JP3854593B2 - Speech synthesis apparatus, cost calculation apparatus therefor, and computer program - Google Patents

Description

  The present invention relates to a speech synthesis technology, and more particularly to a speech synthesis technology for synthesizing speech that is close to a natural utterance by selecting and connecting speech units from a speech unit database.

  As a technique for realizing a human-machine interface (man-machine interface), a technique for performing simple information transmission to / from a machine or by inputting / outputting an acoustic signal from the machine has been used for a long time. In recent years, computer technology and the like have been developed, and information transmitted between machines and humans has become abundant and sophisticated. Accordingly, a man-machine interface using an acoustic signal is required to be capable of transmitting higher-level information.

  Of the information transmission technologies using speech, speech recognition technology that converts speech spoken by humans into linguistic information that can be processed by machines is actively researched and used as a technology for transmitting information from humans to machines. Increased. On the other hand, as a technology for transmitting information from machines to humans, research is also progressing on speech synthesis technology that synthesizes and outputs speech similar to human speech based on language information such as text data to be transmitted. , Has been used in various machines.

In speech synthesis technology,
(1) synthesizing speech signals that enable humans to accurately understand linguistic information;
(2) synthesizing a voice signal that can be heard in natural speech for humans; and
(3) synthesizing audio signals based on arbitrary language information;
Is required.

  In these respects, there is a speech synthesis technology that uses uttered speech as a technology capable of realizing more accurate speech synthesis. In this speech synthesis technology, actual speech speech is recorded and stored in a database, a suitable portion is selected from the recorded speech speech data in accordance with a synthesis target, and a series of speech signals is synthesized by connecting them.

  FIG. 5 shows a block diagram of a configuration of a conventional general speech synthesis system using such a technique. Referring to FIG. 5, a conventional speech synthesis system 40 records a speech of a natural utterance by a human and stores a speech segment of the speech (hereinafter referred to as “speech segment”) in advance. 42 and a segment selection unit 44 for selecting and connecting speech segments from the segment database 42 in accordance with the synthesis target 62 and outputting an output segment sequence 64.

  The speech synthesis system 40 further uses a value called “cost”, which is a reference when the segment selection unit 44 selects a speech unit, to select a speech unit stored in the already selected speech unit and the segment database 42. A cost calculation unit 46 for calculating based on the physical feature amount of the piece is included.

  The synthesis target 62 performs language processing such as morphological analysis and dependency analysis on the input text 60 that is the language information of the speech to be synthesized, converts the input text 60 into phoneme symbols, accent symbols, and the like, and further results of the language processing Based on the above, create phoneme (basic unit of utterance voice. In Japanese, it is almost equivalent to the utterance voice of one character when written in alphabet). To be prepared.

FIG. 6 shows an example of the composition target composition 62. 6, synthesis target 62 includes synthesis targets 82, 84,... For each phoneme used by segment selection unit 44 to select a speech unit for each phoneme. These synthesis targets 82, 84,... For each phoneme include the phoneme symbol for specifying the phoneme, the phoneme prosody command, the phoneme duration (phoneme duration) , the spectrum information for each phoneme, etc. Information indicating features.

  In the segment database 42 shown in FIG. 5, speech segments are stored in advance together with data representing their physical characteristics.

  When the synthesis target 62 is given, the segment selection unit 44 selects from the segment database 42 some speech segments that match the phoneme specified by the synthesis target 62. The selected speech unit is a speech unit candidate used for speech synthesis. The unit selection unit 44 gives a value indicating the physical characteristics of each candidate speech unit to the cost calculation unit 46 together with a value indicating the physical characteristics of the already selected speech unit.

  The cost calculation unit 46 calculates a value called “cost” for each candidate speech unit based on the given physical characteristics, and gives the unit to the unit selection unit 44. “Cost” is an evaluation criterion as to whether or not the speech unit is appropriate as a speech unit to be connected to the previous speech unit. Ideally, this cost should be zero, but it is usually difficult to do so.

  The unit selection unit 44 determines a unit of speech suitable for use in speech synthesis by obtaining a unit sequence that minimizes the sum of given costs. In this way, speech segments corresponding to the speech specified by the synthesis target 60 are extracted. The output segment series 64 composed of the extracted speech segments are connected to each other, and a speech waveform of synthesized speech corresponding to the synthesis target 62 is created.

  In order to obtain high-quality speech using speech synthesis technology for performing speech synthesis in this way, speech units that are sufficiently low in cost are stored in the unit database 42 and cost calculation is performed. It is necessary that the cost calculated by the unit 46 has good affinity with human perception.

