JP4533255B2 - Speech synthesis apparatus, speech synthesis method, speech synthesis program, and recording medium therefor - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a waveform-connecting voice synthesizier/method that can obtain synthesized voices having natural intonations. <P>SOLUTION: This synthesizier stores a voice waveform database and the voice information database, consisting of the entry of voice F<SB>0</SB>pattern detail information and the rhythm information containing the voice F<SB>0</SB>pattern general information in a memory, creates the rhythm information A containing F<SB>0</SB>pattern information from the phoneme sequence obtained by analyzing the inputted text; then calculates the distance scale following the phoneme sequence between the rhythm information A and the rhythm information B of the entry in the voice information database (including the calculation of the cost of the F0 pattern information in the rhythm information A and the F0 pattern general information in the rhythm information B.); chooses the entry having the rhythm information making the calculation result minimum from the voice information database; reads the voice waveform data from the voice waveform database following this chosen entry; and finally synthesizes the voice by connecting them together. <P>COPYRIGHT: (C)2007,JPO&amp;INPIT

Description

  The present invention relates to a speech synthesizer, a speech synthesis method, a speech synthesis program, and a recording medium for inputting text and outputting speech corresponding to the text, and more specifically, by selecting and connecting speech waveform data. The present invention relates to a waveform connection type speech synthesis technology for speech synthesis.

  In recent speech synthesis technology, a speech waveform database is constructed from a large amount of real voice data of several tens of minutes to several tens of hours, and a speech waveform of an appropriate length is generated from the speech waveform database on an appropriate basis according to the input text. A waveform-connected speech synthesis method has been proposed in which synthesized speech is created by selecting and connecting them (see Patent Document 1).

A configuration example of a speech synthesizer in such a waveform connection type speech synthesis method is shown in FIG.
The speech synthesizer (1) includes an external storage device (2) such as a hard disk, a text analysis unit (10), a prosody generation unit (11), a speech waveform selection unit (12), and a speech synthesis unit (13). Yes.

  Describe in more detail. The speech synthesizer (1) is a speech synthesizer that receives text as input and outputs synthesized speech. The external storage device (2) stores a speech waveform database (3) and a speech information database (4). The speech waveform database (3) performs a well-known A / D conversion on speech data read out from a word or sentence, and is cut out by a suitable synthesis unit (for example, phoneme) when assembling synthesized speech (speech waveform element) Audio waveform data as a piece) and stored in the storage area of the external storage device (2).

For example, as shown in FIG. 2, the speech information database (4) has a data structure (table) composed of entries in which a unit (synthesis unit) appropriate for assembling synthesized speech is set as a phoneme and various information is associated with the phoneme. And is stored in the storage area of the external storage device (2). Each entry of the speech information database (4) shown in FIG. 2 includes a speech waveform segment number that is a serial number of the speech waveform segment, phoneme label information that indicates the utterance content, phoneme duration information that indicates the speech duration of the phoneme, and phonemes. Power information obtained by normalizing the average power of the section, F ₀ pattern information representing the time transition of the phoneme pitch, and information indicating the position of the speech waveform data in the speech waveform database (3) (hereinafter, speech Called waveform data position information).
An entry of the voice information database (4) and each voice waveform data (as a voice waveform segment) in the voice waveform database (3) are associated with each other by voice waveform data position information in the voice information database (4).
F ₀ pattern fine information of each entry represents a F ₀ pattern was kept fine variations in F ₀ pattern having a real voice.

  The text analysis unit (10) performs morphological analysis on the input text and outputs a phoneme string and an accent type corresponding to the input text.

Prosody generation unit (11) as input information text analyzer (10) is outputted, phonemes per voice F ₀ pattern (fundamental frequency pattern), the phoneme duration (the length of the phoneme uttered), power Estimate information (sound volume) and output it. Here, “estimation” means that information necessary for speech synthesis (F ₀ pattern, phoneme duration, power information) is determined to be specific.

Speech waveform selector (12) in accordance with a sequence of phoneme string text analyzer (10) has outputted, and outputs in the prosody generation unit (11), F ₀ pattern of phonemes each speech phoneme duration length, power information A combination of speech waveform segments (optimal speech waveform segments) that minimize distortion with these targets and minimize the connection distortion between speech waveform segments when speech waveform segments are connected. Column) is selected from the speech information database (4), and each speech waveform segment number of the optimal speech waveform segment sequence (corresponding to the sequence of phoneme sequences output by the text analysis unit (10)) is output. To do. Dynamic programming or the like is used to determine the optimum speech waveform segment sequence.

The speech synthesis unit (13) receives each speech waveform segment number of the optimum speech waveform segment sequence selected by the speech waveform selection unit (12) as input, and each speech waveform segment number of this optimum speech waveform segment sequence. Is read from the speech waveform database (3) (refer to the speech waveform data position information), and the speech waveform data is sequentially connected to generate continuous speech, which is output as synthesized speech. .
Japanese Patent No. 2761552

The speech waveform data stored in the speech waveform database is a real voice. The real voice F ₀ pattern fluctuates finely, and as shown in the schematic diagram of FIG. 3, the F ₀ pattern (reference numeral 101 in FIG. 3) particularly in the consonant part (corresponding to the part of / R / in FIG. 3). In many cases, it has a fine structure such that

On the other hand, F ₀ pattern of the target obtained by the prosody generation part, because it is not a reflection of the fine variations in F ₀ pattern of real voice, in the selection process of the speech waveform segments in the speech waveform selector 12, the target F _{Between the 0} pattern (not reflecting the fine fluctuation in the real voice F ₀ pattern) and the F ₀ pattern of the selected speech waveform segment (reflecting the fine fluctuation in the real voice F ₀ pattern). May be inconsistent.

Therefore cause sound quality degradation due to a gap or the like of the F ₀ patterns in unnatural and connecting points of the speech waveform element pieces intonation, synthesized speech having audibility appropriate intonation there is a problem that not generated.

