JP2005181998A - Speech synthesizer and speech synthesizing method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech synthesizer which can output natural synthesized speech. <P>SOLUTION: The synthesizer comprises a feature parameter DB 106 for storing foreign origin word attributes of voice unit and voice unit data indicating sound characteristics for every voice unit; a language analyzer 104 and rhythm estimation 109 acquire text data, and estimate the foreign origin word attributes and sound characteristics of the voice unit with respect to each of multiple voice unit for representing the text of text data; a voice unit selector 108 for selecting the voice unit data indicating the contents similar to the foreign origin word attributes of estimated voice unit from the feature parameter DB 106 for every voice unit representing the text; and a speech synthesizer 110 for generating and outputting the synthesized speech by using the selected multiple voice unit data. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a speech synthesis apparatus and speech synthesis method for converting a given character string (text) into speech.

  A conventional speech synthesizer selects a segment sequence from a segment database according to a minimum cost criterion using a cost function based on acoustic features, and generates a synthesized speech using the selected segment sequence (for example, a patent) Reference 1).

FIG. 17 is a block diagram showing the configuration of the conventional speech synthesizer.
The speech analysis unit 10 performs labeling on speech data stored in the speech waveform database 21 using a text database 22 and a phoneme HMM (Hidden Markov Model) 23, and acoustically for each phoneme unit (unit unit). Extract features. Here, the acoustic features include, for example, fundamental frequency, power, duration length, and cepstrum coefficient by cepstrum analysis. Information indicating the extracted acoustic features is stored as a segment in the feature parameter 30 which is the segment database. The speech unit selection unit 12 searches for a segment that is acoustically closest to the target segment (target segment) with reference to the feature parameter 30 that holds information indicating acoustic features. If there are a plurality of target segments, a plurality of segments are searched as a segment sequence. Here, the speech unit selection unit 12 selects a unit sequence in consideration of the error from the target unit with respect to the above-described fundamental frequency, power, and duration, and distortion generated when the units are combined. The speech synthesizer 13 generates a synthesized speech by acquiring and connecting a plurality of speech data corresponding to the segment series selected by the speech unit selector 12 from the speech waveform database 21.
Japanese Patent No. 3050832

  However, the conventional speech synthesizer described above has a problem that unnatural synthesized speech is output. That is, since the conventional speech synthesizer selects a segment based only on acoustic features, it cannot select an appropriate segment, and as a result, a synthesized speech generated using the segment. Will become unnatural. Furthermore, when the above-mentioned conventional synthesized speech apparatus cannot correctly extract the acoustic features of the target segment, the influence of the target segment selection greatly affects the selection of the segment. End up.

  Therefore, the present invention has been made in view of such a problem, and an object thereof is to provide a speech synthesizer and a speech synthesis method capable of outputting natural synthesized speech.

  In order to achieve the above object, a speech synthesizer according to the present invention is a speech synthesizer that acquires text data and converts the text indicated by the text data into speech, and whether the speech unit belongs to a foreign language. A foreign word attribute about whether or not, and voice unit data indicating the acoustic features of the voice unit, storage means for storing for each voice unit, a plurality of text data is obtained and a plurality of texts representing the text of the text data For each speech unit, feature estimation means for estimating the foreign word attributes and acoustic features of the speech unit, and a speech unit foreign word estimated by the feature estimation means for each speech unit representing the text Selection means for selecting sound unit data showing contents similar to attributes and acoustic features from the storage means, and a plurality of sounds selected by the selection means Characterized in that it comprises an audio output means for generating and outputting a synthesized speech by using the position data. For example, the selecting means preferentially selects speech unit data indicating belonging to a foreign word as a foreign word attribute when the feature estimating means estimates that the foreign word belongs to the foreign word as the foreign word attribute of the speech unit. To do.

  Thus, for example, when the speech unit of the text data belongs to a foreign word, the speech unit data that shows the characteristic of the foreign word is selected for the speech unit, so the synthesized speech that seems to be a foreign word is generated as indicated by the text data. Can be output. That is, conventionally, even if the speech unit of the text data belongs to a foreign word, the speech unit data is selected based only on the acoustic features of the speech unit, so that synthesized speech that does not look like a foreign word is output. However, in the present invention, natural synthesized speech can be output as shown by the text data.

  Further, the speech unit data further indicates a final particle attribute as to whether or not the speech unit belongs to a final particle, and the feature estimation means is provided for each of a plurality of speech units representing the text of the text data. , Estimating the foreign word attribute, acoustic feature, and final particle attribute of the speech unit, and the selecting means includes the foreign word attribute, acoustic feature, and final particle of the speech unit estimated by the feature estimating means. Voice unit data showing contents similar to the attribute may be selected from the storage means. For example, the selection means preferentially selects speech unit data indicating that it belongs to a final particle as a final particle attribute when the feature estimation means estimates that it belongs to a final particle as a final particle attribute of the speech unit. To do.

  As a result, for example, when a speech unit of text data belongs to a final particle, speech unit data representing a sense of question corresponding to the final particle is selected for the speech unit. It is possible to generate and output synthesized speech that expresses a feeling or the like.

