JP3495275B2 - 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 voice synthesizing apparatus for converting input arbitrary text into voice.

[0002]

2. Description of the Related Art Speech synthesis technology, especially rule synthesis technology, is a technology for presenting text consisting of arbitrary character strings after converting it into speech. Information services by voice, reading of e-mails, reading devices for persons with disabilities, newspaper review It is used for such purposes.

The structure of a text-to-speech conversion system based on a general rule synthesizing technique is Furui "Digital Speech Processing".
(Tokyo University Press 1985), page 146. The system consists of three modules, a sentence analysis unit, a voice rule synthesis unit, and a voice synthesis unit.

The sentence analysis unit performs a morpheme analysis process for dividing a Japanese text input with reference to a dictionary into units called morphemes. Readings, accent types, parts of speech, etc. are added to the morphemes.

The voice rule synthesizing section further comprises an acoustic processing section and a prosody generation section. Based on the readings and accents obtained by the sentence analysis unit, the acoustic processing unit generates an acoustic parameter composed of a feature amount obtained by analyzing the speech waveform of LSP, mel cepstrum, or the like, or the speech waveform itself, The prosody generation unit generates prosody parameters such as pitch, pause, and duration based on the prosody rules.

The voice synthesizer generates and outputs a synthesized voice based on the acoustic parameters and the prosody parameters. When the acoustic parameter is a feature amount such as LSP, synthetic speech is generated based on an analysis and synthesis method called a vocoder, and when it is a speech waveform, PSOLA (Pitch Synchronous).
Synthesis is performed by a method called OverLap-and-Add).

The PSOLA method enables highly natural speech synthesis by directly using a waveform, but since the waveform information has a large amount of information compared to acoustic parameters such as LSP, the storage capacity becomes large. Therefore, a method of compressing and storing the waveform information by vector quantization has been studied. As such a conventional technique, JP-A-5-73100 is used.
Japanese Patent Laid-Open Publication "Voice Synthesis Method and Apparatus Thereof" (hereinafter referred to as Document 1).

FIG. 17 shows an example of the configuration of a speech synthesizer based on the method described in Document 1 (hereinafter referred to as the first conventional example). 1 is a text, 2 is a language processing unit, 3 is a phoneme string, 4 is accent information, 5 is a prosody generation unit, 6 is a pitch frequency,
11 is a speech unit database, 24 is a codebook, 1
0 is a unit reading unit, 7 is a voice unit, 16 is a codebook reference unit, 13 is a synthesis parameter, 14 is a synthesis unit, and 15 is a synthesis unit.
Is a synthetic voice.

The speech synthesis apparatus according to the first conventional example will be described below. The language processing unit 2 obtains the phoneme string 3 and the accent information 4 representing the reading from the input text 1. The phoneme reading unit 10 reads the phoneme 7 from the phoneme database 11 according to the phoneme sequence 3. The read speech unit 7 is obtained by vector quantization in advance, and a code string corresponding to a waveform series or a spectrum series is recorded.

The codebook 24 is composed of patterns (code vectors) and code sets obtained by dividing the spectrum data of the voice waveform by the existing clustering method. The code vector may be either spectral information or original waveform information. The codebook reference unit 16 refers to the codebook 24 from the code string in the preceding speech unit 7 to obtain the synthesis parameter 13 composed of spectrum information or waveform information. The prosody generation unit 5 uses the accent information 4
The pitch 6 of the synthesized voice is generated according to the rule based on

The synthesizer 14 produces a synthetic voice 15 from the synthesis parameter 13 and the pitch 6. The synthesizing unit 14 is based on a synthesizing method using a digital filter if the synthesizing parameter 13 is spectrum information, and based on the PSOLA method if it is waveform information.

[0012]

According to the first conventional technique, it is possible to obtain a synthesized voice having a high naturalness as compared with the vocoder system. However, since the same pitch waveform is repeated in the same frame in the voiced voice section, there is a problem that the naturalness of the synthesized voice is impaired. This is because the fluctuation component of the natural voice waveform is lost due to the repetition of a single waveform, and the buzzer property of the synthetic sound becomes remarkable and becomes unnatural. In order to solve such a problem, a method of separating a speech waveform into a stationary periodic component and a non-stationary component including fluctuation and controlling them separately to generate a synthesized speech waveform is also proposed. As such a conventional technique, Japanese Patent Application No. 4-
There is a technique disclosed in "Speech synthesizer" (hereinafter referred to as Document 2) described in Japanese Patent No. 358200.

FIG. 18 shows an example of the configuration of a speech synthesizer based on the method disclosed in Document 2 (hereinafter referred to as the second conventional example). 30 is a 1 pitch waveform storage unit, 31 is a pitch storage unit,
32 is a non-stationary waveform storage unit, 33 is a 1-pitch waveform, 6 is a pitch, 34 is a non-stationary waveform, 35 is a moving addition unit, 36 is a steady synthesized voice, 37 is a simple addition unit, and 15 is a synthesized voice.

A second conventional speech synthesizer will be described below. The 1-pitch waveform storage unit 30 stores the low-frequency voice waveform obtained by band-dividing the natural voice by 1 for each pitch.
It is stored as the pitch waveform 33. Unsteady waveform storage unit 32
Is the unsteady waveform 3 of the high-frequency speech waveform obtained by the previous band division.
Remember as 4. The moving addition unit 35 is the pitch storage unit 31.
The 1-pitch waveform 33 is moved and added at intervals of 6 pitches obtained from. The simple addition unit 37 adds the stationary synthesized speech 36 obtained by the moving addition unit 35 and the unsteady waveform 34 read from the unsteady waveform storage unit 32, and outputs this as the synthesized speech 15.

The second prior art aims at voice synthesis by utilizing a waveform focusing on the fluctuation component of the voice waveform. However, there is a problem that a high-quality synthesized sound cannot be obtained because the waveform processing synchronized with the pitch is performed only in the low frequency band, and the storage capacity is large because the entire waveform information is stored.

The object of the present invention is to overcome such problems and
By controlling the pitch waveform by separating it into low-frequency components and high-frequency components, it is possible to generate fluctuation components that affect the naturalness of synthesized speech, and a codebook-based method enables high-quality synthesized speech with a small storage capacity. It is to provide a speech synthesizer capable of

[0017]

A speech synthesis apparatus according to the present invention comprises a speech unit database for storing speech units,
A low-frequency component codebook, which is a vector quantization codebook of low-frequency component waveforms when a voiced sound waveform is separated into low-frequency components and high-frequency components, and high-frequency when a voiced sound waveform is separated into low-frequency components and high-frequency components. A high-frequency component codebook, which is a vector quantization codebook of component waveforms, a language processing unit that obtains phoneme strings and accent information from the input text, and a phoneme unit readout that reads out phoneme units from the phoneme unit database according to the phoneme string. Section, and a codebook reference section that selects a low-frequency component waveform and a high-frequency component waveform from each of the low-frequency component codebook and the high-frequency component codebook based on the speech segment read by the segment reading unit. An adding unit that adds a low-frequency component waveform and a high-frequency component waveform to obtain a synthesis parameter, and a pitch according to the accent information. A prosody generation unit for generating a switch frequency,
A synthesizing section for generating synthetic speech based on the synthesizing parameter and the pitch frequency is provided.

The codebook reference section selects different high frequency component waveforms for each pitch from the high frequency component codebook.

Further, when arranging the low frequency component waveform on the time axis, the adding section changes the arrangement position of the high frequency component waveform for every pitch with respect to the reference position so as to change the low frequency component waveform and the high frequency component waveform. This is a moving addition unit that adds and.

Further, the adding section calculates an average change width of the arrangement position of the high frequency component waveform with respect to the reference position of the low frequency component waveform,
It is a prosody control moving addition unit that adds a low frequency component waveform and a high frequency component waveform by changing the input pitch or power according to the input pitch or power.

The addition unit is an amplitude ratio control addition unit that changes the amplitude ratio of the low frequency component waveform and the high frequency component waveform to be added for each pitch and adds the low frequency component waveform and the high frequency component waveform. is there.

The addition section changes the amplitude ratio of the low frequency component waveform and the high frequency component waveform according to the input pitch or power, and adds the low frequency component waveform and the high frequency component waveform. It is a department.

Further, another speech synthesizer according to the present invention is a speech unit database with frequency information for storing speech units and codebook appearance frequencies, a codebook which is a voiced sound wave vector quantization codebook, A language processing unit that obtains a phoneme string and accent information from the input text, a phoneme unit reading unit that reads out the phoneme units described in the phoneme unit database with frequency information together with the appearance frequency, and a phoneme unit, A codebook reference unit that refers to a codebook according to the appearance frequency to obtain a synthesis parameter, a prosody generation unit that generates a pitch frequency according to accent information, and a synthesis that generates a synthesized voice based on the synthesis parameter and the pitch frequency. And a section.

The codebook reference section obtains the appearance ratio of the code vector from the appearance frequency described in the speech unit, and selects a plurality of codes from the highest appearance ratio among the codes described in the speech unit. A codebook reference unit that refers to a codebook, weights the waveforms obtained by reading, and adds the waveforms according to the appearance ratio to obtain a synthesis parameter.

