JP5366919B2 - Speech synthesis method, apparatus, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To maintain continuity of intonation of a synthetic speech to select a candidate phoneme so that the synthetic speech becomes high quality even for such a target that voiced sounds are connected to each other via voiceless sounds. <P>SOLUTION: A speech synthesis method includes: a preceding voiced sound candidate phoneme search sub-step of searching a candidate phoneme which is selected as a candidate phoneme suitable for a target to be the (i-k)-th voiced sound in which k is 2 or more and k is the minimum (hereafter, called as the (i-k)-th candidate phoneme) when the i-th target (i is a natural number equal to or more than 3) is the voiced sound, and the (i-1)-th target is a voiceless sound; a connection F0 sub-cost calculation sub-step of calculating a connection F0 sub-cost from an F0 value at a terminal position of the (i-k)-th candidate phoneme and F0 value groups at the head positions of the respective candidate phonemes of the i-th candidate phoneme group; and a selection sub-step of selecting one candidate phoneme suitable for the i-th target from the i-th candidate phoneme group based on the connection F0 sub-cost. <P>COPYRIGHT: (C)2012,JPO&amp;INPIT

Description

  The present invention relates to a speech synthesis method, apparatus, and program that are used, for example, in a speech waveform segment selection process in waveform-connected speech synthesis and enable speech synthesis with high quality intonation.

  In the field of speech synthesis in recent years, a large amount of tens of minutes to several tens of hours of real voice data is stored in a large-capacity storage device as a speech waveform database. A speech synthesis method and speech synthesizer have been proposed in which speech waveforms of appropriate lengths are cut out to form speech waveform segments and connected to create synthesized speech (Patent Document 1). Hereinafter, a conventional speech synthesizer 90 will be described in detail with reference to FIGS. FIG. 1 is a block diagram showing a configuration of a conventional speech synthesizer 90. FIG. 2 is a diagram illustrating a speech information index 952 provided in the conventional speech synthesizer 90. The conventional speech synthesizer 90 includes a text analysis unit 91, a prosody generation unit 92, a candidate segment selection unit 93, a speech synthesis unit 94, and a speech waveform database 95. The speech waveform database 95 includes speech waveform data 951 and a speech information index 952. The speech synthesizer 90 is a device that receives text information and outputs synthesized speech. The audio waveform database 95 is composed of a storage medium such as a hard disk, and stores audio waveform data 951 and an audio information index 952. The speech waveform data 951 stores speech waveform segments obtained by subjecting speech data read out from words and sentences to AD conversion. Since the speech waveform segments stored in the speech waveform data 951 are segments that are candidates for speech synthesis in speech synthesis, these speech waveform segments are hereinafter referred to as candidate segments. When two or more candidate segments are called together, it is called a candidate segment group. The audio information index 952 is a table composed of entries in units of phonemes as illustrated in FIG. Each entry corresponds to the speech waveform data 951, a segment number that is a serial number of the candidate segment, a phoneme label that classifies the phoneme of the candidate segment, a phoneme duration (ms) that indicates the duration of the candidate segment, F0 pattern information (Hz) representing the time transition of the pitch of the candidate segment, and segment data positions that are storage location information of each candidate segment in the speech waveform data 951 are provided. The segment data position specifically indicates a memory address of a hard disk that stores the voice waveform data 951 or the like. For example, the candidate segment with segment number 1 is data classified into the phoneme label “a”. In other words, the candidate element with the element number 1 is one of candidate elements classified into the phoneme label “a”. The phoneme duration length of the candidate unit number 1 is 85 (ms), and its F0 changes over time from 300 → 302 → 303 →... → 301 (Hz). Also, the candidate element with the element number 1 is stored in the element data position 0. Specifically, the candidate element with the element number 1 is stored at the memory address 0 of the hard disk that stores the audio waveform data 951. In addition, the candidate element of the element number 2 is one of candidate elements classified into the phoneme label “s”, and the phoneme “s” is an unvoiced sound without accompanying vocal cord vibration. When the unvoiced sound is uttered, the vocal cord vibration stops, so the unvoiced sound has no fundamental frequency (F0). For this reason, for example, -1 is stored as a numerical value indicating that F0 information does not exist in the F0 pattern information (Hz) of an entry that is an unvoiced sound, such as unit number 2.

