JPH09222898A - Regular voice synthesizer - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a regular sound synthesizer capable of implementing the sound synthesizing process via a smaller quantity of calculation with a memory device having a smaller memory capacity than the conventional example. SOLUTION: An acoustic parameter network connected with multiple arcs via nodes for each phoneme based on a sound data base and added with acoustic feature parameters corresponding to the arcs is stored in a memory 21, the connection state of at least three phonemes is expressed based on it, and a connection point distortion table indicating the connection state on each process phoneme is stored in a memory 22. A unit selection section 1 searches the arc series corresponding to each phoneme of the phoneme series in reference to the table based on the inputted phoneme series and reads out and outputs the acoustic feature parameters corresponding to the searched arc series in reference to the network. A parameter time series generation section 2 generates and outputs the time series of the acoustic feather parameters corresponding to the phoneme series, and a sound synthesis section 3 generates the synthesized sound signal corresponding to the phoneme series.

Description

Detailed Description of the Invention

[0001]

TECHNICAL FIELD The present invention relates to a regular speech synthesizer.

[0002]

2. Description of the Related Art Conventionally, in a regular voice synthesizing apparatus for connecting voice units extracted from voice data and synthesizing voices, it is possible that the voice units used have a great influence on the quality of the synthesized voice together with the prosody given during the voice synthesis. It is known and several methods for optimizing voice units have been proposed in the past. For example, Reference 1: “Nakajima et al.,“ Rule Synthesis Method for Automatically Generating Synthesis Units ”, IEICE Technical Report, SI
-87-15, 1987 "
The method attempts to represent acoustic variations of phonemes by clustering learning data based on phoneme environments. A unit selection method that minimizes the distortion of the connection part has also been proposed (see, for example, Reference 2 “Hirokawa et al.,“ Waveform Selection Method Considering Spectral Continuity in Waveform Editing Type Synthesis Method ”), Acoustical Society of Japan. Shu, 2-6-10, 1
See September 990 ”. ).

On the other hand, the present applicant has proposed a non-uniform unit connection speech synthesis method using a phoneme chain of various lengths contained in a speech database as a synthesis unit, and has shown its effectiveness. (For example, Reference 3 “Iwahashi et al.,“ Composite voice unit selection method based on acoustic scale ”, IEICE technical report, SP91-5, 1991 May 5
Month ". ). According to the method of Document 3 (hereinafter, referred to as the method of the first conventional example), in a rule-based speech synthesizer that connects voice units, the phoneme is minimized by narrowing down an appropriate candidate according to the phoneme chain in the usage environment. A method of selecting an optimum voice unit as a unit is used. In this conventional method, when a phoneme chain that matches a part of the target phoneme sequence exists in the database, it can be used as a relatively long unit, so that the acoustic features including the influence of articulatory coupling can be used. Has been confirmed to be reproducible and highly natural.

Reference 4 "Nakamine et al.," Rule Synthesis Method Based on Phonological Environment Clustering ", IEICE Transactions, D-II, Vol. J72-D-II,
N0.8, pp. 1174-1179, 1989 8
"Month" discloses a rule synthesis method by clustering based on a phoneme environment (hereinafter, referred to as a second conventional example method). In this method, a synthesis unit is automatically generated by a statistical method based on learning phonetic data to which phonological symbols and boundaries are added. Here, a cluster, which is a set of segments to which the same phonological symbols are added on the learning data, is sequentially divided into subclasses according to the preceding and following phonological environments. The phonological environment at the time of partitioning is selected so as to minimize the spatial spread of the clusters after partitioning, and as a result of such sequential partitioning processing, finally the center of gravity matrix of the clusters is synthesized. It is characterized by accumulating as a unit.

[0005]

However, in the methods of the first and second prior art examples, candidates are not narrowed down by calculation of a scale based on actual voice data in selection of voice units, which is relatively large. There is a problem that a storage device having a storage capacity and a large amount of calculation are required.

An object of the present invention is to solve the above problems, and to use a storage device having a smaller storage capacity as compared with the conventional example to perform a voice synthesis process with a smaller amount of calculation. To provide a device.

