence coefficient is multiplied by α (α: 0 <α <1) to reduce the degree of unvoiced occurrence.

The phoneme sequence unvoiced occurrence coefficient has, for example, the following characteristics.

1. The vowel is a high tongue vowel (/ i /, / u /) and the consonant preceding this vowel is unvoiced fricative (/ s /, / sh /, / h /) or unvoiced affricate (/ ch / , / ts /) and the subsequent consonant is a voiceless plosive (/ p /, / t /, / k /, / py /, / ky /) or a voiceless affricate (/ ch /, / t
s /), the consonants before and after are silent plosives (/ p /, / t /, / k
/, And subsequent consonants include / py /, / ky /).

2. The vowel is a high tongue vowel (/ i /, / u /), the consonants before and after are unvoiced plosives (/ p /, / t /, / k /, and the following consonants are / py /, / ky / Including)
In the case of, the consonants before and after are different types of unvoiced fricatives (/ s /,
/ sh /, / h /, and the following consonants include / hy /).

3.The vowel is a high tongue vowel (/ i /, / u /), the consonants before and after are different types of unvoiced fricatives (/ s /, / sh /, / h /, and the following consonants are / hy /
), The preceding and following consonants have the same unvoiced fricative (/ s
/, / sh /, / h /), the degree of silence is higher.

4. If the vowel is a high tongue vowel (/ i /, / u /) and the preceding and following consonants are the same unvoiced fricative (/ s /, / sh /, / h /), one of the preceding and following consonants Is more unvoiced than when it is not a voiceless consonant.

5. If the vowel is not a high tongue vowel (/ i /, / u /), do not devoice.

Table 1 shows an example of a phonological sequence unvoiced occurrence coefficient classified by consonants before and after a high tongue vowel. The vertical is the preceding consonant and the horizontal is the succeeding consonant, and the larger the coefficient, the higher the degree of unvoicedness. The numbers in parentheses indicate the case where the preceding and following consonants are the same.

Next, an accent pattern unvoiced occurrence coefficient is determined for each vowel from the accent pattern input from the character string analysis unit 1 and the accent pattern unvoiced occurrence coefficient stored in the rule file 4.

The degree of silence occurrence for each accent pattern has the following characteristics.

a. Vowels in the accent nucleus are hardly devoiced.

The bO type utterance has a higher degree of vowel devoicing than the type 1 or higher utterance.

c. Except for Type 1 utterances, the degree of devoicing is high.

d. Vowels immediately before the accent nucleus are more devoiced than vowels in the accent nucleus, but less devoiced than vowels after the accent nucleus.

Table 2 shows an example of the degree of unvoiced occurrence for each accent pattern. The vertical indicates the accent type, the horizontal indicates the mora position of the target vowel, and the larger the coefficient, the higher the degree of unvoicedness. Here, the accent type is classified according to the position of the syllable that is higher among the syllables forming the word. For example, “n-type” means that the syllables from the second syllable to the nth are sung higher, Syllable and all syllables after the (n + 1) -th syllable indicate that it is to be chanted low.

Further, from the phonological sequence unvoiced occurrence coefficient and the accent pattern unvoiced occurrence coefficient, the unvoiced occurrence degree ρ (n) (n = 1, 2, 1 ,, WM + 1 WM for each vowel is calculated according to the following equation. Mora number, and WM + 1 represents silence at the end).

ρ (n) = phonetic sequence unvoiced occurrence coefficient × accent pattern unvoiced occurrence coefficient In step S ₂ , a voiceless determination threshold for determining voicelessness determined for each voice speed stored in the rule file 4, From the utterance speed input from the character string analysis unit 1, a threshold θ as a criterion for de-voicing is obtained as follows. That is, a threshold value for a normal utterance speed is determined, and when the utterance speed is higher than that, the threshold value is lowered, and when the utterance speed is lower than that, the threshold value is raised.

As an example, Table 3 shows threshold values for each utterance speed when the threshold value of the normal utterance speed is 6.

Further, a voicing determination threshold value θ2 when vowels that are likely to be voicing occur is obtained by the following equation.

[theta] 2 = [theta] * [beta] ([beta] is a real number satisfying 1 <[beta]) Next, a determination of voicelessness is started for each vowel.

In step S _3, n in order to perform the determination of one mora th
= 1.

In Step S _4, devoicing degree of the current mora (n Mora) [rho and (n) and unvoiced assessment threshold θ is compared. As a result, ρ (n) ≧ θ current mora vowel when the process proceeds to step S ₅ as a possible devoicing, ρ (n) <θ step S ₁₁ it is determined that no devoicing when the Proceed to.

In step S _5, the devoicing degree follows mora (n + 1) ρ (n + 1), and devoicing degree [rho of the current mora (n) (n) are compared. As a result, ρ (n) ≧ ρ (n
+1) That case towards the current mora high degree of devoicing the vowel current mora although silent and proceeds to step S ₈ as there are cases where vowels following mora not unvoiced. On the other hand, if ρ (n) <ρ (n + 1) i.e. towards the current mora degree of devoicing is low, the flow proceeds to step S ₆ as there are cases where vowels current mora not unvoiced.

This is because if vowels to be unvoiced continue, one vowel with a low degree of unvoicedness is not unvoiced in order to avoid unclear pronunciation.