  Next, referring to FIG. 3 in conjunction with FIG. 1 and FIG. 2 described above, the speech synthesis operation performed by the conventional speech synthesis apparatus 90 will be described in detail. FIG. 3 is a flowchart showing the operation of the conventional speech synthesizer 90. The description will proceed on the assumption that the text to be synthesized is input to the speech synthesizer 90. The text analysis unit 91 acquires the text input to the speech synthesizer 90, morphologically analyzes the acquired text, generates a phoneme string and an accent type from the morpheme analysis result, and generates the phoneme string and the accent. The type is output to the prosody generation unit 92 (S91). The prosody generation unit 92 acquires the phoneme string and accent type output from the text analysis unit 91, estimates the F0 pattern and phoneme duration for each phoneme, and calculates the estimated F0 pattern and phoneme duration. The length is numbered for each phoneme, and is output to the candidate segment selection unit 93 as a target (S92). The candidate segment selection unit 93 acquires the targets output by the prosody generation unit 92, and selects candidate segments that have a small distortion with these targets and a small connection distortion when the candidate segments are connected to each other. A combination is selected from the speech information index 952, and the segment number of the selected candidate segment is output to the speech synthesizer 94 (S93). The distance measure that defines the above distortion is called cost. The candidate segment selection unit 93 determines a combination of candidate segments that minimizes the cost. Dynamic programming or the like is used to determine the combination of candidate segments that minimizes the cost. Details of the cost calculation performed by the candidate segment selection unit 93 will be described later. The speech synthesizer 94 acquires the segment number of the candidate segment output by the candidate segment selector 93, reads the candidate segment corresponding to the segment number from the speech waveform data 951, and extracts these candidate segments. Connected to generate voice and output as synthesized voice (S94).