This typical example will be described with reference to FIGS. 4 and 5 which are schematic diagrams. 4 and 5, reference numeral 102 denotes a target _F0 pattern. Here, as shown in FIG. 4, the speech waveform segments with a continuous _{F 0} pattern (reference numeral 103a, 103b, 103c) is desirably selected. This is because the connection gap is not smooth, and reproduces the F ₀ pattern having a fine structure conforming to real voice, because speech having natural intonation is considered a possible synthesis. However, according to conventional speech synthesis method, F ₀ pattern of the target, because it is not a reflection of the fine variations in F ₀ pattern of real voice, the consonant portion (real voice) distance between F ₀ pattern is large becomes therefore, as shown in FIG. 5, although the strain of _{F 0} pattern is small, _{F 0} pattern discontinuous voice waveform segments (code 104a, 104b, 104c) from being selected.

  In view of the above problems, an object of the present invention is to provide a waveform-connected speech synthesizer, a speech synthesis method, a speech synthesis program, and a recording medium for obtaining synthesized speech having natural intonation.

In order to solve the above-described problems, the present invention provides a speech waveform database in which speech waveform data is collected, speech F ₀ pattern information (F ₀ pattern fine information holding minute fluctuations in the real voice F ₀ pattern, and F _0. A speech information database comprising entries indicating correspondence between prosodic information including speech waveform data in speech waveform database and F ₀ pattern outline information in which fine variation portions in pattern fine information are corrected. Storing and analyzing the input text to generate a phoneme sequence, generating prosody information A including the F ₀ pattern information of speech for each synthesis unit from the phoneme sequence, and then prosodic information according to the phoneme sequence A distance measure (cost) between A and the prosodic information B of the entry in the speech information database is calculated (F in prosodic information A Including cost calculation of ₀ pattern information and F ₀ pattern outline information in prosodic information B.), an entry having prosodic information that minimizes the calculation result is selected from the speech information database, and the selected entry Then, the voice waveform data is read from the voice waveform database, and the voice waveform data is connected to synthesize voice.

Further, rather than leave it generates advance F ₀ pattern envelope information, each time the speech synthesis process of generating a synthesized speech from the text, may be generated a F ₀ pattern approximate shape information.

Further, in addition to the costs of prosodic information A and prosodic information B, the cost between each entry is calculated (the cost between each entry is calculated by calculating the cost between at least the F ₀ pattern fine information in each entry. In addition, an entry having prosodic information that minimizes the calculation result may be selected from the speech information database.

  The computer can be operated as a speech synthesizer by a speech synthesis program that causes the speech synthesizer of the present invention to function on the computer. A computer-readable program recording medium that records this speech synthesis program enables other computers to function as a speech synthesizer or distribute the speech synthesis program.

According to the present invention, in the calculation of the distance measure (cost) between the prosodic information A generated from the phoneme sequence obtained by text analysis and the prosodic information B of the entry in the speech information database, the F ₀ pattern information in the prosodic information A When, by including the cost of operation of the F ₀ pattern envelope information in prosody information B, and mismatch F ₀ pattern between F ₀ pattern of the speech waveform segments are selected as target of the F ₀ pattern Since it can be avoided, it is possible to obtain a synthesized speech having natural intonation in which sound quality deterioration caused by intonation, particularly accent-type unnaturalness is reduced.

In addition to the cost of the prosodic information A and the prosody information B, it calculates the costs between each entry, in the calculation of the costs between the respective entry, the inclusion of operation costs between F ₀ pattern fine information in each entry Thus, it is possible to avoid the F ₀ gap at the connection location between the speech waveform segments, and thus it is possible to obtain a synthesized speech having a natural intonation with reduced sound quality degradation caused by discontinuity of the F ₀ pattern. Become.

<First Embodiment>
Hereinafter, a first embodiment of a speech synthesizer and method according to the present invention will be described.
FIG. 6 is a hardware configuration diagram illustrating a hardware configuration of the speech synthesizer according to the first embodiment.
FIG. 7 is a functional configuration diagram illustrating the functional configuration of the speech synthesizer according to the first embodiment.
FIG. 8 is a diagram showing a processing flow of speech synthesis according to the first embodiment.
FIG. 9 is a diagram showing a data configuration of the voice information database according to the first embodiment.
FIG. 10 is a diagram (part 1) illustrating an example of a method for generating F ₀ pattern outline information.
FIG. 11 is a diagram (part 2) illustrating an example of a method for generating F ₀ pattern outline information.
FIG. 12 is a diagram (No. 3) illustrating an example of a method for generating F ₀ pattern outline information.

  As illustrated in FIG. 6, the speech synthesizer (500) communicates with an input unit (51) to which a keyboard or the like can be connected, an output unit (52) to which a liquid crystal display or the like can be connected, and the speech synthesizer (500). A communication unit (53) to which a possible communication device (for example, a communication cable) can be connected, and a CPU (Central Processing Unit) (54) [may include a cache memory or the like. ], A RAM (55) as a memory, an external storage device (57) as a ROM (56) and a hard disk, and an input unit (51), an output unit (52), a communication unit (53), a CPU (54), The bus (58) is connected so that data can be exchanged between the RAM (55), the ROM (56), and the external storage device (57). If necessary, the speech synthesizer (500) may be provided with a device (drive) that can read and write a storage medium such as a CD-ROM.

  The text input to the speech synthesizer (500) may be input from the input unit (51). In this embodiment, it is assumed that the text is stored in advance in the external storage device (57). . In the present invention, the type of text is not particularly limited. In this embodiment, Japanese text mixed with kanji and kana is used.

  The external storage device (57) of the speech synthesizer (500) stores and stores a program for speech synthesis, data necessary for processing of this program, and the like. Further, data obtained by the processing of these programs is appropriately stored and stored in a RAM or an external storage device.

More specifically, in a predetermined storage area of the external storage device (57), as data necessary for processing of the program, a synthesis unit (in this embodiment, a phoneme. In addition, for example, a pitch or CV of a voice is used. it is also possible to syllables or the like as a unit.) speech waveform database a collection of audio waveform data for each (571) and the voice waveform data in prosody information and speech waveform database including F ₀ pattern information of phonemes each voice A voice information database (572) consisting of entries indicating the correspondence with is stored.

  Speech waveform data (as a speech waveform segment) in the speech waveform database (571) performs known A / D conversion on real speech data read out from words and sentences, and appropriately applies a window function, etc. Is extracted in units of phonemes.

