JPH09230893A - Regular speech synthesis method and device therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To synthesize speech similar to natural speech by enhancing a matching degree of phoneme environment, etc., in synthesizing in a speech synthesis device selecting and connecting an optimal element piece belonging to a synthesis unit. SOLUTION: A speech synthesis device is configured comprising a pre- processing part 12 to break down an input character string into phoneme units, a plurality of waveform element pieces cut out of natural speech in the units of utterance and a plurality of single syllables, storing means 192 and 193 to store prosodic parameters of each single syllable, an element piece selection part 14, and element piece connecting means 15, 16, and 17 to generate synthesized speech by connecting the selected waveform element pieces in order of the input character string. The element piece selection part 14 selects optimal element piece corresponding to phoneme units to the metrical environment, referred to both of the extracted environment and the prosodic parameter.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a ruled speech synthesis technique for generating synthetic speech from an arbitrary input character string, and more particularly to a method for selecting a waveform segment as a synthesis parameter and a method for transforming a selected waveform segment. .

[0002]

2. Description of the Related Art Speech synthesis technology is widely used, for example, for announcements at stations and for reading sentences by machines. In speech synthesis, it is required that the synthesized speech be clear and comfortable.

FIG. 9 shows an example of the configuration of functional blocks of a conventional general speech synthesizer. In the speech synthesizer having the configuration of FIG. 9, an input character string is input from the input terminal 91 and then divided into phoneme units by the preprocessing unit 92 with reference to the Japanese dictionary 981. The prosody setting unit 93 sets a prosody parameter for each phoneme. The segment selection unit 94 uses the segment dictionary 9
With reference to 82, the waveform segment closest to the reference value of the set prosody parameter is selected. The segment transforming unit 95 unconditionally transforms each segment from the segment file 983 storing voice data of sentences and word texts so that the selected segment matches the reference value. To do.
The segment connecting unit 96 connects the deformed segments in the order of the input character strings, and outputs them through the output terminal 97. As a result, a synthetic voice corresponding to the input character string is obtained.

[0004]

In the conventional speech synthesizer described above, the element selection unit 94 does not consider factors such as the phonological environment at the time of synthesis, and the selection result is
It strongly depends only on the reference value of the set prosody parameter. Therefore, when selecting a waveform segment corresponding to a phoneme unit close to the end of a word, there is a problem that a waveform segment near the beginning of a word occurs, and synthetic speech is often unnatural. Further, conventionally, the segment transforming unit 95 performs the transforming process so as to match the previously selected reference value regardless of the selection result of the waveform segment, so that the quality of the synthesized voice is deteriorated. There were things.

An object of the present invention is to provide a regular speech synthesizing method capable of selecting an optimum waveform segment to prevent deterioration of naturalness during synthesis. Another object of the present invention is to provide a speech synthesizing device having a configuration that suppresses quality deterioration due to excessive deformation of waveform segments and that can obtain synthetic speech with higher naturalness than conventional devices.

[0006]

According to a method for synthesizing a rule of speech of the present invention which solves the above-mentioned problems, a segment selection process of selecting an optimal segment belonging to a phoneme unit and a selected optimal segment are connected in a predetermined order. In the method having a step of generating a synthetic sound, the element selection step, a step of determining whether or not the subsequent phoneme of one waveform element as a synthesis parameter matches the phonological environment at the time of synthesis, Selecting at least one optimal segment candidate using an extraction error representing a difference between the utterance environment of the synthesis target vocabulary and the extraction environment of the waveform segment with respect to the waveform segment determined to be the subsequent phoneme match; A step of determining an optimum segment based on a selection error representing a difference between the prosody information of the selected optimal segment candidate and a predetermined selection reference value.

As a more specific mode of the step of determining the optimum segment, for example, the extraction error and the selection error are ranked for each of the optimum segment candidates, and
An optimum segment candidate having a minimum sum of a first value obtained by multiplying the rank of the extraction error by a predetermined coefficient and a second value obtained by multiplying the rank of the selection error by a predetermined coefficient is selected. It is determined as the optimum segment.

The rule speech synthesizing method of the present invention further comprises the step of deriving a segment transformation rate based on the determined prosodic information of the optimal segment and the selection reference value, and the derived segment transformation rate. When the fragment deformation rate is less than or equal to the deformation threshold, the selected optimal piece is connected as it is, and the fragment deformation exceeds the deformation threshold. Transforms the segment deformation rate of the optimum segment to a deformation rate threshold value or less.