The codebook is a low-frequency component codebook which is a vector quantization codebook of a low-frequency component waveform when a voiced sound waveform is separated into a low-frequency component and a high-frequency component. And a high-frequency component codebook, which is a vector quantization codebook of high-frequency component waveforms when separated into high-frequency components, and a speech element database with frequency information, together with speech elements, the appearance frequency and high-frequency code of low-frequency code vectors. The codebook reference unit stores the appearance frequency of the vector, and determines the appearance ratio of the low-frequency code vector and the high-frequency code vector from the appearance frequency described in the speech unit, and the codebook reference unit appears in the code described in the speech unit. Waveforms obtained by reading out the low-frequency codebook and high-frequency codebook by selecting multiple from the highest ratio It is a codebook reference section that obtains synthesis parameters by weighting shapes according to the appearance ratio and adding them.

Another speech synthesizer according to the present invention comprises a linear sum expression speech unit database in which a speech unit string is stored in a code vector linear sum representation of a codebook, and a voiced sound wave vector quantization codebook. A codebook, a language processing unit that obtains a phoneme string and accent information from the input text, a phoneme unit reading unit that reads out a phoneme unit string from the linear sum expression phoneme unit database according to the phoneme string, and a phoneme unit string. On the other hand, a linear sum is obtained from the coefficients stored in the linear sum expression speech unit database and the waveform obtained by referring to the codebook, and the pitch frequency is determined according to the codebook reference section for obtaining the synthesis parameter and the accent information. It includes a prosody generating section for generating and a synthesizing section for generating synthesized speech based on the synthesizing parameter and the pitch frequency.

Further, the codebook reference unit further has a random number generator for generating a random number, and the codebook reference unit adds the random number generated by the random number generator to the coefficient stored in the linear sum expression speech unit database to obtain the coefficient and the codebook. Is a random-number-based codebook reference unit that obtains a linear sum from the waveform obtained by referring to.

The codebook is a low-frequency component codebook which is a vector quantization codebook of a low-frequency component waveform when a voiced sound waveform is separated into a low-frequency component and a high-frequency component. And a high-frequency component codebook, which is a vector quantization codebook of high-frequency component waveforms when separated into high-frequency components. The code vector of the book is stored as a linear sum expression, and the code book reference section is a linear sum from the coefficients stored in the linear sum expression speech unit database and the waveform obtained by referring to the low frequency component code book and the high frequency component code book. The sum is calculated and the synthesis parameter is obtained.

The codebook is a low-frequency component codebook which is a vector quantization codebook of a low-frequency component waveform when a voiced sound waveform is separated into a low-frequency component and a high-frequency component. And a high-frequency component codebook, which is a vector quantization codebook of high-frequency component waveforms when separated into high-frequency components, and the random-number-based codebook reference unit uses random numbers for the coefficients stored in the linear sum expression speech unit database. Random numbers generated by the generator are added, a linear sum is obtained from coefficients and waveforms obtained by referring to the low-frequency component codebook and the high-frequency component codebook, and a synthesis parameter is obtained.

Further, the codebook is a low frequency component codebook which is a vector quantization codebook of a low frequency component waveform when the voiced sound waveform is separated into a low frequency component and a high frequency component. A high-frequency component codebook, which is a vector quantization codebook of high-frequency component waveforms when separated into a high-frequency component and a high-frequency component, and the random-number-based codebook reference unit is a high-frequency component stored in the linear sum expression speech unit database. A random number generated by a random number generator is added to the coefficient of the component code, and a linear sum is obtained from the coefficient and the waveform obtained by referring to the low frequency component codebook and the high frequency component codebook to obtain a synthesis parameter.

[0031]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiment 1. FIG. 1 is a block diagram showing a speech synthesizer of the present invention. In the figure, 1
Is text, 2 is a language processing unit, 3 is a phoneme string, 4 is accent information, 5 is a prosody generation unit, 6 is a pitch frequency, 11 is a speech unit database, 8 is a low frequency component codebook,
9 is a high frequency component code book, 10 is a segment reading unit,
Reference numeral 7 is a voice unit, 12 is a codebook reference addition unit as a codebook reference unit and an addition unit, 13 is a synthesis parameter, 14 is a synthesis unit, and 15 is a synthesized voice.

Next, the operation will be described. The language processing unit 2 outputs a phoneme sequence 3 and accent information 4 representing reading from the input text 1. The element reading unit 10
The speech unit 7 is read from the speech unit database 11 according to the phoneme string 3. The read voice unit 7 is VCV
(Vowel-consonant-vowel) or CV (consonant-vowel) units, which are obtained by vector quantization in advance, and code strings corresponding to the low-frequency component waveform series and the high-frequency component waveform series, respectively. Is recorded. A method for creating a voice unit will be described later.

The low frequency component codebook 8 is composed of a pattern (code vector) and a code set obtained by dividing the spectrum data of the low frequency component speech waveform by the existing clustering method. On the other hand, high frequency component codebook 9
Consists of a set of divided patterns (code vectors) and codes obtained by a similar method from the spectrum data of the high frequency component speech waveform. For each code vector of the low frequency component code book 8 and the high frequency component code book 9, a one-pitch-length voice waveform is used. How to create a codebook will be described later. The codebook reference adding unit 12 refers to the low-frequency component codebook 8 and the high-frequency component codebook 9 from the code string in the preceding speech unit 7, and
A low frequency component waveform and a high frequency component waveform for the pitch are extracted. The reference and selection method here will be described later.

Then, the low frequency component waveform and the high frequency component waveform are added to obtain a synthesis parameter 13 composed of waveform information for one pitch. The prosody generation unit 5 generates the pitch 6 of the synthetic speech based on the accent information 4 according to the prosody rule. The synthesizing unit 14 uses the synthesis parameter 13 and the pitch 6 to calculate PSOLA.
The synthetic speech 15 is generated based on the method.

Here, the low-frequency component codebook 8 and the high-frequency component codebook 9, and the speech unit database 1
A method of creating the first item will be described. First, the voiced section voice is cut out for one pitch length. Next, for the 1 pitch waveform,
A certain band division frequency separates a high frequency component waveform and a low frequency component waveform. For example, in the case of audio data of 8 [kHz] sampling, a low-pass filter and a high-pass filter that limit the band by 3 [kHz] are created, and the 1-pitch waveform described above is passed through the low-pass filter and the high-frequency component waveform. And separate. Then, each of the low frequency component waveform and the high frequency component waveform is collected, and the low frequency component codebook 8 and the high frequency component codebook 9 are created by the clustering method. Regarding the codebook size, the low frequency component codebook 8 and the high frequency component codebook 9 have the same value, for example, 250 each.

A method of creating the speech unit database 11 by vector quantization using these codebooks will also be described. Vector quantization is performed separately for low frequency components and high frequency components, and code sequences obtained for each phoneme sequence are described in frame units. Here, for low frequency components, vector quantization is performed only for one pitch of the representative waveform for each frame, but for high frequency components, vector quantization is performed for all pitch waveforms in the frame, and the obtained code string is obtained. Describe every frame.
In this way, one speech unit becomes as shown in FIG.

In addition, 1 in the codebook reference addition unit 12
The element selection for each pitch is performed as follows. First, since there is only one low frequency component code for each frame in the speech segment database 11, the waveform corresponding to that code is selected from the low frequency component code book 8. Next, when there are a plurality of codes for the high frequency component waveform, a code is selected for each pitch from the beginning of the frame, and the corresponding waveform is selected from the high frequency waveform codebook 9.

Referring to FIG. 2, the speech unit / a /
In the first pitch section in the first frame, the high frequency component code 201 is selected, 2 is set for the next pitch section, and 201 is returned for the next pitch section. Then, from the second frame, 102, 513 and 512 are selected in this order. The sum of the low-frequency component waveform and the high-frequency component waveform for each pitch is defined as a 1-pitch waveform. That is, this method makes it possible to generate pitch waveforms that differ only in high-frequency components even within the same frame.

According to the present embodiment, by adopting such a configuration, it is possible to separately control a high-frequency component and a low-frequency component with large fluctuation, and the fluctuation component in the pitch-synchronized waveform-using synthesis method. It becomes possible to provide a synthetic speech with higher naturalness.

Embodiment 2. FIG. 3 is a block diagram showing another example of the speech synthesizer of the present invention. The same components and data as those in the above-described first embodiment are designated by the same reference numerals, and description thereof will be omitted. The speech synthesizer according to the present embodiment has a language processing unit 2 similar to that of the first embodiment,
In addition to the prosody generation unit 5, the unit reading unit 10, the speech unit database 11, the low frequency component codebook 8, the high frequency component codebook 9, and the synthesis unit 14, a codebook reference unit 16 and a moving addition unit 19 are provided. Have a configuration.

The codebook reference section 16 refers to the low-frequency component codebook 8 and the high-frequency component codebook 9 from the code string in the input speech unit 7, and outputs the low-frequency component waveform 17 and the high-frequency component waveform in pitch units. Extract 18. The moving addition unit 19 shifts the position where the high-frequency component waveform 18 is added to the reference position of the low-frequency component waveform 17 every time and adds the result, and outputs it as the synthesis parameter 13 including one pitch waveform information. .

