JP5320341B2 - Speaking text set creation method, utterance text set creation device, and utterance text set creation program - Google Patents

Description

  The present invention relates to an utterance text set creation method, an utterance text set creation apparatus, and an utterance text set creation program for creating a text set read by a speaker when constructing a speech synthesis unit DB.

The waveform connection type speech synthesis system has a speech unit DB for speech synthesis because it is necessary to connect speech units when performing speech synthesis. Note that the speech segment is a short unit of speech data prepared in advance. For example, there are CV, VCV, CVC, [C] V ^*, etc. as speech unit units, which depend on individual TTS (text-to-speech system). However, C represents a consonant, V represents a vowel, [C] represents that there may be no C, and V ^* represents a chain of one or more Vs. When constructing the speech unit DB for speech synthesis, it is necessary for the speaker to read the speech text set in advance and obtain the speech unit using the natural speech data. Note that the natural speech data is speech data that records sounds uttered by a speaker in natural units such as words and sentences, and the speech synthesis speech unit DB is necessary for speech synthesis from natural speech data. It is a database that extracts only speech segments. In order to perform more natural speech synthesis processing, it is desired that the speech synthesis speech unit DB includes more speech units necessary for speech synthesis. For this purpose, it is necessary that the utterance text set is composed of sentences that can efficiently collect speech segments.

  When synthesizing high-quality speech that includes a variety of tone, utterance styles, and rich emotions, speech segments created from speech that includes the desired tone, utterance style, and emotion (hereinafter referred to as “X tone”) It is known from Non-Patent Document 1 that the quality of synthesized speech is higher when a DB is used than when a speech segment DB created from speech uttered in a reading tone is used. This is considered to be caused by a decrease in naturalness and the like caused by a large prosodic deformation amount and a difference in spectrum because prosody and spectral characteristics are different for each variation of X tone. Note that the utterance style refers to the acoustic characteristics that occur depending on the speaker's environment and culture, such as dialects, fast speech, awkward speaking, polite speaking, slow speaking, and how to speak clearly. . In addition, emotions include a sad way of speaking and a pleasant way of speaking. The tone refers to the tone of words when they are put out to the mouth and the way of saying things, and is a concept that includes speech that includes the speech style and emotions. Prosodic features are voice pitch, intonation, rhythm, pose, and the like, and a spectrum is a representation of the strength of each frequency divided into frequency components.

  In general, an utterance text set has been created using an algorithm that maximizes the coverage of phoneme strings and prosodic features of a large amount of Japanese text (see Non-Patent Document 2). The phoneme string is a string of phonemes (phonemes) and is a reading pseudonym. A phoneme refers to a minimum sound unit used for distinction (discrimination) of meaning in an arbitrary individual language, such as a vowel or a consonant. In addition, coverage means that when speech synthesis is performed in a waveform-connected speech synthesis system, the speech segments required for speech synthesis of the text to be processed are the phoneme environment, phoneme duration length, and fundamental frequency pattern. It is the probability of being included in the speech synthesis speech element DB when considered.

Koji Onishi, Takashi Masuko, Takao Kobayashi, "Examination of generation of different utterance styles in HMM speech synthesis", IEICE Technical Report, 2003, 102, 619 (SP2002-17), p17-22 Tsuyoshi Kawai, Yoshio Higuchi, Seiichi Yamamoto, “Creation of waveform segment data set for speech synthesis considering fundamental frequency and duration of phoneme duration”, IEICE Transactions (D-II), 1999 8 Month, Vol. J82-D-II, no. 8, p. 1229-1238

  When creating a text set for utterance, a speech synthesis program is used to estimate phonological sequences from a large amount of Japanese text in a kanji-kana mixed sentence, but a general speech synthesis program is assumed to be read out in a reading tone. doing. Therefore, in the conventional technique, when speaking in X tone, the speaker may not utter as estimated. For example, using a general speech synthesis program that assumes reading tone, create a text set for utterance by estimating the phoneme sequence, and if the speaker utters in X tone, the speech synthesis program estimates It is assumed that there is a difference (reading fluctuation) between the phoneme sequence obtained and the phoneme sequence actually obtained by speaking in X tone. For example, the word “Tomorrow” is “Ashi ^ ta” (^ is a symbol for devoicing) when the phoneme sequence is estimated using a general speech synthesis program, but “Ashi ^ "Tau" and ending vowels. Further, in the case of emphasized utterances, it may be assumed that “Ashita!” And “Shi” are not devoiced.

  As described above, when the phoneme sequence assumed at the time of generating the utterance text set is different from the phoneme sequence obtained by actually uttering in the X tone, Is not realized as expected, which causes a problem that the quality of the synthesized speech is lowered.

