JPH0792997A - Speech synthesizing device - Google Patents

Abstract

PURPOSE:To provide the speech synthesizing device which facilitates the addition, deletion, etc., of selection conditions when a waveform element piece is selected and flexibly copes with them. CONSTITUTION:A preprocessing part 12 divides an input character string into phoneme units. A selection reference parameter setting part 13 sets selection reference parameters, used to select waveform element pieces as synthesis parameters, on the basis of the phoneme units. An element piece selection part 14 calculates squaring errors between the set selection reference parameters and element piece parameters inputted from an element piece parameter table 15, selects the element piece parameters in the increasing order of the squaring errors to generate primary candidates, and determines the element piece corresponding to the element piece parameter which meets element piece environment adaption conditions set in a condition setting part 141 most as an optimum element piece for the phoneme unit. An element piece connection part 17 extracts determined optimum element pieces of the phoneme units from an element piece file 16 and connects them in the phoneme units to generate a synthesized speech.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizing technique, and more particularly to a speech synthesizing apparatus for generating synthetic speech from an input character string by using a waveform segment extracted from natural speech as a synthesizing parameter.

[0002]

2. Description of the Related Art FIG. 6 shows a configuration example of a conventional general speech synthesizer. In FIG. 6, 61 is an input terminal, 62 is a preprocessing unit, 63 is a selection reference parameter setting unit, 64 is a segment selection unit, 65 is a waveform dictionary, 66 is a waveform file, 67 is a segment transformation unit, and 68 is a segment. One-sided connecting portion, 69 is an output terminal.

In the speech synthesizer having this configuration, an input character string consisting of phoneme symbols and accent symbols is input to the input terminal 61 and then divided into phoneme units in the preprocessing unit 62. The selection reference parameter setting unit 63 stores a selection reference parameter (average pitch frequency / pitch inclination / time length / average power (RMS value)) and its waveform segment from the phoneme unit and the accent symbol. Set the fragment file. The selection reference parameter is used as a selection reference for the waveform element which is a synthesis parameter. The segment selection unit 64 uses the following evaluation function formula to select the optimal segment for phonological concatenation.

[0004]

## EQU1 ## P = α × (n'-n) / σn + (1-α) × (W'-W) / σw W = ωv × │ (Vt-Vi) / σv│ ² + ωf × │ (Ft- Fi) / σf│ ² + ωt × │ (Tt-Ti) / σt│ ² + ωa × │ (At-Ai) / σa│ ² n = 1 / e ^N ... (1)

In the above equation (1), α is a balance coefficient relating to the phonological environment and prosody characteristics, n is a phonological environment coefficient, and W is a prosody characteristic coefficient. Further, n ′ and W ′ are the average values of n and W, N is the maximum number of matching phonemes, V is the average pitch frequency (Hz), F is the pitch slope, T is the duration, and A is the average power value (RMS). Value), σn, σw, σv, σf, σ
t and σa are variances of the average pitch frequency and the like with respect to the prosodic parameter of the segment, ωv, ωf, ωt, and ωa are weighting factors related to the average pitch frequency, t is a subscript representing the selection reference parameter, and i is the prosodic parameter of the segment. Is a subscript that represents.

The waveform file 66 stores sentence voice data (waveform information) such as novels and essays as file numbers are assigned to each sentence number as they are, and does not take the form of a set of pieces. Further, the segment transforming unit 67 reads the optimum segment selected by the segment selecting unit 64 from the waveform file 66, and performs a predetermined transforming process for each of the selection criterion parameters so as to match the selection criterion parameter. I am giving it.