  The selecting means quantitatively evaluates the similarity between the foreign word attribute of the speech unit estimated by the feature estimating means and the foreign word attribute of the voice unit data stored in the storage means. A first deriving unit for deriving one sub-cost, a sound unit acoustic feature estimated by the feature estimating unit, and a similarity between the acoustic unit data stored in the storage unit And second derivation means for deriving a second sub-cost by evaluating the cost, and a cost derivation means for deriving the cost using the first and second sub-costs derived by the first and second derivation means And data selection means for selecting voice unit data from the storage means based on the cost derived by the cost derivation means. For example, the cost calculation means derives the cost by adding a weight to each of the first and second sub-costs derived by the first and second derivation means.

  As a result, each of the first and second sub-costs is weighted. Therefore, according to the weight, the ratio of the contribution between the similarity of the acoustic feature and the similarity of the foreign word attribute to the selection of the voice unit data is determined. Can be adjusted.

  Further, the speech synthesizer further specifies a reliability of the acoustic feature estimated by the feature estimation unit, and determines a weight for each of the first and second sub-costs according to the reliability. And the cost deriving unit may add the weight determined by the weight determining unit to the first and second sub-costs. For example, when the reliability of the acoustic feature is low, the weight determination means contributes to the selection of the voice unit data by the data selection means by the similarity of the foreign language attribute rather than the similarity of the acoustic feature. Thus, the weights for the first and second sub-costs are determined.

  As a result, the weight assigned to each of the first and second sub-costs changes according to the reliability of the acoustic feature, so that the similarity of the acoustic feature and the similarity of the foreign word attribute to the selection of the speech unit data It is possible to appropriately change the ratio of contribution.

  The selection means further includes third derivation means for deriving a connection cost by quantitatively evaluating acoustic distortion when a plurality of audio unit data stored in the storage means is connected. The cost deriving unit derives the cost using the first and second sub-costs derived by the first and second deriving units and the connection cost derived by the third deriving unit. This may be a feature.

Thereby, acoustic distortion can be suppressed and more natural synthesized speech can be output.
Here, the data creation device according to the present invention is a data creation device for creating speech unit data used for speech synthesis, the speech storage means for storing speech waveform signals indicating speech as waveforms, and the speech waveform Text storage means for storing text data indicating the text corresponding to the speech of the signal, text data is obtained from the text storage means, the text data text is divided into speech units, and each speech unit is converted to a foreign language Language analysis means for analyzing a foreign word attribute as to whether it belongs or not, obtaining a speech waveform signal from the speech storage means, dividing the speech indicated by the speech waveform signal into speech units, and for each speech unit Acoustic analysis means for analyzing acoustic features, foreign language attributes analyzed by the language analysis means, and acoustic features analyzed by the acoustic analysis means To indicate, characterized in that it comprises a creation means for storing the created to the storage means the voice unit data for each speech unit.

  Thereby, since the speech unit data indicating the foreign word attribute and the acoustic feature for each speech unit is stored in the storage unit, the speech unit data can be selected from the storage unit based on the foreign word attribute and the acoustic feature. it can. That is, the storage means in which the voice unit data is stored can be used for the voice synthesizer. As a result, the speech synthesizer estimates the foreign language attributes and acoustic features of each speech unit from the text of the text data, and selects speech unit data having contents similar to the foreign language attributes and acoustic features from the storage means. Then, natural synthesized speech as shown by the text data can be generated.

  The present invention can be implemented not only as such a speech synthesizer, but also as a method and program for synthesizing speech by the speech synthesizer and a storage medium for storing the program.

  The speech synthesizer according to the present invention has the effect of being able to output natural synthesized speech as indicated by text data.

(First embodiment)
FIG. 1 is a configuration diagram showing the configuration of the speech synthesizer according to the first embodiment of the present invention. This speech synthesizer is a text-to-speech synthesizer that converts input text into speech, and includes a feature parameter database (feature parameter DB) 106, a language analysis unit 104, a prosody estimation unit 109, and a speech unit selection unit 108. A speech synthesizer 110 and a speaker 111.

  The feature parameter DB 106 is a database that holds sound unit data indicating features of a plurality of sound units. The language analysis unit 104 acquires text data 100t indicating text, extracts a linguistic feature of the text from the text data 100t, and outputs linguistic information 104d indicating the linguistic feature.

  The prosody estimation unit 109 estimates the prosody of the text based on the linguistic features extracted by the language analysis unit 104, and generates prosody information 109d indicating the estimation result. The speech unit selection unit 108 selects a series of speech unit data most suitable for text as a speech unit sequence from the feature parameter DB 106 based on the language information 104d and the prosody information 109d input from the language analysis unit 104 and the prosody estimation unit 109. To do. Then, the speech unit selection unit 108 notifies the speech synthesis unit 110 of the selected speech unit sequence.

  The speech synthesizer 110 generates a speech waveform signal representing the feature as a speech waveform based on the feature (for example, formant or sound source information) indicated by the speech unit data selected by the speech unit selector 108. Then, the speech synthesizer 110 connects the speech waveform signals of each speech unit data included in the speech unit sequence to generate a synthesized speech signal. The speaker 111 outputs the synthesized voice signal generated by the voice synthesizer 110 to the outside as an audible sound wave (synthesized voice).