Further, the speech synthesizing apparatus of the present invention which solves the above problem is provided with a segment selecting means for selecting an optimal segment belonging to a phoneme unit, and the selected optimal segments are connected in a predetermined order to generate a synthetic speech. In the device, the segment selection means stores segment information including, for each phoneme unit, segment information including prosodic information of a waveform segment that is a synthesis parameter and subsequent phoneme information that is a connection target. A phoneme unit information table that stores distribution statistical information of phoneme units and information that determines the number of reads at the time of selection, and a phoneme selection unit that retrieves information about the optimum phoneme from the phoneme unit information dictionary and the phoneme unit information table at the time of synthesis. The phoneme selection unit determines whether the subsequent phoneme of one waveform phoneme matches the phoneme environment at the time of synthesis, and synthesizes the waveform phoneme determined to be the following phoneme match. Vocal environment of the target vocabulary and the waveform At least one optimal segment candidate is selected using the extraction error that represents the difference between the segment extraction environment and the selection error that represents the difference between the prosody information of the selected optimal segment candidate and the predetermined selection reference value. It is characterized in that the optimum segment is determined based on this. Here, the distribution statistical information stored in the phoneme unit information table is, for example, a statistical representation of the distribution of prosodic information for each phoneme unit,
It is used when deriving the selection error. The information that determines the number of readings at the time of selection is information that indicates the number of candidates that are scheduled to be selected for each waveform segment.

According to the speech synthesizer having such a configuration,
As the selection criterion of the optimal segment belonging to the phoneme unit, not only the prosody information but also the phoneme unit or waveform segment such as the length of the original utterance unit (or vocabulary) and the position occupied by the waveform segment (or phoneme). Since the environment at the time of extraction is added, the unnaturalness at the time of synthesis is eliminated.

The speech synthesizer of the present invention also modifies the prosodic information of each optimum segment before connecting the optimum segment selected by the segment selection means to make the difference from the selection reference value zero. A segment transformation unit that approaches the value, a unit that derives a segment transformation rate based on the prosodic information of the optimal segment selected by the segment selection unit and the selection reference value, and the derived segment transformation rate is And a unit piece connection determination unit including means for guiding the selected optimum unit unit to the unit unit deformation unit only when the unit size is larger than a predetermined deformation rate threshold value. By configuring in this way, the optimum segment is used as it is without being deformed for the one having a small degree of deformation when transforming the waveform segment, so that the natural sound of the voice due to excessive segment deformation is generated. Sexual deterioration is prevented.

[0012]

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block configuration diagram of a speech synthesis apparatus showing an embodiment of the present invention. In addition,
The functions of the preprocessing unit 12, the prosody setting unit 13, and the Japanese dictionary 191 are basically the same as those of the conventional apparatus shown in FIG. 9, and detailed description thereof will be omitted.

2 and 3 are detailed configuration examples of the segment dictionary 192 in this speech synthesizer, and FIGS.
Is a flow chart showing the operation principle of the segment selection unit 14,
FIG. 8 is a flowchart showing the operating principle of the connection condition determination unit 25. The outline of the speech synthesis processing according to the present embodiment will be described below with reference to these drawings.

In FIG. 1, for example, a character string to be read out is input from the input terminal 11. Preprocessing unit 12
In Japanese, the Japanese dictionary 191 is used to analyze an input character string (input character string), and a phonological unit that is a synthesis unit and accent information regarding the phonological unit are output. The Japanese dictionary used by this embodiment is
It is a dictionary customized for speech synthesis in which readings and accent types in word units are described. Various dictionary structures have been conventionally proposed. The phoneme unit in the present embodiment is preferably / a
In addition to syllables such as / and / ka /, continuous vowels / ai / and compound syllables / aN / are included.

Next, the prosody setting unit 13 sets prosody information in units of phonemes, for example, reference values of prosody parameters. The prosody information (hereinafter referred to as prosody parameter) is stored in the segment selection unit 14
Used in Four types of average pitch frequency, pitch slope, time length, and average power are used as elements of the prosody parameter. Calculation of new parameters by combination of these four types and selection of four types are It can be appropriately changed according to the prosody condition.

The segment selection unit 14 stores a segment value dictionary 1 that stores the reference values set by the prosody setting unit 13 and information about waveform segments.
92 and are used to select the optimum phoneme unit of the phoneme unit. The detailed operation of the element selection unit 14 will be described with reference to FIGS.