For example, as shown in FIG. 9, the voice information database (572) has a data structure (table) including entries in which various information is associated with each phoneme as a unit. Each entry in the speech information database (572) shown in FIG. 9 includes a speech waveform segment number that is a serial number of the speech waveform segment, phoneme label information that indicates the utterance content, phoneme duration information that indicates the speech duration of the phoneme, Power information obtained by normalizing the average power of the phoneme section, F ₀ pattern information representing the time transition of the phoneme pitch (frequency), and information indicating the position of the speech waveform data in the speech waveform database (571) (Hereinafter referred to as voice waveform data position information). The entry of the speech information database (572) and each speech waveform data (as a speech waveform segment) in the speech waveform database (571) are associated with each other by speech waveform data position information in the speech information database (572).

The F ₀ pattern information of each entry in the audio information database (572) is composed of F ₀ pattern fine information and F ₀ pattern outline information. The F ₀ pattern fine information represents an F ₀ pattern that retains the fine variation of the F ₀ pattern possessed by the real voice as it is. On the other hand, F ₀ pattern envelope information indicates F ₀ pattern obtained by correcting the fine portion of the change in the F ₀ pattern in F ₀ pattern fine information.

Here, an example of a method of generating a _{F 0} pattern envelope information with reference to FIG. 10, 11 and 12 will be described. F ₀ pattern envelope information is generated by correcting the fine portion of the change in the F ₀ pattern in F ₀ pattern fine information, more specifically, F ₀ according to the consonants F ₀ pattern in F ₀ pattern fine information It is generated by correcting the fine variation portion of the pattern.

Hereinafter, as an example, by removing the variable fine movement of F ₀ pattern consonant segment (here / R / interval), it describes the processing of obtaining the F ₀ pattern in F ₀ pattern approximate shape information. The code | symbol 201 of FIG. 10 has shown the _F0 pattern of phoneme / A // R // U / of a certain audio | voice.

First, each vowel section of each side of the consonant segment (here / A / and / U / intervals) obtaining a peak point indicating the value of the highest F ₀ pattern within. The peak point can be obtained by reference to the F ₀ pattern fine information and the phoneme duration in the entry. 11, a / A / interval point indicated by reference numeral 202, the / U / of interval point indicated by reference numeral 203, a peak point of the value of the highest F ₀ pattern for each vowel section.

Next, linear interpolation between the peak points obtained in each vowel section is performed. In this example, the broken line indicated by reference numeral 204 in FIG. 11 shows an F ₀ pattern obtained by performing linear interpolation. Although the correction method is linear interpolation, the correction method is not limited to this, and may be corrected by, for example, spline interpolation. Through the above processing, it is possible to obtain _{F 0} pattern as shown in FIG. 12 (reference numeral 205a, 205b, 205c) a. _{F 0} pattern obtained here (reference numeral 205a, 205b, 205c) is an _{F 0} pattern envelope information for each phoneme.

Incidentally, as is apparent from the above description, F ₀ pattern to be corrected, rather than just F ₀ pattern consonant portion, a portion of the F ₀ pattern vowel portions (As far in the above example, / A / of F ₀ pattern portion extending the end point from the peak point, / U / F ₀ pattern portion extending the peak point from the beginning of) must also be noted that it may be corrected.

In addition, the external storage device (57) of the speech synthesizer (1) analyzes at least speech for each phoneme from a program for realizing a text analysis unit that analyzes input text and generates a phoneme sequence, and a phoneme sequence. A program for realizing a prosody generation unit that generates prosody information including F ₀ pattern information, a distance measure (cost) between the prosody information and the prosodic information of the entry in the speech information database is calculated, and the calculation result is minimized. A program for realizing a speech waveform selection unit that sequentially selects entries having prosodic information from the speech information database, reads speech waveform data from the speech waveform database according to the sequentially selected entries, and connects these speech waveform data to generate speech A program or the like for realizing a speech synthesizer for synthesizing is stored and stored. In addition, a control program for controlling processing based on these programs is also stored as appropriate.

  In the speech synthesizer (500) according to the first embodiment, each program stored in the external storage device (57) and data necessary for processing each program are read into the RAM (55) as necessary. Interpretation is executed and processed by the CPU (54). As a result, the CPU (54) realizes predetermined functions (text analysis unit, prosody generation unit, speech waveform selection unit, speech synthesis unit), thereby realizing speech synthesis.

Next, the flow of speech synthesis in the speech synthesizer (500) will be sequentially described with reference to FIGS.
The speech synthesis apparatus (500) of the first embodiment includes a text analysis unit (541), a prosody generation unit (542), a speech waveform selection unit (543), and a speech synthesis unit (544) (see FIG. 7). .

  First, the text analysis unit (541) reads the text stored in the external storage device (57), morphologically analyzes the read text, and stores the phoneme string, accent type, and exhalation paragraph (phrase) position corresponding to the text. A phoneme sequence representing the above is output (step S1).

The outline of the morphological analysis will be described. The text analysis unit (541) stores a word model, a kanji-kana conversion model, a kana phoneme conversion model, etc. (these are also stored in the external storage device (57) as necessary). Refer to and convert text to phoneme string. In addition, when the text is in Japanese, if a plurality of words are gathered to form a phrase or the like, a phenomenon such as an accent moving or disappearing occurs. Therefore, these rules (accent combining rules) are stored in advance as data, for example, an external storage device (57) and the text analysis unit (541) determines the accent type of the text according to the accent combination rule. Further, when the text is in Japanese, there is a characteristic tendency that one accent is attached to each semantic or grammatical unit, so these rules (phrase rules) are stored in advance as data in, for example, the external storage device (57). In addition, the text analysis unit (541) determines, as an exhalation paragraph, a combination of several groups with one accent according to this phrase rule.
The outline of the morpheme analysis described here is an example of morpheme analysis, and is not intended to exclude other morpheme analysis methods. In the speech synthesizer / method of the present invention, various morphological analysis can be used, but since these are achieved by known techniques, the details thereof are omitted.

Prosody generation unit (542) is input with information text analyzer (541) is output (phoneme sequence), F ₀ pattern (fundamental frequency pattern) of phonemes each speech phoneme duration (phonemes uttered length The power information (sound volume) is estimated and output (step S2). The phoneme duration length and the power information are appropriately set according to the phoneme position in the exhalation paragraph, the utterance speed, the phoneme environment before and after the phoneme, etc., which are regulated in advance. As for the F ₀ pattern, determined by the so-called Fujisaki model. As described above, “estimation” means that information necessary for speech synthesis (F ₀ pattern, phoneme duration length, power information) is determined to be specific. In the speech synthesizer / method of the present invention, the prosodic information can be generated by using a known prosodic information generating method, and the details thereof are omitted.