Next, the operation will be described. The language processing unit 2 obtains the phoneme string 3 and the accent information 4 representing the reading from the input text 1. The unit reading unit 10 stores the phonemes 7 according to the phoneme string 3 in the phoneme database 1.
Read from 1. The read speech unit 7 is composed of units such as VCV or CV, and is obtained by vector quantization in advance. Code strings corresponding to the low-frequency component waveform series and the high-frequency component waveform series are recorded. Has been done. However, one low frequency code and one high frequency code are used for each frame. That is, vector quantization is performed in frame synchronization with respect to one representative waveform in a frame.

The low frequency component codebook 8 and the high frequency component codebook 9 are constructed and created in the same manner as in the first embodiment. In the codebook reference unit 16, the speech unit 7
With reference to the low-frequency component codebook 8 and the high-frequency component codebook 9, the low-frequency component waveform 17 and the high-frequency component waveform 18 for one pitch are extracted from the code string in. The moving adder 19 adds the low-frequency component waveform 17 and the high-frequency component waveform 18 to obtain a synthesis parameter 13 consisting of waveform information. The addition method will be described later.

The prosody generation unit 5 generates the pitch 6 of the synthetic voice based on the accent information 4 according to the prosody rule. Synthesis part 1
4 generates a synthetic voice 15 from the synthesis parameter 13 and the pitch 6 based on the PSOLA method.

Here, the operations of the codebook reference unit 16 and the moving addition unit 19 will be described. First, a waveform corresponding to the low frequency component code for each frame of the speech segment database is selected from the low frequency component code book. Next, the same code is used for the high frequency component waveform, and the corresponding waveform is selected from the high frequency component codebook.

Explaining the addition method with reference to FIG. 4, the low-frequency component waveform of one pitch length is arranged at the pitch synchronization reference position at pitch intervals. This is shown in the figure by arranging the upper waveforms at pitch period (T1, T2, T3) intervals. On the other hand, the high-frequency component waveform of one pitch length is arranged in pitch synchronization at a position displaced by several samples from the previous reference position. For example, the first high-frequency component waveform of the frame is the reference position + t1, the second pitch waveform is the reference position + t2, and the third waveform is -t3. In this way, a waveform for one frame is obtained. That is, due to the difference in the arrangement of high frequency components even in the same frame by this method,
It is possible to generate a waveform fluctuation between pitch waveforms.

According to the present embodiment, by adopting such a configuration, it is possible to separately control a high-frequency component having a large fluctuation and a low-frequency component, and it is possible to further control in a pitch-synchronized waveform utilizing synthesis method. It is possible to provide synthetic speech with high naturalness.

Embodiment 3. FIG. 5 is a block diagram showing another example of the speech synthesizer of the present invention. The same components and data as those in the first and second embodiments described above are designated by the same reference numerals and the description thereof will be omitted. The speech synthesis apparatus according to the present embodiment has the same language processing unit 2, prosody generation unit 5, and segment reading unit 1 as in the first embodiment.
0, the speech unit database 11, the low frequency component codebook 8, the high frequency component codebook 9 and the synthesis unit 14, and the codebook reference unit 16 similar to the second embodiment,
It has a configuration including a prosody control moving addition unit 20.
The prosody control moving adder 20 adds the high-frequency component waveform 18 to the reference position of the low-frequency component waveform 17,
The average fluctuation width is changed according to the pitch 6 obtained by the prosody generation unit 5, and the values are shifted and added every time, and the sum is output as a synthesis parameter 13 including one pitch waveform information.

Next, the operation will be described. The language processing unit 2 obtains the phoneme string 3 and the accent information 4 representing the reading from the input text 1. The phoneme reading unit 10 uses the phoneme string 3
Then, the speech unit 7 is read from the speech unit database 11. The read voice unit 7 is VCV or CV.
Is obtained by prior vector quantization, and the code strings corresponding to each of the low frequency component waveform series and the high frequency component waveform series are recorded.
However, one low frequency code and one high frequency code are used for each frame. That is, vector quantization is performed in frame synchronization with respect to one representative waveform in a frame.

The low frequency component codebook 8 and the high frequency component codebook 9 are constructed and created in the same manner as in the first embodiment. In the codebook reference unit 16, the speech unit 7
With reference to the low-frequency component codebook 8 and the high-frequency component codebook 9, the low-frequency component waveform 17 and the high-frequency component waveform 18 for one pitch are extracted from the code string in. In the prosody control moving adder 20, the low frequency component waveform 1
7 and the high-frequency component waveform 18 are added by controlling the average change width of the added position by controlling the pitch 6 to obtain a synthesis parameter 13 consisting of 1-pitch waveform information. The addition method will be described later. The synthesizer 14 produces a synthetic voice 15 from the synthesis parameter 13 and the pitch 6 based on the PSOLA method.

Here, the operation of the prosody control moving addition section 20 will be described. First, the waveform corresponding to the low frequency component code for each frame of the speech segment database is selected from the low frequency waveform codebook 8. Next, the same code is used for the high frequency component waveform, and the corresponding waveform is selected from the high frequency component code book 9. And 1
The arrangement position of the high frequency component waveform in FIG. 4 with respect to the reference position of the low frequency component waveform of the pitch length is controlled by the pitch 6 generated by the prosody generation unit 5. For example, when the value of the pitch 6 is α [Hz], the change width between the reference position and the high frequency component waveform arrangement position is multiplied by 1 / α. This is based on the finding that the lower the pitch, the larger the waveform fluctuation.

Explaining the addition method with reference to FIG. 4, the low-frequency component waveform of one pitch length is arranged at the pitch synchronization reference position at the pitch interval. On the other hand, the high-frequency component waveform of one pitch length is arranged in pitch synchronization at a position displaced by several samples from the previous reference position. When the average change width is k, for example, the deviations t1, t2, t3 of the first, second, and third high-frequency component waveforms from the reference position are k / α, respectively.
The value is proportional to. In this way, a waveform for one pitch is obtained. That is, according to this method, it is possible to generate a waveform fluctuation between pitch waveforms due to a difference in arrangement of high frequency components even in the same frame.

According to the present embodiment, by adopting such a configuration, it is possible to control the high-frequency component and the low-frequency component with large fluctuations separately, and the fluctuation component in the pitch-synchronized waveform-using synthesis method. Is generated and the fluctuation amount of fluctuation is controlled by the pitch having a high correlation, it becomes possible to provide a synthetic speech having a higher naturalness.

Fourth Embodiment FIG. 6 is a block diagram showing another example of the speech synthesizer of the present invention. The same components and data as those in the first and second embodiments described above are designated by the same reference numerals and the description thereof will be omitted. The speech synthesis apparatus according to the present embodiment has the same language processing unit 2, prosody generation unit 5, and segment reading unit 1 as in the first embodiment.
0, the speech unit database 11, the low frequency component codebook 8, the high frequency component codebook 9 and the synthesis unit 14, and the codebook reference unit 16 similar to the second embodiment,
It has a configuration including an amplitude control waveform adder 21.
The amplitude ratio control addition unit 21 changes the addition ratio of the amplitudes of the low frequency component waveform 17 and the high frequency component waveform 18 and adds them, and outputs the result as a synthesis parameter 13 consisting of 1 pitch waveform information.

Next, the operation will be described. The language processing unit 2 obtains the phoneme string 3 and the accent information 4 representing the reading from the input text 1. The phoneme reading unit 10 uses the phoneme string 3
Then, the speech unit 7 is read from the speech unit database 11. The read voice unit 7 is VCV or CV.
Is obtained by prior vector quantization, and the code strings corresponding to each of the low frequency component waveform series and the high frequency component waveform series are recorded.
However, one low frequency code and one high frequency code are used for each frame. That is, vector quantization is performed in frame synchronization with respect to one representative waveform in a frame.

The low frequency component codebook 8 and the high frequency component codebook 9 are constructed and created in the same manner as in the first embodiment. In the codebook reference unit 16, the speech unit 7
With reference to the low-frequency component codebook 8 and the high-frequency component codebook 9, the low-frequency component waveform 17 and the high-frequency component waveform 18 for one pitch are extracted from the code string in. In the amplitude ratio control adder 21, the low frequency component waveform 1
7 and the high frequency component waveform 18 are added by changing their amplitude ratio for each pitch, and a synthesis parameter 13 consisting of 1 pitch waveform information is obtained. The synthesizer 14 produces a synthetic voice 15 from the synthesis parameter 13 and the pitch 6 based on the PSOLA method.

According to the present embodiment, by adopting such a configuration, it is possible to separately control a high-frequency component having a large fluctuation and a low-frequency component, and the high-frequency component in the pitch-synchronized waveform utilizing synthesis method. By controlling the fluctuations by changing the amplitude ratio of the waveform and the low-frequency component waveform for each pitch, it is possible to provide synthetic speech with a higher naturalness.