  In order to solve the above-described problem, the utterance text set creation technology according to the present invention uses the parameter distribution obtained from the natural speech data of reading tone and the parameter distribution obtained from the natural speech data of target X tone. A parameter distribution conversion function to be converted into a pre-stored, speech synthesis processing is performed by the speech synthesis program using the utterance text set candidate, a predetermined parameter is obtained from the speech synthesis data, and the parameter distribution conversion function is used, The obtained parameter distribution is converted, and the utterance text set candidates are evaluated using the converted parameter distribution.

  The present invention produces an effect that it is possible to generate an utterance text set that maximizes the coverage of phoneme strings and prosodic features even when uttered in a tone other than reading tone.

The block diagram of the utterance text set preparation part. The figure which shows the processing flow of the utterance text set preparation part. The block diagram of a conversion function preparation part. The figure which shows the processing flow of a conversion function preparation part. (A-1) Segment distribution in reading tone, (a-2) Segment distribution in X tone, (b-1) Duration distribution in reading tone, (b-2) Duration distribution in X tone (C-1) The F0 distribution in reading tone, (b-2) The figure which shows F0 distribution in X tone. The block diagram of a text set preparation part. The figure which shows the processing flow of a text set preparation part.

  Hereinafter, embodiments of the present invention will be described in detail.

<Speaking text set creation device 1000>
An utterance text set creation apparatus 1000 according to the first embodiment will be described with reference to FIGS. 1 and 2. The utterance text set creation apparatus 1000 creates a text set read by the utterer when the speech synthesis speech segment DB is constructed.

  The utterance text set creation apparatus 1000 includes an input / output interface unit 101, a conversion function creation unit 100, a storage unit 203, and a text set creation unit 200.

  When the utterance text set creation apparatus 1000 receives a conversion function creation instruction from the utterance text set creator (hereinafter referred to as “user”) via the input / output interface unit 101, the conversion function creation unit 100 performs offline processing. , A parameter distribution conversion function that converts a parameter distribution obtained from natural speech data of reading tone to a parameter distribution obtained from natural speech data of target X tone (for example, a segment distribution conversion function, a continuation length described later) Distribution conversion function and F0 distribution conversion function) are created (s100), and stored in the storage unit 203 before the utterance text set candidate to be described later is created.

  Furthermore, when the utterance text set creation device 1000 receives a text set creation instruction and tone designation information from the user via the input / output interface unit 101 (s101), the text set creation unit 200 performs an X tone by online processing. A text set for maximizing the coverage when uttered in step S200 is created (s200) and output to the user via the input / output interface unit 101. The processing contents of each unit will be described below.

<Input / output interface unit 101>
The input / output interface unit 101 receives input from a user and outputs information to the user. For example, an input interface (for example, a keyboard, a mouse, etc.) for inputting data and an output interface (for example, a display, a printer, etc.) for outputting data, or input / output terminals for these input / output interfaces are included. In addition, when the utterance text set creation apparatus 1000 is a server on the network and the user accesses via the network, the input / output interface unit 101 is a communication unit for transmitting and receiving data to and from the user. May be.

<Storage unit 203>
The storage unit 203 stores / reads each input / output data and each data of the calculation process one by one. Thereby, each calculation process is advanced. However, the data need not necessarily be stored in the storage unit 203, and data may be directly transferred between the units. Note that an element distribution conversion function DB 234, a continuation length distribution conversion function DB 236, and an F0 distribution conversion function DB 238 described later may be part of the storage unit 203.

<Conversion function creation unit 100>
The conversion function creation unit 100 uses, for example, natural speech uttered by a human in X tone, and uses a difference between reading tone tone and X tone (difference in frequency distribution of speech units included in the phoneme sequence, each speech unit The difference between the appearance frequency distributions of the phoneme duration lengths and the difference in the appearance frequency distributions of the basic frequency patterns for each speech unit) is extracted, and a parameter distribution conversion function is obtained and stored in the storage unit 203.

  The conversion function creation unit 100 will be described with reference to FIGS. 3 and 4. The conversion function creation unit 100 includes a natural speech DB 110, a phonological labeling unit 111, a first parameter distribution extraction unit 120, and a parameter distribution conversion function calculation unit 130.

(Natural speech DB 110 and phonological labeling unit 111)
The natural voice DB 110 stores read-tone natural voice data and X-tone natural voice data in advance. For example, each natural voice data is obtained by reading the same text (for example, “turn”) in a reading tone and an X tone. Note that the X tone may have various variations, and natural voice data is created and stored for each variation.

  When receiving the conversion function creation instruction via the input / output interface unit 101, the phonological labeling unit 111 acquires the reading-tone natural sound data and the X-tone natural sound data from the natural sound DB 110 (s110), and each natural sound data. The phoneme label (for example, / seNkaisuru / etc.) is assigned manually or automatically (s111), and the acquired natural speech data and the corresponding label data are output to the segment distribution extraction unit 123.