  In order to satisfy the former, in the current speech synthesis technology, a large-scale speech corpus in which speech speech for several tens of hours is recorded may be used as the segment database 42. When the unit database 42 becomes large-scale, the options for selecting a speech unit by the unit selection unit 44 shown in FIG. 5 increase. Therefore, the unit selection unit 44 can determine a speech unit suitable for connection from among the many options, and the possibility that the sound quality of the synthesized speech is improved is increased.

As a technique for satisfying the latter, a cost calculation method using a sub-cost function is proposed in Non-Patent Document 1 described later. In the technique described in Non-Patent Document 1, for each physical quantity used for segment selection, a sub-cost function that describes the relationship with the perceptual characteristics is considered, and the cost function calculated by the cost calculation unit 46 is a linear sum of the sub-cost functions. Express. The cost calculation unit 46 combines the synthesis target t _i (i indicates the order of the speech units in the synthesis target), and the speech unit u _i-1 selected by the unit selection unit 44 in the previous selection operation. Based on the above, the cost C (u _i , t _i ) related to the segment as a selection candidate is calculated by the following equation.

この式において、ｗ_pro、ｗ_typ、ｗ_env、ｗ_spec、及びｗ_F0は、それぞれサブコストＣ_pro、Ｃ_typ、Ｃ_env、Ｃ_spec、及びＣ_F0に対応する重みである。非特許文献１に記載の技術では、これらの重みは、各コストの主観評価実験のスコアから重相関分析により推定した定数などを用いる。

In this equation, w _pro , w _typ , w _env , w _spec , and w _F0 are weights corresponding to the sub-costs C _pro , C _typ , C _env , C _spec , and C _F0 , respectively. In the technique described in Non-Patent Document 1, these weights use constants estimated by multiple correlation analysis from the scores of subjective evaluation experiments for each cost.

  Since speech synthesized by connecting speech units selected based on such costs uses speech units selected using a scale that takes human speech perception into account, so-called “mechanical sound” is used. It is expected to be a relatively natural voice that does not make you feel “like”.

Tomoki Toda, Tsune Kawai, Minoru Tsuzaki, Kiyohiro Shikano, “Fragment Selection Based on Phoneme Units and Diphone Units in Segment-Connected Japanese Text Speech Synthesis”, IEICE Transactions, Vol. J85-D-II. , No. 12, pp. 1760-1770, Dec. 2002.

  The cost function in the technique described in Non-Patent Document 1 is calculated by a linear sum of sub cost functions. As a result, it is expected that a more natural voice having good affinity with the perceptual characteristic can be synthesized. However, even if the cost calculation method in the technology described in Non-Patent Document 1 is used, or even when the largest unit database currently available is used, it seems to affect perception. In some cases, a speech unit having a large error must be selected. As a result, the quality of the synthesized speech is insufficient.

  This is considered to be based on the following reasons. That is, in the technique described in Non-Patent Document 1, a speech unit is selected based on the sum of sub costs. However, for some sub-costs, the impact on perception in certain cases may be small compared to other sub-costs. In such a case, even if the sub-cost is particularly small, if the value of the other sub-cost is large, the influence on perception becomes large, and the quality of the synthesized speech is deteriorated. On the other hand, in the sub cost where the influence on the perception is particularly large in a specific case, it is necessary to reduce the value particularly in comparison with other sub costs.

  In the technique described in Non-Patent Document 1, such a problem relating to the sub-cost is not recognized. As a result, when this technique is simply used, there remains a possibility that the quality of the synthesized speech will be insufficient.

  SUMMARY OF THE INVENTION Therefore, an object of the present invention is to provide a speech synthesizer that synthesizes speech by selecting speech segments according to cost, and synthesizes speech speech that synthesizes high-quality speech speech with a more natural impression on perception. Is to provide.

  Another object of the present invention is to provide a speech synthesizer capable of selecting speech segments so that the quality of synthesized speech is increased and a sense of incongruity does not occur in the speech synthesizer.

  Still another object of the present invention is to reduce the influence of a connection unit between speech units on perception in selecting speech units in a speech synthesizer, and to improve the quality of synthesized speech as a whole. It is to provide a speech synthesizer.

  A speech synthesizer according to a first aspect of the present invention is a speech synthesizer that synthesizes speech using speech segments selected from a predetermined speech segment database in accordance with a speech synthesis target. The speech synthesis target is described by predetermined feature information of speech. This apparatus is configured to detect, based on predetermined feature information, a portion of a speech synthesis target where predetermined feature information satisfies a predetermined condition, and a predetermined acoustic feature is detected by the detection unit. A cost calculation means for performing cost calculation based on a synthesis target using a predetermined function different from each other at a location where it is detected that the condition is satisfied, and a cost calculated by the cost calculation means Includes a speech segment selection means for selecting speech segments that satisfy a predetermined condition from the speech segment database.

