JP3081300B2 - Residual driven speech synthesizer - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a ruled speech synthesizer capable of uttering an arbitrary word, and more particularly to a residual drive type ruled speech synthesizer for performing residual drive.

[0002]

2. Description of the Related Art In recent years, research on a rule synthesis method for synthesizing speech from an arbitrary sentence has been actively conducted. is there.

A rule synthesizing apparatus for synthesizing speech from an arbitrary sentence, for example, performs a sentence analysis on a text input to determine a reading or an accent, and determines a phonetic rule from a phonetic rule.
A speech unit (for example, a CVC unit) which is a necessary synthesis unit is determined and combined, a pitch of voice is determined from a prosody rule, a time series of voice parameters and a pitch pattern are generated, and these parameters are generated. Then, a synthetic speech is generated by forming a sound source and a digital filter.

The speech parameters used in such a speech synthesis method are generally parameters of a linear prediction system such as LPC and LSP, formants, and the like. On the other hand, as a sound source, the memory is reduced and the processing is simple. For this purpose, impulse and white noise were used.

[0005] It is known that in a speech synthesis technique of a linear prediction system such as LPC and LSP, a synthesized speech close to the original speech can be obtained by using a prediction residual as a driving sound source.

Accordingly, it is expected that a high-quality synthesized sound can be obtained from the above-described principle by inputting a residual as a driving sound source also in rule synthesis in which an arbitrary voice can be uttered by inputting characters or the like. Such a residual-driven rule-based speech synthesizer has been proposed (Japanese Patent Application No. 2-2494).
No. 93), which is outlined below.

FIG. 1 shows a configuration of a previously proposed residual-driven rule-based speech synthesizer. According to this device, when a character string to be uttered is input to a character string buffer (1), The phoneme symbol string generation unit (2) converts the input character string into a phoneme symbol string. For example, when a character string of "** ni *" (* is a symbol indicating an accent position) is input, the character string is converted to a phoneme symbol string of "tabenikitita".

Next, the selection circuit 1 (3) sequentially determines and selects necessary speech units from the phoneme symbol string, and selects necessary speech units from the speech units stored in the speech unit memory (4). A piece
They are connected by a voice unit connection part (5). The unit of the speech unit is CVC (consonant + vowel + consonant) or CV
A combination of (consonant + vowel) and VC (vowel + consonant) may be used, but for simplicity, CV (consonant +
Vowels), that is, syllables,
"a, be, ni, ki, ta" are sequentially selected and connected as necessary speech units. The speech unit connected here is stored as a speech parameter in the speech parameter buffer (6), and given as a coefficient to the synthesis filter (11).

On the other hand, a symbol representing intonation such as an accent position is also input together with a character string to be uttered,
When supplied from the character string buffer (1) to the pitch pattern generator (7), the pitch pattern generator (7)
Determines the pitch (pitch) of the entire utterance. For example,
In the case of input of "** ni **", "*" and "*"
Since accent is present, the pitch pattern shown in Figure 2. In the pitch pattern generator (7),
A phrase component with a pitch drop in the entire sentence and an accent component with a higher pitch at the accent position are added,
A pitch pattern is generated.

As shown in FIG. 3, a residual waveform for use as a driving sound source is stored in the residual waveform memory (8) corresponding to each speech unit.
A necessary residual waveform is selected corresponding to the speech unit determined and selected in (3). In the example, "ta, be, ni,
The corresponding residual waveform is selected in the order of “ki, ta”. Further, as shown in FIG. 3, the residual waveform memory (8) stores a plurality of residual waveforms having different pitches for each speech unit, and the residual waveform is generated by the pitch pattern generator (7). The selection circuit 2 (9) selects and determines a residual waveform having an appropriate pitch in accordance with the pitch thus selected, and the driving sound source generator (1)
0). Lastly, the selected residual waveform is subjected to pitch change in the driving sound source generating section (10) to a value corresponding to the pitch generated by the pitch pattern generating section (7), and the residual waveform having a desired pitch is obtained. Is generated.

The residual waveform generated in this manner is input to the synthesis filter (11) as a driving sound source, and the synthesis filter (11) generates a synthesized voice. Synthesized voice is D
The signal is output from the speaker (13) via the / A converter (12).

The operation of the driving sound source generator (10) of the proposed device will be further described below.

FIG. 3 shows an example of waveform data of a residual signal stored in the residual waveform memory (8). The waveform data in the memory (8) corresponds to a case where a syllable having a CV (consonant + vowel) configuration is used as a unit of a speech unit, and three pitch periods are different for each speech unit. That is, a treble residual waveform (residue with a short pitch cycle), a medium residual waveform (medium residual with a pitch cycle),
A low-frequency residual waveform (residual having a long pitch cycle) is stored. The waveform data of the residual signal stored in the residual waveform memory (8) is processed as follows by the driving sound source generation unit (10) of FIG.