(First parameter distribution extraction unit 120)
The first parameter distribution extraction unit 120 obtains predetermined parameters from the reading-tone natural voice data obtained by reading a predetermined document in the reading tone and the X-tone natural voice data obtained by reading the same document in the X-tone, respectively. Is extracted (s120).

  For example, the first parameter distribution extraction unit 120 includes an all speech unit variation storage unit 122, a unit distribution extraction unit 123, a duration distribution extraction unit 125, and an F0 distribution extraction unit 127.

{All speech segment variation storage unit 122 and segment distribution extraction unit 123}
The segment distribution extraction unit 123 refers to the all speech unit variation storage unit 122 and assigns a speech unit number to each speech unit obtained from each natural speech data (s123a), so that the speech unit appears. A frequency distribution (hereinafter referred to as “segment distribution”) is extracted (s123b).

  The speech unit variation storage unit 122 stores speech units (or feature values obtained from speech units, label data corresponding to speech units, etc.) and speech unit numbers for each speech unit. . However, the speech unit stored in the all speech unit variation storage unit 122 depends on the text speech synthesis system to be developed.

  The segment distribution extraction unit 123 receives each natural speech data and label data, searches the entire speech unit variation storage unit 122 using a speech unit obtained from the natural speech data as a key, and performs speech for each speech unit. Get the segment number. Based on the number of speech unit numbers obtained (appearance frequency), the appearance frequency for every type of all speech units is obtained, and the segment distribution is extracted. The unit distribution is transmitted to the unit distribution conversion function calculating unit 133, and each natural speech data and the speech unit number associated therewith are transmitted to the duration distribution extracting unit 125 and the F0 distribution extracting unit 127, and the duration distribution extracting unit Further, the label data attached to each natural voice data is also transmitted to 125.

{Continuation length distribution extraction unit 125}
The continuation length distribution extracting unit 125 receives the label data and the speech unit number, and calculates the phoneme duration for each speech unit using the label data and the speech unit number (s125a), and the phoneme duration for each speech unit. Is extracted (s125b) and transmitted to the duration distribution conversion function calculation unit 135. The phoneme continuation length is calculated as vector data. For example, when the duration length of each phoneme of the speech unit “KAS” is 12 ms for K, 22 ms for A, and 11 ms for S, the vector data is (12, 22). 11). However, the phoneme duration for each speech unit may be represented by other conventional techniques.

{F0 distribution extraction unit 127}
The F0 distribution extraction unit 127 receives the natural speech data, the speech unit number, and the label data, and extracts the fundamental frequency pattern for each speech unit using these (s127a), A frequency distribution of basic frequency patterns (hereinafter referred to as “F0 distribution”) is extracted (s127b) and transmitted to the F0 distribution extraction unit 127. The fundamental frequency pattern is calculated as vector data. For example, the average value of the fundamental frequency pattern frequency of each phoneme of the speech unit “ASU” is 120 Hz and the average value of S is 0 Hz (since S is an unvoiced consonant and has no fundamental frequency), When the average value of U is 220 Hz, the vector data is (120, 0, 220). However, there are various basic frequency pattern designation methods other than this method, and the fundamental frequency pattern for each speech unit may be represented by other conventional techniques. For example, vector data consisting of the average value of the fundamental frequency pattern of the speech element, the frequency variance, the start point frequency, and the end point frequency may be extracted, or the average value may be taken for each phoneme. Alternatively, the temporal change pattern of the fundamental frequency may be approximated by a three-point broken line.

(Parameter distribution conversion function calculation unit 130)
The parameter distribution conversion function calculation unit 130 calculates a parameter distribution conversion function for converting the parameter distribution obtained from the natural speech data in the reading tone into the parameter distribution obtained from the natural speech data in the X tone (s130).

  For example, the parameter distribution conversion function calculation unit 130 includes an element distribution conversion function calculation unit 133, a duration distribution conversion function calculation unit 135, and an F0 distribution conversion function calculation unit 137.

{Element distribution conversion function calculation unit 133}
The segment distribution conversion function calculation unit 133 receives the segment distribution obtained from each natural speech data, and converts the segment distribution (FIG. 5 (a-1)) obtained from the natural speech data of the reading tone to the X tone. A segment distribution conversion function to be converted into the segment distribution (FIG. 5 (a-2)) obtained from the natural speech data is calculated (s133), transmitted to the segment distribution conversion function DB 234, and registered. The upper part of FIG. 5 shows an image of the segment distribution conversion function f for converting the segment appearance frequency distribution from the reading tone to the X tone. N speech unit numbers are arranged in order from the left on the horizontal axis of (a-1) and (a-2). The vertical axis represents the appearance frequency. The conversion function f is a function for converting the left distribution into the right distribution. As a result, the difference between the two tones relating to the reading swing can be incorporated into the conversion function f.