[0007]

A rule-based speech synthesizer according to a first aspect of the present invention is created in advance on the basis of a predetermined voice database in which acoustic characteristic parameters for a plurality of phoneme strings are stored in advance, and the phoneme-by-phoneme is defined for each phoneme. First storage means for pre-storing an acoustic parameter network in which a plurality of arcs are connected via nodes and acoustic feature parameters corresponding to each arc are attached, and created in advance based on the acoustic parameter network And a connection state of at least three phonemes including the preceding phoneme, the processed phoneme, and the subsequent phoneme, and the connection point distortion table representing the connection state for a plurality of processed phonemes is stored in advance in the second storage. Based on the means and the input phoneme sequence,
After referring to the connection point distortion table stored in the second storage means to search for a series of arcs corresponding to each phoneme of the phoneme string, the acoustic parameter network stored in the first storage means is searched. Based on the acoustic feature parameter data output from the unit selection means for reading out and outputting the acoustic feature parameter data corresponding to the searched arc series with reference to the input phoneme string Generating means for generating and outputting time-series data of acoustic feature parameters corresponding to, and based on the time-series data of acoustic feature parameters corresponding to the input phoneme sequence generated by the generating means, the input And a voice synthesizing unit for generating and outputting a voice signal of a synthetic voice corresponding to the phoneme sequence.

The rule speech synthesizer according to a second aspect is the rule speech synthesizer according to the first aspect, wherein the acoustic parameter network is based on the voice database and is among a plurality of phonemes constituting a phoneme. In the first storage means, the segment having the highest degree of concentration is stored in the first storage means, and it is determined whether or not the part of the other segment that has not become an arc can be shared with the part of the arc. It is characterized in that it is created in advance by replacing the part as an arc and storing it in the first storage means.

Furthermore, the rule-based speech synthesizing device according to claim 3 is the rule-based speech synthesizing device according to claim 1 or 2.
Based on the acoustic parameter network, the connection point distortion table searches for a sequence of arcs that are filled by transitioning arcs of all phonemes from the start phoneme to the end phoneme of each phoneme, and searches for the second arc. It is characterized in that it is created in advance by storing it in the storage means.

[0010]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of a regular speech synthesizer that is an embodiment according to the present invention.
In FIG. 1, the voice database memory 11 stores, for example, acoustic characteristic parameter data obtained by performing A / D conversion and LPC analysis based on a voice signal uttered by one speaker for a plurality of text data, for example. A speech database labeled by a phoneme unit, an accent phrase unit and a phrase unit is stored in advance by performing analysis using a known critical braking model analysis method and detecting a pitch frequency pattern and a model pattern. The rule speech synthesis device of this embodiment stores a long unit speech database generation unit (hereinafter, referred to as a generation unit) 10 in FIG. 1 and an acoustic parameter network generated by the generation unit 10 based on the speech database. Inside the acoustic parameter network memory 21 based on the input phoneme sequence, and the acoustic distortion network memory 21 that stores the acoustic distortion network table 21 that stores the acoustic distortion table generated by the generation unit 10 based on the acoustic parameter network. Of the acoustic parameter network and the connection distortion table in the connection distortion table memory 22 to select a phoneme sequence which is a voice unit sequence, and the acoustic characteristic parameters corresponding to the phoneme sequence are transmitted to the parameter time series generation unit 2. Having a unit selection unit 1 for outputting And it features.

In order to solve the above-mentioned problems of the conventional example, a unit connection voice synthesis system (subphonet (SUBPHON
ET) method) is used. Here, a network of the generated acoustic parameter series is called a subphonet. The basic idea of this method is as follows. (1) An acoustic parameter network is used in which a plurality of arcs are connected via a node for each phoneme, and acoustic parameters are attached to each arc. The node is defined as follows. A node is defined as a point where the spectral distance between acoustic parameter sequences is equal to or greater than a threshold value, and an arc is defined by acoustic parameter sequences separated by the nodes. Here, as the acoustic parameter, for example, a 1st to 30th order cepstrum coefficient is used, and the spectral distance is obtained by summing the square of the difference between the two cepstrum coefficients for all coefficients and the sum for all frames. It was taken. (2) These are obtained by processing all acoustic parameter sequences included in the same phoneme. The network size for each phoneme is given in advance from the outside. (3) A synthesis unit longer than a phoneme is represented by a combination of arc numbers. Therefore, the influence of articulation coupling is efficiently saved, and the storage data size is smaller than that of the conventional example.