Next, each component of the speech synthesizer shown in FIG. 1 will be described in detail.
FIG. 2 is a configuration diagram illustrating an example of an internal configuration of the language analysis unit 104.

  The language analysis unit 104 includes a morpheme analysis unit 301, a syntax analysis unit 302, a reading provision unit 303, and an accent phrase estimation unit 304.

  The morpheme analysis unit 301 performs morpheme analysis on the text indicated by the text data 100t. The syntax analysis unit 302 analyzes the dependency relationship for each morpheme analyzed by the morpheme analysis unit 301. Hereinafter, such analysis is referred to as syntax analysis. When there are a plurality of readings in the morpheme analyzed by the morpheme analysis unit 301, the reading assigning unit 303 assigns an appropriate reading to the morpheme. The accent phrase estimation unit 304 performs accent phrase division and accent combination processing on each morpheme analyzed by the morpheme analysis unit 301.

  When the language analysis unit 104 acquires the text data 100t, the language analysis unit 104 performs morphological analysis and syntax analysis, and gives appropriate reading, and outputs language information 104d as shown in FIG. 3, for example.

FIG. 3 is a diagram illustrating an example of the content of the language information 104d.
The language information 104d output from the language analysis unit 104 includes text, a series of phonemes (phoneme notation) corresponding to the text, morphemes included in the text, clauses included in the text, and part of speech of each morpheme. , The position in the morpheme, the position in the accent phrase, and the relationship position. The text is, for example, “It is fine today” in Japanese. Each morpheme is “today” or “of” indicated by a vertical broken line in the text of FIG. Each phrase is “today's”, “weather is”, and “sunny” shown in the text of FIG. 3 separated by a vertical bar “|”. The part of speech for the morpheme “today” is “common noun”, and the part of speech for the morpheme “no” is “particle”. A particle is one of the parts of speech in Japanese, and is a word that is not used among words that are always used after other words. In addition, particles have the function of indicating how the previous word is related to other words, adding a certain meaning such as the emotion of the speaker, and completing the sentence. In addition, the part of speech in the present embodiment indicates an attribute as to whether or not a morpheme is a foreign word, and also indicates an attribute as to whether or not a morpheme is a final particle. Here, the final particle refers to a particle that expresses the meaning of questions, prohibition, mourning, excitement, etc. when used at the end of a sentence or phrase.

  The destination position indicates the destination of each phrase. For example, the number “1” indicating the position of the destination corresponding to the phrase “today” in FIG. 3 relates to the phrase “weather is” after the phrase “today”. It is shown that. The same applies to other clauses. The position in the morpheme indicates the position of each phoneme included in each morpheme, and the position in the accent phrase indicates the position of each phoneme included in each accent phrase.

  The phoneme notation indicates text with phonemes, and further indicates an accent phrase, a sentence head, and a sentence end. For example, one accent phrase is between the symbols “/” and “/” in the phonetic notation of FIG. The symbol “^” in the phoneme notation indicates the beginning of a sentence, and the symbol “$” indicates the end of the sentence.

  The language information 104d may indicate parts of speech hierarchically. For example, in the example shown in FIG. 3, the part of speech of the morpheme “no” is “particle”, and the morpheme “no” belongs to “case particle” among the “particles”. Therefore, the language information 104d indicates “participant” and “case particle” as part of speech for the morpheme “no”. Note that a case particle is one of the classifications of Japanese particles, and is a particle that mainly indicates a case and indicates a case relationship between the word and another word.

  Further, the language analysis unit 104 may be configured to estimate a field (sports, news, entertainment, etc.) to which the text belongs. For example, the text data 100t may have information about a field to which the text belongs, or a field may be estimated by extracting a keyword from the text.

  In addition, the language analysis unit 104 may be configured to estimate emotions such as joy, anger, sadness, and comfort that are appropriate for the text simultaneously with the estimation of the above field. For example, text data 100t may be preliminarily provided with information indicating emotions desired to be expressed in the text (voice XML or the like exists as a standard that can realize this).

  The prosody estimation unit 109 estimates the maximum likelihood prosody for the text indicated by the text data 100t based on the language information 104d sent from the language analysis unit 104, and generates prosody information 109d as the estimation result. Here, the prosody information 109d indicates at least the duration length, the fundamental frequency, and the power for each phoneme unit. In addition to the phoneme unit, the prosody estimation unit 109 may be designed to estimate the duration, the fundamental frequency, and the power for each mora unit or syllable unit. The prosody estimation unit 109 may perform any method of estimation. For example, estimation may be performed by a method based on quantification type I that is generally used.

  Here, the prosodic information 109d indicates the duration, fundamental frequency, and power for each phoneme unit, but may also indicate the reliability of the prosodic result at the time of prosodic estimation.

  The feature parameter DB 106 stores a plurality of audio unit data. The speech unit data includes acoustic feature information indicating acoustic features of speech units and linguistic feature information indicating linguistic features of speech units.

FIG. 4 is a diagram showing the contents of the acoustic feature information.
The acoustic feature information 106a may indicate at least a fundamental frequency, a duration, power, and the like as an acoustic feature for each voice unit, and may further indicate a cepstrum coefficient by cepstrum analysis.