  Preferably, the predetermined feature information includes information related to the prosody of the speech that is a synthesis target, and the detection means detects a prosody condition for detecting a part of the speech synthesis target that satisfies a predetermined condition of the information related to the prosody Including detection means.

  More preferably, the information about the prosody includes information about the accent of the speech that is the synthesis target, and the prosody condition detection means includes an accent change section detection means for detecting a section where the accent changes in the synthesis target of the voice. Including.

  The cost calculating means includes a plurality of sub cost function calculating means for calculating values of a plurality of sub cost functions respectively defined according to a plurality of types of voice feature information including predetermined feature information of the voice according to a voice synthesis target. The first preparation means for preparing the first constant group, the second preparation means for preparing the second constant group different from the first constant group, and the detection result of the detection means Selecting means for selecting one of the first constant group prepared by the first preparation means and the second constant group prepared by the second preparation means, and the first constant selected by the selection means Means for calculating the cost by a linear sum of the sub-costs calculated by the plurality of sub-cost function calculating means using the first or second constant group as a coefficient may be included. Constants corresponding to predetermined feature information included in the first constant group and the second constant group are different from each other.

  Preferably, the selection means selects the first constant group in the section detected by the accent change section detection means as the section in which the accent of the speech to be synthesized changes, and the second constant in the other sections. Select a constant group.

  For example, the value of the constant corresponding to the predetermined feature information included in the first constant group is larger than the value of the corresponding constant included in the second constant group.

  For example, the value of the constant corresponding to the predetermined feature information included in the second constant group is smaller than the value of the corresponding constant included in the first constant group.

  More preferably, the segment selection unit includes a unit for selecting a speech unit from the speech unit database so that the cost calculated by the cost calculation unit is minimized.

  The speech synthesized by the speech synthesizer may be Japanese speech.

  More preferably, the information about the prosody includes accent height information about the height of the accent of the speech that is the synthesis target, and the accent change section detecting means includes a vowel part of the mora immediately before the location where the accent height changes, and Means for detecting a section constituted by a mora immediately after the location where the height of the accent changes as an accent change section.

  The speech synthesizer may further include a speech unit database.

  When the computer program according to the second aspect of the present invention is executed by a computer, it causes the computer to operate as any of the speech synthesizers according to the first aspect of the present invention.

  A cost calculation apparatus according to a third aspect of the present invention is a speech synthesis apparatus that synthesizes speech using speech units selected from a predetermined speech unit database according to a speech synthesis target. It is a cost calculation device for calculating the cost. The speech synthesis target is described by predetermined feature information of the speech. The cost calculating apparatus detects a part of the speech synthesis target for detecting a point where the predetermined feature information satisfies a predetermined condition, and the detection unit detects that the predetermined feature information satisfies the condition. Cost calculation means for performing cost calculation based on the synthesis target using predetermined functions different from each other and the other locations.

  If low-cost segment selection is performed in a part that is not perceptually important, and conversely, a high-cost segment selection is performed in a part that is perceptually important, resulting in a decrease in the quality of the synthesized speech. If so, the cost function calculation needs to be improved. Therefore, as one method, it is conceivable to change the cost function with time. As a result, the quality of the whole sentence is improved by selecting an element that allows an error in an unimportant part and reduces the cost of an important part. For example, if attention is given to the accent of a sentence, it is considered that the change in prosodic features has a greater influence on perception where the change in accent occurs. Thus, where accent changes occur, the sub-cost corresponding to the prosodic features is considered more important than the other sub-costs, so that the impression of the synthesized speech is improved. In this embodiment, attention is paid to the relationship between the change in accent and the prosodic feature.

  For this purpose, information on the accent (accent information) is included in the synthesis target. For speech synthesis for the purpose of text-to-speech conversion, language analysis is performed in advance. Therefore, it is possible to include the accent information in the synthesis target without any problem. Accent information regarding individual words can be obtained by referring to a general accent dictionary or the like.

  Hereinafter, an embodiment in which the present invention is applied to a Japanese speech synthesis technique will be described with reference to the drawings. In this specification, “accent” indicates an accent in Japanese or the like. In other words, this “accent” is a pitch accent generated by changing the fundamental frequency of the spoken voice, unlike the stress accent often found in Indo-European languages. However, it is needless to say that the same handling is possible with respect to phonetic features other than high and low accents by referring to the following description.

  An outline of Japanese accent will be described with reference to FIG. Referring to FIG. 1, the pronunciation of the word 70 “collectively” is expressed in katakana and romaji as a pronunciation notation 72. The horizontal line 74 on the phonetic notation 72 above the notation “Limatome” is an accent symbol indicating the accent of the word 70 “collect”. This accent symbol 74 indicates that when the word 70 is pronounced, the “limatome” portion of the phonetic notation 72 is pronounced high. On the contrary, the part without the horizontal line is pronounced low.