The subphonet method used in the present invention will be described. The method described here has a high appearance frequency, and is effective for phonemes that are redundant in storing all phonemes as they are. Here, one phoneme is composed of a plurality of phonemes corresponding to each frame. The subphonet method is divided into the following 6 stages. The first two stages are the process for preparation for composition, and the rest is the actual composition process. (1-1) Generation of subphonet, that is, acoustic parameter network. (1-2) The acoustic parameter sequence of the learning word is mapped to the arc number sequence forming the subphonet. (2-1) Determination of synthesis unit. (2-2) An arc number sequence is generated using the acoustic parameter network. (2-3) Generation of acoustic parameter sequence. (2-4) Synthesis of voice waveform.

The above (2-1), (2-3), (2-4)
The same method as the method of the first conventional example can be used as it is. The remaining four steps are described below.

(1-1) Generation of Subphonet (or Acoustic Parameter Network) In order to generate a subphonet, a predetermined threshold distance DL, which is the upper limit of the allowable connection distortion, is used and selected. The portion whose spectral distance from the arc is the threshold distance DL or less is replaced with the arc. Here, Ct representing the degree of concentration or proximity of another segment to the i-th segment
(I) is used as an evaluation function. In this embodiment, the evaluation function Ct (i) of the following equation is used. The evaluation function becomes large when the two pieces are close to each other, and becomes small when the two pieces are far from each other.

[0015]

## EQU1 ## Ct (i) = exp (-sd _i ).

Here, sd _i is the Euclidean distance between two pieces. (1-1-1) A phoneme piece having the same phoneme is extracted from the database, and the length is normalized to the longest phoneme piece. (1-1-2) Ct (i) is calculated for all the pieces and C
The segment with the maximum t (i) is registered in the first arc. Next, consider a pipe of radius DL, and let the center be the arc that has just been selected. If other pieces enter or exit this pipe, split it into an inner part and an outer part. Then, the part included in the pipe is replaced with an arc. For pieces that come out of the pipe once and then enter the pipe again, replace all the spaces between them with the pipe. The remaining part outside the pipe is registered as a new segment and becomes a candidate for the selection of the next arc. This process continues until the total amount of registered arcs is above the threshold (see Figure 5). There are two ways to give the threshold distance DL: (a) fixed value or (b) value inversely proportional to power.

In the example of FIG. 5, in the stage 1, the same phoneme unit Seq. # 1, Seq. # 2, Se
q. The evaluation function value Ct (i) of the acoustic parameter # 3 is calculated, and the segment having the maximum evaluation function value Ct (i) is selected as the arc # 1. Then consider a pipe with a radius DL around arc # 1. The acoustic parameters of other segments included in the pipe are deleted, and the rest are stored as new segment candidates. Next, in stage 2, the segment Seq. # 4, Seq. # 5 is Seq. #
3 is a newly registered segment deleted by the arc # 1 selected in stage 1. Element Seq. # 1,
Seq. # 4, Seq. The evaluation function value Ct (i) of the acoustic parameter # 5 is calculated, and the segment having the maximum evaluation function value Ct (i) is selected as the arc # 2. And
Consider a pipe with a radius DL around arc # 2. A part of the acoustic parameter segments in the pipe is deleted, and the rest is stored as segment candidates. Next, in stage 3, the segment Seq. # 1 and Seq. # 4
The evaluation function value Ct (i) of the acoustic parameter of
The segment having the largest evaluation function value Ct (i) is selected as arc # 3. And around arc # 3, radius D
Consider a pipe with L. A part of the acoustic parameter segments in the pipe is deleted, and the rest is stored as segment candidates. Furthermore, in stage 4, the element Se
q. # 1, Seq. # 2, Seq. # 3 is the arc number {Arc # 1}, {Arc # 1}, and {Ar, respectively.
c # 3, Arc # 1, Arc # 2}.

(1-1-3) As described above, a part of the acoustic parameter sequence may be selected as the arc. In this case, since each arc has a backward pointer or a forward pointer, it is possible to connect to another arc when the connection distortion at the start point or the end point of the arc is the threshold distance DL or less.