FIG. 5 is a diagram showing the contents of the linguistic feature information.
As shown in FIG. 5, the linguistic feature information 106b indicates phonological environment, morpheme information, accent phrase information, and syntax information as linguistic features. The phoneme environment indicates a phoneme currently targeted (target phoneme), a phoneme immediately before the target phoneme (forward phoneme), and a phoneme immediately behind the target phoneme (back phoneme). Note that such a phoneme environment is information that has been used conventionally. The morpheme information indicates a morpheme including the target phoneme. Specifically, the morpheme information indicates the notation and part of speech of the morpheme. The parts of speech indicated by the morpheme information are subdivided (hierarchized) as necessary. When the part of speech is utilized, the utilization form is also indicated by morpheme information. The accent phrase information is information indicating the position of the target phoneme in the accent phrase. Specifically, the accent phrase information indicates the distance from the accent phrase head, the distance to the accent phrase end, and the distance from the accent nucleus. The syntax information indicates dependency relation information of a phrase including the target phoneme.

  FIG. 6 is a diagram showing the contents of one piece of audio unit data stored in the feature parameter DB 106.

For example, when the speech unit is a phoneme, the feature parameter DB 106 holds speech unit data indicating the feature of the phoneme for each phoneme by vector display as shown in FIG. The speech unit data includes the above-described linguistic feature information 106b and acoustic feature information 106a corresponding to the phoneme. For example, the speech unit data indicates the notation “today”, the phoneme “ky”, the part of speech “common noun”, and the like as features of the phoneme u _i . Note that the voice unit data may indicate the emotion of the speaker when the voice unit is uttered and the field (domain) to which the text that is the basis of the voice unit belongs.

FIG. 7 is a diagram illustrating a specific configuration example of the audio unit selection unit 108.
The voice unit selection unit 108 includes a voice unit candidate extraction unit 401, a search unit 402, and a cost calculation unit 403.

  The speech unit candidate extraction unit 401 is used for speech synthesis in consideration of the prosodic information 109d sent from the prosody estimation unit 109 for each speech unit (phoneme) indicated by the language information 104d sent from the language analysis unit 104. A set of speech unit data that can be candidates for speech unit data is extracted from the feature parameter DB 106. The search unit 402 searches the candidate extracted by the speech unit candidate extraction unit 401 for the most likely speech unit data in the language information 104d sent from the language analysis unit 104 and the prosody information 109d sent from the prosody estimation unit 109. To do. Note that the search unit 402 searches a series of speech unit data arranged in a time series corresponding to the phoneme notation of the language information 104d as a speech unit sequence at a time.

  The cost calculation unit 403 calculates a cost as a reference for the search unit 402 to search for the most likely speech unit sequence. Such a cost calculation unit 403 includes a target cost calculation unit 404 and a connection cost calculation unit 405.

  The target cost calculation unit 404, for each speech unit (phoneme) indicated by the language information 104d, the language information 104d and prosodic information 109d, the candidate linguistic feature information 106b and the sound extracted by the speech unit candidate extraction unit 401. The consistency with the characteristic feature information 106a is calculated as a cost (target cost).

  Cost calculation based on bilingual features indicated by the linguistic information 104d and the linguistic feature information 106b is performed for each of the part of speech, the position in the morpheme, the position in the accent phrase, the syntax information, the phonological environment, and the morpheme notation. This is based on the degree of coincidence. The degree of coincidence of part of speech is the degree of coincidence between the part of speech of the morpheme to which the phoneme indicated by the language information 104d belongs and the part of speech indicated by the linguistic feature information 106b. The degree of coincidence of positions in the morpheme is the phoneme morpheme position indicated by the language information 104d and the phoneme morpheme position indicated by the linguistic feature information 106b (the distance from the morpheme head and the distance to the morpheme end). Is the degree of coincidence. The degree of coincidence of positions in the accent phrase refers to the position in the accent phrase of the phoneme indicated by the language information 104d and the position in the accent phrase of the phoneme indicated by the linguistic feature information 106b (the distance from the beginning of the accent phrase or the end of the accent phrase). Degree of coincidence). The coincidence degree of the syntax information is a coincidence degree between the relation destination position of the phrase including the phoneme indicated by the language information 104d and the relation destination distance indicated by the syntax information included in the linguistic feature information 106b. The degree of coincidence of the phoneme environment is the degree of coincidence between the phoneme indicated by the language information 104d and the phonemes before and after the phoneme, and the front phoneme, the target phoneme, and the rear phoneme indicated by the linguistic feature information 106b.

  As shown in FIG. 3, the degree of coincidence of the parts of speech can be made higher by considering the degree of coincidence of the finely classified parts of speech as shown in FIG. 3.

  The calculation of the cost based on both acoustic features indicated by the prosodic information 109d and the acoustic feature information 106a is performed based on the degree of coincidence of the duration time, the fundamental frequency, and the power. The degree of coincidence of the duration length is the degree of coincidence between the duration length of the phoneme indicated by the prosody information 109d and the duration length indicated by the acoustic feature information 106a. The coincidence degree of the fundamental frequency is the coincidence degree between the fundamental frequency of the phoneme indicated by the prosody information 109d and the fundamental frequency indicated by the acoustic feature information 106a. The power coincidence is the degree of coincidence between the phoneme power indicated by the prosody information 109d and the power indicated by the acoustic feature information 106a.