Embodiment 5. FIG. 7 is a block diagram showing another example of the speech synthesizer of the present invention. The same components and data as those in the first and second embodiments described above are designated by the same reference numerals and the description thereof will be omitted. The speech synthesis apparatus according to the present embodiment has the same language processing unit 2, prosody generation unit 5, and segment reading unit 1 as in the first embodiment.
0, the speech unit database 11, the low frequency component codebook 8, the high frequency component codebook 9 and the synthesis unit 14, and the codebook reference unit 16 similar to the second embodiment,
The configuration is provided with the amplitude ratio pitch control addition unit 22. The amplitude ratio pitch control addition unit 22 changes the addition ratio of the amplitudes of the low-frequency component waveform 17 and the high-frequency component waveform 18 according to the pitch 6 obtained by the prosody generation unit 5, and adds the addition ratio from the 1-pitch waveform information. Is output as the synthesis parameter 13.

Next, the operation will be described. The language processing unit 2 obtains the phoneme string 3 and the accent information 4 representing the reading from the input text 1. The phoneme reading unit 10 uses the phoneme string 3
Then, the speech unit 7 is read from the speech unit database 11. The read voice unit 7 is VCV or CV.
Is obtained by prior vector quantization, and the code strings corresponding to each of the low frequency component waveform series and the high frequency component waveform series are recorded.
However, one low frequency code and one high frequency code are used for each frame. That is, vector quantization is performed in frame synchronization with respect to one representative waveform in a frame.

It is assumed that the low frequency component codebook 8 and the high frequency component codebook 9 are constructed and created in the same manner as in the first embodiment. In the codebook reference unit 16, the speech unit 7
With reference to the low-frequency component codebook 8 and the high-frequency component codebook 9, the low-frequency component waveform 17 and the high-frequency component waveform 18 for one pitch are extracted from the code string in. The amplitude ratio pitch control adder 22 controls the amplitude ratio of the low frequency component waveform 17 and the high frequency component waveform 18 by the pitch 6, and adds them to obtain a synthesis parameter 13 consisting of one pitch waveform information. The addition method will be described later. The synthesizing unit 14 uses the synthesis parameter 13 and the pitch 6
From the PSOLA system to generate a synthetic voice 15.

The operation of the amplitude ratio pitch control adder 22 will be described. First, when the amplitude ratio of the selected high frequency component waveform of one pitch length is β (0 <β <1), the low frequency waveform is xl, and the high frequency component waveform is xh,
The weighted addition of the waveform represented by the equation (1) is performed. (1-β) · xl + β · xh (1) At this time, the relationship between the value of the pitch 6 and β is set to be inversely proportional. This is based on the finding that the lower the pitch of the voice waveform is, the larger the waveform fluctuation is. Therefore, the ratio of the high frequency component is increased. In this way, a waveform for one pitch is obtained. That is, according to this method, the waveform fluctuation can be controlled by the pitch value.

According to the present embodiment, by adopting such a configuration, it is possible to separately control a high-frequency component having a large fluctuation and a low-frequency component, and it is possible to control the amplitude ratio in the pitch-synchronized waveform-using synthesis method. It is possible to provide synthetic speech with higher naturalness by controlling the pitch with a pitch highly correlated with the fluctuation.

Sixth Embodiment FIG. 8 is a block diagram showing another example of the speech synthesizer of the present invention. The same components and data as those in the first embodiment and the conventional example 1 described above are designated by the same reference numerals and description thereof will be omitted. The speech synthesis apparatus according to the present embodiment has a language processing unit 2, a prosody generation unit 5 and a synthesis unit 14 similar to those of the first embodiment, a codebook reference unit 16 and a codebook 24 similar to those of the conventional example 1, Speech element database with frequency information 23
And a configuration including the element reading unit 10. The speech element database with frequency information 23 is composed of a codebook code described at the time of vector quantization in advance and a sequence of its appearance frequency. Element reading section 10
Determines the code string to be written in the speech unit 7 from the frequency of the code in the phoneme string 3.

Next, the operation will be described. The language processing unit 2 obtains the phoneme string 3 and the accent information 4 representing the reading from the input text 1. The phoneme reading unit 10 uses the phoneme string 3
Accordingly, the voice element database 23 with frequency information is referred to. The frequency-information-attached speech unit database 23 has a structure as shown in FIG. 9, and reads a sequence of frequency-information-attached speech units corresponding to the phoneme string 3. The frequency information is a vector quantization of a 1-pitch waveform between the entire phoneme sequence in the original speech database and the codebook,
It is realized by recording the number of times the code vector in the codebook is selected in the same phoneme sequence. In the example of FIG. 9, the frequency of appearance of the code 101 is 30, the code 100 is 15, and the code 54 is 5 in the first frame of the speech unit / a /.

Next, in the voice unit reading section 10, the code for each frame in the voice unit is uniquely determined and is output as the voice unit 7. As a determination method, the code appearance rate is obtained from the code appearance frequency for each frame, and if there are a plurality of codes whose rates exceed a threshold value, the code is changed every pitch.

The codebook reference section 16 refers to the codebook 24 from the code string in the preceding speech unit 7 to obtain the synthesis parameter 13 consisting of waveform information. Based on the accent information 4, the prosody generation unit 5 produces a pitch 6 of synthesized speech.
Is generated according to the prosody rule. The synthesizer 14 produces a synthetic voice 15 from the synthesis parameter 13 and the pitch 6 based on the PSOLA method.

According to the present embodiment, by adopting such a configuration, in the pitch-synchronized waveform-utilizing synthesis method, by preparing the speech units in consideration of the frequency of appearance of the codebook, it is possible to perform simple processing for each frame. Compared to a voice unit of one code, it is possible to synthesize a wider variety of waveforms, and it is possible to provide a synthesized voice with higher naturalness.

Embodiment 7. FIG. 10 is a block diagram showing another example of the speech synthesizer of the present invention. The same components and data as those in the first and sixth embodiments described above are designated by the same reference numerals and the description thereof will be omitted.
The speech synthesis device according to the present embodiment has a language processing unit 2, a prosody generation unit 5, and a synthesis unit 14, which are similar to those of the first embodiment.
The codebook reference unit 16, the codebook 24, and the speech segment database with frequency information 2 similar to those in the sixth embodiment.
In addition to 3, the configuration has a configuration including a unit reading weighting selection unit 25. Element reading weighting selection unit 25
Determines the weighting coefficient from the frequency of the chords in the phoneme string 3, and outputs the code string described in the speech unit 7 together with the weighting.

Next, the operation will be described. The language processing unit 2 obtains the phoneme string 3 and the accent information 4 representing the reading from the input text 1. The unit readout weighting selection unit 25 refers to the frequency unit-attached speech unit database 23 according to the phoneme sequence 3. The structure of the speech unit database with frequency information 23 is similar to that of the sixth embodiment. All the codes for each frame in the speech unit are selected in the unit reading weighting selection unit 25. Then, the appearance rate determined from the appearance frequency of each code is used as a weighting coefficient, and each code is multiplied. Then, it is output as the speech unit 7.

The codebook reference section 16 refers to the codebook 24 from the code string in the speech unit 7 above,
A plurality of pitch waveforms are multiplied by a weighting factor obtained from the appearance rate and added. A synthesis parameter 13 including the waveform information of one pitch length thus obtained is obtained. The prosody generation unit 5 generates the pitch 6 of the synthetic speech based on the accent information 4 according to the prosody rule. The synthesizer 14 produces a synthetic voice 15 from the synthesis parameter 13 and the pitch 6 based on the PSOLA method.

According to the present embodiment, by adopting such a configuration, it is possible to perform more detailed processing than in the case of a speech element having a single code for each frame by performing the element selection and weighted addition considering the appearance frequency. It is possible to synthesize various waveforms and provide more natural synthesized speech.

Embodiment 8. FIG. 11 is a block diagram showing another example of the speech synthesizer of the present invention. The same components and data as those in the above-mentioned first, sixth and seventh embodiments are designated by the same reference numerals and the description thereof will be omitted. The speech synthesizer in the present embodiment is
The language processing unit 2, the prosody generation unit 5, the low frequency component codebook 8, the high frequency component codebook 9 and the synthesis unit 14 which are the same as those in the first embodiment, and the codebook reference unit 16 and the codebook 24 which are the same as those in the sixth embodiment. In addition to the frequency-information-added speech unit database 23, a unit reading weighting selection unit 25 is provided. The unit reading weighting selection unit 25 determines a weighting coefficient from the frequency of the low frequency component code and the high frequency component code in the phoneme sequence 3, and outputs the code sequence described in the speech unit 7 together with the weighting.

Next, the operation will be described. The language processing unit 2 obtains the phoneme string 3 and the accent information 4 representing the reading from the input text 1. Element reading weighting selection unit 2
Reference numeral 5 refers to the phoneme database with frequency information 23 according to the phoneme sequence 3. The structure of the speech unit database with frequency information 23 is similar to that of the sixth embodiment. All the codes for each frame in the speech unit are selected in the unit reading weighting selection unit 25. Then, the appearance rate determined from the appearance frequency of each code is used as a weighting coefficient, and each code is multiplied. Then, it is output as the speech unit 7.