In the speech synthesizer having such a configuration, the coefficients α and ω in the evaluation function formula (formula (1)) used for selecting the phoneme are empirically set variables and are not constant values. Further, each coefficient is a numerical value tuned only for a specific speaker (hereinafter, the first speaker) used for creating the waveform file. Therefore, when a new waveform file 66 is created by using a third party (hereinafter, second speaker) different from the first speaker, the above coefficients are retuned for the second speaker. There was a drawback that had to be made. In the case of pitch transformation processing, the transformation processing in the segment transformation unit 67 is performed by setting a pitch change ratio between the selection reference parameter and the optimum segment selected by the segment selection unit 64. However, when the pitch is changed, if this change ratio is large, the sound quality of the original optimum segment is deteriorated. Therefore, there is also a drawback that a high-level transformation process is required for the segment transformation process. This is also related to the file structure in which the optimum segment including the corresponding phonological concatenation does not exist in the waveform file 66.

Therefore, the present inventors have previously proposed a speech synthesizer capable of eliminating the above-mentioned drawbacks (see Japanese Patent Application No. 5-59385).
Proposed. This speech synthesizer is, in short, a particular prosody parameter of a waveform segment containing phonological concatenation decomposed from an input character string that corresponds to the one with the smallest error (for example, squared error) from the selection reference parameter. By selecting a waveform segment and sequentially connecting the selected waveform segments for each phonological concatenation to generate a synthetic voice, it is possible to cope with a change in speaker data.

[0009]

According to the speech synthesizer according to the above-mentioned proposal, even if the second speaker data is different from the first speaker data, the evaluation function expression (the above expression (1) is used). ) It becomes unnecessary to retune each coefficient. Moreover, in this speech synthesizer, the influence of the prosody parameter of the vowel is well taken into consideration, and a fairly natural synthesized speech is obtained. However, as a result of the multifaceted verification thereafter, it was found that a more natural synthesized speech can be obtained by considering the change of the phonological environment and other conditions when selecting the waveform segment. It is difficult to select a waveform segment by adding such a new condition with the configuration of the speech synthesizer described above, and in particular, a condition that is difficult to be reflected in the selection reference parameter such as pitch inclination code is added. I couldn't. The present invention has been made under the above background, and an object of the present invention is to provide a speech synthesizing device capable of more flexibly responding to a selection condition of a waveform segment.

[0010]

A speech synthesizer according to the present invention which achieves the above object, comprises a preprocessing unit for decomposing an input character string corresponding to a speech into phoneme units, a plurality of waveform segments and each waveform segment. Waveform information storage means for storing the prosody parameters of the, and the prosody parameters of the waveform segment corresponding to the decomposed individual phonemes are extracted from the waveform information storage means, and among the extracted prosody parameters, preset parameters are set. Segment selection means for selecting a waveform segment corresponding to a segment that satisfies the environmental suitability condition and has a minimum error from the selection reference parameter set as the selection criterion of the waveform segment, and the selected waveform. And a segment connecting unit for connecting the segments in the order of the input character string to generate a synthesized voice.

In the speech synthesizer having the above structure, specifically, the waveform information storage means cuts out waveform units of a plurality of phonological environments including accent information for the phonological unit from natural speech and stores them individually. An element storage means,
It consists of a segment parameter table that stores prosody parameters corresponding to each waveform segment, and the error is, for example, a value obtained by dividing the difference between the selection reference parameter and each prosody parameter by the variation width of each prosody parameter. The sum of squares of

Another structure of the present invention that achieves the above object is characterized in that the element selection means includes condition changing means for adding, changing, or deleting the element environment suitability condition. It should be noted that the segment environment suitability condition can be set arbitrarily, for example, the degree of agreement between the phonological environment before and after the phonological unit and the phonological environment of the waveform segment corresponding to each prosodic parameter in the selection candidate group. Alternatively, it is preferable to use the consistency of the sign of the pitch slope in the phoneme unit and each prosody parameter in the selection candidate group.

[0013]

In the speech synthesizer of the present invention, the prosody parameter of the waveform segment corresponding to each phoneme decomposed by the preprocessor is extracted from the waveform information storage means. In this case, a plurality of prosody parameters are extracted for each phoneme unit. In the segment connecting section, of the extracted prosody parameters, those having the smallest error (for example, squared error) from the selection reference parameter are arranged in ascending order to form a selection candidate group (hereinafter, primary candidate) in phoneme units. create. Then, from the selection candidate group, the waveform element corresponding to the element element environmental suitability condition set in advance and having the minimum error is selected.
Since the selected waveform segment is the optimum waveform segment of the phoneme unit, it is connected in the order of the input character string to obtain the synthesized speech.

