JP3186263B2 - Accent processing method of speech synthesizer - Google Patents

Abstract

PURPOSE:To improve the naturalness of a synthesized speech whose pitch pattern is generated corresponding to a multistage accent pattern by using accent basic tables of two-stage accents. CONSTITUTION:Phoneme rain data and an N-stage accent pattern corresponding to an input sentence from a host computer 1 are inputted as to the pertinent phoneme and phonemes before and after it and multistage accents are quantized 15 into two stages to set 16 a basic accent from the two-stage accent basic table 14; and the accent component is performed in accent component variation quantity control 19, movement control 21, and mixing control 23 according to tables 20, 22, and 24 in consideration of the phonemes before and after it, thereby synthesizing a pitch pattern.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an accent processing method in a prosody processing for converting a character code inputted from a host computer or the like into a prosody parameter in a rule-synthesizing speech synthesizer.

[0002]

2. Description of the Related Art A rule-based speech synthesizer divides an input character string into words and phrases by syntactic analysis and morphological analysis, determines intonation and accent, decomposes words and phrases into syllables and phonemes, Alternatively, parameters of a sound source wave and an articulation filter are obtained for each phoneme, and a synthesized speech is obtained as a response output of the articulation filter to the sound source wave.

The prosody processing includes pitch pattern parameter generation in which the pitch frequency (basic frequency) of each syllable or phoneme is adjusted according to intonation, accent, and the pitch pattern to be generated, and energy pattern parameter generation in which the intensity of each sound is adjusted. Generates a relay time length parameter for adjusting the intensity of each sound.

FIG. 3 shows a prosody processing method using a conventional rule speech synthesis method. For a sentence mixed with kanji or kana given from the host computer 1 or the like, the Japanese processing unit 2 performs syntax analysis and morphological analysis to determine accents including intonation and generate phoneme string data.

The prosody processing unit 3 includes a pitch pattern generation unit 3A for determining a pitch frequency from accent data for each phoneme of the phoneme sequence data, and an energy pattern generation unit 3B for determining basic energy data for each phoneme from pitch pattern parameters. And a time length calculator 3C for determining the duration of each phoneme.

The synthesizing section 4 obtains a synthesized sound signal from each parameter as a result of the prosody processing through an articulation filter or the like, and obtains a synthesized sound from the sound output device 5.

[0007]

SUMMARY OF THE INVENTION In the conventional system,
An accent change pattern for an input sentence is given in two stages of high / low.

Here, the prosody pattern of the real voice that becomes a human utterance changes continuously with respect to the accent change pattern, and the two-stage accent change is too coarse to improve the voice quality.

However, with the current Japanese processing technology, it is not possible to convert a sentence mixed with kanji or kana into an accent of three or more levels. Therefore, in order to add three or more levels of accent to a target sentence, it is necessary for a human to intentionally give it.

On the other hand, a pitch pattern which controls the pitch of a synthesized voice is generated by referring to a corresponding table value from an input accent pattern or the like from an accent basic table. However, the basic accent table obtained from the two-stage accent environment cannot be used as it is for the N-stage input.

An object of the present invention is to provide an accent processing method for generating a pitch pattern corresponding to a multi-stage accent pattern using a two-stage accent basic accent table to enhance the naturalness of a synthesized sound.

[0012]

SUMMARY OF THE INVENTION In order to solve the above-mentioned problems, the present invention provides a speech synthesizing apparatus using a rule synthesizing method, comprising means for obtaining phoneme string data from an input sentence and multi-stage accent amounts for each phoneme. Means for obtaining two basic accent data of the phoneme represented in two stages from the phoneme of the phoneme sequence and the phonemes before and after the phoneme, and at least one of the respective accent amounts of the phoneme and the phonemes before and after the phoneme is different. Means for determining whether or not to mix the two basic accent data at a predetermined ratio depending on whether or not the two are different from each other; and an accent component change amount table for the basic accent data based on the accent patterns of the phoneme and the phonemes before and after the phoneme. Accent component change amount control means for changing the accent component according to the value of A movement control unit for adding the accent component with a value of a movement amount table according to the accent pattern; and a mixing control unit for mixing the movement-controlled basic accent data at a predetermined ratio by the determination unit. It is characterized by synthesizing a pitch pattern in which basic accent data of each phoneme for a sentence is mixed with the accent component change and movement.