The codebook reference section 16 refers to the low-frequency component codebook 8 and the high-frequency component codebook 9 from the code string in the preceding speech unit 7, and refers to a plurality of 1-pitch-long low-frequency component waveforms and high-frequency component waveforms. Is multiplied by the weighting factor obtained from the appearance rate and added. Synthesis parameter 1 consisting of waveform information of one pitch length thus obtained
Get 3. The prosody generation unit 5 generates the pitch 6 of the synthetic speech based on the accent information 4 according to the prosody rule. Synthesizer 14
Generates a synthesized voice 15 from the synthesis parameter 13 and the pitch 6 based on the PSOLA method.

According to the present embodiment, by adopting such a configuration, it is possible to perform more detailed processing as compared with a speech element having a single code for each frame by performing the element selection and weighted addition in consideration of the appearance frequency. By separating the low-frequency component and the high-frequency component, it becomes possible to perform waveform synthesis in consideration of fluctuations, and it is possible to provide synthetic speech with a higher naturalness.

Ninth Embodiment FIG. 12 is a block diagram showing another example of the speech synthesizer of the present invention. The same components and data as those in the first embodiment and the conventional example 1 described above are designated by the same reference numerals and description thereof will be omitted. The speech synthesis device according to the present embodiment includes a language processing unit 2, a prosody generation unit 5, a segment reading unit 10 and a synthesis unit 14 similar to those of the first embodiment, and a codebook 24 similar to that of the first conventional example. It has a configuration including a linear sum expression speech unit database 26 and a codebook reference unit 16. The linear sum representation speech unit database 26 is composed of a sequence in which the codes of the codebook described at the time of the vector quantization in advance are represented by the linear sum. The codebook reference unit 16 determines the codebook 2 from the code string described in the speech unit 7.
The code vector of No. 4 is referred to, and the linear sum of the code vectors is output as the synthesis parameter 13.

Next, the operation will be described. The language processing unit 2 obtains the phoneme string 3 and the accent information 4 representing the reading from the input text 1. The phoneme reading unit 10 uses the phoneme string 3
Accordingly, the linear sum expression speech unit database 26 is referred to. The linear sum expression speech unit database 26 is shown in FIG.
The structure is as shown in (1), and the series of speech unit in the linear sum expression format corresponding to the phoneme string 3 is read out. The linear sum expression is (αX + βY) for two speech vectors X and Y in the codebook for the speech unit database.
Is performed by performing vector quantization of a one-pitch waveform that minimizes quantization distortion and recording two codes and coefficients α and β. In FIG. 15, the code corresponding to X is 101 and the code corresponding to Y is 100 in the first frame of the speech unit / a /, and the coefficients α and β are 0.
It becomes 6 and 0.4.

The codebook reference unit 16 refers to the codebook 24 from the code string in the preceding speech unit 7, reads out the two code vectors X and Y, and uses the above coefficients to call (αX + βY). A weighted 1-pitch waveform is obtained, and this is obtained as the synthesis parameter 13.
The prosody generation unit 5 generates the pitch 6 of the synthetic speech based on the accent information 4 according to the prosody rule. The synthesizer 14 produces a synthetic voice 15 from the synthesis parameter 13 and the pitch 6 based on the PSOLA method.

According to this embodiment, by adopting such a configuration, as compared with the voice unit having a single code for each frame,
It enables more detailed waveform synthesis and provides more natural synthesized speech.

Embodiment 10. FIG. 13 is a block diagram showing another example of the speech synthesizer of the present invention. The same components and data as those in the first and ninth embodiments described above are designated by the same reference numerals, and description thereof will be omitted. The speech synthesis apparatus according to the present embodiment includes a language processing unit 2, a prosody generation unit 5, a segment reading unit 10 and a synthesis unit 14 similar to those of the first embodiment, a codebook 24 similar to that of the ninth embodiment, and a linear sum. In addition to the expression voice unit database 26, a random number generator 27, a random number using codebook reference unit 29
It has a configuration including. The random number-using codebook reference unit 29 adds the random number 28 generated by the random number generator 27 to the coefficient described in the speech unit 7 to create a synthesis parameter.

Next, the operation will be described. The language processing unit 2 obtains the phoneme string 3 and the accent information 4 representing the reading from the input text 1. The phoneme reading unit 10 uses the phoneme string 3
Accordingly, the linear sum expression speech unit database 26 is referred to. The structure of the linear sum expression speech unit database 26 is similar to that of the ninth embodiment.

The random number-based codebook reference unit 29 refers to the codebook 24 from the code string in the speech unit 7 and reads out two code vectors X and Y. Then, the random number 28 generated by the random number generator 27 is used as the speech segment 7
To the new coefficients α ',
β ′ and weighted as (α′X + β′Y) 1
A pitch waveform is obtained and this is obtained as the synthesis parameter 13. The prosody generation unit 5 generates the pitch 6 of the synthetic speech based on the accent information 4 according to the prosody rule. The synthesizer 14 produces a synthetic voice 15 from the synthesis parameter 13 and the pitch 6 based on the PSOLA method.

According to the present embodiment, by adopting such a configuration, as compared with the voice unit having a single code per frame,
By enabling more detailed waveform synthesis and using random numbers, it becomes possible to provide more natural synthesized speech that incorporates the effects of fluctuations.

Eleventh Embodiment FIG. 14 is a block diagram showing another example of the speech synthesizer of the present invention. The same components and data as those in the first embodiment and the conventional example 1 described above are designated by the same reference numerals and description thereof will be omitted. The speech synthesis apparatus according to the present embodiment has the same language processing unit 2, prosody generation unit 5, and segment reading unit 1 as in the first embodiment.
0, the low frequency component codebook 8, the high frequency component codebook 9 and the synthesizing unit 14, the linear sum expression speech unit database 26 and the codebook reference unit 16 similar to those in the ninth embodiment. The speech unit database with frequency information 23 is composed of a series of linear sums of the codes of the low frequency component waveform and the high frequency component waveform described in the previous vector quantization. The codebook reference unit 16 refers to the code vectors of the low frequency waveform codebook 8 and the high frequency component codebook 9 from the code string described in the speech unit 7, and outputs the linear sum of the code vectors as the synthesis parameter 13.

Next, the operation will be described. The language processing unit 2 obtains the phoneme string 3 and the accent information 4 representing the reading from the input text 1. The phoneme reading unit 10 uses the phoneme string 3
Accordingly, the linear sum expression speech unit database 26 is referred to. The linear sum expression speech unit database 26 is shown in FIG.
The structure as shown in (3) is taken for both the low frequency component waveform and the high frequency component waveform. In the codebook reference section 16,
Referring to the low-frequency waveform codebook 8 and the high-frequency component codebook 8 from the code string in the above speech unit 7,
The low-frequency component waveform and the high-frequency component waveform of one pitch length are obtained in the weighting format of (αX + βY), and this is obtained as the synthesis parameter 13. The prosody generation unit 5 uses the accent information 4
Based on, the pitch 6 of the synthetic speech is generated by the prosody rule.
The synthesizing unit 14 calculates the PSO from the synthesis parameter 13 and the pitch 6.
The synthetic speech 15 is generated based on the LA method.

According to the present embodiment, by adopting such a configuration, as compared with the voice unit having a single code for each frame,
More detailed waveform synthesis is possible, and by separating the low-frequency component and the high-frequency component, it is possible to perform waveform synthesis in consideration of fluctuations, and it is possible to provide synthetic speech with higher naturalness.

Twelfth Embodiment FIG. 16 is a block diagram showing another example of the speech synthesizer of the present invention. The same components and data as those in the first, tenth, and eleventh embodiments described above are designated by the same reference numerals, and description thereof will be omitted. The speech synthesis apparatus according to the present embodiment has the same language processing unit 2, prosody generation unit 5, segment reading unit 10 and synthesis unit 14, low frequency component codebook 8, high frequency component codebook 9, as in the first embodiment. Linear sum representation speech unit database 26 in the ninth embodiment and random number utilizing codebook reference unit 29 and random number generator 27 in the tenth embodiment.
It has a configuration including. The random-number-based codebook reference unit 29 refers to the code vectors of the low-frequency component codebook 8 and the high-frequency component codebook 9 from the code string described in the speech unit 7, and uses the coefficients described in the speech unit 7 as the coefficient. The random number 28 generated by the random number generator 27 is added, and a linear sum is output as the synthesis parameter 13.

Next, the operation will be described. The language processing unit 2 obtains the phoneme string 3 and the accent information 4 representing the reading from the input text 1. The phoneme reading unit 10 uses the phoneme string 3
Accordingly, the linear sum expression speech unit database 26 is referred to. The linear sum expression speech unit database 26 is shown in FIG.
The structure as shown in is taken for both low frequency components and high frequency components. In the codebook reference section 29 using random numbers,
Referring to the low-frequency component codebook 8 and the high-frequency component codebook 8 from the code string in the speech unit 7,
Read out two code vectors from both codebooks.
Then, the random number 28 generated by the random number generator 27 is added to the coefficients α and β described in the speech unit 7, and a new coefficient α ′,
β ′ and 1 in the weighting format of (α′X + β′Y)
A low-frequency component waveform and a high-frequency component waveform of the pitch length are obtained, and this is obtained as the synthesis parameter 13. The prosody generation unit 5 generates the pitch 6 of the synthetic speech based on the accent information 4 according to the prosody rule. The synthesizer 14 produces a synthetic voice 15 from the synthesis parameter 13 and the pitch 6 based on the PSOLA method.