  At the start point portion 76 of the horizontal line 74, the pitch of the voice that is pronounced increases. Further, at the end point portion 78 of the horizontal line 74, the pitch of the voice to be pronounced decreases. A typical phonetic symbol string 80 represents such a change in voice pitch. When the word 70, “collect”, is pronounced, the “tri (/ t // o // r // i /)” part, the “t”, and the “mail (/ m // e / The prosody changes greatly between “Me” and “Le” in the “/ r / u /)” portion. In Japanese, an accent is formed by the rise and fall of the fundamental frequency of such speech.

  In the following description, a portion where the pitch of the voice greatly increases is referred to as “rise of (accent)”. The part where the pitch of the voice falls is called “(accent) falling”.

  In Japanese accents, such rise and fall of accents occur at the boundary between successive “mora”. “Mora” is a unit for measuring the phonological length of a word. Usually, the length of time required for sound generation for one mora is the length of time for generating one set of a set consisting of a consonant and one short vowel. However, the sound produced by one mora is not limited to a set of consonants and one short vowel. The second half of long vowels (the part marked with “-” in katakana notation), the second part of double vowels (for example, the “i” part of the word “Kansai”), and the sound (for example, “development (hattatsu) ) ”) And sound repellent (for example,“ N ”part of the word“ Kansai ”) also form one independent mora.

FIG. 2 is a block diagram showing the functional configuration of the speech synthesis system according to the present embodiment. Referring to FIG. 2 , this speech synthesis system 90 includes a segment database 42 similar to that of the prior art shown in FIG.

  The speech synthesis system 90 is further calculated using the synthesis target 102 including the accent information as described above and a different function based on the accent information as described later, instead of the segment selection unit 44 shown in FIG. A segment selection unit 120 for selecting a speech segment suitable for use in speech synthesis from the segment database 42 based on the cost of the segment database 42 is included.

  Further, the speech synthesis system 90 replaces the cost calculation unit 46 shown in FIG. 5 with different functions based on the accent information based on the physical features given from the unit selection unit 120 for the candidate speech units. A cost calculation unit 122 is included for calculating the cost by using it. Specifically, the cost calculation unit 122 calculates a plurality of sub-costs in the same manner as described in Non-Patent Document 1, and calculates the cost by their linear sum using different weight sets selected according to accent information. calculate.

  The speech synthesis system 90 further selects, based on the synthesis target 102 including the accent information, the weight set used when the cost calculation unit 122 calculates the linear sum of the sub-costs from the two set in advance, and the cost calculation unit A weight selection unit 100 for giving to 122 is included.

  The weight selection unit 100 has a function of giving the element selection unit 120 a completion signal indicating that the change of the weight used when the cost calculation unit 122 calculates the cost is completed. In addition, the unit selection unit 120 has a function of starting the selection of a speech unit in response to the completion signal given from the weight selection unit 100.

Weight selection section 100, first and weight set holding unit 104, a second weight set for holding the second weight set W _B different from the first weight set W _A which holds the first weight set W _A Holding part 106. The first weight set W _A and the second weight set W _B comprises a weight shown in the following formula cost calculation unit 122 is corresponding to each sub-cost in calculating the cost.

これら重みはいずれも定数である。これらの重みセットのうち、音素の適合性に関するサブコストの重みｗ_typ、音素環境代替に関するサブコストの重みｗ_env、スペクトルの不連続に関するサブコストの重みｗ_spec、及び基本周波数Ｆ₀に関するサブコストの重みｗ_F0は、非特許文献１でのコスト計算に用いられている重みと同様である。また、第２の重みセットＷ_Bの韻律に関するサブコストの重みｗ_proは、非特許文献１でのコスト計算に用いられるものと同様である。

These weights are all constants. Of these weight sets, sub-cost weight w _typ for phoneme suitability, sub-cost weight w _env for phoneme environment substitution, sub-cost weight w _spec for spectrum discontinuity, and sub-cost weight w _F0 for fundamental frequency F _0. Is the same as the weight used for the cost calculation in Non-Patent Document 1. The weight w _pro metrical about subcosts of the second weight set W _B is the same as that used in the cost calculation in the non-patent document 1.

In the present embodiment, the weight w _proA in the first weight set W _A and the weight w _pro in the second weight set W _{B with} respect to the sub-cost C _pro related to the prosody have the relationship shown in the following equation.