[0013]

[Action] For each phoneme, the multi-stage accent amount including the preceding and succeeding phonemes is quantized into two stages, and two basic accent data expressed in two stages are extracted. The pitch pattern of each phoneme is synthesized by mixing the two basic accent data, changing the accent component, and moving.

[0014]

FIG. 1 is a flowchart showing a pitch pattern generating process according to an embodiment of the present invention. The host computer 1 gives a phoneme symbol and an N-stage accent pattern for each phoneme for pitch pattern generation by Japanese processing of the input sentence.

The three-phoneme window multiplier 11 extracts the phoneme to be processed, the preceding phoneme and the subsequent phoneme data from the given phoneme string data. On the other hand, a 5-mora window 12 and a 3-mora window 13
Is the amount of N stages of the mora and the N stages of two mora before and after the mora (preceding mora, preceding mora, subsequent mora, and subsequent mora) or one mora before and after (the preceding mora, the following mora) Is extracted as the accent environment of the mora.

The accent basic table 14 stores the types of the phonemes to be processed (10 types in Table 1 below), the types of preceding phonemes (long sounds, other), and the types of subsequent phonemes (promotion sounds,
A typical accent pattern is selected for each group from the pitch pattern data for each group created by voiced consonants, unvoiced consonants, vowels, end of sentences and phrases, and seven values (consonants) representing the basic components of this accent pattern Part 3
(Points, vowels 4 points) are stored.

[0017]

[Table 1]

The two-stage quantization 15 divides the five-mora N-stage accent pattern extracted by the five-mora window multiplier 12 into two-stage accent patterns.

The basic accent data set 16 is composed of two patterns from the basic accent table 14 of the phoneme data from the three-phoneme window unit 11 using the two-stage accent from the two-stage quantization 15 as an accent environment. Extract.

The mixing control determining unit 17 determines whether or not to perform the mixing control based on the N-stage accents of the preceding mora, the succeeding mora, and the mora extracted in the three-mora window frame 13. In this mixing control determination, it is determined whether or not two patterns of accent data extracted in the basic accent data set 16 are mixed at a predetermined ratio (mixing coefficient).

The mixing control flag for this is obtained from the mixing control flag table 18, and the actual mixing control is performed by a mixing control 23 and a mixing coefficient table 24 described later.

The mixing control flag table 18 and the mixing coefficient table 24 are exemplified in Table 2 below.

[0023]

[Table 2]

This table shows a case in which each of the three mora has five accent levels (N = 5). The accent amounts (1 to 5) of the preceding mora and the subsequent mora, and the case where the preceding mora is the beginning of a phrase. The case where the succeeding mora is the end of a phrase is also shown. In each combination, the mixing control flag (1, 0) and the mixing coefficients α and β are shown in five stages for the mora. The mixing control flag is set to 1 when mixing control is performed, is set to 0 when mixing control is not performed, and the mixing coefficients α and β are set to 0.5 and 0.5 only when the mixing control flag is set to 1.

For example, the amount of accent of the preceding mora is 2,
When the amount of accent of the following mora is 3 and the amount of accent of the mora is 4, the mixing control flag is set to 1 to perform mixing control, and the mixing coefficients α and β are set to 0.5 and 0.5, respectively. Become mixed.

The determination as to whether or not to perform the above-described mixing control is made by determining whether the preceding mora and the accent amounts a, b, and c of the mora and the subsequent mora have the flag = 1 when a ≠ b, b ≠ c, and c ≠ a. When a = b or b = c or c = a, it is determined according to the flag = 0.

Table 3 shows an example of calculating a mixed accent in the mixing control 23.