For example, when the number of types of speech segments is N, the segment distribution of reading tone {u _1w , u _2w ,..., N _Nw } and the segment distribution of X tone {u _1x , u _{2x for} each speech unit. ,..., N _Nx }, the difference {u _1w −u _1x , u _2w −u _2x ,..., N _Nw −u _Nx } is obtained and stored. In the element distribution conversion unit 233 described later, the element distribution conversion function converts the distribution by subtracting this difference from the input element distribution. Further, for example, the segment distribution conversion function may be converted by multiplying the input segment distribution by the ratio between the reading tone segment distribution and the X tone segment distribution for each speech segment. The reading tone segment distribution may be converted into an X tone segment distribution by other methods. The element distribution conversion function calculation unit 133 calculates the element distribution conversion functions for the number of variations of the X tone, and transmits and registers them in the element distribution conversion function DB 234.

{Duration distribution conversion function calculation unit 135}
The continuous length distribution conversion function calculation unit 135 receives the continuous length distribution for each speech unit of each natural speech data, and the continuous length distribution for each speech unit obtained from the natural speech data of the reading tone (FIG. 5 (b- 1)) is calculated into a continuous length distribution conversion function (FIG. 5 (b-2)) for each speech unit obtained from natural speech data of X tone (s135), and a continuous length distribution conversion is performed. It transmits to function DB236 and registers. Accordingly, the continuous length distribution conversion function DB 236 is registered with continuous length distribution conversion functions for (the number of X tone variations) × (the number N of speech segment types). The middle part of FIG. 5 shows an image of a duration distribution conversion function for converting the appearance frequency distribution of the phoneme duration from the reading tone to the X tone. Appearance frequency of phoneme duration length of a speech unit i in the reading tone on the left (Mi is the number of variations of the phoneme duration vector), and appearance of the phoneme duration of a phoneme i in the X tone The frequency is shown, and Mi phoneme continuation length vectors are arranged in order from the left on the horizontal axis of (b-1) and (b-2). The vertical axis represents the appearance frequency. The conversion function gi is a function for converting the left distribution into the right distribution.

For example, when the number of variations of the phoneme duration length vector for a certain speech unit i is Mi, the continuous length distribution {u _1w , u _2w ,..., N _Miw } for each speech unit and the continuation of the X tone the length distribution _{{u 1x, u 2x, ...} , n Mix} and the difference _{{u 1w -u 1x, u 2w} -u 2x, ..., n Miw -u Mix} sought and stored. In the continuation length distribution conversion unit 235 described later, the continuation length distribution conversion function converts the distribution by subtracting this difference from the input continuation length distribution. This process is performed for all speech segments. Further, the reading tone tone duration distribution may be converted into the X tone duration duration distribution by other methods. The F0 distribution conversion function calculation unit 137 and the F0 distribution conversion unit 237 described later can also obtain the F0 distribution conversion function and convert the F0 distribution by the same processing.

{F0 distribution conversion function calculation unit 137}
The F0 distribution conversion function calculation unit 137 receives the F0 distribution for each speech unit of each natural speech data, and the F0 distribution for each speech unit obtained from the natural speech data of the reading tone (FIG. 5 (c-1)). Is converted into an F0 distribution conversion function for each speech unit (FIG. 5 (c-2)) obtained from natural speech data of X tone (s137), and transmitted to the F0 distribution conversion function DB 238. sign up. In the F0 distribution conversion function DB 238, F0 distribution conversion functions for (the number of X tone variations) × (the number N of speech segment types) are registered. The lower part of FIG. 5 shows an image of a function for converting the F0 distribution from the reading tone to the X tone. The left side shows the frequency of appearance of the fundamental frequency pattern of a certain speech unit i in reading tone (the number of variations of the fundamental frequency pattern vector is Li), and the right side shows the frequency of appearance of the fundamental frequency pattern of a unit i in X tone Li basic frequency pattern vectors are arranged in order from the left on the horizontal axes of (c-1) and (c-2). The vertical axis represents the appearance frequency. The conversion function hi is a function for converting the left distribution into the right distribution.

<Text set creation unit 200>
The text set creation unit 200 will be described with reference to FIGS. The text set creation unit 200 includes an utterance text set candidate creation unit 210, a large-volume Japanese DB 211, a second parameter distribution extraction unit 220, a parameter distribution conversion unit 230, an evaluation unit 250, and an end determination unit 260. Have. In FIG. 6, the parameter distribution conversion unit 230 and the conversion function DBs 234, 236, and 238 are portions added by the present invention, and the other portions may be subjected to an iterative process equivalent to the conventional technique ( For example, Non-Patent Document 2). The iterative processing includes “exchange method”, “greedy algorithm”, and the like, but FIGS. 6 and 7 show the exchange method as an example.