The segment dictionary 192 used in this embodiment is
As shown in FIG. 2, it has a segment information dictionary 22 that stores individual information of waveform segments that are synthesis parameters, and a phoneme unit information table 23 that records statistical information of each phoneme unit. The input unit 21 and the output unit 24 connected to the element selection unit 14 are connected in parallel. The segment information dictionary 22 is specifically configured as shown in FIG. 3A, and includes a phoneme unit 31, a file number 32, a subsequent phoneme 33,
It includes a plurality of dictionary elements 34, 35. Dictionary element 34
Is composed of, for example, average pitch, pitch slope, RMS power, start sample and end sample, and the dictionary element 35 is composed of, for example, an utterance unit length and an utterance unit position.

The phoneme unit information table is constructed as shown in FIG. 3B and includes a phoneme unit 31, a start index 36, an end index 37 and a phoneme unit element 38. The start index 36 and the end index 37 are
This is to define the number of waveform segment read corresponding to a phoneme unit. The phoneme unit element 38 defines distribution statistical information for each phoneme unit, and includes, for example, an average pitch, a pitch slope, a time length, and a maximum value and a minimum value of the RMS power. These maximum value and minimum value are fixed values, and the information in this range is used for the calculation of the selection error. Note that FIG.
In FIG. 3, “B” indicates a byte, but the illustrated number of bytes is an example, and the present invention is not limited to this.

The operating principle of the segment selection section 14 will be described below with reference to FIG.
~ It demonstrates with reference to FIG. Referring to FIG. 4, the phoneme selection unit 14 provides information about the phoneme unit to the phoneme information dictionary 22 (corresponding to the phoneme unit 31 to the dictionary element 35 in FIG. 3) and the phoneme unit information table 23 (phoneme in FIG. 3). The unit 31 and the start index 36 to the phoneme unit element 38) are read (S (processing step, the same applies hereinafter) 101). The number of readings is calculated from the start index 36 and the end index 37 of the phoneme unit information table 23.

Next, a candidate which is an optimum segment is selected from the waveform segments belonging to the phoneme unit, and an extraction error which is one of a plurality of selection criteria is calculated (S102-).
S104). Specifically, as a proper check regarding the phoneme environment, it is checked whether or not the subsequent phoneme of the waveform segment matches the phoneme environment at the time of synthesis (S102). In this case, a phoneme whose articulation style is the same as that of the succeeding phoneme at the time of synthesis is also regarded as a match. For example, if the phonological environment at the time of synthesis is / k /, / t / and / p /, which have the same articulation style, are also regarded as "match". Next, an extraction error is calculated for the waveform segment determined to be the subsequent phoneme match in S102 (S103). Here, the extraction error is a distance measure between the utterance environment of the vocabulary to be synthesized and the extraction environment of the waveform segment, and in this embodiment, the following expressions (1) to
It was calculated as in (3).

[0021]

[Equation 1] Extraction error = weight coefficient w1 × moor difference in utterance unit + weight coefficient w2 × positional difference in unit ... (1) Mora difference in utterance unit = utterance unit length−phoneme unit number ... = Utterance unit position-phoneme unit position (3)

Here, the utterance unit length and the utterance unit position relate to the extraction environment of the waveform segment, and are variables indicating from which position of the utterance unit the waveform segment is extracted. . The utterance unit is a unit representing voice information uttered during one breath. Similarly, the phoneme unit number and the phoneme unit position are variables indicating to which position in the synthetic vocabulary of which length the selected phoneme unit belongs. An example of the extraction error calculation is shown in FIG.

FIG. 5 shows the second phonological unit 51 (/ ta /) of the word "warm (warm)" to be synthesized.
The method for calculating the extraction error when selecting is described. In FIG. 5, the unit with a border line corresponds to one phoneme unit. In this case, the utterance unit length is “8” and the utterance unit position is “4” for the waveform segment 52, and the phonological unit number is “5” and the phonological unit position is “2” for the phoneme unit 51. From equations (2) and (3),
The mora difference in the utterance unit is “8” − “5” = “3”, the position difference in the unit is “4” − “2” = “2”, and the extraction error regarding the waveform segment 52 is expressed by the equation (1). From w1 × 3 +
w2 × 2.