According to this embodiment, by adopting such a configuration, as compared with the voice unit having a single code for each frame,
By enabling more detailed waveform synthesis, separating low-frequency components and high-frequency components, and using random numbers, it is possible to provide more natural synthesized speech that incorporates the influence of fluctuations.

Thirteenth Embodiment The present embodiment has the same configuration as that of the twelfth embodiment. The operation of this embodiment will be described. The language processing unit 2 obtains the phoneme string 3 and the accent information 4 representing the reading from the input text 1. The unit reading unit 10 refers to the linear sum expression speech unit database 26 according to the phoneme sequence 3. The linear sum expression speech unit database 26 has a structure as shown in FIG. 15 for both low frequency components and high frequency components. The codebook reference unit 29 using random numbers refers to the low-frequency component codebook 8 and the high-frequency component codebook 8 from the code string in the preceding speech unit 7, and reads out two code vectors from both codebooks. Then, the random numbers 28 generated by the random number generator 27 are added to the coefficients α and β of the speech unit 7 for the high frequency component to obtain new coefficients α ′ and β ′, and (αX + β) for the low frequency component.
Y), for the high frequency component, the low frequency component waveform and the high frequency component waveform of one pitch length are obtained in the weighting format of (α′X + β′Y), and this is obtained as the synthesis parameter 13.
The prosody generation unit 5 generates the pitch 6 of the synthetic speech based on the accent information 4 according to the prosody rule. The synthesizer 14 produces a synthetic voice 15 from the synthesis parameter 13 and the pitch 6 based on the PSOLA method.

According to this embodiment, by adopting such a configuration, as compared with the voice unit having a single code for each frame,
By enabling more detailed waveform synthesis, separating the low-frequency component and the high-frequency component, and using the random number for the high-frequency component having a strong influence of fluctuation, it becomes possible to provide synthetic speech with a higher naturalness.

The low-frequency component codebook 8 and the high-frequency component codebook 9 or the codebook 24 in the first to thirteenth embodiments has a spectrum thereof instead of the one-pitch long speech waveform, or LSP, mel cepstrum, or the like. The spectrum parameter may be a code vector, and the synthesizing unit 14 may also perform vocoder-based speech synthesis.

The first to fifth, eighth, eleventh to thirteenth embodiments are also provided.
The low-frequency component codebook 8 and the high-frequency component codebook 9 in FIG. 2 do not fix the band limiting frequency at the time of creating the codebook, but can variably change the band-limiting frequency according to the pitch length of the original voice.

Furthermore, Embodiments 1 to 5, 8 and 11 to 1
The low-frequency component codebook 8 and the high-frequency component codebook 9 in 3 do not need to have the same codebook size, and are different in that the size of the high-frequency component codebook 9 is smaller than that of the low-frequency component codebook 8. You may comprise.

Further, in the high frequency component codebook 9 in the first to fifth embodiments, each code vector is recorded together with the representative pitch, and in the codebook reference, the high frequency component code whose representative pitch is closest to the target pitch may be selected. it can.

Further, the determination of the high frequency component code read by the segment reading unit 10 in the first to fifth embodiments can also be realized by randomly performing the same code so as not to be continuously selected.

Further, the speech unit database 11 according to the first embodiment performs vector quantization on all in-frame pitch waveforms for low frequency components at the time of creation, and describes all obtained code strings for each frame. It is also possible for the segment reading unit 10 to select the low frequency component waveform for each pitch as well as the high frequency component waveform.

Furthermore, the speech unit database 11 in the first to fifth embodiments does not describe a high frequency component code, but separately includes a matrix of phoneme types and high frequency component codes associated therewith, and The one-sided reading unit 10 searches the matrix by the input phoneme,
It is also possible to randomly determine the high-frequency component code among the phoneme-corresponding codes.

Further, the speech unit database 11 in the first to fifth embodiments does not describe a high frequency component code, but separately includes a matrix with a high frequency component code associated with a phoneme type and a representative pitch, The phoneme reading unit 10 can also search the matrix by the input phoneme and the target pitch and determine the high frequency component code.

Further, in the speech segment database 11 in the second to fifth embodiments, the description of the high frequency code can be a plurality of codes for each frame, and the codebook reference part has different high frequencies for each pitch. You can also select the waveform.

Furthermore, as a change of the amplitude ratio in the fourth embodiment, control by a random number generator can be considered.

Further, the amplitude ratio control addition unit 21 in the fourth embodiment and the amplitude ratio pitch control addition unit 22 in the fifth embodiment can also determine the addition ratio of the amplitudes according to the phoneme type of the input speech unit. is there.

Furthermore, the prosody control moving addition unit 20 of the third embodiment can be determined by the power determined by the prosody rule in the prosody generation unit 5 instead of the pitch frequency 6. That is, the average fluctuation rate of the addition position of the high frequency component waveform can be increased as the power is decreased.

Furthermore, the amplitude ratio pitch control addition unit 22 of the fifth embodiment can be determined by the power determined by the prosody rule in the prosody generation unit 5 instead of the pitch frequency 22. That is, the smaller the power, the larger the amplitude ratio of the high frequency waveform component can be configured.

Further, in the speech segment database with frequency information 23 in the sixth to eighth embodiments, when the appearance frequency of a certain code is below a certain threshold value, or the appearance frequency of a certain code is constant with respect to the total appearance frequency. Below the value,
The code can also be omitted from the description.

Furthermore, the linear sum expression speech segment database 26 in the ninth to thirteenth embodiments may include not only the sum of the two terms but also the one expressed in the linear form of the sum of three or more terms.

[0107]

The speech synthesizer according to the present invention has a speech unit database for storing speech units, and a vector quantization code for a low frequency component waveform when a voiced sound waveform is separated into a low frequency component and a high frequency component. A low-frequency component codebook that is a book, a high-frequency component codebook that is a vector quantization codebook of a high-frequency component waveform when a voiced sound waveform is separated into a low-frequency component and a high-frequency component, and a phoneme sequence from the input text and A low-frequency component codebook and a high-frequency component based on a language processing unit for obtaining accent information, a phoneme reading unit for reading a phoneme unit from a phoneme unit database according to a phoneme sequence, and a phoneme unit read out by the unit unit reading unit. A codebook reference section for selecting low-frequency component waveforms and high-frequency component waveforms from each of the component codebooks, and the selected low-frequency component waveforms. And a high-frequency component waveform to obtain a synthesis parameter, a prosody generation unit that generates a pitch frequency according to the accent information, and a synthesis unit that generates a synthetic voice based on the synthesis parameter and the pitch frequency. There is. Therefore, in a pitch-synchronized waveform-based synthesis method, by separating a conventional waveform codebook into a low-frequency component and a high-frequency component, it is possible to separately control a high-frequency component and a low-frequency component with large fluctuations, It is possible to provide a more natural synthetic voice.

The codebook reference section selects a different high frequency component waveform for each pitch from the high frequency component codebook. Therefore, in the pitch-synchronized waveform-using synthesis method, it is possible to generate a fluctuation component by selecting one pitch waveform of a high-frequency component for each pitch section, and it is possible to provide a more natural synthesized voice.

Further, when arranging the low frequency component waveform on the time axis, the adding section changes the arrangement position of the high frequency component waveform for every pitch with respect to the reference position so that the low frequency component waveform and the high frequency component waveform are arranged. This is a moving addition unit that adds and. Therefore, in a pitch-synchronized waveform-based synthesis method, it is possible to generate fluctuation components by varying the placement position of the high-frequency component waveform with respect to the low-frequency component waveform for each pitch section, and to provide a more natural synthesized speech. Is possible.

Further, the adding section calculates the average change width of the arrangement position of the high frequency component waveform with respect to the reference position of the low frequency component waveform,
It is a prosody control moving addition unit that adds a low frequency component waveform and a high frequency component waveform by changing the input pitch or power according to the input pitch or power. Therefore, in the pitch-synchronized waveform-using synthesis method, it is possible to generate a fluctuation component by changing the arrangement position of the high-frequency component waveform with respect to the low-frequency component waveform for each pitch section, and to fluctuate the fluctuation amount of the arrangement position. It is possible to provide synthetic speech with a higher degree of naturalness by controlling the pitch with a high correlation.

The adder is an amplitude ratio control adder for changing the amplitude ratio of the low frequency component waveform and the high frequency component waveform to be added for each pitch and adding the low frequency component waveform and the high frequency component waveform. is there. Therefore, in the pitch-synchronized waveform-based synthesis method, it is possible to generate fluctuations by changing the amplitude ratio of the high-frequency component and the low-frequency component for each pitch, and it is possible to provide a more natural synthesized voice. Become.