In this way, by selecting the prosodic parameters in ascending order of the error from the selection reference parameter, the waveform segment corresponding to each prosodic parameter is not constrained by the vocal sound of the specific speaker and is preset. By selecting the one with the smallest error among the above-mentioned conditions for suiting the segment environment, a waveform segment considering various segment environments can be obtained. In particular, it is difficult to evaluate the degree of coincidence of phonological environments and the consistency of pitch tendency as errors with the above-mentioned selection reference parameters, but by adopting these as the conditions for segment environment suitability It can be evaluated and the synthesized speech is more like a natural speech.

[0015]

Embodiments of the present invention will now be described in detail with reference to the drawings. FIG. 1 is a block diagram of a speech synthesizer according to an embodiment of the present invention, in which 11 is an input terminal, 12 is a preprocessing unit, 13 is a selection reference parameter setting unit, 14 is a segment selection unit, and 141 is a condition. A setting unit, 15 is a segment parameter table, 16 is a segment file, 17 is a segment connection unit, and 18 is an output terminal. The condition setting unit 141 sets various element environment suitability conditions used in the element selection process in the element selection unit 14, and is configured to freely add, change, or delete these setting conditions.

FIG. 2 is a conceptual configuration diagram of the segment parameter table 15 and the segment file 16. The segment file 16 is for extracting the corresponding waveform segment from the natural speech according to the presence or absence of accent on the phoneme unit and the central phoneme of each phoneme unit and storing it individually. The parameter 16 and the prosody parameter of the waveform segment belonging to the segment parameter 16 (hereinafter, segment parameter) are stored.

A plurality of sets of the phoneme parameter table 15 and the phoneme file 16 are prepared for each waveform phoneme in the cut out natural voice, and as shown in FIG. 2, a phoneme unit 22 and an accent symbol indicating the presence or absence of an accent. Specific items are set according to the number 23. Here, the phoneme unit 22 is, in addition to the usual phonemes of / a / and / k /, in consideration of the smoothness of phoneme connection, continuous vowels such as / ai / and / a
It includes compound phonemes such as N /.

Next, the operation of this embodiment will be concretely described by taking a voice synthesis process for a character string input of "# karewa #" as an example. It is assumed that the input character string includes the phoneme symbol and the accent symbol 23. First, the character string (#
karewawa #) in the preprocessing unit 12 to / ka /, / re
It is decomposed in phoneme units such as / and / wa /. In the following description, for convenience, / ka / is defined as a first phoneme unit, / re / is defined as a second phoneme unit, and / wa / is defined as a third phoneme unit.
Note that # is an additional symbol indicating that the phoneme unit is at the beginning / end of the word. Therefore, # karewa # (he) is divided into phoneme units of # ka-, re-, and wa #-, respectively. Also, # ka- is a natural voice (word /
It means that the phoneme at the beginning of (sentence) is / or / and that no accent is given. When accented, the phoneme expression is # ka +. This also applies to other phonemes.

In the selection reference parameter setting section 13, a selection reference parameter used for selecting a waveform segment as a synthesis parameter and a segment file 16 for storing the waveform segment are selected from each phoneme unit and the accent symbol 23. Set.
This process is performed as follows.

First, the average pitch frequency Vt, pitch slope Ft, time length Tt, and average power (RMS value) At are determined using the first phoneme unit and the accent symbol. The parameter having these elements is the first selection criterion parameter. Similarly, for each of the second, third, ... Phoneme units, the corresponding second, third, ... Selection reference parameters are obtained.
Next, the segment files corresponding to the first selection criterion parameters are the first selection file, and the segment files corresponding to the second, third ... Set as.