The speech waveform selection unit (543), according to the arrangement order of the phoneme sequence output by the text analysis unit (541), the F ₀ pattern of the speech for each phoneme output by the prosody generation unit (542), the phoneme duration length, A combination of speech waveform segments (optimal speech) that uses power information as a target and minimizes distortion with these targets and minimizes connection distortion between speech waveform segments. (Speech segment sequence) is selected from the speech information database (572), and each speech waveform number of the optimal speech waveform segment sequence (corresponding to the order of the phoneme sequence output by the text analysis unit (541)) is selected. Output (step S3). Hereinafter, the distance measure defined from the distortion is referred to as cost. Dynamic programming or the like is used to determine the optimum speech waveform segment sequence with the lowest cost.

  The selection of the optimum speech waveform segment sequence in the speech waveform selection unit (543) will be described in further detail. Here, a case where the optimum speech waveform segment sequence is selected in units of phonemes will be described. Further, it is assumed that not all phoneme strings (for text) output by the text analysis unit (541) are input to the speech waveform selection unit (543) for each phoneme string corresponding to one exhalation paragraph. This is because there is no need to consider the connection between the end phoneme of a certain exhalation paragraph and the start point phoneme of the next exhalation paragraph connected to this exhalation paragraph. Of course, depending on the length of the text and the like, all the phoneme strings (for the text) output by the text analysis unit (541) may be input to the speech waveform selection unit (543).

Hereinafter, the speech waveform selector (543) phoneme sequence (hereinafter, the target phoneme string and called.) Input within a represents the i-th target phoneme t _i, read from the speech information database (572) Of the speech waveform segment sequences of the entries (hereinafter referred to as candidate segment sequences), the i-th candidate segment is represented by u _i . Note that the candidate segment u _i read from the speech information database (572) has the same phoneme as the phoneme of the target phoneme t _i with reference to the phoneme label item of the speech information database (572).

The speech waveform selection unit (543) is a combination of candidate segments that minimizes the distortion between the target phoneme sequence and the candidate segment sequence and the connection distortion between the candidate segment sequences to be connected in one whole exhalation paragraph. In order to determine (optimum candidate segment sequence), a distance measure representing the distortion between the target phoneme t _i and the candidate segment u _i is obtained as the cost C (t _i , u _i ) for each phoneme.

As an example, the cost _{C (t i, u} i) and, as a weighted sum of the later-described various _{sub-cost, C (t i, u i} ) = Wtf · Stf (t i, u i) + Wtdur · Stdur (t i, u _i ) + Wtpow · Spow (t _i , u _i ) + Wcf · Scf (u _i−1 , u _i ) + Wcpow · Scpow (u _i−1 , u _i ) + Wcenv · Scenv (u _i−1 , u _i ).・ ・ Defined as (1).

Stf _(t i, _{u i)} is the target phoneme _{t i} represents the distortion between _{F 0} patterns in _{F 0} pattern envelope information _{F 0} pattern and the candidate segment _{u i} of the _{F 0} pattern _{t i} Ft (t _i), when the _{(F 0} pattern envelope) _{F 0} patterns in information _{u i} Fu and _{(u i),} the difference of the squares Stf of Ft _{(t i)} and _{Fu (u i) (t i} , u i ) = {Ft (t _i ) −Fu (u _i )} ² Hereinafter, this is referred to as target F ₀ sub-cost.
In the prior art, Fu _{(u i)} is _{F 0} pattern in _{F 0} pattern fine information _{u i,} in the present invention, noted that it is _{F 0} pattern in _{F 0} pattern envelope information _{u i} To do.

Stdur (t _i , u _i ) represents the distortion of the duration between the target phoneme t _i and the candidate segment u _i , the duration of t _i is DURt (t _i ), and the duration of u _i Let DURu (u _i ) be the square of the difference between DURt (t _i ) and DURu (u _i ) Stdur (t _i , u _i ) = {DURt (t _i ) −DURu (u _i )} ² . Hereinafter, this is referred to as a target duration long sub-cost.

Stpow _{(t i, u} i) represents the distortion of the power between the target phoneme _{t i} and the candidate segment _{u i,} power the POWt _(t i) of _{t i,} POWu power of _{u i (u i)} , It is assumed that the square of the difference between POWt (t _i ) and POWu (u _i ) is Spow (t _i , u _i ) = {POWt (t _i ) −POWu (u _i )} ² . Hereinafter, this is referred to as a target power sub-cost.

Scf (u _i−1 , u _i ) represents the distortion of the F ₀ pattern (in the F ₀ pattern fine information, respectively) at the connection point between the candidate element u _i and the preceding candidate element u _i−1. the value of _{F 0} pattern of the start point of the _{i FSu} (u _i), when the value of _{F 0} pattern of the end point of the _{u i-1} and _{fEu (u i-1),} FSu (u i) and feu _{(u i −1} ) squared difference Scf (u _i−1 , u _i ) = {FSu (u _i ) −FEu (u _i−1 )} ² Hereinafter, this is referred to as connection F ₀ sub-cost.
Incidentally, _{F 0} pattern in this connection _{F 0} subcosts is noted that it is _{F 0} pattern in _{F 0} pattern fine information.

Scpow (u _i−1 , u _i ) represents the power distortion at the connection point between the candidate element u _i and the preceding candidate element u _i−1 , and the power of the starting point of u _i is represented by POWSu (u _i ). , U _{i−1 where} the end point power is POWEu (u _i−1 ), the square of the difference between POWSu (u _i ) and POWEu (u _i−1 ) Scow (u _i−1 , u _i ) = { POWSu (ui) −POWEu (u _i−1 )} ² . This is hereinafter referred to as connection power sub-cost.

Scenv and _{(u i-1, u i)} is the candidate segment _{u i,} represents the difference between the preceding candidate segment _{u i-1} of the phoneme environment, the _{u i} preceding phoneme and _{u i-1} of the subsequent phoneme and the It is defined from acoustic similarity (for example, spectral similarity). Hereinafter, this is referred to as a connected phoneme environment sub-cost. For example, if the subsequent phonemes and matching of the preceding phoneme and _{u i-1} of _{u i,} is _{Scenv (u i-1, u} i) = 0. For example, these values are previously ordered as acoustic similarity database, if necessary, from the acoustic similarity database, acoustic similarity between subsequent phoneme preceding phoneme and u _i-1 of u _i Read the value corresponding to.