Further, the adding section changes the amplitude ratio of the low frequency component waveform and the high frequency component waveform according to the input pitch or power, and adds the low frequency component waveform and the high frequency component waveform. It is a department. for that reason,
In a pitch-synchronized waveform-based synthesis method, fluctuations can be generated by changing the amplitude ratio of high-frequency components and low-frequency components for each pitch, and the amplitude ratio is controlled by a pitch highly correlated with fluctuations. By doing so, it becomes possible to provide synthetic speech with higher naturalness.

Another speech synthesizer according to the present invention is a speech unit database with frequency information for storing speech units and codebook appearance frequencies, a codebook which is a voiced sound vector quantized codebook, A language processing unit that obtains a phoneme string and accent information from the input text, a phoneme unit reading unit that reads out the phoneme units described in the phoneme unit database with frequency information together with the appearance frequency, and a phoneme unit, A codebook reference unit that refers to a codebook according to the appearance frequency to obtain a synthesis parameter, a prosody generation unit that generates a pitch frequency according to accent information, and a synthesis that generates a synthesized voice based on the synthesis parameter and the pitch frequency. And a section. Therefore, in a pitch-synchronized waveform-based synthesis method, by preparing a speech unit in consideration of the frequency of appearance of a codebook, it is possible to synthesize more diverse waveforms than a speech unit having a single code per frame, It is possible to provide a more natural synthetic voice.

Further, the codebook reference unit obtains the appearance ratio of the code vector from the appearance frequency described in the speech unit and selects a plurality of codes from the highest appearance ratio among the codes described in the speech unit. A codebook reference unit that refers to a codebook, weights the waveforms obtained by reading, and adds the waveforms according to the appearance ratio to obtain a synthesis parameter.
Therefore, in the pitch-synchronized waveform-based synthesis method,
Prepare speech units considering the frequency of codebook appearance,
Furthermore, by performing weighted addition of 1-pitch waveforms according to the frequency of appearance, more detailed waveform synthesis is possible compared to speech units of a single code for each frame, and more natural synthesized speech can be provided. It will be possible.

The codebook is a low-frequency component codebook which is a vector quantization codebook of a low-frequency component waveform when a voiced sound waveform is separated into a low-frequency component and a high-frequency component. And a high-frequency component codebook, which is a vector quantization codebook of high-frequency component waveforms when separated into high-frequency components, and a speech element database with frequency information, together with speech elements, the appearance frequency and high-frequency code of low-frequency code vectors. The codebook reference unit stores the appearance frequency of the vector, and determines the appearance ratio of the low-frequency code vector and the high-frequency code vector from the appearance frequency described in the speech unit, and the codebook reference unit appears in the code described in the speech unit. Waveforms obtained by reading out the low-frequency codebook and high-frequency codebook by selecting multiple from the highest ratio It is a codebook reference section that obtains synthesis parameters by weighting shapes according to the appearance ratio and adding them. Therefore, in the pitch-synchronized waveform-using synthesis method, a speech unit considering the appearance frequency of the codebook is prepared for the high-frequency component and the low-frequency component, and the 1-pitch waveform is weighted and added according to the appearance frequency. This makes it possible to synthesize waveforms that are more detailed and separate the low-frequency component and high-frequency component, taking into account fluctuations, as compared to a single-code speech segment for each frame, providing a more natural synthesized speech. Is possible.

Another speech synthesis apparatus according to the present invention is a linear sum representation speech unit database in which a speech unit string is stored in a code vector linear sum representation of a codebook, and a voiced sound waveform vector quantization codebook. A codebook, a language processing unit that obtains a phoneme string and accent information from the input text, a phoneme unit reading unit that reads out a phoneme unit string from the linear sum expression phoneme unit database according to the phoneme string, and a phoneme unit string. On the other hand, a linear sum is obtained from the coefficients stored in the linear sum expression speech unit database and the waveform obtained by referring to the codebook, and the pitch frequency is determined according to the codebook reference section for obtaining the synthesis parameter and the accent information. It includes a prosody generating section for generating and a synthesizing section for generating synthesized speech based on the synthesizing parameter and the pitch frequency. Therefore, in a pitch-synchronized waveform-based synthesis method, by preparing a speech unit consisting of a linear sum expression of multiple code vectors, more detailed waveform synthesis is possible compared to a speech unit with a single code per frame. Therefore, it becomes possible to provide more natural synthetic speech.

Further, the codebook reference unit further has a random number generator for generating random numbers, and the codebook reference unit adds the random number generated by the random number generator to the coefficient stored in the linear sum expression speech unit database to obtain the coefficient and the codebook. Is a random-number-based codebook reference unit that obtains a linear sum from the waveform obtained by referring to. Therefore, in a pitch-synchronized waveform-based synthesis method, by preparing a speech unit consisting of a linear sum expression of multiple code vectors, more detailed waveform synthesis is possible compared to a speech unit with a single code per frame. By using random numbers, it is possible to provide more natural synthesized speech that incorporates the effects of fluctuations.

The codebook is a low-frequency component codebook which is a vector quantization codebook of a low-frequency component waveform when a voiced sound waveform is separated into a low-frequency component and a high-frequency component. And a high-frequency component codebook, which is a vector quantization codebook of high-frequency component waveforms when separated into high-frequency components, and the linear sum representation speech unit database is a low-frequency component codebook and a high-frequency component code for a speech unit string. The code vector of the book is stored as a linear sum expression, and the code book reference section is a linear sum from the coefficients stored in the linear sum expression speech unit database and the waveform obtained by referring to the low frequency component code book and the high frequency component code book. The sum is calculated to obtain the synthesis parameter. Therefore, in a pitch-synchronized waveform-based synthesis method, by preparing a speech unit consisting of a linear sum expression of multiple code vectors, more detailed waveform synthesis is possible compared to a speech unit with a single code per frame. By separating the low-frequency component and the high-frequency component, it is possible to perform waveform synthesis in consideration of fluctuations, and it is possible to provide synthetic speech with a higher naturalness.

The codebook is a low frequency component codebook which is a vector quantization codebook of low frequency component waveforms when a voiced sound waveform is separated into a low frequency component and a high frequency component. And a high-frequency component codebook, which is a vector quantization codebook of high-frequency component waveforms when separated into high-frequency components, and the random-number-based codebook reference unit uses random numbers for the coefficients stored in the linear sum expression speech unit database. Random numbers generated by the generator are added, a linear sum is obtained from coefficients and waveforms obtained by referring to the low-frequency component codebook and the high-frequency component codebook, and a synthesis parameter is obtained. Therefore, in a pitch-synchronized waveform-based synthesis method, by preparing a speech unit consisting of a linear sum expression of multiple code vectors, more detailed waveform synthesis is possible compared to a speech unit with a single code per frame. By separating the low-frequency component and the high-frequency component and further using random numbers, it is possible to provide a more natural synthesized speech that incorporates the influence of fluctuations.

Further, the codebook is a low frequency component codebook which is a vector quantization codebook of a low frequency component waveform when the voiced sound waveform is separated into a low frequency component and a high frequency component. A high-frequency component codebook, which is a vector quantization codebook of high-frequency component waveforms when separated into a high-frequency component and a high-frequency component, and the random-number-based codebook reference unit is a high-frequency component stored in the linear sum expression speech unit database. A random number generated by a random number generator is added to the coefficient of the component code, and a linear sum is obtained from the coefficient and the waveform obtained by referring to the low frequency component codebook and the high frequency component codebook to obtain a synthesis parameter. Therefore, in a pitch-synchronized waveform-based synthesis method, by preparing a speech unit consisting of a linear sum expression of multiple code vectors, more detailed waveform synthesis is possible compared to a speech unit with a single code per frame. Then, by separating the low-frequency component and the high-frequency component and using the random number for the high-frequency component which is strongly influenced by the fluctuation, it is possible to provide a more natural synthetic speech.

[Brief description of drawings]

FIG. 1 is a block diagram showing a speech synthesizer of the present invention.

FIG. 2 is a diagram showing an example of contents of a speech unit database.

FIG. 3 is a block diagram showing another example of the speech synthesizer of the present invention.

FIG. 4 is an explanatory diagram showing an example of an addition method of a moving addition unit and a prosody control moving addition unit.

FIG. 5 is a block diagram showing another example of the speech synthesizer of the present invention.

FIG. 6 is a block diagram showing another example of the speech synthesizer of the present invention.

FIG. 7 is a block diagram showing another example of the speech synthesizer of the present invention.

FIG. 8 is a block diagram showing another example of the speech synthesizer of the present invention.

FIG. 9 is a diagram showing an example of contents of a voice segment database with frequency information.

FIG. 10 is a block diagram showing another example of the speech synthesizer of the present invention.

FIG. 11 is a block diagram showing another example of the speech synthesizer of the present invention.

FIG. 12 is a block diagram showing another example of the speech synthesizer of the present invention.

FIG. 13 is a block diagram showing another example of the speech synthesizer of the present invention.

FIG. 14 is a block diagram showing another example of the speech synthesizer of the present invention.

FIG. 15 is a diagram showing an example of contents of a linear sum expression speech unit database.