That is, the switch 1 (101) is selected by the selection signal from the selection circuit 1 (3) in FIG. 4, and the residual waveform corresponding to the required speech unit is stored in the residual waveform memory (8).
And is stored in the residual waveform buffer 1 (102). In the example of "** ni **", first, "ta"
Is read out. Next, the selection circuit 2 (9)
The switch 2 (103) is selected by the selection signal from, and a residual having an appropriate pitch is selected and stored in the residual waveform buffer 2 (104). In the example, as shown in FIG. 2, since the pitch of “ta” is as high as 400 Hz, the residual waveform of “ta” for the high tone is selected, and the residual waveform buffer 2
(104). Finally, the pitch change circuit (105) changes the pitch cycle of the residual waveform so that the pitch is determined by the pitch pattern generation unit (7). For example, the residual waveform of “ta” for treble is 380H
If it is extracted from the sound of z, the frequency is set to 20 Hz so that the frequency becomes 400 Hz determined by the pitch pattern generator (7).
A change is made such that the pitch is increased by Hz (the pitch period is shortened ).

The pitch change of the residual waveform performed by the pitch change circuit (105) is performed, for example, when lowering the pitch, zero data is inserted in the middle to increase the pitch cycle and increase the pitch. In such a case, a "zero padding truncation method" is used in which data in the middle is deleted to shorten the pitch period.

If the pitch change is made drastically, the sound quality deteriorates. Therefore, in the example of the residual drive type rule synthesizing apparatus described above,
Residuals having three different pitches for high, medium, and low pitches are stored in advance in a residual waveform memory, and a residual waveform having a pitch cycle close to a desired pitch is used. We are devising to reduce it.

As described above, the residual waveform corresponding to the selected speech unit and changed to a desired pitch is generated as a driving sound source, so that a synthesized speech with a high naturalness of pronunciation can be obtained. It is.

According to the residual driving type speech synthesizer outlined above, residual waveforms having different pitches stored in the residual waveform memory (8) are conventionally created by the method shown in FIG. Was common.

That is, for example, when three types (three stages) of residual waveforms are created, as shown in FIG.
By analyzing the treble input voice, a treble residual waveform having a short pitch cycle is extracted. FIG.
As shown in (c), a low-pitched input speech is analyzed to extract a low-pitched residual waveform having a long pitch cycle. As for the middle tone, as shown in FIG.
By analyzing the input sound of the middle sound, a residual waveform for a low sound having an average length of the pitch period is extracted.

[0020]

As described above, in the conventional residual driving rule synthesizing apparatus, when a driving sound source is generated using residual waveforms having different pitches, the driving sound sources are stored as speech units. Since the parameters such as LPC and LSP and the residual waveform used as the driving sound source are analyzed and extracted from different voices, the stored residual waveform is used as the driving sound source, and the stored speech unit is used. However, the original voice cannot be reproduced even if the parameter passes through the synthesis filter as a coefficient, and the generated synthesized voice deteriorates.

It is considered that if a speech segment extracted from a speech of the corresponding pitch is used for each residual waveform, the deterioration will be eliminated. In this case, however, the speech segment corresponding to each pitch is removed. It is necessary to store a plurality of them, and the amount of memory increases.

[0022]

A residual-drive speech synthesizer according to the present invention uses speech parameters in speech unit memory of a speech unit memory as coefficients of a speech analysis filter comprising an inverse filter of a speech synthesis filter. The waveform data of the residual signals of the voices having different pitch periods obtained by inputting the voices having different pitch periods into the analysis filter is extracted, and the waveform data is stored in a residual waveform memory.

[0023]

According to the residual driving type speech synthesizer of the present invention, no matter what cycle the residual waveform is used as the driving sound source for the speech unit selected for synthesizing the speech,
Waveform data for all of the residual signal that may be selected are those which are extracted by the inverse filter with coefficient speech parameters themselves to be subjected to actual synthesis, regardless of the change of pitch, and the parameters of the synthesis Extract residual signal
Since the parameters used at the time of performing are the same , a more natural and high-quality synthesized speech can be obtained.

[0024]

FIG. 6 shows a method of creating waveform data of a residual signal used in the residual drive type speech synthesizer of the present invention. Each of H (z) in FIGS. 6A, 6B, and 6C represents the transfer characteristic of a speech analysis filter composed of an inverse filter of a speech synthesis filter in a description of Z-transform. As shown in FIGS. 6 (a), 6 (b) and 6 (c), LPCs (linear prediction coefficients, linear prediction coefficients,
By applying an inverse filter that uses audio parameters such as autocorrelation coefficients and LSP as coefficients, residual waveforms for high-, medium-, and low-frequency sounds are generated, respectively.

The residual-drive speech synthesizer of the present invention stores, for each speech unit, a residual waveform created by the same method as shown in FIG. 6 in a residual waveform memory shown in FIG. In the configuration shown in FIG. 1, the residual driving type rule synthesis is performed.

That is, when a character string to be uttered is input to the character string buffer (1), the phoneme symbol string generation unit (2) converts the input character string into a phoneme symbol string. For example,
When a character string of "** ni **" (* is a symbol indicating an accent position) is input, "tabenik" is input.
ita ".