[0028]

[Table 3]

In this example, the result of quantizing the accent environment in three stages is represented by L / M / H. That is, 3
If the mora accent amounts are a, b, and c, abc => HML when a>b> c, abc => HHL when a = b> c, abc => HLL a>c> when a> b = c abc => HLM when b = abc => HLH when b = a> c abc => MHL when b>a> c abc => LHL when b> a = c => LHM when b>c> a When b = c> a, abc => LHH When c>a> b, abc => MLH When c> a = b => LLH When a = b = c => HHH or LLL where HHH is the preceding When the quantization accent is H, when the preceding accent is a change in LM, there is no preceding mora, and when the succeeding is L. LLL is when the preceding accent is L, when the preceding is an ML change, when there is no preceding mora, and when the following is H.

In such an accent environment quantized in three stages, when M is included, mixing control is performed, and the target environment is HML → HLL, HHL HLM → HLL, HLH MHL → LHL, HHL MLH → LLH , HLH LHM → LHL, LHH LMH → LHH, LLH.

[0031] Taking as an example the first term of Table 3, when the accent environment 1,2,3, to calculate the pitch pattern data, the accent pattern data Pitch ₁₁₃ alpha (=
0.5) multiplied by accent pattern data Pitch
It is obtained by adding the value obtained by multiplying ₁₃₃ by β (= 0.5).

Next, the accent component change amount control 19
The data in the accent component change amount table 20 is extracted from the mora N-stage accent pattern, and the amount of change is used to correct the accent amount extracted in the basic accent data set 16.

The accent component change amount is calculated by analyzing a voice generated by a person, and is stored in the table 20. In this calculation, as shown in FIG. 2, a pitch pattern is generated by pitch analysis for a voice with a large accent change amount, a normal voice, and a small voice of a human generated sound, and a time axis normalization that matches each time axis is performed. After performing the conversion, windowing of three moras, and averaging, the amount of change in the accent component is calculated.

Table 4 shows an example of calculation of the change amount in the case of a five-stage accent pattern, and Table 5 shows an accent component change amount table.

[0035]

[Table 4]

[0036]

[Table 5]

[0037] Among the above Table 4, HHL (L) shows accent environment HHL, a string of accents amount small, Pitch ₄₄₁
(J) shows the j-th extracted accent data value in the accent environment ₄₄₁ .

The general formula for calculating the amount of change in accent is as follows:
The change amount Pabc with respect to the basic accent data in the accent environment abc (preceding, relevant, succeeding) is represented by the following equation.

[0039] _{Pabc = Pitch abc-Pitch ABC .........} (1) However, · a = b or b = c or c = a (a, b, c = 1~N) · Pitch ABC when in accent amount Max (A, B, C) = N in pitch pattern data (average processed)
/ 2 + 1 (truncated below the decimal point). (Ma
x (A, B, C) are A, B, the maximum value of C) A = B or B = C or _{C = A · Pitch A 'B} ' C ' pitch pattern data when the accent maximum (average processed ), Max (A ′, B ′, C ′) = max (A, B, C) * 2.

A '= B' or B '= C' = A 'Pitch _A ", _B ", _C "are pitch pattern data (average processed) when the accent is minimum, and max (A", B " , C ″) = 2.

A "= B" or B "= C" or C "= A" max (a, b, c) -min (a, b, c) <N / 2 +
1, max (a, b, c)> 2 Pitcha'b'c '= (Pitch _ABC + Pitch _A " _B " _C ")
* X / Z X = max (a, b, c) -2 Z = N / 2-1 (decimal point rounded down) • max (a, b, c) -min (a, b, c)> N / 2 +
1, max (a, b, c) <N Pitcha ', b', c '= (Pitch _ABC + Pitc
_{h A 'B' C ')} * Y / Z Y = max (a, b, c) - (N / 2 + 1) Z = N / 2-1 ( rounded down below the decimal point) Here, max (a, b, c ) = N and N is odd, the above expression does not apply. Therefore, only in this case, the change amount is calculated by the equation of Pabc = 2 * Pitcha'b'c '(2).

Returning to FIG. 1, after the control of the change amount of the accent component, the movement control 21 using the movement amount table 22 is performed. Table 6 shows an example of this movement amount table in the case of a five-stage accent pattern.

[0043]

[Table 6]

When the N-stage accent patterns a, b, and c are input, the amount of movement is expressed as follows: min (a, b, c) = 1 and a = b or b = c or c = f (3) When the condition is satisfied, there is no need to move, but when the accent pattern def is input, min (d, e, f)> 1 and d = e or e = f or f = d (4) When the condition is satisfied, the movement amount S _def is uniformly added to the pitch pattern value of the basic table. Here, the movement amount S
The calculation of _def is determined under the following conditions.