FIG. 16 is a block diagram showing another example of the speech synthesizer of the present invention.

FIG. 17 is a block diagram showing a configuration example of a conventional speech synthesizer.

FIG. 18 is a block diagram showing another example of a conventional speech synthesizer.

[Explanation of symbols]

2 language processing unit, 5 prosody generation unit, 8 low frequency component codebook, 9 high frequency component codebook, 10 unit reading unit, 11 speech unit database, 12 codebook reference addition unit (codebook reference unit), 14 synthesis Section, 16 codebook reference section, 19 moving addition section, 2
0 prosody control moving addition unit, 21 amplitude ratio control addition unit,
22 amplitude ratio pitch control addition unit, 23 frequency information-added voice unit database, 24 codebook, 25 unit readout weighting selection unit, 26 linear sum expression voice unit database, 27 random number generator.

─────────────────────────────────────────────────── ─── Continuation of the front page (56) Reference JP-A-7-141000 (JP, A) JP-A-9-204192 (JP, A) JP-A-10-247097 (JP, A) JP-A-8- 335096 (JP, A) JP 10-143196 (JP, A) JP 5-73100 (JP, A) JP 7-20894 (JP, A) (58) Fields investigated (Int.Cl. ⁷ , DB name) G10L 13/08 G10L 13/06

Claims (13)

(57) [Claims] 1. A speech element database for storing speech elements, and a low frequency component codebook which is a vector quantization codebook of low frequency component waveforms when a voiced sound waveform is separated into low frequency components and high frequency components. , A high-frequency component codebook that is a vector quantization codebook of high-frequency component waveforms when a voiced sound waveform is separated into low-frequency components and high-frequency components, and a language processing unit that obtains phoneme strings and accent information from input text, A low-frequency component codebook and a high-frequency component code based on the phoneme reading unit that reads the phoneme from the phoneme database according to the phoneme sequence, and the phoneme read by the phoneme read unit. A codebook reference section that selects low-frequency component waveforms and high-frequency component waveforms from each of the books, and the selected low-frequency component waveforms. An addition unit that adds the split waveform and the high-frequency component waveform to obtain a synthesis parameter, a prosody generation unit that generates a pitch frequency according to the accent information, and a synthetic voice based on the synthesis parameter and the pitch frequency. A voice synthesizing device comprising: a synthesizing unit. 2. The speech synthesizer according to claim 1, wherein the codebook reference section selects different high frequency component waveforms for each pitch from the high frequency component codebook. 3. The low frequency component waveform, wherein the adding section changes the high frequency component waveform arrangement position with respect to a reference position for each pitch when arranging the low frequency component waveform on a time axis. The speech synthesizer according to claim 1 or 2, wherein the speech synthesizer is a moving addition unit that adds the high frequency component waveform and the high frequency component waveform. 4. The adding unit changes an average change width of an arrangement position of the high frequency component waveform with respect to a reference position of the low frequency component waveform according to an input pitch or power,
The speech synthesizer according to claim 1 or 2, which is a prosody control moving addition unit that adds the low-frequency component waveform and the high-frequency component waveform. 5. The amplitude for adding the low frequency component waveform and the high frequency component waveform by changing the amplitude ratio of the added low frequency component waveform and the high frequency component waveform for each pitch The speech synthesis apparatus according to claim 1 or 2, which is a ratio control addition unit. 6. An amplitude for adding the low-frequency component waveform and the high-frequency component waveform by changing the amplitude ratio of the low-frequency component waveform and the high-frequency component waveform according to the input pitch or power. The speech synthesizer according to claim 1 or 2, wherein the speech synthesizer is a ratio pitch control moving addition unit. 7. A speech unit and voiced sound vector quantization
And frequency information with speech unit database that stores the code frequency codebook with a codebook is voiced waveform vector quantization codebook, and the language processing unit to obtain a phoneme and accent information from the input text , a segment reading unit for reading the speech units that are described in the speech unit database with the frequency information along with the frequency of occurrence, with the above speech unit, by referring to the code book,
A codebook reference unit that obtains a synthesis parameter based on the appearance frequency, a prosody generation unit that generates a pitch frequency according to the accent information, and a synthesis unit that generates a synthesized voice based on the synthesis parameter and the pitch frequency. A voice synthesizer comprising: 8. The codebook reference unit obtains an appearance ratio of a code vector from the appearance frequency described in the speech unit, and selects a plurality of code vectors from the highest appearance ratio among the codes described in the speech unit. 8. A codebook reference unit for selecting individual pieces and referring to the codebook, weighting and adding each waveform obtained by reading according to an appearance ratio, and obtaining a synthesis parameter. The described speech synthesizer. 9. The codebook is a low-frequency component codebook which is a vector quantization codebook of low-frequency component waveforms when a voiced sound waveform is separated into low-frequency components and high-frequency components, and voiced sound waveforms at low-frequency components. Composed of a high-frequency component codebook, which is a vector quantization codebook of high-frequency component waveforms when separated into a high-frequency component and a high-frequency component, and the speech element database with frequency information mentioned above, together with the speech element, the frequency of appearance of low-frequency code vectors. And the appearance frequency of the high-frequency code vector is stored, and the codebook reference unit obtains the appearance ratios of the low-frequency code vector and the high-frequency code vector from the appearance frequency described in the speech unit, and describes them in the speech unit. Among the codes, select a plurality from the top of the appearance ratio and refer to the above low frequency codebook and high frequency codebook. And, read each waveform obtained by adding to a weighting according to the appearance ratio, the speech synthesizing apparatus according to claim 7, characterized in that the codebook reference unit for obtaining a synthesis parameter. 10.The speech unit string is converted to the codebook codebook.
Coutle Linear sum expression stored in linear sum expression Speech segment data
Base, Codeb, a voiced sound vector quantized codebook
And Phoneme strings and accent information from the entered text
A language processing unit to obtain, According to the phoneme string, the phoneme string string is converted into the linear sum table.
A unit reading unit that reads from the current speech unit database
When, For the speech unit string, the linear sum expression speech unit data
Refer to the coefficients and codebook stored in the database
The linear sum is obtained from the waveforms obtained by
Reference section Generates pitch frequency according to the above accent information
A prosody generation unit that Based on the synthesis parameters and the pitch frequency,
And a synthesis unit that generates a synthesized voice, Further having a random number generator for generating random numbers, The codebook reference section is for the linear sum expression speech unit
The coefficient stored in the database is
Waves obtained by adding random numbers and referring to the coefficient and codebook
A random-number-based
It is a readbook reference sectionSoundVoice synthesizer
Place 11.The speech unit string is converted to the codebook codebook.
Coutle Linear sum expression stored in linear sum expression Speech segment data
Base, Codeb, a voiced sound vector quantized codebook
And Phoneme strings and accent information from the entered text
A language processing unit to obtain, According to the phoneme string, the phoneme string string is converted into the linear sum table.
A unit reading unit that reads from the current speech unit database
When, For the speech unit string, the linear sum expression speech unit data
Refer to the coefficients and codebook stored in the database
The linear sum is obtained from the waveforms obtained by
Reference section Generates pitch frequency according to the above accent information
A prosody generation unit that Based on the synthesis parameters and the pitch frequency,
And a synthesis unit that generates a synthesized voice, The above codebook shows the voiced sound waveform as low frequency components and high frequency components.
Vector quantization coefficient of low-frequency component waveform when separated into components
Low frequency component codebook and voiced sound wave
High frequency component when the shape is separated into low frequency component and high frequency component
A high-frequency component code, which is a vector quantization codebook for waveforms.
And a book, The above linear sum expression speech unit database stores speech unit strings.
The low frequency component codebook and the high frequency component code
Dobook's code vector linear sum expression, The codebook reference section is for the linear sum expression speech unit
The coefficient and low frequency component code block stored in the database.
Waveforms obtained by referring to the clock and high frequency component codebooks
It is characterized in that a linear sum is obtained from
SoundVoice synthesizer. 12. The codebook is a low-frequency component codebook which is a vector quantization codebook of a low-frequency component waveform when a voiced sound waveform is separated into a low-frequency component and a high-frequency component. And a high-frequency component codebook, which is a vector quantization codebook of high-frequency component waveforms when separated into high-frequency components, and the random-number-use codebook reference section is stored in the linear sum expression speech unit database. adding a random number to the coefficient by the random number generator, the coefficients and the low frequency component codebook and a high frequency component codes obtained by reference to the book waveform of claim 10, wherein to obtain the synthesis parameter obtaining the linear sum Speech synthesizer. 13. The low frequency component codebook, which is a vector quantization codebook of a low frequency component waveform when the voiced sound waveform is separated into a low frequency component and a high frequency component. The high-frequency component codebook, which is a vector quantization codebook of high-frequency component waveforms when separated into low-frequency components and high-frequency components, is stored in the linear sum expression speech unit database. The random number generated by the random number generator is added to the coefficient of the high frequency component code, and a linear sum is obtained from the coefficient and the waveform obtained by referring to the low frequency component codebook and the high frequency component codebook to obtain a synthesis parameter. The speech synthesizer according to claim 10 .