Next, the cost calculated by the candidate segment selection unit 93 will be specifically described. In the phoneme string output by the text analysis unit 91, the target of the jth phoneme (j is an integer equal to or greater than 1) is assumed to be t (j). As described above, the target includes F0 pattern information for each phoneme and information about the phoneme duration. Of the speech waveform segments stored in the speech waveform data 951, a candidate segment group corresponding to the target t (j) is represented as U (j). The candidate segment group U (j) represents all entries whose phoneme labels match the target t (j) among the entries stored in the speech information index 952. Arbitrary one candidate segment in the candidate segment group U (j) is represented as u (j) and used for the following description. A distance measure representing the distortion of t (j) and u (j) is set as a target cost Ct (t (j), u (j)). The target cost Ct (t (j), u (j)) is a weighted sum of various sub-costs described later.
Ct (t (j), u (j)) = Wtf ・ Stf (t (j), u (j)) + Wtdur ・ Stdur (t (j), u (j))
It is defined as Further, the connection cost representing the connection distortion between u (j-1) and u (j) is Cc (u (j-1), u (j)). The connection cost Cc (u (j-1), u (j)) is a weighted sum of various subcosts to be described later.
Cc (u (j-1), u (j)) = Wcf ・ Scf (u (j-1), u (j)) + Wcenv ・ Scenv (u (j-1), u (j))
It is defined as Wtf is the weight for Stf (t (j), u (j)), Wtdur is the weight for Stdur (t (j), u (j)), Wcf is Scf (u (j-1), u (j)) Wcenv is a weight for Scenv (u (j-1), u (j)). Stf (t (j), u (j)) represents the distortion of the F0 pattern between the target t (j) and the candidate segment u (j), and the F0 pattern of the target t (j) is expressed as Ft (t ( j)), when the F0 pattern of u (j) is Fu (u (j)), the square of the difference between Ft (t (j)) and Fu (u (j))
Stf (t (j), u (j)) = {Ft (t (j))-Fu (u (j))} ²
And This is hereinafter referred to as target F0 sub cost. If the candidate segment is an unvoiced sound, the target F0 subcost cannot be obtained because it does not have an F0 pattern, so the value of Stf (t (j), u (j)) is a constant value. (For example, 0). Stdur (t (j), u (j)) represents the distortion of the duration between the target phoneme t (j) and the candidate unit u (j), and the duration of t (j) is expressed as DURt ( t (j)), when the duration of u (j) is DURu (u (j)), the square of the difference between DURt (t (j)) and DURu (u (j))
Stdur (t (j), u (j)) = {DURt (t (j))-DURu (u (j))} ²
And Hereinafter, this is referred to as a target duration long sub-cost. Scf (u (j-1), u (j)) represents the distortion of F0 at the connection point between the candidate element u (j) and the preceding candidate element u (j-1), and u (j) Where Fu (u (j)) and FEu (u (j ()) are Fu (u (j)) and Fu (u (j-1)) is Fu (u (j-1)). -1)) squared difference
Scf (u (j-1), u (j)) = {FSu (u (j))-FEu (u (j-1))} ²
And Hereinafter, this is referred to as connection F0 sub-cost. Note that if either u (j-1) or u (j) or both are unvoiced sounds, the F0 sub-cost cannot be obtained because it does not have an F0 value, so Scf (u ( The values of j-1) and u (j)) are set to constant values (for example, 0). Scenv (u (j-1), u (j)) represents the difference between the phoneme environment before and after the candidate unit u (j) and the preceding candidate unit u (j-1), and the target t (j) And the phonetic label of the phoneme label that precedes u (j-1) in the phoneme information index 952 and the phoneme label that precedes u (j) in the target t (j-1) and the phoneme information index 952 Defined from acoustic similarity. Hereinafter, this is referred to as a connected phoneme environment sub-cost. The higher the acoustic similarity between the subsequent phonemes of the target t (j) and u (j-1) and the preceding phonemes of the target t (j-1) and u (j), the smaller the value of the sub-cost, For example, the phoneme label that follows u (j-1) in the phoneme information index 952 matches t (j), and t (j-1) precedes u (j) in the phoneme information index 952 If the phoneme labels match, Scenv (u (j-1), u (j)) = 0. Among these sub-costs, Stf (t (j), u (j)) and Stdur (t (j), u (j)) are F0 possessed by the candidate segment group for the target estimated by the prosody generation unit 92. This is a sub-cost consisting of the difference in pattern and phoneme duration. Scf (u (j-1), u (j)) and Scenv (u (j-1), u (j)) are sub-costs consisting of differences in F0 patterns and phoneme environments between candidate segments. is there. The F0 pattern and duration length of u (j) necessary for the calculation of the sub-cost can be obtained from the audio information index 952. When there are two or more candidate segments u (j) in the candidate segment group U (j) (that is, when there are two or more candidates for the same phoneme), the above calculation is repeated for the number of candidate segments. . Total cost C for the entire sentence to be synthesized

(Where N is the number of phonemes in the sentence to be synthesized), the candidate segment combination that minimizes C is determined by a method such as dynamic programming. The optimal candidate segment is determined.