S _def = I (H, a, b, c) (5) I (H, a, b, c): Value where pattern abc is determined by accent difference H H = max (d, e, f) -max (a, b, c) (6) where d-H = a, e-H = b, and f-H = c The pitch for N-stage accent pattern input using the above five tables The pattern is generated for each phoneme of the input phoneme sequence. Unless the pitch pattern generation processing for the final phoneme is completed (25), the window shift to the next phoneme of the phoneme (26) and its N-stage accent are performed. A window shift to a pattern (27) is performed, and pitch pattern synthesis of all moras is performed (28) after completion of the pitch pattern generation processing for all phonemes, thereby obtaining pitch pattern parameters.

[0046]

As described above, according to the present invention, the control and movement of the accent component change amount of the basic accent data from the phoneme sequence data and the multi-stage accent amount for the input sentence are considered in consideration of the phoneme and the phonemes before and after the phoneme. The control and the mixing control have the following effects.

(1) In a rule speech synthesizer,
Although the pitch of the voice was controlled by two levels of accents, high or low, the pitch of the voice can be controlled by multiple levels of accents.

(2) A fundamental frequency pattern representing the pitch of a voice can be continuously generated.

(3) Accents in three or more stages, which could not be obtained by using the result of the conventional Japanese processing, can be added to the target sentence.

(4) Conventional pitch pattern data of an accent basic table obtained from a two-stage accent environment can be used for multi-stage input.

(5) Since a pitch pattern of a voice uttered by a human is used as a basis, a pitch pattern closer to a human pitch pattern can be generated.

(6) A natural synthesized voice similar to a voice uttered by a human can be obtained.

[Brief description of the drawings]

FIG. 1 shows a pitch pattern generation process according to an embodiment of the present invention.

FIG. 2 is a diagram illustrating an example of calculating an accent component change amount according to an embodiment.

FIG. 3 shows a conventional rule speech synthesis method.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Host computer 11 ... Over 3 phoneme windows 12 ... 5 Mora window over 13 ... 3 Mora window over 14 ... Accent basic table 15 ... Two-stage quantization 16 ... Basic accent data set 17 ... Mixed control determination 18 ... Mixed control table 19 ... Accent component change amount control 20 ... Accent component change amount cable 21 ... Movement control 22 ... Movement amount table 23 ... Mixing control 24 ... Mixing coefficient table

──────────────────────────────────────────────────続 き Continuation of front page (56) References JP-A-4-46396 (JP, A) JP-A-64-61795 (JP, A) JP-A-64-61796 (JP, A) JP-A-2- 48700 (JP, A) JP-A-61-57997 (JP, A) JP-A-63-85797 (JP, A) JP-A-4-134499 (JP, A) JP-A-2-197897 (JP, A) JP-A-4-83298 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 13/06 G10L 13/08

Claims (2)

(57) [Claims] 1. A speech synthesizing apparatus using a rule synthesizing method, comprising: means for obtaining phoneme string data from a sentence and multi-stage accent amounts for each phoneme; Means for obtaining two basic accent data of the phoneme represented, and mixing the two basic accent data at a predetermined ratio depending on whether at least one of the accent amounts of the phoneme and the phonemes before and after the phoneme is different. Determination means for determining whether or not, the accent component change amount control means for changing the accent component according to the value of the accent component change amount table for the basic accent data from the accent pattern of the phoneme and the phonemes before and after the phoneme, Movement table according to the accent pattern with respect to the basic accent data in which the accent component is changed A movement control unit that adds the accent component with a value of: and a mixing control unit that mixes the movement-controlled basic accent data at a predetermined ratio by the determination unit, wherein the basic accent data of each phoneme for the input sentence is An accent processing method for a speech synthesizer characterized by synthesizing a pitch pattern mixed with a change in accent component and movement. 2. The accent processing method for a speech synthesizer according to claim 1, wherein said accent component change amount table is obtained by analyzing pitch patterns of three-stage voices uttered by a human.