These set selection files show the waveform segment corresponding to each designated phoneme unit and its characteristics. For example, it corresponds to a phoneme unit of # ka- (phoneme / or / at the beginning of a natural voice (word / sentence), which means that no accent is given, and # ka + is accented). The waveform element is #k
It can be cut out from the waveform of a word such as akaeru #, # kakageru #, etc., and the average pitch frequency Vt, pitch slope Ft, duration T can be extracted depending on the word to be extracted.
t etc. are different.

The segment parameter table 15 identifies each extracted waveform segment by assigning a file number and identifies the extracted word as a segment extraction environment. One-side parameter (average pitch frequency Vi, pitch inclination Fi, time length Ti, average power (RMS
The value) Ai) is stored. Furthermore, the average value, variance value, maximum value, and minimum value of each segment parameter are also stored. FIG. 3 shows an explanatory diagram of the contents.

FIG. 3 shows an example of contents relating to the first phoneme unit / ka /. 31 is a table name, 32 is a segment extraction environment, 33 is an average pitch frequency (HZ) Vi, 34 is a pitch slope Fi, 35 is duration Ti, 36 is average power value A
i and 37 are maximum values of the respective parameters 33 to 36, 38 is a minimum value of the respective parameters, and 39 is a file number which is symmetrical in selection in the segment selecting unit 14. Average pitch frequency 3
3, pitch inclination 34, time length 35, average power (RMS
A fragment parameter is configured with (value) 36 as an element.

The table name 31 functions as an index of the segment extraction environment 32, for example. In the example shown,
Indexes for expressing # kakaeru # to # kagami # are described. Also, each parameter 3
Numerical values such as 3 to 36 and the maximum value 37 are values analyzed at the time of waveform cutting and are fixed numerical values.

Returning to FIG. 1, the segment selection unit 14 calculates a squared error between each selection criterion parameter and the segment parameter belonging to the selected file set corresponding to these selection criterion parameters. , A plurality of segment parameters are selected in ascending order of the error value, and these are used as primary candidates for the first phoneme unit. Similarly, primary candidates are selected for the second, third, ... Phoneme units. Furthermore, the condition setting unit 1
With reference to the setting conditions from No. 41, the phoneme environment, the average pitch, the code of the parameter having the highest characteristic likelihood such as the code of the pitch inclination is selected from each of the primary candidates, and the first, second, ... The final optimum segment for each phoneme unit is determined.
This principle will be described with reference to FIG.

FIG. 4 is a flow chart showing a detailed processing procedure of the segment selecting section 14, and S is a step number of each processing. Referring to FIG. 4, the segment selection unit 14 reads the first selection criterion parameter set by the selection criterion parameter setting unit 13 (S41), and the segment parameter table 1
5, the corresponding segment parameters (average pitch frequency 33, pitch slope 34, duration 35, average power 36 in FIG. 3), and their maximum value 37 and minimum value 38 are read (S42). Then, the difference between the maximum value 37 and the minimum value 38 of each parameter is calculated, and this is output to the subsequent stage as the parameter fluctuation range (S43). Next, the difference between the first selection criterion parameter and the segment parameter in the first selection file is divided by the fluctuation range of the parameter, and the sum of the squared values (square error) of each parameter is segmented. Is calculated (S44). This squared error is taken as a selection extraction scale SC (Selection Cr
iteria). This calculation process is expressed by a mathematical formula as shown in the following formula (2).

[0027]

## EQU2 ## SC = (Vt-Vi / wideV) ² + (Ft-Fi / wideF) ² + (Tt-Ti / wideT) ² + (At-Ai / wideA) ² ... (2) (2 ≦ i ≦ n)

Here, variable names wideV, wideF, wideT,
wideA is the average pitch frequency V obtained in S43,
It is a parameter variation width of pitch inclination F, duration T, and average power value A, and n represents the number of segment parameters belonging to the selected table. When the central phoneme of the phoneme concatenation is the unvoiced vowel, the term regarding the pitch component of the square error shown in the above equation (2) is set to “0” and calculated by the following equation.