(1-2) Mapping to arc number series using subphonets All words included in the learning data are converted to arc number series in order to efficiently store the influence of articulatory coupling. (1-2-1) Normalize the continuation length of the learning word to the length of the subphonet, and replace each frame with the number of the closest arc. All arc numbers used here are called a set of arc numbers for this word. (1-2-2) Approximate error is calculated for all combinations of the set of word arc numbers. The combination that minimizes this approximation error is selected as the output. Here, the transfer of the arc is restricted by the backward pointer and the forward pointer. Due to this limitation, some arc numbers cannot be combined. Therefore, the number of selectable paths can be very small. Furthermore, it is possible that there is no route from the start point to the end point of this phoneme.

(2-2) Generation of Arc Number Sequence Using Subphonet All synthesis units used here each have an arc number sequence in the subphonet. Therefore, the change is easy.

The long unit voice database generation process executed by the long unit voice database generation unit shown in FIG. 1 will be described in detail below.

FIG. 2 is a flow chart showing the long unit voice database generation processing executed by the long unit voice database generation unit of FIG. As shown in FIG. 2, first, in step S1, the voice database memory 20
The cepstrum data is cut out into phonemes based on the internal speech database. Next, in step S2, the same phonemes are collected, and the time length is normalized according to the length of the longest unit by using, for example, dynamic programming (DP method), and the normalized plurality of phonemes and each phoneme are normalized. The cepstrum data and phoneme data are stored in the phoneme data memory 23 as phoneme data. Then, in step S3, the acoustic parameter network generation process is executed, and finally, the connection distortion table generation process is executed.

FIG. 3 is a flowchart of the acoustic parameter network generation processing which is the subroutine of FIG. In FIG. 3, first, in step S11,
One phoneme is selected from the phoneme data in the phoneme data memory 23, and information on the start frame and the end frame of the phoneme is obtained. Next, in step S12, one phoneme is composed of a plurality of n units (also referred to as unit pieces), and the unit having the highest concentration degree is obtained from the n units. Then, in step S13, the unit having the highest degree of concentration is given an arc number and stored in the acoustic parameter network memory 21,
In step S14, it is checked whether the other (n-1) unit parts that have not become arcs can be shared with the arc part, and the sharable part is replaced with an arc having a new number, and the acoustic parameter is changed. Stored in the network memory 21.

When the total number of arcs stored and registered in the acoustic parameter network memory 21 in step S15 is equal to or smaller than the predetermined threshold byte number in the memory 21, the step S17 Then, another unprocessed phoneme to be examined next is selected, and after the information of the start frame and the end frame of the phoneme of the phoneme is obtained, the process returns to step S12 and the processes of step S12 and thereafter are repeated. On the other hand, in step S15, when the total number of arcs stored and registered in the acoustic parameter network memory 21 is larger than the predetermined threshold byte number in the memory 21, the total number of arcs is determined in step S16. It is determined whether or not the processing of step 1 has been executed, and if the processing is not, the processing returns to step S11 and the above processing is repeated. On the other hand, if YES in step S16, it is determined that acoustic parameter networks have been created for all phonemes and stored in the acoustic parameter network memory 21, and the process returns to the main routine. Therefore, the acoustic parameter network stored in the acoustic parameter network memory 21 is composed of a plurality of segments for each phoneme and is composed of one or a plurality of arcs, for example, as shown in stage 4 of FIG. And includes data of acoustic feature parameters such as cepstrum coefficients for each arc.

FIG. 4 is a flowchart of the connection distortion table generation process of FIG. In FIG. 4, first, in step S21, based on the acoustic parameter network in the acoustic parameter network memory 21, 1
Search for one or more root arcs. Next, in step S22, an entrance number table and an output number table are created for each route arc and stored in the connection strain table 22. Here, the components of the entrance number table are (a) the preceding phoneme, (b) the starting arc number,
(C) Spectral distance is included. For the spectral distance, set the value obtained as follows. The average of the cepstrum distances between all the data with the same preceding phoneme and all the starting arcs is calculated. Then, the starting arc number and the distance having the shortest distance are set in this entrance number table. However, the cepstrum distance is obtained only for the first frame segment and the second frame segment. The weights for setting the above elements are as follows: 1st frame segment to 2nd frame segment = 1 to 0.
5 is assumed. The constituent elements of the exit number table include (a) subsequent phonemes, (b) end arc number, and (c) spectral distance. For the spectral distance, set the value obtained as follows. The average of the cepstrum distances between all data with the same succeeding phoneme and all ending arcs is calculated.
Then, the end arc number and the distance having the shortest distance are set in this exit number table. However, the distance is calculated only for the end frame and the frame one frame before the end frame (hereinafter referred to as the end second frame). The weight when setting to the above elements is set to the segment of the end frame versus the segment of the end second frame = 1: 0.5.