(Speech text set candidate creation unit 210 and mass Japanese DB 211)
Upon receiving the text set creation instruction via the interface unit 101, the utterance text set candidate creation unit 210 extracts a predetermined number (for example, 500) of sentences from the large-volume Japanese sentence DB 211, and the first utterance text set A candidate (hereinafter referred to as “T”) is created (s210) and transmitted to the second parameter distribution extraction unit 220. Information specifying the number of sentences extracted from the large-volume Japanese sentence DB 211 (hereinafter referred to as “extraction number specifying information”) may be added to the text set creation instruction. The extraction number designation information is a value that is first designated and input by the user.

(Second parameter distribution extraction unit 220)
The second parameter distribution extraction unit 220 performs speech synthesis processing using the speech synthesis program using T, obtains predetermined parameters from the speech synthesis data, and extracts the obtained parameter distribution (s220).

  For example, the second parameter distribution extraction unit 220 includes a phoneme string, a fundamental frequency pattern, a phoneme duration extraction unit 221, a segment distribution extraction unit 223, a duration distribution extraction unit 225, and an F0 distribution extraction unit 227. Prepare.

{Phoneme sequence, fundamental frequency pattern, phoneme duration extraction unit 221}
The phoneme sequence, fundamental frequency pattern, and phoneme duration length extraction unit 221 receives the utterance text set candidate, uses this to perform speech synthesis processing by a speech synthesis program, and uses the speech synthesis data to generate the phoneme sequence and fundamental frequency pattern. Then, the phoneme duration length is estimated, extracted (s221), and transmitted to the segment distribution extraction unit 223.

{Element distribution extraction unit 223}
The segment distribution extraction unit 223 receives the phoneme sequence, the fundamental frequency pattern, and the phoneme duration, obtains the appearance frequency of each speech segment using the phoneme sequence, extracts the segment distribution (s223), The data is transmitted to the distribution conversion unit 233. Further, the speech unit and the phoneme duration associated with it are transmitted to the duration distribution extraction unit 225, and the speech unit and the fundamental frequency pattern associated with it are transmitted to the F0 distribution extraction unit 227.

{Continuation length distribution extraction unit 225}
The continuation length distribution extracting unit 225 receives the segment distribution and the phoneme continuation length for each speech unit, obtains the phoneme duration for each speech unit, extracts the continuation distribution from the appearance frequency (s225), It transmits to the continuation length distribution conversion part 235.

{F0 distribution extraction unit 227}
The F0 distribution extraction unit 227 receives the unit distribution and the fundamental frequency pattern for each speech unit, obtains the fundamental frequency pattern for each speech unit, extracts the F0 distribution from the appearance frequency (s227), and converts the F0 distribution. To the unit 237.

  The segment distribution extraction unit 223 extracts one segment distribution, and the duration distribution extraction unit 225 and the F0 distribution extraction unit 227 extract F0 distributions and duration distributions corresponding to the number of variations of the speech segment.

(Parameter distribution converter 230)
The parameter distribution conversion unit 230 receives the tone designation information via the input / output interface unit 101, extracts the parameter distribution conversion function from the storage unit 203 based on the tone designation information, and uses the parameter distribution conversion function to perform speech synthesis. The parameter distribution obtained from the data is converted (s230). For example, the parameter distribution converter 230 includes an element distribution converter 233, a duration distribution converter 235, and an F0 distribution converter 237.

{Element distribution conversion unit 233}
The segment distribution conversion unit 233 receives the tone designation information and the segment distribution, searches the segment distribution conversion function DB 234 in the storage unit 203 using the tone designation information as a key, and extracts the corresponding segment distribution conversion function. Using this, the received segment distribution (obtained from the combined data) is converted (s233), and the converted segment distribution is transmitted to the evaluation unit 250.

{Continuation length distribution conversion unit 235}
The duration distribution converter 235 receives the tone designation information and the duration distribution, searches the duration distribution conversion function DB 236 in the storage unit 203 using the tone designation information as a key, and extracts the corresponding duration distribution conversion function. Using this, the received duration distribution (obtained from the speech synthesis data) is converted (s235), and the converted duration distribution is transmitted to the evaluation unit 250.

{F0 distribution conversion unit 237}
The F0 distribution conversion unit 237 receives the tone designation information and the F0 distribution, searches the F0 distribution conversion function DB in the storage unit 203 using the tone designation information as a key, extracts the corresponding F0 distribution conversion function, and uses this The received F0 distribution (obtained from the speech synthesis data) is converted (s237), and the converted F0 distribution is transmitted to the evaluation unit 250.