  Next, the operation performed by the candidate segment selection unit 93 will be described in detail with reference to FIGS. FIG. 4 is a block diagram showing details of the candidate segment selection unit 93 provided in the conventional speech synthesizer 90. FIG. 5 is a flowchart showing the operation of the candidate segment selection unit 93 provided in the conventional speech synthesizer 90. The candidate segment selection unit 93 includes a target F0 sub cost calculation unit 931, a target duration length sub cost calculation unit 932, a connection F0 sub cost calculation unit 934, a connected phoneme environment sub cost calculation unit 935, a search hypothesis expansion unit 936, Selection means 937. The target F0 sub cost calculating unit 931 calculates the target F0 sub cost using the F0 pattern of the jth target t (j) and the F0 pattern of the jth candidate segment group U (j) (S931). The target duration length sub-cost calculating means 932 uses the duration length of the j-th target t (j) and the duration length group of the j-th candidate segment group U (j) to calculate the target duration length sub-cost. Is calculated (S932). Here, the search hypothesis group corresponding to the target t (j) is represented as H (j). Also, any one search hypothesis in H (j) is represented as h (j). The connection F0 sub-cost calculating means 934 calculates the end point F0 of the candidate unit u (j-1) of the search hypothesis group H (j-1) and the start point F0 of the jth candidate unit group U (j). And calculate the connection F0 sub-cost (S934). The connected phoneme environment sub-cost calculating means 935 calculates the acoustic similarity between the candidate unit u (j-1) of the search hypothesis group H (j-1) and the jth candidate unit group U (j). Calculate as the environmental sub-cost (S935). Next, the search hypothesis expanding means 936 assumes that one search hypothesis having the lowest cost is assumed when the calculated sub-cost is added to each hypothesis h (j-1) of the search hypothesis group H (j-1). Candidate segment u (j) is added to h (j-1) to obtain a new search hypothesis h (j) (S936). In this way, Step S931 to Step S936 are repeatedly executed for each candidate unit u (j) of the candidate unit group U (j), and the sub cost calculation and search hypothesis development are performed (S93b, S93c). ). Further, steps S931 to S936 are repeatedly executed for each target (S93b, S93c), and a search hypothesis group corresponding to the candidate segment group of each target is developed (S93a, S93d). Next, the selection means 937 refers to the expanded search hypothesis group, and finally outputs the unit numbers of candidate segments included in the path of the search hypothesis with the lowest cost to the speech synthesizer 94. The operation of the speech synthesizer 94 that acquired the segment number is as described above. In this way, the speech synthesizer 90 selects a candidate segment that is optimal for the target for each phoneme generated from the input text, and supports the input text by connecting the selected candidate segments. A synthesized speech can be generated.

Japanese Patent No. 2761552

  In the connection F0 sub cost calculation means 934, if either or both of the adjacent candidate segments are unvoiced sounds, the connection segment F0 sub cost cannot be calculated because there is no F0 pattern for these unvoiced sound candidate segments. In this case, as described above, the value of the connection F0 sub-cost is assumed to be a constant value, for example, 0. For this reason, since the distance of F0 between candidate segments is not evaluated as in the second phoneme / S / in the phoneme string / A / − / S / − / U /, for example, the candidate element selected by the selection unit 936 There is a problem that the continuity of a piece of intonation is not always guaranteed. This typical example will be specifically described with reference to FIG. FIG. 6 is a diagram illustrating candidate segments selected by the candidate segment selection unit 93 of the conventional speech synthesizer 90. In the graph of FIG. 6, the horizontal axis represents time (ms) and the vertical axis represents frequency (Hz). A curve represented by a broken line in FIG. 6 is the target F0 pattern generated by the prosody generation unit 92. The target F0 pattern is divided into phoneme segments and compared with candidate segments. In FIG. 6, the curve represented by a broken line is referred to as a target 31, a target 32, a target 33, a target 35, and a target 36 for a range divided for each phoneme. Specifically, the range delimited by the phoneme / E / of the target F0 pattern is the target 31, the range delimited by the phoneme / S / is the target 32, the range delimited by the phoneme / A / is the target 33, and the phoneme / K / The range to be delimited is called a target 35, and the range delimited by phonemes / I / is called a target 36. It is assumed that the candidate segment corresponding to the target 31 is the candidate segment 21. It is assumed that the candidate segment group corresponding to the target 33 is the candidate segment 23 and the candidate segment 24, and has two candidate segments. It is assumed that the candidate segment corresponding to the target 36 is the candidate segment 26. The candidate segments of the target 32 and the target 35 are not shown.