Returning to FIG. 4, the above steps S102 to 103 are repeated for all the read pieces (S1).
04). After that, a selection error, which is a distance measure from the reference value, is calculated for the optimal segment candidate selected through S104 (S105), and the optimal segment is determined from the extraction error and the selection error (S106). The selection error is calculated as in Expressions (4) to (6) using the reference value and the prosodic parameter of the optimum segment candidate in the phoneme unit information table 23.

[0025]

[Equation 2] Selection error = average pitch error + pitch inclination error + time length error + RMS power error (4) Each parameter error = (normalized error value) × weighting coefficient (5) Normalized error value regarding average pitch = (Reference value-unit value) / (unit maximum value-unit minimum value) (6)

In the equation (6), when the denominator is "0" when the normalized error value is calculated (when there is only one waveform segment belonging to the phoneme unit), the normalized error value is "
Note that there is no particular limitation on the method of calculating the selection error, and various methods other than the methods of the above formulas (4) to (6) may be used.

The details of the operation of S106 will be described with reference to FIGS. 6 and 7. FIG. 6 is a flow chart showing the processing procedure in S106. First, the extraction errors of the optimum segment candidates are ranked in the smallest order for each candidate number (S201). Next, the selection errors are ranked similarly to determine the selection error rank (S20).
2). Then, based on the rank of each error, a combination score that is a criterion of optimality is calculated by the following equation (7) (S
203).

[0028]

## EQU00003 ## Combined score = selection error rank.times.weighting coefficient w1 '+ extraction error rank.times.weighting coefficient w2' ... (7)

Further, the combined scores of the optimum segment candidates are ranked in the smallest order (S204), and the candidate having the smallest combined score is determined as the optimum segment (S20).
5).

FIG. 7 shows an example of the ranking result for a specific phoneme unit, and an example in which 12 optimal segment candidates indicated by candidate numbers 0 to 11 are selected at the end of S104. It is shown. In the illustrated example, w1 ′ =
2, w2 '= 1 is set. By repeating the above-described processing of S101 to S105 and S106 described in detail in S201 to S205 for all the phoneme units for these optimal segment candidates (S10
7), the candidate number "1" indicated by the diagonal lines in FIG. 7 is determined as the optimum segment. As a result, for the phoneme unit close to the beginning of the word, the waveform segment at the position close to the beginning of the word is surely selected.

Returning to FIG. 1, the connection condition determination unit 15 determines whether or not to deform the optimum segment based on the processing result in the segment selection unit 14, and the modification flag value corresponding to the determination result. To set. FIG. 8 shows a detailed operation procedure of the connection condition determination unit 15.

Referring to FIG. 8, the connection condition determination unit 15 reads the reference value set by the prosody setting unit 13 (S301), and further reads the prosody parameter of the optimum segment (S302). Next, the segment transformation rate is calculated for each prosody parameter by the equation (8) (S303).

[0033]

## EQU00004 ## Fragment deformation rate = {(reference value / segment value) -1} × 100 (8)

Next, the fragment transformation rate and the transformation rate threshold value are compared for each prosody parameter (S304), and a transformation flag setting process is performed on the comparison result (S305 /
S306). The transformation rate threshold value is a numerical value that can be changed for each phoneme unit. Further, in this example, the method of deforming the waveform segment is one type, but it is also possible to generate synthetic speech with less deterioration by changing the modification technique according to the deformation rate of the waveform segment. S301 to S305 or S described above
The process of 306 is performed for all phoneme units (S30).
7). Finally, the optimum segment is cut out from the segment file 193 and written in the synthesized voice buffer (S308).

For each prosody parameter, when the transformation flag value is "1", the segment transformation unit 16 transforms the optimum segment according to the segment transformation rate calculated by the equation (8). Although any method can be considered as the transformation process, in the case of time length transformation, for example, there is a process of thinning / interpolating the waveform sample of the waveform segment according to the transformation rate. The transformed waveform sample is separately stored in a waveform transformation buffer (not shown). When the segment flag value is 0, the transformation process is not performed.

In the segment connecting section 17 of FIG. 1, the waveform samples stored in the synthetic speech buffer and the waveform modification buffer (not shown) are sequentially combined to generate synthetic speech. Through the processing described above, the synthesized voice corresponding to the input character string is output to the output terminal 18.

[0037]

As is apparent from the above description, according to the present invention, the optimum segment is selected by a plurality of selection criteria including the extraction environment of the waveform segment and the phoneme unit, and thus the speech to be synthesized is synthesized. Has the effect of improving the naturalness of. In addition, it is determined whether or not the transformation processing of the optimal segment is necessary before connecting the optimal segment, and the transformation process is performed only when the transformation process is highly necessary. It is effective in avoiding deterioration.