Next, the selection circuit 1 (3) sequentially determines and selects necessary speech units from the phoneme symbol string, and selects necessary speech units among the speech units stored in the speech unit memory (4). A piece
They are connected by a voice unit connection part (5). The unit of the speech unit is CVC (consonant + vowel + consonant) or CV
A combination of (consonant + vowel) and VC (vowel + consonant) may be used, but for simplicity, CV (consonant +
Vowels), that is, syllables,
"a, be, ni, ki, ta" are sequentially selected and connected as necessary speech units.

The connected speech segments are stored in the speech parameter buffer (6) as speech parameters, and given to the synthesis filter (11) as coefficients.

On the other hand, a symbol representing intonation such as an accent position is also input together with a character string to be uttered,
When supplied from the character string buffer (1) to the pitch pattern generator (7), the pitch pattern generator (7)
Determines the pitch (pitch) of the entire utterance. For example,
In the case of the input "ta * be ni * ta", the pitch pattern shown in FIG. 2 is obtained because accents exist in "ta" and "ki".

The residual waveform memory (8) stores residual waveforms to be used as a driving sound source corresponding to each speech unit, and is determined and selected by the selection circuit 1 (3). The necessary residual waveform is selected in accordance with the speech unit.
In the case of the example, the residual waveform is selected in the order of “ta, be, ni, ki, ta”.

Further, the residual waveform memory (8) stores a plurality of residual waveforms having different pitches created by the method proposed in the present invention for each speech unit. According to the pitch generated in (7), the selection circuit 2 (9) selects and determines a residual waveform having an appropriate pitch,
It is stored in the driving sound source generation unit (10). Lastly, the selected residual waveform is subjected to pitch change in the driving sound source generating section (10) to a value corresponding to the pitch generated by the pitch pattern generating section (7), and the residual waveform having a desired pitch is obtained. Is generated.

The residual waveform generated in this manner is input to the synthesis filter (11) as a driving sound source, and the synthesis filter (11) generates a synthesized voice. Synthesized voice is D
The signal is output from the speaker (13) via the / A converter (12).

[0033]

According to the present invention, the residual driving type speech synthesizer is capable of synthesizing parameters and residual signals regardless of a change in pitch.
Since the parameters used when extracting the numbers are the same ,
As long as the residual waveform of any pitch is used as the driving sound source, a high-quality synthesized voice close to the original voice can be obtained as long as the residual waveform is selected corresponding to the speech unit to be connected.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a residual driving rule synthesis device,

FIG. 2 is a schematic view of a pitch pattern,

FIG. 3 is a schematic diagram of a residual waveform memory,

FIG. 4 is a schematic diagram of a driving sound source generation unit,

FIG. 5 is an explanatory diagram of a conventional residual waveform creation method,

FIG. 6 is an explanatory diagram of a residual waveform creation method used in the residual drive type speech synthesizer of the present invention.

[Explanation of symbols]

 (1) ... character string buffer, (2) ... phoneme symbol string generation unit, (3) ... selection circuit 1, (4) ... speech unit memory, (5) ... speech Unit connection unit, (6) voice parameter buffer, (7) pitch pattern generation unit, (8) residual waveform memory, (9) selection circuit 2, (10)・ ・ Drive sound generator, (11) ・ Synthesis filter, (12) ・ D / A converter, (13) ・ Speaker, (101) ・ Switch 1, (102) ・ Residual waveform buffer 1, (103) Switch 2, (104) Residual waveform buffer 2, (105) Pitch change circuit

──────────────────────────────────────────────────続 き Continued on the front page (58) Fields surveyed (Int. Cl. ⁷ , DB name) G10L 11/00-21/06 JICST file (JOIS)

Claims (1)

(57) [Claims] 1. A speech unit memory storing speech feature parameters of a linear prediction system such as LPC and LSP for each speech unit, and a phoneme symbol sequence generation unit for generating a symbol sequence indicating a speech unit of a speech to be uttered. And a speech feature parameter control means comprising a speech unit connection unit for sequentially connecting speech units read from the speech unit memory based on the symbol string generated by the phoneme symbol sequence generation unit, for each speech unit Residual waveform memory storing waveform data groups of a plurality of residual signals having different pitch periods, a pitch pattern generating unit for generating a pitch pattern indicating a pitch period change of a voice to be uttered, and the residual waveform A residual selection circuit for selecting residual waveform data according to a pitch cycle at each time point determined by the pitch pattern generation unit from a residual signal waveform data group corresponding to a speech unit in a memory. A driving sound source control means provided with a residual selected by the residual selection circuit as a driving sound source, and a linear synthesizing voice using coefficients of voice parameters of the voice unit connected by the voice unit connection unit. In a residual drive type speech synthesizer comprising a speech synthesizer provided with a speech synthesis filter of a prediction system, the waveform data of a plurality of residual signals having different pitch periods to be stored in the residual waveform memory correspond to this. The speech parameter of the speech unit of the speech unit memory
A residual-drive speech synthesizer characterized in that speech is obtained by inputting speech having different pitch periods to a speech analysis filter comprising an inverse filter of the speech synthesis filter.