  The target cost calculation unit 404 calculates the final cost (target cost) by adding the cost based on the linguistic feature calculated as described above and the cost based on the acoustic feature.

  The connection cost calculation unit 405 calculates a distortion that occurs when candidates are combined as a connection cost.

  Here, the operation of the speech synthesizer in the present embodiment shown in FIG. 1 will be described in detail. Here, an operation when the text data 100t indicating the text “Today's weather is sunny” shown in FIG. 3 is input will be described. In the following description, phonemes are used as an example of speech units, but the present invention does not limit speech units to phonemes.

  First, the language analysis unit 104 represents the text indicated by the text data 100t in phoneme notation, and divides the phoneme notation into morphemes. In addition, the language analysis unit 104 performs syntax analysis to obtain syntax information (information indicating the relationship destination position). Further, the language analysis unit 104 performs reading and accent phrases. As a result, language information 104d as shown in FIG. 3 is generated.

  The prosody estimation unit 109 estimates the duration, fundamental frequency, and power for each phoneme based on the language information 104d shown in FIG. As a result, prosodic information 109d is generated.

The speech unit candidate extraction unit 401 of the speech unit selection unit 108 uses the acquired target language t _i having the language information 104d and the prosodic information 109d as elements in speech units (like the speech unit data in vector notation shown in FIG. Here, each phoneme) is configured. In this case, since the phoneme “ky” is an unvoiced consonant, there is no information regarding the fundamental frequency. However, when the phoneme is a voiced sound, the duration of the phoneme is divided into four, and the fundamental frequency at the intermediate point in each section is represented by four points (basic frequencies 1 to 4), thereby changing the fundamental frequency. features also can be expressed in the audio unit data and the target vector t _i. The present invention is not limited to the expression of the fundamental frequency as described above.

  Next, the speech unit candidate extraction unit 401 extracts a set of candidate speech unit data from the feature parameter DB 106. Specifically, the speech unit candidate extraction unit 401 extracts all speech unit data indicating the same phoneme as the selection target phoneme indicated in the language information 104d.

  Note that when the speech unit data stored in the feature parameter DB 106 is sufficiently large, the speech unit candidate extraction unit 401 may acquire candidates by adding restrictions based on the phoneme environment (forward phoneme and backward phoneme).

The target cost calculation unit 404 calculates the degree of coincidence between the target vector t _i and the candidate u _i as the target cost vector C _i ^t .

FIG. 8 is a diagram illustrating target vectors, candidates, and target cost vectors.
For example, as shown in FIG. 8, when the candidate u _i and the target vector t _i are given, the target cost calculation unit 404 calculates the degree of coincidence for each element of the vector, and the result is the target cost vector C _i. ^t . The target cost calculation unit 404 calculates a target cost from the target cost vector C _i ^t . That is, each element shown in the target cost vector C _i ^t is set as a sub cost, and the target cost is calculated by adding the weights to the sub costs.

  The weight of each sub-cost may be set based on experience. For example, the target cost calculation unit 404 may be configured to be determined by the following method. For example, the target cost calculation unit 404 performs multiple regression analysis using the cost value calculated by each parameter and the distance from the target with a representative phoneme, and uses the coefficient of each cost value of the regression model as a weight. . Further, the cepstrum distance may be used for estimating the distance from the target. Also, other weights such as equal weights may be used.

  Note that the sub-cost weights for linguistic features may be reduced in the order of morpheme information, accent phrase information, and syntax information. That is, the priority that contributes to the selection of the voice unit data decreases in the order of the matching degree of the morpheme information, the matching degree of the accent phrase information, and the matching degree of the syntax information. In the accent phrase information, the weight of the sub cost may be decreased in the order of the distance from the accent nucleus, the distance to the end of the accent phrase, and the distance from the beginning of the accent phrase. That is, the priority that contributes to the selection of audio unit data decreases in the order of the degree of coincidence of distance from the accent core, the degree of coincidence of the distance to the end of the accent phrase, and the degree of coincidence of the distance from the beginning of the accent phrase.

In the case of the example shown in FIG. 8, disagreement out with "erasing" of the candidate u _i with the "today" of the target vector t _i with respect to representation, "^ (beginning of a sentence)," said candidate u of the target vector t _i with respect to the front phoneme disagreement out with a "u" of _i, match out with "o" with the "o" in the target vector t _i candidate u _i with respect to the rear phoneme, the target vector t _i with respect to part-of-speech "common noun" and the candidate u _i Disagrees with “sa kin noun”. The target cost calculation unit 404 sets the match as the sub cost “0” and the mismatch as the sub cost “1” in the items related to the non-numeric elements. In addition, the target cost calculation unit 404 sets an absolute value of a difference between elements as a sub cost in an item related to a numerical element. Therefore, the sub-cost for the distance from the morpheme head is 4-1 = 3, the sub-cost for the distance to the morpheme head is 4-3 = 1, the sub-cost for the distance from the accent head is 4-1 = 3, The sub-cost for distance is 8-5 = 3, the sub-cost for duration is 32-25 = 7, and the sub-cost for power is 3000-2910 = 90. The target cost calculation unit 404 calculates the total target cost by adding the above-mentioned weights determined by experience, for example, to each sub cost.