[0029]

[Equation 3] SC = (Tt-Ti / wideT) ² + (At-Ai / wideA) ² ... (3) (2≤i≤n)

After the squared error SC is calculated for each of the segment parameters in the first selected file as described above, m (0≤m≤n) waveform segments are arranged in ascending order of SC value. Is selected as the primary candidate of the waveform segment corresponding to the first phoneme unit (S45). Note that the value of m may be set arbitrarily and may always be n = m, but if the number n of waveform segments is too large, a large amount of calculation is required when selecting the optimum waveform segment. Therefore, it is preferable to set m to an appropriate value in advance. Similarly, the operations of S41 to 45 are performed for each of the second, third, ... Phoneme units to select primary candidates for all phoneme units (S4).
6).

Next, for the primary candidate for the first phoneme unit, a phoneme suitability check based on the following phoneme environment suitability conditions read from the condition setting unit 141 is performed. In this check, specifically, the waveform segment that satisfies all of these conditions and has the smallest SC value is determined as the optimal segment for the first phoneme unit.

Element environment suitability condition (1): The phoneme environment in which the phoneme is extracted matches the phoneme environment at the time of synthesis or the articulation method is similar. Element environment suitability condition (2): The magnitude relationship of the average pitch matches the magnitude relationship of the selection criterion parameter. Fragment environmental suitability condition (3): Sign of pitch inclination (positive / negative /
0) matches the sign of the selection criterion parameter. The above conditions are also applied to the primary candidates for the second, third, ... Phoneme units to determine optimum segments (S47),
The file number of each determined optimum segment is output (S48).

Here, the processing in the fragment suitability check (S47) will be described in more detail. Figure 5 shows the input string #
It is a processing content figure at the time of finishing the primary candidate selection processing (S45) for each phoneme unit for karewa #. As shown in FIG. 5, the waveform segment of each phoneme unit is S
The C values are arranged in ascending order. In the process of determining the optimum segment, the segment with the smallest SC value (file number 076
(Corresponding to each unit of 0, 0050, 0391)).

For example, in the case of #ka which is the first phoneme unit, the SC value becomes the smallest at the file number 076.
It is a waveform element of 0. This corrugated piece satisfies the element environment suitability condition (2) and the element piece environment suitability condition (3) described above, but does not satisfy the element piece environment suitability condition (1).
That is, the phoneme environment following #ka is / r /, but the phoneme environment following in the waveform segment of file number 0760 is / k /, and the phoneme environments do not match. Therefore, the waveform segment of file number 0746, which has a segment extraction environment that completely matches the phoneme environment at the time of synthesis and has the smallest SC value, is determined as the optimum segment for the first phoneme unit.

When there is no waveform segment having a segment extraction environment that matches the phonological environment / r / at the time of synthesis, / w / and / j /, which have a similar articulation method to / r /, are used as the extraction environment. The corrugated segment that has is selected as the optimal segment. In this way, when there is no waveform segment satisfying the above three conditions, a waveform segment having a high degree of satisfaction with the conditions is appropriately selected.

With respect to the second phoneme unit re- and the third phoneme unit wa #-, the waveform segment having the smallest SC value satisfies all of the above three conditions. It is decided as it is as the optimum fragment. Therefore, the file number 005 is set as the optimum segment corresponding to the second phoneme unit.
The waveform segment of file number 0391 is determined as the waveform segment of 0 and the optimal segment corresponding to the third phoneme unit.

In the unit segment connection unit 17, the file number of each optimum unit output from the unit selection unit 14 is used as a search key to read out the corresponding optimum unit from the unit file 16 and the optimum unit units are read. A synthetic voice is generated by sequentially combining the.

Through the above-mentioned processing, the output terminal 18 outputs the synthesized voice corresponding to the input character string # carewa #.