Next, in step S23, the acoustic parameter network in the acoustic parameter network memory 21 generated in step S3 and step S2
2 based on the ingress number table and the egress number table stored in the connection distortion table 22, all frames from the start frame to the end frame having the normalized phoneme length are set to the acoustic parameter network. A series of arc numbers to be filled in is searched by transitioning the arcs registered in. Then, in step S24, a connection distortion table in which the arc number sequence number is uniquely determined when at least the parameters of the preceding phoneme, the processed phoneme and the subsequent phoneme are given based on the searched arc number sequence. It is created and stored in the connection strain table memory 22.

In the above embodiment, the generation unit 10
The unit selection unit 1 and the parameter sequence generation unit 2 are configured by, for example, a digital computer that is a control device.

The operation of the regular voice synthesizing apparatus shown in FIG. 1 and configured as described above will be described. Speech database memory 20, phoneme data memory 23, acoustic parameter network memory 21, connection distortion table memory 22
Is connected to the long unit voice database generation unit 10,
The acoustic parameter network memory 21 and the connection strain table memory 22 are connected to the unit selection unit 1. The phoneme sequence to be speech-synthesized is input to the unit selection unit 1, and in response thereto, the unit selection unit 1 refers to the connection distortion table in the connection distortion table memory 22 based on the input phoneme sequence. , A sequence of arc numbers corresponding to each phoneme of the above phoneme sequence, so that the maximum value of the connection distortion referenced in the connection distortion table is minimized, while searching for the optimal path, acoustic parameter network By referring to the acoustic parameter network in the memory 21, acoustic characteristic parameter data (for example, cepstrum coefficient) corresponding to the searched arc number sequence is read out and output to the parameter time series generation unit 2. In response to this, the parameter time-series generation unit 2 generates time-series data of acoustic feature parameters corresponding to the input phoneme sequence, and outputs the time-series data to the speech synthesis unit 3. Then, the voice synthesizer 3
For example, a pulse generator, a noise generator, a variable gain amplifier, and a filter are provided, and an audio signal is generated based on the time-series data of the input acoustic feature parameter and output to the speaker 5, thereby inputting the audio signal. The synthesized speech for the selected phoneme sequence is generated.

In the rule synthesizing apparatus configured as described above, the connection relation of each voice unit, which is, for example, a phoneme, is used as a connection strain table in advance.
Based on the combination of long-unit phonemes that are stored in No. 2 and have at least three longer phonemes (preceding phonemes, relevant processed phonemes, subsequent phonemes), an optimal phoneme sequence for the input phoneme sequence is generated. Decide and select. Therefore,
The optimal phoneme sequence can be determined while maintaining the long-term naturalness. As a result, the table memory 22 and the network memory 21 are not deteriorated.
It is possible to reduce the storage capacity of the above compared to the conventional example and to significantly reduce the amount of calculation related to voice synthesis. Therefore, the voice synthesis process can be executed at high speed.

[0030]

As described in detail above, according to the present invention, a plurality of arcs are preliminarily created on the basis of a predetermined speech database in which acoustic feature parameters for a plurality of phoneme strings are stored in advance, and each phoneme is transmitted through a node via a node. First storage means for storing in advance an acoustic parameter network in which acoustic feature parameters corresponding to each of the arcs are attached, and a preceding phoneme, which is created in advance based on the acoustic parameter network. A connection state of at least three phonemes including a processed phoneme and a subsequent phoneme is represented,
Second storage means for storing in advance a connection point distortion table representing the connection state for a plurality of processed phonemes, and a connection point distortion table stored in the second storage means based on an input phoneme sequence. After referring to a sequence of arcs corresponding to each phoneme of the phoneme string, the acoustic feature parameter corresponding to the sequence of arcs searched by referring to the acoustic parameter network stored in the first storage means. Based on the data of the acoustic feature parameter output from the unit selection unit and the unit selection unit for reading and outputting the data of the above, the time series data of the acoustic feature parameter corresponding to the input phoneme sequence is generated and output. Based on the time-series data of the acoustic feature parameter corresponding to the input phoneme sequence generated by the generating means, Serial generates a sound signal of the corresponding synthesized speech to the input phoneme sequence and outputs and a speech synthesis means.