  The connection cost calculation unit 405 calculates the distortion when the two audio unit data are connected as the connection cost. The calculation method may be anything, but for example, the connection cost calculation unit 405 uses the cepstrum distance in the connection frame of two voice unit data as the connection cost.

  The search unit 402 selects optimal speech unit data from the candidates extracted by the speech unit candidate extraction unit 401 using the target cost and the connection cost. Specifically, the search unit 402 searches for an optimal speech unit sequence based on the following mathematical formula 1.

数１において、ｎはテキスト（言語情報１０４ｄの音素表記）に含まれる音素の数（音素数）である。例えば、テキスト「今日の天気は晴れですよ。」に対してｎは２１である。ｕは候補として挙げられる音声単位データ、ｔは目標ベクトル、Ｃtは目標コスト、Ｃcは接続コストを示す。

In Equation 1, n is the number of phonemes (phoneme number) included in the text (phoneme representation of the language information 104d). For example, n is 21 for the text “Today's weather is sunny.” u represents voice unit data listed as candidates, t represents a target vector, Ct represents a target cost, and Cc represents a connection cost.

  The search unit 402 identifies a speech unit sequence that minimizes the cost C in the entire text and notifies the speech synthesis unit 110 of it.

  Next, a specific operation when the speech synthesizer of this embodiment acquires text data 100t indicating a foreign word will be described.

  For example, the language analysis unit 104 of the speech synthesizer acquires text data 100t indicating the Japanese text “This is a ground.”. The part of speech of “ground” in the above text is a foreign word.

  The language analysis unit 104 that has acquired such text data 100t generates language information 104d based on the text data 100t.

  FIG. 9 is a diagram showing the contents of the language information 104d generated from the text data 100t indicating a foreign word.

  Similar to the language information 104d shown in FIG. 3, the language information 104d includes a text, a series of phonemes (phoneme notation) corresponding to the text, morphemes included in the text, clauses included in the text, A morpheme part of speech, a morpheme position, an accent phrase position, and a dependency point position are shown. The language information 104d indicates a foreign word as a part of speech of the morpheme “ground”.

  The speech unit selection unit 108 selects optimal speech unit data for each phoneme indicated in the language information 104d from the feature parameter DB 106.

  FIG. 10A is an explanatory diagram for explaining a phoneme to be selected for voice unit data.

  For example, the speech unit selection unit 108 selects speech unit data optimal for the phoneme “u” of the vowel part of the foreign word “G” of “ground”.

Specifically, the speech unit selection unit 108 first generates a target vector t _i for the phoneme “u” and selects candidates u ₁ and u ₂ corresponding to the phoneme “u” from the feature parameter DB 106.

FIG. 10B shows a target vector t _i and candidates u ₁ and u ₂ for the phoneme “u”. The part-of-speech of the notation “Tadaomi” of the candidate u ₁ is a proper noun. The part of speech of the notation “glass” indicated by the candidate u ₂ is a foreign word, and the notation “glass” means “glass #”.

Speech unit selection unit 108, among the candidate u _1, u _2, selects the closest candidate to the target vector t _i, as the optimum speech unit data used for speech synthesis.

Here, the conventional speech synthesizer uses the phoneme environment (front phoneme, target phoneme, and rear phoneme) and the acoustic features (duration, power, fundamental frequency, etc.), so that candidates u ₁ and u Among the _two , candidate u ₁ is selected. The reason for the selection is that in the acoustic feature, the candidate u ₁ is closer to the target vector t _i than the candidate u ₂ . However, there is a difference between the phoneme “u” included in the Japanese proper noun and the phoneme “u” included in the foreign word that cannot be expressed by the acoustic feature described above. As a result, since the conventional speech synthesizer selects inappropriate speech unit data and uses it for speech synthesis, it outputs an unnatural synthesized speech.

On the other hand, the speech synthesizer of the present embodiment can select optimal speech unit data by using the part of speech that is one of linguistic features. That is, if the part of speech of the target vector t _i is a foreign word, the speech unit selection unit 108 of the speech synthesizer selects the candidate u ₂ that indicates the foreign word as a part of speech in consideration of the foreign word. As a result, the speech synthesizer of the present embodiment can convert a foreign word indicated by the text data 100t into a natural synthesized speech that seems to be a foreign word.

  Next, a specific operation when the speech synthesizer of this embodiment acquires text data 100t indicating a final particle will be described.

  For example, the language analysis unit 104 of the speech synthesizer acquires the text data 100t indicating the Japanese text “This is a ground, isn't it?”. The part of speech of “Ne” in the above text is a final particle.

  The language analysis unit 104 that has acquired such text data 100t generates language information 104d based on the text data 100t.

  The speech unit selection unit 108 selects optimal speech unit data for each phoneme indicated in the language information 104d from the feature parameter DB 106.

  FIG. 11A is an explanatory diagram for explaining a phoneme to be selected for speech unit data.