Further, Wtf is Stf _{(t i, u} i) weight for, Wtdur the Stdur _{(t i, u} i) weight for, Wtpow the Stpow _{(t i, u} i) weight for, Wcf is Scf _{(u i-1,} The weight for u _i ), Wcpow is the weight for Scoop (u _i−1 , u _i ), and Wcenv is the weight for Scenv (u _i−1 , u _i ).

Among the sub-costs described above, Stf (t _i , u _i ), Stdur (t _i , u _i ), and Stpow (t _i , u _i ) are the target information (F ₀ pattern) obtained by the prosody generation unit (542), Phoneme duration length, power information) and the sub cost obtained from the difference between the F ₀ pattern, phoneme duration length, and power information of the candidate unit u _i .

Scf (u _i−1 , u _i ), Scpow (u _i−1 , u _i ), and Scenv (u _i−1 , u _i ) are F ₀ patterns, power information, and phonemes between candidate segments. This is a sub-cost required for environmental differences.

The F ₀ pattern, duration length, and power information of the candidate segment u _i necessary for the calculation of each sub-cost can be obtained from the voice information database (572). And the cost C with respect to the phoneme string of one whole exhalation paragraph is calculated | required by Formula (2). Here, N represents the number of phonemes in one exhalation paragraph.

  The speech waveform selection unit (543) selects an optimal speech waveform segment sequence for a target of one expiratory paragraph by obtaining an optimal candidate segment sequence that minimizes C by a method such as dynamic programming. Then, each speech waveform segment number of the optimal speech waveform segment sequence (corresponding to the arrangement order of the phoneme sequences input to the speech waveform selection unit (543)) is obtained. Then, the speech waveform selection unit (543) performs the same processing for all expiratory paragraphs (that is, all phoneme sequences for text), and optimal speech waveform segment sequences (optimal speech waveforms) corresponding to all phoneme sequences. The number of each speech waveform segment in the segment sequence) is output.

  The speech synthesis unit (544) receives the speech waveform segment number sequence of the optimum speech waveform segment sequence selected by the speech waveform selection unit (543) as input, and outputs speech waveform data corresponding to each number (by each number). A voice is read from the voice waveform database (571) with reference to the voice waveform data position information of the identified entry, and the voice waveform data is sequentially connected to generate a continuous voice, which is output as a synthesized voice (step S). ). The speech synthesis method in the speech synthesis unit (544) is realized by a known technique such as a waveform superposition method.

By using the F ₀ pattern in the F ₀ pattern outline information of the candidate segment as the target F ₀ sub cost of the cost calculation by the speech waveform selection unit (543), as shown in the schematic diagrams of FIG. 13 and FIG. Speech waveform segments having F ₀ patterns (reference numerals 302a, 302b, and 302c in FIG. 14) in the F ₀ pattern outline information close to the F ₀ pattern (reference numeral 301 in FIG. 13) can be easily selected.

Moreover, as in this embodiment, selectively used type of F ₀ pattern information in accordance with sub-cost (i.e., the target F ₀ subcost using F ₀ pattern in F ₀ pattern envelope information, connection F ₀ to subcost F The advantage of using the ₀ pattern fine information F ₀ pattern) will be described with reference to FIGS. 15 and 16. FIG. 15 and 16, respectively, and setting the candidate of the speech waveform segment for F ₀ pattern of the same target. Further, FIG. 15, both 16, 401 denotes the _{F 0} pattern of the target.

Code 402a, 402b, 402c in FIG. 15 is a _{F 0} pattern in _{F 0} pattern fine information of a voice waveform segments, code 403a, 403b, 403c, respectively, reference numerals 402a, 402b, _{F 0} pattern outline of 402c This is the _F0 pattern in the shape information. Further, a _{F 0} pattern in the code 404a, 404b, 404c is _{F 0} pattern fine information of the speech waveform segments with 16, reference numeral 405a, 405 b, 405c, respectively, reference numerals 404a, 404b, _F of 404c ₀ pattern it is _{F 0} pattern in approximate shape information.

In the case of using _{F 0} pattern in _{F 0} pattern fine information in connection _{F 0} sub-cost, reference numeral 402a, 402b, 402c, 404a, 404b, F 0 pattern shown in 404c is used in the calculation of the sub-costs. On the other hand, in the case of using the _{F 0} pattern in _{F 0} pattern envelope information in the connection _{F 0} sub-cost, consists codes 403a, 403b, 403c, 405a, 405 b, _{that F 0} pattern shown in 405c is used in the calculation of the sub-costs .

In this example, when the connection F ₀ sub-cost is obtained using the F ₀ pattern in the F ₀ pattern outline information (in FIG. 15, the connection point between the reference numerals 403a and 403b and the connection point between the reference numerals 403b and 403c. In FIG. 16, the connection point between reference numeral 405a and reference numeral 405b, and the connection point between reference numeral 405b and reference numeral 405c.) As can be seen from FIGS. 15 and 16, the values are similar, but the / U / part the distortion of the _{F 0} pattern for the target for smaller for voice waveform segments shown in FIG. 15, _{F 0} pattern shown in FIG. 15 (reference numeral 403a, 403b, 403c) voice waveform segment from being selected for Probability is high.

Then, the speech synthesis unit, reference numeral 403a, 403b, _{F 0} pattern (which is a _{F 0} pattern in _{F 0} pattern approximate shape information.) Of 403c speech waveform data corresponding to the entry having the _{F 0} pattern information It will read and connect. However, these sound waveform data, the code 402a, 402b, _{F 0} pattern (which, _F is ₀ pattern. At _{F 0} pattern fine information) 402c where with properties corresponding to, / A / and / R / There is a significant connection distortion between the phonemes, and the synthesized speech synthesized with such speech waveform data becomes unnatural with a loss of smoothness. That is, in this case, if the speech waveform segment shown in FIG. 16 is selected, the difference in the connection point of the real voice F ₀ pattern is smaller, and therefore the quality of the synthesized speech (smoothness and naturalness) becomes higher. Conceivable.

Therefore, in order not to lose the smoothness and naturalness of audibility of synthesized speech, the connection F ₀ sub-cost, it is to be referred to with the F ₀ pattern in F ₀ pattern fine information.