[Brief description of drawings]

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

FIG. 2 is an explanatory diagram showing a configuration example of a segment dictionary in the speech synthesis device of the present embodiment.

FIG. 3 is an explanatory diagram of contents of a phoneme dictionary in the speech synthesis device of this embodiment.

FIG. 4 is a flowchart showing an operation principle of a phoneme selection unit in the speech synthesizer of this embodiment.

FIG. 5 is an explanatory diagram of an example of calculating an extraction error.

FIG. 6 is a flowchart showing a procedure for determining an optimum segment.

FIG. 7 is an explanatory diagram of an example of an optimum segment selection result in the segment selection unit.

FIG. 8 is a flowchart showing a processing procedure of a connection condition determination unit.

FIG. 9 is a block configuration diagram of a conventional speech synthesizer.

[Explanation of symbols]

 12 preprocessing unit 13 prosody setting unit 14 unit selection unit 15 connection condition determination unit 16 unit transformation unit 17 unit connection unit 22 unit information dictionary 23 phoneme unit information table 191 Japanese dictionary 192 unit dictionary 193 unit file

Claims (5)

[Claims] 1. A regular speech synthesizing method comprising: a segment selection process of selecting an optimal segment belonging to a phoneme unit; and a process of connecting the selected optimal segments in a predetermined order to generate a synthesized voice. The selection process includes a step of determining whether or not a subsequent phoneme of one waveform element as a synthesis parameter matches the phoneme environment at the time of synthesis, and a target of synthesis for a waveform element determined to be a subsequent phoneme match. Selecting at least one optimal segment candidate using an extraction error that represents the difference between the vocabulary utterance environment and the waveform segment extraction environment; prosody information of the selected optimal segment candidate and a predetermined selection criterion And a step of determining an optimum segment based on a selection error representing a difference from the value. 2. The step of determining the optimum segment includes ranking the extraction error and the selection error for each optimum segment candidate, and a first coefficient obtained by multiplying the order of the extraction error by a predetermined coefficient. 2. The optimum segment candidate having the minimum sum of the value and the second value obtained by multiplying the rank of the selection error by a predetermined coefficient is determined as the optimal segment. Regular speech synthesis method. 3. A step of deriving a segment variation rate based on the determined prosodic information of the optimal segment and the selection reference value; a derived segment variation rate and a predetermined transformation rate threshold value. If the element deformation rate is less than or equal to the deformation rate threshold value, the selected optimum element pieces are connected as they are, and if the element element deformation rate exceeds the deformation rate threshold value, the element for the optimum element piece is connected. 3. The regular speech synthesis method according to claim 1, further comprising the step of transforming the one-sided deformation rate to be equal to or less than a deformation rate threshold value. 4. A voice synthesizing device, comprising: a unit selection unit for selecting an optimum unit belonging to a phoneme unit, and connecting the selected optimum units in a predetermined order to generate a synthesized voice. Is a phoneme information dictionary that stores, for each phoneme unit, phoneme information including prosodic information of a waveform phoneme that is a synthesis parameter and subsequent phoneme information that is a connection target, distribution statistical information of each phoneme unit, and reading at the time of selection. A phoneme unit information table storing information that determines the number, and a phoneme unit selection unit that retrieves information about the optimum phoneme unit from the phoneme unit information dictionary and the phoneme unit information table at the time of synthesis, and the phoneme unit selection unit, It is determined whether or not the subsequent phoneme of one waveform segment matches the phoneme environment at the time of synthesis, and for the waveform segment determined to be the subsequent phoneme match, the utterance environment of the synthesis target vocabulary and the waveform segment Extraction error that represents the difference with the extraction environment And selecting at least one optimal segment candidate, and determining the optimal segment based on a selection error representing a difference between the prosody information of the selected optimal segment candidate and a predetermined selection reference value. Speech synthesizer. 5. A segment transforming unit for transforming prosodic information of each optimum segment to bring the difference from the selection reference value close to zero value before connecting the optimal segment selected by the segment selection means. A means for deriving a segment deformation rate based on the prosodic information of the optimum segment selected by the segment selection means and the selection reference value, and the derived segment deformation rate being greater than a predetermined transformation rate threshold value. 5. The speech synthesis apparatus according to claim 4, further comprising: a segment connection determination unit having a unit that guides the selected optimal segment to the segment transformation unit only when it is larger.