  Here, the phoneme environments of the candidate segment 23 and the candidate segment 24 are both the pre-environment / S / and the post-environment / K /, and the value of the connected phoneme environment sub-cost is 0. Further, the phoneme durations of the candidate segment 23 and the candidate segment 24 are assumed to be equal, and the target duration length sub-cost values are assumed to be equal to each other. In this case, a selectable combination of candidate segments (speech waveform segment sequence) is a combination A of candidate segment 21 → candidate segment 23 → candidate segment 26 or candidate segment 21 → candidate segment 24 → candidate. One of the combinations B of the element pieces 26. Here, since the continuity of the global F0 pattern is maintained in the combination A, natural intonation can be expected. However, the above-described target F0 sub-cost calculating unit 931 calculates the target F0 sub-cost of the candidate segment 24 to a smaller value than the target F0 sub-cost of the candidate segment 23. Further, since both the candidate segment 23 and the candidate segment 24 are / S / of the preceding phoneme being an unvoiced sound, the connection F0 sub-cost cannot be calculated, and the value thereof is 0. As described above, the values of the connected phoneme environment sub cost and the target duration length sub cost of the candidate segment 23 and the candidate segment 24 are equal to each other. Therefore, when the sum of the target cost and the connection cost is compared, the candidate segment 24 has a smaller cost sum than the candidate segment 23, and therefore, the candidate segment 24 that cannot be expected to be a natural intonation is selected. The present invention maintains the continuity of intonation of synthesized speech even for a target in which voiced sounds are connected via unvoiced sounds as shown in FIG. An object of the present invention is to provide a speech synthesizer capable of selecting a voice.

  The speech synthesizer of the present invention selects candidate segments suitable for a synthesized speech target (target) numbered for each phoneme label from a plurality of candidate segments (candidate segment group) stored in the speech waveform database. Then, the selected candidate segments are connected to generate a synthesized speech. The speech synthesizer of the present invention comprises at least a preceding voiced candidate segment search means, a connected F0 sub-cost calculation means, and a selection means. In the preceding voiced candidate segment search means, when the i-th target (i is a natural number of 3 or more) is a voiced sound and the (i-1) -th target is an unvoiced sound, k is 2 or more. The candidate segment selected as the candidate segment suitable for the target of the (ik) th voiced sound with the smallest k is searched (hereinafter referred to as the (ik) th candidate segment). The connected F0 sub-cost calculating means calculates each candidate segment of the candidate segment group (hereinafter referred to as the i-th candidate segment group) corresponding to the F0 value of the terminal position of the (ik) th candidate segment and the i-th target. The connection F0 sub-cost is calculated from the F0 value group at the head position of. The selection means selects one candidate segment suitable for the i-th target from the i-th candidate segment group based on the connection F0 sub-cost and sets it as the i-th candidate segment.

  According to the speech synthesizer of the present invention, even if a target in which voiced sounds are connected via unvoiced sound, the candidate speech is maintained so that the synthesized speech can be of high quality while maintaining the continuity of the synthesized speech intonation. A piece can be selected.

The block diagram which shows the structure of the conventional speech synthesizer. The figure which illustrates the speech information index with which the conventional speech synthesizer is provided. The flowchart which shows operation | movement of the conventional speech synthesizer. The block diagram which shows the detail of the candidate unit selection part with which the conventional speech synthesizer is provided. The flowchart which shows operation | movement of the candidate segment selection part with which the conventional speech synthesizer is provided. The figure which illustrates the candidate segment which the candidate segment selection part of the conventional speech synthesizer selects. 1 is a block diagram illustrating a configuration example of a speech synthesis device according to Embodiment 1. FIG. FIG. 3 is a block diagram illustrating details of a candidate segment selection unit included in the speech synthesis device according to the first embodiment. 6 is a flowchart illustrating an operation of a candidate segment selection unit included in the speech synthesizer according to the first embodiment. The figure which illustrates the candidate segment which the candidate segment selection part of the speech synthesizer concerning Example 1 selects.

  Hereinafter, embodiments of the present invention will be described in detail. In addition, the same number is attached | subjected to the structure part which has the same function, and duplication description is abbreviate | omitted.