  For example, the speech unit selection unit 108 selects speech unit data that is optimal for the phoneme “e” of the vowel part of the final particle “ne”.

Specifically, the speech unit selection unit 108 first generates a target vector t _i for the phoneme “e” and selects candidates u ₁ and u ₂ corresponding to the phoneme “e” from the feature parameter DB 106.

FIG. 11B shows the target vector t _i and the candidates u ₁ and u ₂ for the phoneme “e”. The part of speech of the notation “root” of the candidate u ₁ is a common noun, and the notation “root” means “root”. The part of speech of the notation “ne” of candidate u ₂ is a final particle.

Speech unit selection unit 108, among the candidate u _1, u _2, selects the closest candidate to the target vector t _i, as the optimum speech unit data used for speech synthesis.

Here, the conventional speech synthesizer uses the phoneme environment (front phoneme, target phoneme, and rear phoneme) and the acoustic features (duration, power, fundamental frequency, etc.), so that candidates u ₁ and u Among the _two , candidate u ₁ is selected. The reason for the selection is that in the acoustic feature, the candidate u ₁ is closer to the target vector t _i than the candidate u ₂ . However, the phoneme “e” included in the Japanese final particle “ne” has a unique feature, and the feature is greatly different from the feature of the phoneme “e” included in the other part of speech “ne”. Therefore, although the speech unit data selected by the conventional speech synthesizer has a high degree of coincidence with the target vector t _{i in} terms of acoustic characteristics, there is a possibility that the speech unit data is not necessarily appropriate as speech unit data used for actual synthesized speech. .

On the other hand, the speech synthesizer of the present embodiment can select optimal speech unit data by using the part of speech that is one of linguistic features. That is, if the part of speech of the target vector t _i is a final particle, the speech unit selection unit 108 of the speech synthesizer selects the candidate u ₂ indicating the final particle as a part of speech considering that it is a final particle. As a result, the speech synthesizer according to the present embodiment can convert the final particle indicated by the text data 100t into a natural synthesized speech that expresses the feeling of doubt indicated by the final particle.

  12 (a), 12 (b), 12 (c), and 12 (d) are explanatory diagrams for explaining the effect of the present invention. These figures show the speech section of the phoneme “yo” (phoneme notation “yo”) belonging to the four words, “Ohtsuka, Kasuya,“ An Improved Speech Analysis-Synthesis Algorithm based on the Autoregressive with Exogenous Input Speech Production Model ” The result of the analysis shown in ", ICSLP2000" is shown. Such analysis is performed by applying a voice signal to a voice generation model and separating it into a vocal cord (sound source) and a vocal tract (filter).

  FIG. 12A shows the analysis result of the adverb “fully” for the phoneme “yo”. FIG. 12B shows the analysis result for the phoneme “yo” of the common noun “yoru” (night). FIG. 12C shows the analysis result for the phoneme “yo” which is a final particle included in one sentence. FIG. 12D shows the analysis result for the phoneme “yo”, which is a final particle included in another sentence.

  These analysis results show the center frequency F1 of the first formant, the center frequency F2 of the second formant, the center frequency F3 of the third formant, and the bandwidth of each formant. In each figure, the bandwidth is represented by a vertical line segment written over the lines indicating the center frequencies F1, F2, and F3. The center frequency of each formant indicates a peak due to resonance of the vocal tract, and the bandwidth indicates the strength of the resonance. A wide bandwidth means a weak resonance and a narrow bandwidth means a strong resonance.

  In the four analysis results shown in each figure, the common form attributed to the phoneme “yo” that the center frequency F1 of the first formant increases and the center frequency F2 of the second formant decreases from the first half to the second half of the time axis. Show features. Therefore, regardless of the part of speech of the morpheme including the phoneme “yo”, the trajectory of the center frequency of each formant for the phoneme “yo” (hereinafter referred to as formant trajectory) is similar.

  Thus, although the formant trajectory of the phoneme “yo” has a common feature, the feeling of hearing the phoneme “yo” with the ear is clearly different depending on the part of speech of the morpheme that contains the phoneme “yo”. . That is, a person has a close impression to the final particle phoneme “yo” shown in FIG. 12 (c) and FIG. 12 (d). On the other hand, a person has a different impression of the phoneme “yo” of the final particle shown in FIG. 12C and the phoneme “yo” included in the morpheme of the adverb shown in FIG. Similarly, a person has a different impression with respect to the phoneme “yo” of the final particle shown in FIG. 12C and the phoneme “yo” included in the morpheme of the common noun shown in FIG.

Such a difference in impression is difficult to explain with a formant trajectory alone.
At the end of the sentence where the final particle is uttered, a relaxed utterance is performed, so that the vocal cords tend to be weakened while closing while vibrating. In such a situation where the glottis (the gap formed between the folds on both sides of the glottis) is wide, it is known that the influence of resonance due to the space below the glottis such as the trachea and lungs (hereinafter referred to as the glottal lower space) appears strongly. ing. This is described in the literature “D. Klatt, L. Klatt,“ Analysis, Synthesis, and Perception of Voice Quality Variations among Female and Male talkers ”, J. Acoust. Soc. Am. 87 (2), February 1990, pp.820. -857 ".