<Second Embodiment>
In the first embodiment, F ₀ pattern envelope information was stored as a component of pre-speech information database (572). On the other hand, in the second embodiment, from the viewpoint of saving the storage capacity of the external storage device, the speech synthesis process for generating synthesized speech from text, instead of generating F ₀ pattern outline information in advance. generating a F ₀ pattern envelope information for every.
The same functions and processes as those in the first embodiment are assigned the same reference numerals and description thereof is omitted, and only differences from the first embodiment are described.

The F ₀ pattern information of each entry in the voice information database (672) of the second embodiment is the F ₀ pattern fine information described in the first embodiment. In the second embodiment, the F ₀ pattern outline information described in the first embodiment is not a component of the F ₀ pattern information of each entry. That is, the voice information database (672) of the second embodiment has a data configuration as shown in FIG.

The external storage device (57) of the speech synthesizer (600) according to the second embodiment, in addition to the program described in the first embodiment, the F ₀ pattern in the F ₀ pattern fine information of each entry, F ₀ pattern A program for obtaining outline information is also stored and stored. In addition, a control program for controlling processing based on these programs is also stored as appropriate.

  In the speech synthesizer (600), each program stored in the external storage device (57) and data necessary for processing each program are read into the RAM (55) as necessary, and interpreted by the CPU (54). Executed and processed. As a result, the CPU (54) realizes predetermined functions (text analysis unit, prosody generation unit, outline information generation unit, speech waveform selection unit, speech synthesis unit), thereby realizing speech synthesis.

Next, the flow of speech synthesis in the speech synthesizer (600) will be sequentially described with reference to FIGS.
FIG. 17 is a functional configuration diagram illustrating the functional configuration of the speech synthesizer according to the second embodiment.
FIG. 18 is a diagram showing a processing flow of speech synthesis according to the second embodiment.
Figure 19 is a diagram depicting a processing flow of generating a F ₀ pattern approximate shape information.

  The speech synthesis apparatus (600) of the first embodiment includes a text analysis unit (541), a prosody generation unit (542), a rough shape information generation unit (645), a speech waveform selection unit (543), and a speech synthesis unit (544). (See FIG. 17).

  Since the process of step S1 and step S2 is the same as that of 1st Embodiment, description is abbreviate | omitted.

After step S2, outline information generating unit (645), reads the _{F 0} pattern information entry of a voice information database (672) _{(F 0} pattern fine information), _{F 0} from the _{F 0} pattern fine information Pattern outline information is generated (step S2a). The generation of the F ₀ pattern outline information is as described in the first embodiment, and will be described accordingly (see FIGS. 10, 11, and 12).

F ₀ pattern envelope information is generated by correcting the fine portion of the change in the F ₀ pattern in F ₀ pattern fine information, more specifically, F ₀ according to the consonants F ₀ pattern in F ₀ pattern fine information It is generated by correcting the fine variation portion of the pattern. So as an example, by removing the variable fine movement of F ₀ pattern consonant segment (/ R / interval), it describes the processing of obtaining the F ₀ pattern in F ₀ pattern approximate shape information.

First, the peak points indicating the value of the highest F ₀ pattern in both sides of each vowel section (/ A / and / U / interval) of consonant segment (/ R / interval) (step S2A1). The peak point can be obtained by reference to the F ₀ pattern fine information and the phoneme duration in the entry. 11, a / A / interval point indicated by reference numeral 202, the / U / of interval point indicated by reference numeral 203, a peak point of the value of the highest F ₀ pattern for each vowel section.

Next, linear interpolation between the peak points of each obtained vowel section is performed (step S2a2). The correction method is not limited to linear interpolation, but may be corrected by, for example, spline interpolation. A broken line 204 in FIG. 11 indicates an F ₀ pattern obtained by performing linear interpolation between the peak points of each vowel section. Through the above processing, it is possible to obtain _{F 0} pattern as shown in FIG. 12 (reference numeral 205a, 205b, 205c) a. _{F 0} pattern obtained here (reference numeral 205a, 205b, 205c) is an _{F 0} pattern envelope information for each phoneme.

Incidentally, as is apparent from the above description, F ₀ pattern (F ₀ pattern fine information) corrected by, instead of only F ₀ pattern consonant portion, a portion of the F ₀ pattern of the vowel portion (the above example as far in, / a / F ₀ pattern portion extending the end point from the peak point of the, / U / F ₀ pattern portion extending the peak point from the beginning of) must also be noted that it may be corrected.

Outline information generating unit (645) the generated _{F 0} pattern envelope information, stored in addition to _{F 0} pattern information of the corresponding phoneme entry (step S2a3).
That at this time, the F ₀ pattern information entry will include F ₀ patterns fine information and F ₀ patterns approximate shape information (see FIG. 9).

Incidentally, reads the F ₀ pattern information of all entries (F ₀ pattern fine information), be from F ₀ pattern fine information to generate F ₀ pattern envelope information is sometimes a tedious for each entry. Therefore, the outline information generation unit (645) refers to the phoneme label item in the speech information database (572) for the phonemes included in the phoneme string generated by the text analysis unit (541), and the F of the entry. It is also possible to read ₀ pattern information (F ₀ pattern fine information) and generate F ₀ pattern outline information for each phoneme.

  Since the process of step S3 and step S4 is the same as that of 1st Embodiment, description is abbreviate | omitted.

The speech synthesizer / method according to the present invention is not limited to the above-described embodiment, and can be appropriately changed without departing from the gist of the present invention. For example, as a sub cost in the cost calculation by the speech waveform selection unit, a sub cost indicating the degree of coincidence of inclination between the target F ₀ pattern and the candidate segment F ₀ pattern is also introduced (added to the right side of Expression (1)). , it may be used F ₀ pattern in F ₀ pattern envelope information in the calculation of its sub-costs. In addition, the processing described in the speech synthesizer / method is not only executed in chronological order according to the order described, but also executed in parallel or individually as required by the processing capability of the device that executes the processing. It is good.

  When the processing functions in the speech synthesizer are realized by a computer, the processing contents of the functions that the speech synthesizer should have are described by a program. Then, by executing this program on a computer, the processing functions in the speech synthesizer are realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording medium, MO (Magneto-Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from the portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, the speech synthesizer is configured by executing a predetermined program on a computer. However, at least a part of the processing contents may be realized by hardware.