すなわち、第１の重みセットＷ_Aを用いる場合には韻律に関するサブコストが重視され、第２の重みセットＷ_Bを用いる場合には韻律に関するサブコストは特に重視はされない（通常と同じ）。第１の重みセットＷ_Aは、アクセントの影響により韻律的な特徴変化の大きな時間的区間（以下、この時間的区間を「韻律変化区間」と呼ぶ。）の音声素片を選択する際に用いられる重みである。第２の重みセットＷ_Bは、それ以外の時間的区間（以下、この区間を「平坦区間」と呼ぶ。）の音声素片を選択する際のコスト計算に用いられる。

That is, in the case of using the first weight set W _A sub-cost is emphasized about prosody, sub-cost is not particularly emphasized about prosody in the case of using the second weight set W _B (usually the same as). The first weight set W _A, a large time interval of prosodic features changes due to the influence of accent (hereinafter, this time interval is referred to as "prosodic change interval".) Used to select the speech units Is the weight to be given. The second weight set W _B is the other time interval (hereinafter, this period is referred to as "flat sections".) Used in the cost calculation in selecting a speech unit of.

Based on the synthesis target 102, the weight selection unit 100 further responds to the detection result of the prosody change interval detection unit 108 for detecting the prosody change interval and the prosody change interval detection unit 108, and the first weight set W _A selection unit 110 for selecting any one of _A and the second weight set W _B and giving the selection to the cost calculation unit 122.

The section around the part where the accent information changes is a section where the prosodic feature change is large due to the influence of the accent. The prosody change section detecting unit 108
(1) The vowel part immediately before the part where the accent information changes, and
(2) The consonant part and vowel part (if there is a consonant part) immediately after the part where the accent information changes, or the vowel part (if there is no consonant part),
Are detected as prosodic change intervals.

  FIG. 3 shows an example of the configuration of the synthesis target 102. Referring to FIG. 3, the synthesis target 102 is created based on the input text 60 in the same manner as the synthesis target 62 in the prior art shown in FIG. 6, and the segment selection unit 120 selects a speech segment. ... Includes synthesis targets 142, 144,.

The synthesis target 142, 144,... For each phoneme is similar to the synthesis target 82, 84,... For each phoneme shown in FIG. 6, the phoneme symbol for specifying the phoneme, the phoneme prosody command, the phoneme duration ( Phoneme duration) and information indicating physical characteristics of the phoneme, such as spectrum information for each phoneme. The synthesis target 142, 144,... For each phoneme further includes accent information 146, 148,... Indicating whether the phoneme is pronounced higher or lower by accents. “H” in the accent information 146, 148,... Indicates that the phoneme is pronounced highly. “L” in the accent information 146, 148,... Indicates that the phoneme is pronounced low. When the accent information about adjacent phonemes is different from the others, there is an accent rise or fall at the boundary portion.

When calculating the cost for the speech unit of prosody change interval, the sub-cost C _pro about prosody weights w _proA held in the first weight set W _A is multiplied. Therefore, the cost function calculated by the cost calculation unit 122 is as follows.

平坦区間の音声素片についてコストを計算する場合、韻律に関するサブコストＣ_proに、第２の重みセットＷ_Bに保持された重みｗ_proが乗算される。従ってコスト関数は、非特許文献１に記載のコスト関数と同様の以下の式となる。

When calculating the cost for the speech unit of the flat sections, a sub-cost C _pro about prosody, weights w _pro held in the second weight set W _B is multiplied. Therefore, the cost function is the following expression similar to the cost function described in Non-Patent Document 1.

即ちコスト計算部１２２は、韻律変化区間内の音声を合成する際の候補となる音声素片についてコストを算出する場合と、平坦区間内の音声を合成する際の候補となる音声素片についてコストを算出する場合とで、それぞれ異なるコスト関数によってコストの計算を行なうこととなる。

That is, the cost calculation unit 122 calculates the cost for speech units that are candidates for synthesizing speech in the prosodic change section and the cost for speech units that are candidates for synthesizing speech in the flat section. In this case, the cost is calculated using different cost functions.

  The speech synthesis system 90 operates as follows. Referring to FIG. 2, it is assumed that a synthesis target 102 has been created from input text 60 in advance. At this time, accent information is created in the synthesis target 102 for each phoneme. The synthesis target 102 is given to the prosody change section detection unit 108 and the segment selection unit 120 of the weight selection unit 100. The segment selection unit 120 temporarily stores the given synthesis target 102.

The prosody change section detector 108 detects the prosody change section and the flat section as follows. That is, the prosody change section detection unit 108 refers to the accent information 146, 148,... (See FIG. 3) of the synthesis target 102, and detects a prosody change section and a flat section for two adjacent phonemes. The prosodic change section detection unit 108 gives the selection unit 110 a detection result for each phoneme. At the same time, the prosodic change section detection unit 108 gives a completion signal indicating that the detection of the section is completed to the segment selection unit 120.