In step S _6, the process proceeds by increasing the ρ (n + 1) as a vowel follows mora is always devoiced to step S _7.

In step S _7, devoicing the degree of the current mora ρ (n)
Is compared with a voicing determination threshold θ2 when vowels that are likely to cause voicing are chained. As a result, ρ (n)
Current Mora of the vowel in the case of ≧ θ2, the process proceeds to step S ₁₀ for a strong degree of devoicing. On the other hand, in the case of ρ (n) <θ2 proceeds to step S ₁₁ since no silent the current mora vowel.

In step S _8, devoicing degree of next mora ρ (n +
1) is compared with a voicing determination threshold value θ2 when vowels that are likely to cause voicing are chained. As a result, ρ (n +
1) In the case of <.theta.2 proceeds vowel follows Mora Step S ₉ determines not to unvoiced. On the other hand, vowel follows Mora For ρ (n + 1) ≧ θ2 is directly proceeds to step S ₁₀ a strong degree of unvoiced.

In step S _9, the vowels following mora in order not to devoiced, the process proceeds to step S ₁₀ as ρ (n + 1) = 0 .

In step S _10, the process proceeds to step S ₁₂ it is determined that the unvoiced the n mora th vowel.

In step S _11, the process proceeds to step S ₁₂ it is determined that the voicing an n mora th vowel.

In step S _12, in order to move to the determination of the vowel next Mora, it is incremented by one n.

In step S _13, n is number of moras phoneme sequences inputted (W
M) It is determined whether or not: As a result, n ≦ WM
For Step S ₁₂ is repeated from the step S _4,
If n> WM, the unvoiced determination routine ends.

As described above, in the present invention, the unvoiced occurrence threshold for each phoneme sequence, the unvoiced occurrence for each accent pattern, and the unvoiced determination threshold set according to the utterance speed are linked to the vowels that are likely to be unvoiced. Based on the unvoiced determination threshold at the time, the voiced / unvoiced voice of the target vowel is determined,
For an input character string, vowels can be devoiced in accordance with the phoneme sequence, accent pattern, and utterance speed of the character string. Therefore, according to the present invention, a more natural synthesized speech can be generated.

<Effects of the Invention> As is apparent from the above description, in the speech synthesis apparatus according to the first aspect of the present invention, the speech synthesis parameter creation unit uses the phonetic sequence of the rule file to generate unvoiced voices and accent patterns to generate voiceless voices. Based on the degree, the degree of unvoiced occurrence for each vowel of the input character string is calculated, and further, based on the unvoiced determination threshold for each vocalization rate, a devoicing determination threshold is set according to the specified vocalization rate Then, voicing / voicing is determined for each vowel of the input character string based on each of the unvoiced occurrence degrees and the voicing determination threshold, and a voice synthesis parameter is generated according to the voicing / voicing determination result. ,
By performing speech synthesis based on these speech synthesis parameters, a target vowel can be devoiced in consideration of a phoneme sequence, an accent pattern, and a utterance speed that greatly affect the vowel devoicing phenomenon. That is, according to the present invention, it is possible to generate a natural synthesized speech that is close to a real speech.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of a voice synthesizing apparatus according to the present invention, FIG. 2 is a flowchart of a voicing / voicing determination routine for vowels in the above embodiment, and FIG. 3 is conventional voicing / voicing. It is a flowchart of a determination routine. 1 ... character string analysis unit, 2 ... word dictionary, 3 ... rule control unit, 4 ... rule file, 5 ... speech synthesizer, 6 ... target feature parameter file, 7 ... time-series feature parameter file , 8 ... Feature parameter file.

──────────────────────────────────────────────────続 き Continued on front page (58) Field surveyed (Int. Cl. ⁶ , DB name) G10L 5/02 G10L 3/00

Claims (1)

(57) [Claims] A character string analysis unit for determining a phoneme sequence and an accent pattern from an input character string;
Speech synthesis including a speech synthesis parameter creation unit (3) for creating speech synthesis parameters in accordance with the rules stored in the accent file and the rule file (4), and a speech synthesis unit (5) for performing speech synthesis based on the speech synthesis parameters. A speech synthesis parameter creation unit (3), based on the phoneme sequence and the degree of unvoiced occurrence for each phonological sequence stored in the rule file (4), a degree of unvoiced occurrence of a phonological sequence for each vowel. First unvoiced occurrence degree setting means (S1) for setting an accent pattern and an unvoiced occurrence degree for each vowel based on the accent pattern and the degree of unvoiced occurrence for each accent pattern stored in the rule file (4). A second unvoiced occurrence degree setting means (S1) for setting the degree; a phonetic sequence unvoiced occurrence degree and an accent pattern unvoiced occurrence degree; Based on the voicing occurrence degree calculating means (S1) for calculating the voicing occurrence degree for each vowel, the specified vocalization rate, and the voicing determination threshold for each vocalization rate stored in the rule file (4). Setting means (S2) for setting a devoicing judgment threshold value; and vowel voiced and unvoiced judgment means (S4) for judging devoicing / voicing for each vowel based on the set devoiced occurrence degree and devoiced judgment threshold value. To S13), and generates a speech synthesis parameter in accordance with the determination result of the vowel voiced and unvoiced determination means (S4 to S13).