In the speech synthesizing device of this embodiment, since the speech synthesis unit is a segment in which the phonological environment of natural speech and accent information are individually considered, the segment transformation unit 5 shown in FIG.
7 becomes unnecessary. Further, the optimum segment is extracted individually from the plurality of segments stored in the segment file 16 so that the above-mentioned squared error has the minimum value. The same waveform is obtained, and tuning for each speaker is unnecessary. Furthermore, after selecting a plurality of waveform segments having a small error from the selection criterion parameter, the phoneme environment, which has a particular influence on the evaluation of the synthetic speech, the magnitude relationship of the average pitch, and the high suitability for the pitch slope code, and Since the waveform element with a small error is the optimum waveform element,
A more natural synthesized voice can be obtained. in this way,
According to the present embodiment, it is possible to realize a speech synthesizer capable of more flexibly handling the selection condition of the waveform segment.

Although the present embodiment is as described above, the present invention is not limited to the constitution of this embodiment, and the constitution can be arbitrarily changed without departing from the gist thereof.
For example, in this embodiment, as the selection scale SC of the optimum segment, the form satisfying the least square error with the reference value has been described, but this is an example suitable in consideration of the direction of each error. , Is not necessarily bound to this configuration. As another configuration, the sum of those obtained by dividing each parameter by the variation width or the sum of the higher-order multiplier errors than the square error may be selected to be the smallest.
Further, other elements may be used as the elements constituting the selection criterion parameter and the segment parameter.

[0041]

As described above, according to the speech synthesizer of the present invention, among the prosodic parameters of the phoneme unit extracted on the basis of the input character string, a predetermined segment environment suitability condition is selected and selected. Since the waveform segment corresponding to the one with the smallest error from the reference parameter is selected and connected in the order of the character strings, substantially the same voice waveform can be obtained even if the speaker data changes, and this voice waveform can be varied. It has the effect of adding the elemental environmental suitability conditions of. Therefore, it is possible to provide a speech synthesizing device that can more flexibly deal with the selection condition when selecting a waveform segment. Also,
Since this element environment suitability condition can be arbitrarily added by the element selection means, it becomes extremely easy to bring the synthesized speech closer to a natural speech, which is much easier than the conventional speech synthesizer of this type. The performance is improved.

[Brief description of drawings]

FIG. 1 is a block diagram of a speech synthesizer according to an embodiment of the present invention.

FIG. 2 is a conceptual configuration diagram of a segment file and a segment parameter table.

FIG. 3 is an explanatory diagram of contents of a segment parameter table.

FIG. 4 is a flowchart showing a process in a segment selection unit.

FIG. 5 is an explanatory diagram of an optimum segment determination process for an input character string # karewa #.

FIG. 6 is a block diagram of a speech synthesizer according to a conventional example.

[Explanation of symbols]

 11, 61 Input terminal 12, 62 Pre-processing unit 13, 63 Selection reference parameter setting unit 14, 64 Element selection unit (element selection means) 141 Condition setting section (element selection means, condition change means) 15 Element parameter Table (waveform information storage means) 16 Element file (waveform information storage means) 17 Element connection section 18 Output terminal

Claims (4)

[Claims] 1. A preprocessing unit for decomposing an input character string corresponding to a voice into phonological units, waveform information storage means for storing a plurality of waveform segments and prosodic parameters of each waveform segment, and the decomposing unit. While extracting the prosody parameter of the waveform segment corresponding to the unit phoneme from the waveform information storage means,
Of the extracted prosody parameters, a segment that selects a waveform segment corresponding to one that satisfies a preset segment environment suitability condition and has a minimum error from the selection criterion parameter set as the selection criterion of the waveform segment. A voice synthesizing apparatus comprising: a selection unit; and a segment connection unit that connects the selected waveform segments in the order of the input character strings to generate a synthesized voice. 2. The waveform information storage means includes a segment storage means for extracting waveform segments of a plurality of phonological environments including accent information for each phonological unit from natural speech and individually storing the segment, and each waveform segment. The speech synthesis apparatus according to claim 1, further comprising: a segment parameter table storing corresponding prosody parameters. 3. The error is a sum of squares of a value obtained by dividing a difference between the selection reference parameter and each prosody parameter by a variation width of each prosody parameter. The described speech synthesizer. 4. The speech synthesis apparatus according to claim 1, wherein the element selection means includes condition changing means for adding, changing, or deleting the element environment suitability condition. A speech synthesizer characterized by.