Referring to FIG. 2, the selector 110, when the detection results given are those representing the prosodic change interval when the first weight set W _A, a flat section a second weight set the W _B, respectively given to the cost calculation unit 122. As a result, the weight of the sub-cost C _pro related to the prosody given from the weight selection unit 100 to the cost calculation unit 122 is w _proA in the prosodic change sections 218 and 220 and w _pro in the flat sections 222 and 224 as shown in FIG. Become.

  Referring to FIG. 2, when a completion signal is given, segment selection unit 120 reads stored synthesis target 102 for one phoneme. The segment selection unit 120 extracts candidate speech segments from the segment database 42 based on the read synthesis target 102, and the acoustic features of the extracted speech segments and the selections made so far Information indicating the acoustic characteristics of phonemes is given to the cost calculation unit 122.

  The cost calculation unit 122 calculates a cost for the given speech unit using the weight set given through the selection unit 110. When calculating the cost for the speech segment in the prosodic change section, the cost calculation unit 122 calculates the cost by the following formula.

平坦区間の音声素片についてコストを計算する場合、コスト計算部１２２は、以下の式によりコストを算出する。

When calculating the cost for the speech segment in the flat section, the cost calculation unit 122 calculates the cost by the following formula.

コスト計算部１２２は、このようにして算出したコストを素片選択部１２０に与える。素片選択部１２０は、与えられたコストをもとに、図５に示す従来の技術における素片選択部４４と同様に、選択された音素列のコストの総和が最小となるような音声素片を選択し出力する。

The cost calculation unit 122 gives the cost calculated in this way to the segment selection unit 120. Based on the given cost, the unit selection unit 120 is similar to the unit selection unit 44 in the prior art shown in FIG. 5, and the speech element that minimizes the sum of the costs of the selected phoneme string is used. Select a piece and output it.

  By repeating the above operation, the segment selection unit 120 outputs an output segment sequence 64 that minimizes the total cost.

4, the synthetic target 102, a prosody changed section and flat sections for detecting the prosody change section detecting unit 108 shown in FIG. 2, the relationship between the weight of subcost C _pro about prosody schematically illustrates. When synthesizing the speech of the word 70 “collect” shown in FIG. 1, the accent of the word 70 is represented by an accent symbol on the phonetic notation 72 shown in FIG. 1. Referring to FIG. 4, a synthesis target 102 corresponding to this word includes accent information 200 created based on this accent symbol.

  The prosody change section detection unit 108 refers to the accent information 200 of the synthesis target 102 and detects a portion where the accent information about two adjacent phonemes changes. For example, in the part 202 where the phonemes “/ o /” and “/ r /” are adjacent, the accent information of the phoneme “/ o /” is “L”, whereas the phoneme “/ r /” adjacent to the phoneme “/ r /” is “L”. "H" is the accent information. Therefore, the portion 202 where the phonemes “/ o /” and “/ r /” are adjacent has a rise in accent. Similarly, the portion 204 where the phoneme “/ e /” and the phoneme “/ r /” are adjacent has a falling edge of the accent.

  In the example shown in FIG. 4, phonemes “/ o /” 206 and “/ e /” 212 correspond to the vowel part immediately before the part where the accent information changes. The phonemes “/ r /” 208 and “/ i /” 210 and the phonemes “/ r /” 214 and “/ i /” 216 correspond to the consonant part and the vowel part immediately after the part where the accent information changes. To do. Sections including these phonemes are prosodic change sections 218 and 220. The other sections are flat sections 222 and 224.

  In the present embodiment, when a portion where the prosody changes greatly due to an accent is detected and the phoneme of the detected portion is selected, the weight of the sub-cost related to the prosody is set to a large value. As a result, the sub-cost related to the prosody for the detected speech segment of the detected portion is evaluated to be relatively heavier than the other sub-costs as the weight is increased. Therefore, the cost for this speech segment is a value that more strongly reflects the sub-cost related to prosody. When a speech unit is selected using such a cost, a speech unit is selected such that the sub-cost related to the prosody becomes smaller than usual in a section where the influence on the perception of the prosody is large. Therefore, in such a section, it is possible to select a segment with less error regarding the prosody, and the influence on perception is reduced. As a result, it is less likely that the person feels uncomfortable when the synthesized speech is heard by a human.

  In the present embodiment, the speech segment is selected using the conventional cost in a section other than the section where the change in the prosody due to the accent has a significant perceptual effect. Therefore, the voice synthesized for this section has quality that can be expected in the conventional technology. Thus, by changing the cost calculation method according to the target voice characteristics, the synthesized voice can be expected to improve in quality as a whole.

  Each functional unit described in the above block diagram format can be realized by computer hardware and a program executed on the computer. As this computer, a computer having general-purpose hardware can be used as long as it has equipment for handling sound. Further, each functional block of the apparatus described above can be realized by a program by those skilled in the art based on the description in this specification. Such a program is also a piece of data and can be stored in a storage medium and distributed.