  The speech synthesizer / method of the present invention is useful for text-to-speech conversion.

The figure which shows the function structural example of the conventional speech synthesizer. The figure which shows an example of an audio | voice information database. F ₀ pattern diagram showing an example of fine information. It illustrates an example of a F ₀ pattern of the speech waveform segment selection result. It illustrates an example of a F ₀ pattern of the speech waveform segment selection result. The hardware block diagram which illustrated the hardware configuration of the speech synthesizer concerning 1st Embodiment. The function block diagram which illustrated the function structure of the speech synthesizer concerning 1st Embodiment. The figure which shows the processing flow of the speech synthesis concerning 1st Embodiment. The figure which shows the data structure of the audio | voice information database concerning 1st Embodiment. It illustrates an example of a generation method of F ₀ pattern outline information (Part 1). It illustrates an example of a generation method of F ₀ pattern outline information (Part 2). It illustrates an example of a generation method of F ₀ pattern outline information (Part 3). The figure which shows an example of _F0 pattern of a target. It illustrates an example of a F ₀ pattern of the speech waveform segment selection result. It shows the relationship between F ₀ pattern targets F ₀ pattern and the voice waveform segments selected result (Part 1). It shows the relationship between F ₀ pattern targets F ₀ pattern and the voice waveform segments selected results (Part 2). The function block diagram which illustrated the function structure of the speech synthesizer concerning 2nd Embodiment. The figure which shows the processing flow of the speech synthesis concerning 2nd Embodiment. It shows a process flow of generating an F ₀ pattern approximate shape information.

Explanation of symbols

101 F ₀ pattern of speech waveform segment (F ₀ pattern fine information)
102 target of the _{F 0} pattern 103a speech waveform segment of the _{F 0} pattern _{(F 0} pattern fine information)
103b F ₀ pattern of speech waveform segment (F ₀ pattern fine information)
103c F ₀ pattern of speech waveform segment (F ₀ pattern fine information)
104a F ₀ pattern of speech waveform segment (F ₀ pattern fine information)
104b F ₀ pattern of speech waveform segment (F ₀ pattern fine information)
104c F ₀ pattern of speech waveform segment (F ₀ pattern fine information)
201 F ₀ pattern of speech waveform segment (F ₀ pattern fine information)
In the example of the method for generating 201 F ₀ pattern outline information, the peak point 203 having the highest value of the F ₀ pattern in the vowel section 203 In the example of the method for generating F ₀ pattern outline information, the F ₀ pattern in the vowel section is generated. Peak point 204 with the highest value 204 F ₀ pattern outline information generation method in one example of linear interpolation result 205a F ₀ pattern of generated speech waveform segment (F ₀ pattern outline information)
205b _{F 0} pattern of the generated voice waveform segments _{(F 0} pattern envelope information)
205c F ₀ pattern of generated speech waveform segment (F ₀ pattern outline information)
301 target of the _{F 0} pattern 302a speech waveform segment of the _{F 0} pattern _{(F 0} pattern envelope information)
302b F ₀ pattern of speech waveform segment (F ₀ pattern outline information)
302c voice演形segment of the _{F 0} pattern _{(F 0} pattern approximate shape information)
401 target of the _{F 0} pattern 402a speech waveform segment of the _{F 0} pattern _{(F 0} pattern fine information)
402b F ₀ pattern of speech waveform segment (F ₀ pattern fine information)
402c speech waveform segment of the _{F 0} pattern _{(F 0} pattern fine information)
403a F ₀ pattern of speech waveform segment (F ₀ pattern outline information)
403b F ₀ pattern of speech waveform segment (F ₀ pattern outline information)
403c F ₀ pattern of speech waveform segment (F ₀ pattern outline information)
404a F ₀ pattern of speech waveform segment (F ₀ pattern fine information)
404b speech waveform segment of the _{F 0} pattern _{(F 0} pattern fine information)
404c F ₀ pattern of speech waveform segment (F ₀ pattern fine information)
405a F ₀ pattern of speech waveform segment (F ₀ pattern outline information)
405b F ₀ pattern of speech waveform segment (F ₀ pattern outline information)
405c F ₀ pattern of speech waveform segment (F ₀ pattern outline information)
500 Speech synthesizer 541 Text analysis unit 542 Prosody generation unit 543 Speech waveform selection unit 544 Speech synthesis unit 571 Speech waveform database 572 Speech information database 600 Speech synthesizer 645 Outline information generation unit 672 Speech information database

Claims (8)