Therefore, for example, a connection relation of each voice unit, which is a phoneme, is stored in advance in the connection distortion table, and a phoneme of at least three longer phonemes (preceding phoneme, the processed phoneme, and subsequent phoneme) is a long unit. Based on the phoneme combination, the optimum phoneme sequence for the input phoneme sequence is determined and selected. Therefore, it is possible to determine the optimum phoneme sequence while maintaining the long-term naturalness. As a result, it is possible to reduce the storage capacities of the first and second storage means as compared with the conventional example and significantly reduce the amount of calculation related to voice synthesis without degrading the sound quality. Therefore, the voice synthesis process can be executed at high speed.

[Brief description of drawings]

FIG. 1 is a block diagram of a regular voice synthesizing device according to an embodiment of the present invention.

FIG. 2 is a flowchart showing a long unit voice database generation process executed by a long unit voice database generation unit in FIG.

3 is a flowchart of an acoustic parameter network generation process which is a subroutine of FIG.

FIG. 4 is a flowchart of a connection distortion table generation process of FIG.

5 is a diagram showing an example of the acoustic parameter network generation process of FIG.

[Explanation of symbols]

 1 ... Unit selection unit, 2 ... Parameter time series generation unit, 3 ... Speech synthesis unit, 4 ... Speaker, 10 ... Long unit voice database generation unit, 20 ... Voice database memory, 21 ... Acoustic parameter network memory, 22 ... Connection distortion Table memory, 23 ... Phoneme data memory.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Yoshio Higuchi Kyoto Prefecture Soraku-gun Seika-cho Osamu Osamu Osamu 5 Hiratani No. 5 ATR Co., Ltd. Speech Translation Communication Research Laboratories (72) Inventor Makoto Hashimoto Kyoto Prefecture Shiraka-gun Seika-cho, Osamu Osamu, Osamu Osamu, No. 5, Mihiraya, ATR Co., Ltd.

Claims (3)

[Claims] 1. A plurality of arcs are created in advance based on a predetermined voice database in which acoustic feature parameters for a plurality of phoneme strings are stored in advance, and a plurality of arcs are connected via a node for each phoneme. First storage means for pre-storing an acoustic parameter network to which acoustic characteristic parameters are attached, and at least three pre-created phonemes including the preceding phoneme, the processed phoneme, and the subsequent phoneme, which are created in advance based on the acoustic parameter network. Second storage means that represents a connection state of phonemes and that stores a connection point distortion table that represents the connection states of a plurality of processed phonemes in advance, and the second storage means based on an input phoneme sequence. After searching the sequence of arcs corresponding to each phoneme of the phoneme string by referring to the stored connection point distortion table, Unit selection means for reading out and outputting acoustic feature parameter data corresponding to the searched arc series by referring to the acoustic parameter network stored in the storage means; and acoustic feature parameter data output from the unit selection means Based on the generating means for generating and outputting time-series data of acoustic feature parameters corresponding to the input phoneme sequence, and of the acoustic feature parameters corresponding to the input phoneme sequence generated by the generating means. A regular voice synthesizing device comprising: a voice synthesizing unit for generating and outputting a voice signal of a synthetic voice corresponding to the input phoneme sequence based on time-series data. 2. The acoustic parameter network stores, in the first storage means, a segment having a maximum concentration degree among a plurality of segments forming a phoneme as an arc based on the speech database. It is determined whether or not the part of the other segment that has not become an arc can be shared with the part of the arc, and the sharable part is replaced as an arc, and the first
The rule-based speech synthesizing apparatus according to claim 1, wherein the rule-based speech synthesizing apparatus is created in advance by storing it in the storage means. 3. The connection point distortion table searches for a sequence of arcs to be filled by transitioning arcs of all phonemes from a start phoneme to an end phoneme of each phoneme, based on the acoustic parameter network. 3. The regular voice synthesizing device according to claim 1, wherein the regular voice synthesizing device is created in advance by storing it in the second storage means.