In the above-described embodiment, the cost calculation is performed by setting the sub-cost related to the prosody to a large value in the prosody change section. However, the present invention is not limited to such an embodiment. For example, the value of the sub cost related to the prosody may be set to w _pro described above in the prosodic change section, and the weight of the sub cost related to the prosody may be set to a value smaller than w _pro in the flat section.

In this way, the sub-cost related to the prosody for the speech segment of this portion is evaluated relatively lightly as compared with the other sub-costs by the amount of the reduced weight. The cost for the speech segment in the flat section reflects the value of the sub-cost related to the prosody more weakly. When a speech unit is selected using such a cost, the speech unit is selected in such a manner that a perceptual error relating to the prosody is allowed in a flat section. As a result, it is possible to indirectly avoid an increase in the sub-cost related to the prosody in the prosodic change section while suppressing the discontinuity of the fundamental frequency F ₀ at the boundary between the flat section and the prosodic change section.

This is due to the following reason. That is, the discontinuity of the fundamental frequency F ₀ is considered to have a great influence on the perception of synthesized speech. Therefore, even in the prosody change interval in prosody flat sections, by evaluating similarly heavier subcost about discontinuity of the fundamental frequency F _0, has a unit selection such discontinuities is reduced fundamental frequency F _0. However, as a result, it is considered that a speech unit having a large sub-cost related to the prosody may be selected particularly in the prosody change section.

The influence of the prosody on the perception of the synthesized speech is small in the flat section where the prosody change is small and conversely large in the prosody change section. Therefore, if a perceptual error related to prosody is allowed in a flat section where the perceptual influence on prosody is small, the number of speech segments that can be selected in the flat section increases. Along with this, the number of combinations of speech units connected at locations adjacent to the flat section increases. For this reason, even if the sub-cost related to the discontinuity of the fundamental frequency F ₀ is heavily evaluated, the number of speech segments that can be selected at this point increases. If the number of speech units that can be selected in this way increases, there is a high possibility that speech units having a small sub-cost related to the prosody are present.

  Therefore, in the prosodic change section that is a section adjacent to the flat section and has a large perceptual influence due to the prosody, there is a high possibility that a speech unit having a small sub-cost related to the prosody is selected. As a result, it is possible to avoid an increase in the sub-cost related to the prosody in the prosodic change section.

  Furthermore, the cost for the speech segment in the flat section reflects the value of the sub-cost related to the prosody more weakly. When a speech unit is selected using such a cost, the speech unit is selected in a form that relatively reflects sub-costs other than the sub-cost related to prosody. Thereby, it can also be expected that the quality of the synthesized speech related to other perceptual factors in this section is improved.

Furthermore, in the prosodic change section, the weight of the sub-cost related to the prosody may be set to a value larger than w _pro , and may be set to a value smaller than w _pro in the flat section. Even in this case, in the prosodic change section, the prosody sub-cost is more important than the other sub-costs, and in the flat section, other sub-costs are more important than the prosody sub-cost, and there is a perceptual effect in the section where the accent changes. It is possible to perform segment selection so as to reduce the number of segments.

  In the above-described embodiment, the weight selection unit 100 shown in FIG. 2 changes only the sub-cost weight related to the prosody. However, the present invention is not limited to this. Other sub-cost weights can also be changed. In this way, it is possible to optimize the cost so as to reduce the influence on the perception in more detail for more acoustic features.

In the embodiment described above, two types of weight sets used for cost calculation are prepared, and the weight set is switched when the cost calculation unit 122 shown in FIG. 2 calculates the cost. However, the present invention is not limited to such an embodiment. The weight selection unit 100 may calculate the weight by some function according to the synthesis target. For example, the change amount of the sub-cost weight w _pro related to the prosody may be calculated based on the information related to the fundamental frequency F ₀ and the information related to the phoneme duration included in the synthesis target 102. Depending on the detection result of the prosody change section and the flat section by the prosody change section detection unit 108, the cost calculation function itself in the cost calculation unit 122 may be completely different. Further, such processing is not limited to only focusing on prosody changes (accent changes), but it goes without saying that other acoustic feature changes may be focused.

  Note that when attention is focused on the change in accent, the form of the accent information included in the synthesis target 102 does not matter. Any type of information may be used as long as the information can identify a portion where the utterance is greatly changed by the accent. The same applies when focusing on acoustic features other than accents.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each claim in the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are intended. Including.

It is the schematic which shows the feature of the accent in Japanese speech. It is a block diagram which shows the functional structure of the speech synthesis system which concerns on one embodiment of this invention. It is a figure which shows an example of the synthetic | combination target which concerns on one embodiment of this invention. And combining the target is a schematic diagram illustrating a prosody changed section and flat sections, the relationship between the weight of subcost C _pro about prosody. It is a block diagram which shows the structure of a common speech synthesis system. It is a figure which shows the structure of the synthetic | combination target based on the background art of this invention.