(Evaluation unit 250)
The evaluation unit 250 calculates an evaluation function using the converted parameter distribution (segment distribution, duration distribution, and F0 distribution), evaluates the text set candidate for utterance (s250), and determines the evaluation result as an end determination unit 260. To the utterance text set candidate creation unit 210. For example, the evaluation function is calculated by the method of Non-Patent Document 2.

For example, the type of all speech units is represented as N, the appearance frequency of speech units appearing in the utterance text set candidate is represented as {u ₁ , u ₂ ,..., _{N N} }, and the relative appearance frequency of u _i is represented as p. _{Let i} . The type of phoneme duration corresponding to u _i is represented by N _i , the respective appearance frequencies are represented as {v _i1 , v _i2 ,..., v _iNi }, and the relative appearance frequency of v _ij is represented by q _ij . The fundamental frequency pattern can also be obtained by the same method as the phoneme duration time.

R _i is introduced as an index representing the degree of coverage of the speech unit u _i . However,

And d _ij (T) is 1 when the utterance text set candidate T includes waveform segments that can realize the fundamental frequency and phoneme duration length of v _ij by deformation within the allowable range of quality degradation. Otherwise, it is a function that takes 0.

  The total coverage of speech units included in the utterance text set candidate T is

Among phoneme duration lengths and fundamental frequency patterns belonging to the same speech segment, the higher the appearance frequency of the phoneme duration length and the fundamental frequency pattern, the greater the contribution to the evaluation criteria for measuring the goodness of covering. . This may be used as an evaluation function. Furthermore, a nonlinear function or the like may be introduced as a mechanism for adjusting the weight between the spread of the phoneme environment and the basic frequency pattern and the spread of the speech unit duration (see Non-Patent Document 2).

(End determination unit 260)
The end determination unit 260 determines whether the end condition is satisfied (s260), and transmits the end determination result to the utterance text set candidate creation unit 210. The termination condition is, for example, that the number of sentences that have been exchanged has reached a predetermined value, or that the size of the evaluation function is greater than or equal to a predetermined value.

[Repetition processing]
The utterance text set candidate creation unit 210 receives the evaluation result and the end determination result. If the end determination result means that the end condition is satisfied (s260), the utterance text set candidate creation unit 210 uses the utterance text set candidate for utterance. A text set is output (s315). If the end determination result means that the end condition is not satisfied (s260), a new utterance text set candidate is created (s210), and the processing (s210 to s260) is repeated.

  It should be noted that a new utterance text set candidate may be created by taking an arbitrary sentence from the mass Japanese DB 211 and replacing it with an arbitrary sentence in the utterance text set candidate. In this case, the utterance text set candidate in which an arbitrary sentence is exchanged and the utterance text set candidate that has not been exchanged are stored in the storage unit 203, and the utterance having a low evaluation according to the evaluation result of the evaluation unit 250 It is good also as a structure which deletes the text set candidate for use. Each process after the second week can be processed efficiently because only the difference needs to be processed.

<Program>
The utterance text set creation device described above can also be operated by a computer. In this case, the program for causing the computer to function as the target device (the device having the functional configuration shown in the drawings in the embodiment) or each process of the processing procedure (shown in the embodiment) is stored in the computer. A program to be executed may be downloaded into a computer from a recording medium such as a CD-ROM, a magnetic disk, or a semiconductor storage device or via a communication line, and the program may be executed.

<Effect>
In the present invention, for each X tone, the utterance difference that is likely to occur between the reading tone and tone is extracted in advance and patterned, so that the coverage of the phoneme sequence and prosodic features is maximized according to the utterance difference. By correcting the utterance text set, the coverage can be maximized even when uttered in a tone other than reading tone. By using the utterance text set created according to the present invention, the utterer utters, and by constructing the speech segment DB based on the natural speech data, deterioration of the synthesized speech in the X tone can be prevented.

[Modification]
The utterance text set 1000 may not include the conversion function creation unit 100. For example, each conversion function created by another device may be stored in the storage unit 203.

  The utterance text set 1000 has three parameter distributions (segment distribution, continuation length distribution, and F0 distribution) as conversion targets, but at least one may be converted. Although the estimation accuracy is reduced, the amount of data and calculation can be reduced. Further, parameter distributions other than the three parameter distributions may be converted.

  In the first embodiment, the utterance text set candidates are evaluated using the method described in Non-Patent Document 2, but may be evaluated using other existing techniques.

  The present invention can be used when creating a text set read by a speaker when constructing a speech unit DB for speech synthesis. A speech unit DB for speech synthesis that maximizes the coverage of phoneme strings and prosodic features in the X tone is constructed by the speaker reading out the text set created by the speech set creation device 1000 of the present invention in the X tone. By using the DB, high-quality synthesized speech in the X tone is made possible.