Generates a phoneme sequence of the text from the input text, reads speech waveform data corresponding to this phoneme sequence from the speech waveform database in an appropriate unit (synthesis unit) for assembling the synthesized speech, and extracts these speech waveform data A speech synthesizer that synthesizes speech by connecting
Text analysis means for analyzing the input text and generating a phoneme sequence of the text;
Prosody generation means for generating prosody information including at least F ₀ pattern information of speech for each synthesis unit from the phoneme sequence generated by the text analysis means;
Speech waveform database were collected sound waveform data for each synthesis unit and stores voice information database of entries indicating the correspondence between the speech waveform data in the prosodic information and the speech waveform database including F ₀ pattern information of the speech for each synthesis unit Storage means for
In accordance with the phoneme sequence generated by the text analysis means, at least the distance measure (cost) between the prosody information generated by the prosody generation means and the prosodic information of the entry in the speech information database is calculated, and the prosody that minimizes the calculation result Voice waveform selection means for selecting an entry having information from a voice information database;
A voice synthesis unit that reads voice waveform data from the voice waveform database according to the entry selected by the voice waveform selection unit, and connects the voice waveform data to synthesize a voice;
The voice _F0 pattern information for each synthesis unit in the voice information database entry is
F ₀ pattern fine information and holding the fine variation of F ₀ pattern of real voice, between the peak point showing the maximum value of F ₀ pattern of each vowel portion of both sides of the consonant portion that put the fine information F ₀ pattern fine information Consists of F ₀ pattern outline information generated by interpolation ,
The cost calculation in the speech waveform selection means includes at least the cost calculation of the F ₀ pattern information in the prosodic information generated by the prosody generation means and the F ₀ pattern outline information in the entry of the speech information database. A featured voice synthesizer. Generates a phoneme sequence of the text from the input text, reads speech waveform data corresponding to this phoneme sequence from the speech waveform database in an appropriate unit (synthesis unit) for assembling the synthesized speech, and extracts these speech waveform data A speech synthesizer that synthesizes speech by connecting
Text analysis means for analyzing the input text and generating a phoneme sequence of the text;
Prosody generation means for generating prosody information including at least F ₀ pattern information of speech for each synthesis unit from the phoneme sequence generated by the text analysis means;
Speech waveform database were collected sound waveform data for each synthesis unit and a voice information database of entries indicating the correspondence between the speech waveform data in the prosodic information and the speech waveform database including F ₀ pattern fine information of the speech for each synthesis unit Storage means for storing;
F ₀ pattern by interpolating between peak point showing the maximum value of F ₀ pattern fine information of each vowel portion of both sides of the consonant portions in the F ₀ pattern fine information of speech each synthesis unit in the entry of the audio information database Outline information generating means for generating outline information;
In accordance with the phoneme sequence generated by the text analysis means, at least the distance measure (cost) between the prosody information generated by the prosody generation means and the prosodic information of the entry in the speech information database is calculated, and the prosody that minimizes the calculation result Voice waveform selection means for selecting an entry having information from a voice information database;
A voice synthesis unit that reads voice waveform data from the voice waveform database according to the entry selected by the voice waveform selection unit, and connects the voice waveform data to synthesize a voice;
The cost of the operation in the audio waveform selecting means, at least, a F ₀ pattern information in prosody information generated by the prosody generation means, the cost of operation of the F ₀ pattern envelope information generated by the envelope information generator A speech synthesizer comprising: The voice waveform selection means
According to the phoneme sequence generated by the text analysis means, the distance measure (cost) between the prosodic information generated by the prosody generation means and the prosodic information of the entry in the speech information database and the cost between each entry are calculated. The entry having the minimum prosodic information is selected from the speech information database,
The calculation of the cost between each entry, speech synthesis apparatus according to claim 1 or 2, characterized in that it comprises the calculation of the costs between F ₀ pattern fine information in at least each entry. Generates a phoneme sequence of the text from the input text, reads speech waveform data corresponding to this phoneme sequence from the speech waveform database in an appropriate unit (synthesis unit) for assembling the synthesized speech, and extracts these speech waveform data Is a speech synthesis method for synthesizing speech by connecting
The storage unit, and the speech waveform database were collected sound waveform data for each synthesis unit, consisting of an entry showing the correspondence between the speech waveform data in the prosodic information and the speech waveform database including F ₀ pattern information of the speech for each synthesis unit Memorize voice information database,
A text analysis step in which the text analysis means analyzes the input text to generate a phoneme sequence of the text;
A prosody generation step in which prosody generation means generates prosody information including at least F ₀ pattern information of speech for each synthesis unit from the phoneme sequence generated in the text analysis step;
The speech waveform selection means, according to the phoneme sequence generated in the text analysis step, at least a distance measure (cost) between the prosody information generated in the prosody generation step and the prosodic information of the entry in the speech information database stored in the storage means A speech waveform selection step of selecting an entry having prosodic information that minimizes the computation result from a speech information database stored in the storage means;
The voice synthesis means includes a voice synthesis step of reading voice waveform data from the voice waveform database stored in the storage means according to the entry selected in the voice waveform selection step, and synthesizing voice by connecting the voice waveform data. ,
The voice F ₀ pattern information for each synthesis unit in the voice information database entry stored in the storage means is:
F ₀ pattern fine information and holding the fine variation of F ₀ pattern of real voice, between the peak point showing the maximum value of F ₀ pattern of each vowel portion of both sides of the consonant portion that put the fine information F ₀ pattern fine information Consists of F ₀ pattern outline information generated by interpolation ,
The cost calculation in the speech waveform selection step includes at least the cost calculation of the F ₀ pattern information in the prosody information generated in the prosody generation step and the F ₀ pattern outline information in the entry of the speech information database. A featured speech synthesis method. Generates a phoneme sequence of the text from the input text, reads speech waveform data corresponding to this phoneme sequence from the speech waveform database in an appropriate unit (synthesis unit) for assembling the synthesized speech, and extracts these speech waveform data Is a speech synthesis method for synthesizing speech by connecting
Storage means, and speech waveform database were collected sound waveform data for each synthesis unit, consisting of an entry showing the correspondence between the speech waveform data in the prosodic information and the speech waveform database including F ₀ pattern fine information of the speech for each synthesis unit Memorize voice information database,
A text analysis step in which the text analysis means analyzes the input text to generate a phoneme sequence of the text;
A prosody generation step in which prosody generation means generates prosody information including at least F ₀ pattern information of speech for each synthesis unit from the phoneme sequence generated in the text analysis step;
Envelope information generating means, the maximum value of the audio of F ₀ pattern of each vowel portion of both sides of the consonant portions in the fine information F ₀ pattern fine information for each synthetic unit in the entry of the audio information database stored in the storage means An outline information generation step for generating F ₀ pattern outline information by interpolating between peak points indicating
The speech waveform selection means, according to the phoneme sequence generated in the text analysis step, at least a distance measure (cost) between the prosody information generated in the prosody generation step and the prosodic information of the entry in the speech information database stored in the storage means ) And selecting an entry having prosodic information that minimizes the calculation result from the speech information database;
The voice synthesis means includes a voice synthesis step of reading voice waveform data from the voice waveform database stored in the storage means according to the entry selected in the voice waveform selection step, and synthesizing voice by connecting the voice waveform data. ,
The cost of the operation in the speech waveform selecting step, at least, a F ₀ pattern information in prosody information generated in the prosody generation step, the cost of operation of the F ₀ pattern outline information generated at the approximate shape information generating step A speech synthesis method comprising: The voice waveform selection step
In accordance with the phoneme sequence generated in the text analysis step, the distance measure (cost) between the prosodic information generated in the prosody generation step and the prosodic information of the entry in the speech information database and the cost between each entry are calculated. The entry having the minimum prosodic information is selected from the speech information database,
The cost of operation between each entry, speech synthesis method according to claim 4 or 5, characterized in that it comprises the calculation of the costs between F ₀ pattern fine information in at least each entry. A speech synthesis program for causing a computer to function as the speech synthesizer according to any one of claims 1 to 3 . A computer-readable program recording medium on which the speech synthesis program according to claim 7 is recorded.

Images