Explanation of symbols

40, 90 Speech synthesis system, 42 Segment database, 44, 120 Segment selection unit, 46, 122 Cost calculation unit, 62, 102 Synthesis target, 100 Weight selection unit, 104, 106 Weight set holding unit, 108 Prosody change section Detection unit, 110 selection unit, 146, 148 Accent information

Claims (8)

A speech synthesizer that synthesizes speech using speech units selected from a predetermined speech unit database according to a speech synthesis target, wherein the synthesis target includes a plurality of speech features including predetermined feature information of speech Described by information ,
Detecting means for detecting a portion of the speech synthesis target that satisfies a predetermined condition of the predetermined feature information;
Using the predetermined functions different from each other for the location where the predetermined feature information is detected by the detection means to satisfy the condition , the synthesis target and the speech segment database are used. Cost calculation means for performing cost calculation based on each speech unit included ;
A speech synthesizer, comprising: a speech segment selection unit for selecting from the speech segment database a speech unit that minimizes the cost calculated by the cost calculation unit. The predetermined feature information includes accent information relating to an accent of a voice to be synthesized,
The condition includes an accent condition that the accent information of the voice that is the synthesis target changes,
The speech synthesis apparatus according to claim 1, wherein the detection unit includes an accent condition satisfaction section detection unit for detecting a section in which the accent information satisfies the accent condition among the speech synthesis targets. The cost calculated by the cost calculation includes a plurality of sub-costs calculated by a plurality of sub-cost functions defined so as to describe a relationship with the perceptual characteristics with respect to the feature information of the plurality of sounds.
The cost calculation means includes
The value of the sub-cost function before Kifuku number, and support Bukosu preparative calculation detecting means for de San follow the feature amount of each speech unit included in the composite target and the speech unit database of the speech,
First preparation means for preparing a first group of constants;
Second preparation means for preparing a second constant group different from the first constant group;
In the section where the accent condition of the speech to be synthesized is detected as a section that satisfies the accent condition by the accent condition satisfaction section detecting means, the first constant group is selected, and in other sections Selecting means for selecting the second constant group ;
As the coefficient of the first or second constant group selected by said selecting means, the linear sum of sub-cost calculated by the pre-hexa Bukosu preparative calculation detecting means and means for calculating the cost,
The voice according to claim 2, wherein a value of a constant corresponding to a sub-cost related to the prosody included in the first constant group is larger than a value of a corresponding constant included in the second constant group. Synthesizer. The cost calculated by the cost calculation includes a plurality of sub-costs calculated by a plurality of sub-cost functions defined so as to describe a relationship with the perceptual characteristics with respect to the feature information of the plurality of sounds.
The cost calculation means includes
Sub-cost calculating means for calculating the values of the plurality of sub-cost functions according to the speech synthesis target and the feature value of each speech unit included in the speech unit database;
First preparation means for preparing a first group of constants;
Second preparation means for preparing a second constant group different from the first constant group;
In the section where the accent condition of the speech to be synthesized is detected as a section that satisfies the accent condition by the accent condition satisfaction section detecting means, the first constant group is selected, and in other sections Selecting means for selecting the second constant group;
Means for calculating the cost by a linear sum of the sub-costs calculated by the sub-cost calculation means, using the first or second constant group selected by the selection means as a coefficient,
The value of a constant corresponding to a sub-cost other than the sub-cost related to prosody included in the first constant group is smaller than a value of a corresponding constant included in the second constant group. The speech synthesizer described. The accent condition satisfaction section detecting means is composed of a vowel part of a mora immediately before a point where the accent information of the speech as the synthesis target changes, and a mora immediately after a point where the accent information of the speech as the synthesis target changes. a section that is comprises means for detecting a period which satisfies the accent condition, the speech synthesizing apparatus according to claim 3 or claim 4. The speech synthesizer according to any one of claims 1 to 5 , further including the speech segment database. A computer-executable computer program that, when executed by a computer, causes the computer to operate as the speech synthesizer according to any one of claims 1 to 6 . In a speech synthesizer that synthesizes speech using speech units selected from a predetermined speech unit database according to a speech synthesis target, a cost calculation device for calculating costs for selecting speech units, The synthesis target is described by a plurality of types of speech feature information including predetermined feature information of speech ,
Based on the predetermined feature information, detection means for detecting a portion of the speech synthesis target that satisfies a predetermined condition of the predetermined feature information;
Using the predetermined functions different from each other for the location where the predetermined feature information is detected by the detection means to satisfy the condition , the synthesis target and the speech segment database are used. A cost calculation device comprising: cost calculation means for performing cost calculation based on each phoneme included .

Images