1000 Spoken Text Set Creation Device 100 Conversion Function Creation Unit 101 Input / Output Interface Unit 110 Natural Speech DB
111 Phonological labeling unit 120 First parameter distribution extraction unit 130 Parameter distribution conversion function calculation unit 200 Text set creation unit 203 Storage unit 210 Utterance text set candidate creation unit 220 Second parameter distribution extraction unit 230 Parameter distribution conversion unit 234 Segment distribution Conversion function DB
236 Continuous length distribution conversion function DB
238 F0 distribution conversion function DB
250 Evaluation Unit 260 End Determination Unit

Claims (9)

A speech text set creation method for creating a text set read out by a speaker when constructing a speech synthesis unit DB for speech synthesis,
It is assumed that a parameter distribution conversion function for converting a parameter distribution obtained from natural speech data in reading tone into a parameter distribution obtained from natural speech data in target X tone is stored in the storage unit in advance. ,
A utterance text set candidate creation step of extracting a predetermined number of sentences randomly from a large volume Japanese sentence DB and creating a utterance text set candidate;
A second parameter distribution extraction step of performing speech synthesis processing by a speech synthesis program using the utterance text set candidate, obtaining a predetermined parameter from the speech synthesis data, and extracting a distribution of the obtained parameter;
A parameter distribution conversion step of taking out the parameter distribution conversion function from the storage unit and converting the parameter distribution obtained from the speech synthesis data using the parameter distribution conversion function;
Evaluating the utterance text set candidates using the converted parameter distribution, and
A method for generating a text set for speech. The utterance text set generation method according to claim 1,
A first parameter distribution for obtaining predetermined parameters from the reading-tone natural voice data read out from the predetermined document in the reading-tone and the X-tone natural voice data read out from the same document in the X-tone, and extracting the parameter distribution, respectively. An extraction step;
A parameter distribution conversion function calculating step for calculating a parameter distribution conversion function for converting the parameter distribution obtained from the natural speech data of the reading tone into the parameter distribution obtained from the natural speech data of the X tone;
Storing the parameter distribution conversion function in the storage unit before creating a text set candidate for utterance,
A method for generating a text set for speech. A speech text set creation method for creating a text set read out by a speaker when constructing a speech synthesis unit DB for speech synthesis,
The storage unit includes an appearance frequency distribution of each speech unit obtained from natural speech data of reading tone, a distribution of phoneme durations for each speech unit (hereinafter referred to as “continuation length distribution”), and a basic unit for each speech unit. The frequency distribution (hereinafter referred to as “F0 distribution”) is the frequency distribution of each speech unit, the duration distribution for each speech unit, and the F0 for each speech unit obtained from natural speech data of the intended X tone. It is assumed that an element distribution conversion function, a duration distribution conversion function, and an F0 distribution conversion function to be converted into a distribution are stored in advance,
A utterance text set candidate creation step of extracting a predetermined number of sentences randomly from a large volume Japanese sentence DB and creating a utterance text set candidate;
Using the utterance text set candidates, the speech synthesis process is performed by a speech synthesis program, and the phoneme sequence, the fundamental frequency pattern, and the phoneme duration time length are extracted from the speech synthesis data. Steps,
A second segment distribution extraction step of obtaining an appearance frequency of each speech segment from the phoneme sequence and extracting a segment distribution;
A second phoneme duration length and F0 distribution extraction step for obtaining a phoneme duration for each speech unit and extracting a duration distribution, obtaining a fundamental frequency pattern for each speech unit and extracting an F0 distribution;
A segment distribution conversion function, a duration distribution conversion function, and an F0 distribution conversion function are extracted from the storage unit, and the segment distribution, duration distribution, and F0 distribution obtained from the speech synthesis data using these distribution conversion functions, respectively. A parameter distribution conversion step for converting
An evaluation function that calculates an evaluation function using the segment distribution, the duration distribution, and the F0 distribution after the conversion, and evaluates the utterance text set candidate.
A method for generating a text set for speech. A utterance text set generation method according to claim 3,
A phonological labeling step of assigning a phonological label to the reading-tone natural voice data of a predetermined document read out in a reading-tone style and the X-tone natural voice data of the same document read out in an X tone;
A first segment distribution extraction step of referring to the all speech segment variation storage unit and assigning a speech unit number to a speech unit obtained from each natural speech data and extracting an appearance frequency distribution of the speech unit When,
A first phoneme duration and F0 distribution extraction step of calculating a phoneme duration for each speech unit, extracting a duration distribution, extracting a fundamental frequency pattern for each speech unit, and extracting an F0 distribution;
A segment distribution conversion function for converting a segment distribution, duration distribution, and F0 distribution obtained from natural speech data in reading tone into a segment distribution, duration distribution, and F0 distribution obtained from natural speech data in X tone, A parameter distribution conversion function calculating step for calculating a duration distribution conversion function and an F0 distribution conversion function;