  With reference to FIG. 7, the speech synthesizer 10 according to the first embodiment will be described in detail. FIG. 7 is a block diagram illustrating a configuration example of the speech synthesizer 10 according to the present embodiment. The speech synthesizer 10 of this embodiment includes a text analysis unit 91, a prosody generation unit 92, a candidate segment selection unit 13, a speech synthesis unit 94, and a speech waveform database 95. The speech waveform database 95 includes speech waveform data 951 and a speech information index 952. Since each component other than the candidate segment selection unit 13 (indicated by a thick frame in the figure) numbered differently from the prior art speech synthesizer 90 performs the same operation as the prior art speech synthesizer 90, The description is omitted. Next, with reference to FIGS. 8 and 9, the candidate segment selection unit 13 different from the prior art will be described in detail. FIG. 8 is a block diagram illustrating details of the candidate segment selection unit 13 included in the speech synthesizer 10 according to the present embodiment. FIG. 9 is a flowchart illustrating the operation of the candidate segment selection unit 13 included in the speech synthesizer 10 according to the present embodiment. The candidate segment selection unit 13 includes a target F0 sub cost calculation unit 931, a target duration length sub cost calculation unit 932, a preceding voiced candidate segment search unit 133, a connection F0 sub cost calculation unit 134, and a connected phoneme environment sub cost calculation unit. 935, search hypothesis expansion means 936, and selection means 937. For each component other than the preceding voiced candidate segment searching means 133 and the connected F0 sub cost calculating means 134 (indicated by a thick frame in the figure) numbered differently from the prior art speech synthesizer 90, the conventional speech synthesizer Since the operation is exactly the same as 90, the description thereof is omitted.

  The preceding voiced candidate segment searching means 133, when the i-th target (i is a natural number of 3 or more) is a voiced sound and the (i-1) -th target is an unvoiced sound (S13aY), A candidate segment ((ik) th candidate segment) selected as a candidate segment suitable for the target of the (ik) th voiced sound that is 2 or more and has the smallest k is (i-1). The search hypothesis path of the first candidate segment is traced in the direction of t (1) to search (S133). When there is a preceding voiced sound candidate segment (S13bY), the connected F0 sub-cost calculating unit 134 selects the F0 value of the terminal position of the (ik) th candidate segment and the candidate corresponding to the i th target. The connection F0 sub-cost is calculated from the F0 value group at the head position of each candidate element of the element group (i-th candidate element group) (S134). On the other hand, when the condition of step S13a is not satisfied (for example, when u (i) is an unvoiced sound, u (i) is a voiced sound, but the preceding phoneme is also a voiced sound, S13aN), the process proceeds to step S134. As in the conventional speech synthesizer 90, the connection F0 is determined from the F0 value at the end position of the (i-1) th candidate segment and the F0 value group at the start position of each candidate segment of the ith candidate segment group. The sub cost is calculated (S134). Further, when the condition of step S13b is not satisfied (for example, as a result of the search, an unvoiced sound comes to the beginning of the sentence and there is no preceding voiced sound, etc., S13bN), as in the prior art, the (i-1) th The connection F0 sub-cost is calculated from the F0 value at the end position of the candidate element and the F0 value group at the head position of each candidate element of the i-th candidate element group (S134).

  Next, with reference to FIG. 10, as a typical example in which the speech synthesizer of the present embodiment exhibits an effect superior to that of the conventional speech synthesizer 90, the example of FIG. 6 will be described in detail. . FIG. 10 is a diagram illustrating candidate segments selected by the candidate segment selection unit 13 of the speech synthesizer 10 according to the present embodiment. The difference from FIG. 6 is the boundary between phonemes / E / and phonemes / S /. Is the position L, the boundary between the phoneme / S / and the phoneme / A / is shown as the position T, the end of the candidate segment 21 is * 21a, the beginning of the candidate segment 23 is ○ 23a, and the candidate segment 24 This is only the point indicated by Δ24a at the beginning. When the preceding voiced candidate segment searching unit 133 of the candidate segment selecting unit 13 of the present embodiment determines the candidate segment of the phoneme / A /, the preceding phoneme of the phoneme / A / is the unvoiced sound / S /. Then, a candidate segment of voiced sound preceding phoneme / A / is searched (S133). Since there is a candidate segment 21 selected as suitable for the target of the phoneme / E / as the preceding candidate segment, the connection F0 sub-cost calculating means 134 uses the F0 value of the end position L of the candidate segment 21 ( The connection F0 sub-cost is calculated from * 21a) in the figure and the F0 value (◯ 23a in the figure) at the leading position T of the candidate segment 23 (S134). Similarly, the connection F0 sub-cost calculation means 134 calculates the connection F0 sub-cost from the F0 value (* 21a in the figure) of the end position L of the candidate segment 21 and the F0 value (Δ24a in the figure) of the leading position T of the candidate segment 24. Is calculated (S134). Using the connection F0 subcost calculated in this way, the target F0 subcost, the target duration length subcost, and the connection phoneme environment subcost calculated in the same manner as before, a comparison of the sum of the target cost and the connection cost can be made. Since the cost of the candidate segment 23 is smaller than 24, the candidate segment 23 from which natural intonation can be expected is selected. In this way, according to the speech synthesizer 10 according to the present embodiment, the candidate segment so that the synthesized speech has a natural intonation even for a target in which voiced sounds are connected via unvoiced sounds. Can be selected.