Storing the segment distribution conversion function, the continuation length distribution conversion function, and the F0 distribution conversion function in the storage unit before creating the utterance text set candidate.
A method for generating a text set for speech. An utterance text set creation device for creating a text set read by a speaker when constructing a speech segment DB for speech synthesis,
A storage unit that stores in advance a parameter distribution conversion function for converting the parameter distribution obtained from the natural speech data of the reading tone into the parameter distribution obtained from the natural speech data of the target X tone;
A utterance text set candidate creation unit that randomly extracts a predetermined number of sentences from a large amount of Japanese sentence DB and creates utterance text set candidates;
A second parameter distribution extraction unit that performs speech synthesis processing by a speech synthesis program using the utterance text set candidate, obtains a predetermined parameter from speech synthesis data, and extracts a distribution of the obtained parameter;
A parameter distribution conversion unit that takes out the parameter distribution conversion function from the storage unit and converts the parameter distribution obtained from the speech synthesis data using the parameter distribution conversion function;
An evaluation unit that evaluates the utterance text set candidate using the converted parameter distribution;
An utterance text set generation device characterized by the above. The utterance text set generation device according to claim 5,
A first parameter distribution for obtaining predetermined parameters from the reading-tone natural voice data read out from the predetermined document in the reading-tone and the X-tone natural voice data read out from the same document in the X-tone, and extracting the parameter distribution, respectively. An extractor;
A parameter distribution conversion function calculation unit for calculating a parameter distribution conversion function for converting the parameter distribution obtained from the natural speech data of the reading tone into the parameter distribution obtained from the natural speech data of the X tone,
The storage unit stores the parameter distribution conversion function before generating a text set candidate for utterance.
An utterance text set generation device characterized by the above. An utterance text set creation device for creating a text set read by a speaker when constructing a speech segment DB for speech synthesis,
Appearance frequency distribution of each speech segment (hereinafter referred to as “segment distribution”) obtained from natural speech data of reading tone, and distribution of appearance frequency of phoneme duration length for each speech segment (hereinafter referred to as “continuation length distribution”) Distribution of frequency of appearance of fundamental frequency patterns for each speech unit (hereinafter referred to as “F0 distribution”), frequency distribution of speech units obtained from natural speech data of target X tone, speech unit A storage unit in which a duration distribution for each unit, a unit distribution conversion function for converting into a F0 distribution for each speech unit, a duration distribution conversion function, and an F0 distribution conversion function are stored in advance;
A utterance text set candidate creation unit that randomly extracts a predetermined number of sentences from a large amount of Japanese sentence DB and creates utterance text set candidates;
Using the utterance text set candidates, the speech synthesis process is performed by a speech synthesis program, and the phoneme sequence, the fundamental frequency pattern, and the phoneme duration time length are extracted from the speech synthesis data. And
A second segment distribution extraction unit for obtaining an appearance frequency of each speech segment from the phoneme sequence and extracting a segment distribution;
A second duration distribution extraction unit for obtaining a phoneme duration for each speech unit and extracting a duration distribution;
A second F0 distribution extraction unit for obtaining a fundamental frequency pattern for each speech unit and extracting an F0 distribution;
A segment distribution conversion function, a duration distribution conversion function, and an F0 distribution conversion function are extracted from the storage unit, and the segment distribution, duration distribution, and F0 distribution obtained from the speech synthesis data using these distribution conversion functions, respectively. A parameter distribution conversion unit for converting
An evaluation function that calculates an evaluation function using the segment distribution after conversion, duration distribution, and F0 distribution, and evaluates the utterance text set candidate,
An utterance text set generation device characterized by the above. The utterance text set generation device according to claim 7,
A phonological labeling unit that assigns a phonological label to the reading-tone natural voice data that reads a predetermined document in a reading-tone and the X-tone natural voice data that reads the same document in an X-tone;
A first unit distribution extraction unit that refers to the whole speech unit variation storage unit, assigns a speech unit number to a speech unit obtained from each natural speech data, and extracts a unit distribution;
A first duration distribution extractor for calculating a phoneme duration for each speech unit and extracting a duration distribution;
A first F0 distribution extraction unit that extracts a fundamental frequency pattern for each speech unit and extracts an F0 distribution;
A segment distribution conversion function for converting a segment distribution, duration distribution, and F0 distribution obtained from natural speech data in reading tone into a segment distribution, duration distribution, and F0 distribution obtained from natural speech data in X tone, A parameter distribution conversion function calculation unit for calculating a continuation length distribution conversion function and an F0 distribution conversion function,
The storage unit stores the element distribution conversion function, the duration distribution conversion function, and the F0 distribution conversion function before creating the utterance text set candidate.
An utterance text set generation device characterized by the above.   A utterance text set generation program for causing a computer to function as the utterance text set generation device according to any one of claims 5 to 8.

Images