Claims (5)

Select a candidate segment suitable for a synthesized speech target (hereinafter referred to as a target) numbered for each phoneme label from a plurality of candidate segments (hereinafter referred to as a candidate segment group) stored in the speech waveform database, A speech synthesis method for generating synthesized speech by connecting the selected candidate segments,
When the i-th target (i is a natural number of 3 or more) is a voiced sound and the (i-1) -th target is an unvoiced sound, k is 2 or more and k is minimized (ik) A preceding voiced candidate segment search substep for searching for a candidate segment (hereinafter referred to as (ik) th candidate segment) selected as a candidate segment suitable for the target that is the th voiced sound;
The F0 value of the end position of the (ik) th candidate element and the F0 of the leading position of each candidate element of the candidate element group corresponding to the i th target (hereinafter referred to as the i th candidate element group) A connection F0 subcost calculation substep for calculating a connection F0 subcost from a value group; and
A candidate element having a selection sub-step of selecting one candidate element suitable for the i-th target from the i-th candidate element group based on the connection F0 sub-cost and making it the i-th candidate element A speech synthesis method comprising a single selection step. The speech synthesis method according to claim 1,
The candidate segment selection step includes:
Using the F0 pattern of the i th target and the F0 pattern group of the i th candidate segment group, a target F0 sub cost calculation substep for calculating a target F0 sub cost;
A target duration length sub-cost calculation sub-step for calculating a target duration length sub-cost using the duration length of the i-th target and the duration length group of the i-th candidate segment group;
(I-1) a connected phoneme environment sub-cost calculation substep for calculating an acoustic similarity between the i th candidate segment and the i th candidate segment group as a connected phoneme environment sub cost;
Have
The selection sub-step is suitable for the i-th target from the i-th candidate segment group based on the target F0 sub-cost, the target duration length sub-cost, and the connected phoneme environment sub-cost in addition to the connection F0 sub-cost. A speech synthesis method, wherein one candidate segment is selected as the i-th candidate segment. The speech synthesis method according to claim 1 or 2,
A text analysis step of acquiring text, performing morphological analysis on the acquired text, generating a phoneme string and an accent type from the morpheme analysis result, and outputting the phoneme string and the accent type;
Obtain the output phoneme string and accent type, estimate the F0 pattern and phoneme duration for each phoneme, and number the estimated F0 pattern and phoneme duration for each phoneme. And a prosody generation step of outputting the target as the target. Select a candidate segment suitable for a synthesized speech target (hereinafter referred to as a target) numbered for each phoneme label from a plurality of candidate segments (hereinafter referred to as a candidate segment group) stored in the speech waveform database, A speech synthesizer that connects the selected candidate segments to generate synthesized speech,
When the i-th target (i is a natural number of 3 or more) is a voiced sound and the (i-1) -th target is an unvoiced sound, k is 2 or more and k is minimized (ik) A preceding voiced candidate segment search means for searching for a candidate segment selected as a candidate segment suitable for the target to be the voiced sound (hereinafter referred to as (ik) th candidate segment);
The F0 value of the end position of the (ik) -th candidate element and the F0 of the leading position of each candidate element of the candidate element group corresponding to the i-th target (hereinafter referred to as the i-th candidate element group) A connection F0 subcost calculating means for calculating a connection F0 subcost from the value group;
A candidate unit comprising: a selecting unit that selects one candidate unit suitable for the i-th target from the i-th candidate unit group based on the connection F0 sub-cost and sets it as the i-th candidate unit A speech synthesizer comprising a selection unit.   The program which gives the instruction | command which should perform the speech synthesis method in any one of Claim 1 to 3 with respect to a computer.

Images