  According to “D. Results III: Tracheal coupling” (p832) in this document, the phenomenon of the appearance of poles and zeros and the phenomenon of an increase in the bandwidth of the first formant occur due to the influence of resonance in the glottal space. In the analysis result shown in FIG. 12C, a weak peak 141 having a frequency different from that of the formant appears. Also, in the analysis result shown in FIG. 12D, a weak peak 142 having a frequency different from that of the formant appears as described above. In the above document, since it is described that a pole appears in the vicinity of 1700 Hz in the case of a woman, the peaks 141 and 142 are presumed to be the appearance phenomenon of the pole. In addition, the bandwidth of the first formant shown in the analysis result of FIG. 12C and the bandwidth of the first formant shown in the analysis result of FIG. Have.

  On the other hand, the two phenomena described above do not appear greatly in the analysis results shown in FIGS. 12 (a) and 12 (b).

  In this way, the impression that a person hears each phoneme varies greatly depending on the part of speech to which each phoneme belongs, even if the acoustic features indicated by the formant trajectory of each phoneme are similar. In particular, it largely differs depending on whether the part of speech is a final particle or whether the part of speech is a foreign word.

  Therefore, the speech synthesizer according to the present embodiment outputs a natural synthesized speech for a phoneme included in a morpheme, in order to select speech unit data corresponding to the part of speech of the morpheme (particularly a final particle or a foreign word). Can do it. That is, the speech synthesizer of this embodiment can output natural synthesized speech as shown in the text of the text data 100t.

  In the speech synthesizer according to the present embodiment, not only the acoustic features but also linguistic features such as whether or not a foreign word or a final particle is taken into account, so that the accuracy of the estimation of the acoustic features by the prosody estimation unit 109 is improved. Even when it is not sufficient, it is possible to select speech unit data with higher reliability based on the linguistic features of the speech unit data stored in the feature parameter DB 106.

  The speech synthesizer according to the present invention is useful as a car navigation system, a reading device in the entertainment field, and the like.

(Modification 1)
The speech synthesizer 110 of the above embodiment generates a synthesized speech signal based on a series of speech unit data held in the feature parameter DB 106. The voice synthesizer according to the present modification acquires a signal indicating a voice waveform corresponding to each voice unit data, and generates a synthesized voice signal by connecting the signals.

FIG. 13 is a configuration diagram showing the configuration of the speech synthesizer according to this modification.
The speech synthesizer according to this modification includes a feature parameter DB 106, a language analysis unit 104, a prosody estimation unit 109, a speech unit selection unit 108, a speech synthesis unit 110a, a speaker 111, and a speech waveform signal DB 101. Prepare.

  The voice waveform signal DB 101 holds a voice waveform signal indicating a voice waveform corresponding to each voice unit data stored in the feature parameter DB 106.

  When the voice synthesizer 110a identifies a series of voice unit data (voice unit series) selected by the voice unit selector 108, the voice synthesizer 110a acquires a voice waveform signal corresponding to each voice unit data from the voice waveform signal DB 101. Then, the voice synthesizer 110a generates a synthesized voice signal by connecting the voice waveform signals.

(Modification 2)
The cost calculation unit 403 of the above embodiment calculates a target cost by performing a predetermined weight on each sub-cost and integrating the weighted sub-costs. The cost calculation unit according to this modification is characterized in that the weight is changed.

FIG. 14 is a configuration diagram illustrating an internal configuration of the cost calculation unit according to the present modification.
The cost calculation unit 403a according to this modification includes a target cost calculation unit 404, a connection cost calculation unit 405, and a weight determination unit 501.

  The weight determination unit 501 changes the weight of the linguistic feature and the weight of the acoustic feature based on the reliability of the prosody information 109 d output from the prosody estimation unit 109 during the cost calculation in the target cost calculation unit 404. . Then, the weight determination unit 501 notifies the target cost calculation unit 404 of the changed weight. The target cost calculation unit 404 calculates the target cost based on the weight notified from the weight determination unit 501.

For example, when the reliability of the prosodic information 109d is low, the weight determination unit 501 increases the weight for the sub-cost of the linguistic feature and decreases the weight for the sub-cost of the acoustic feature. As a result, the target cost calculation unit 404 calculates the target cost based on the linguistic feature coincidence rather than the acoustic feature coincidence. Then, the search unit 402 selects speech unit data in consideration of the linguistic feature coincidence rather than the acoustic feature coincidence. That is, the search unit 402 does not select a candidate for a target vector t _i if the candidate matches with the acoustic feature but does not match with the linguistic feature. Make a choice.

  As described above, the cost calculation unit 403a according to the present modification changes the weight for the sub-cost according to the reliability of the prosody information 109d that is the estimation result of the prosody estimation unit 109. When it is difficult to select, robust voice unit data can be selected by focusing on the degree of matching of linguistic features without relying on direct fundamental frequency, duration length, power estimation results, etc. .

  For example, the prosody estimation unit 109 acquires language information 104d indicating a foreign word as shown in FIG. 9, and generates prosody information 109d based on the language information 104d. In this case, the weight determining unit 501 determines that the reliability of the prosodic information 109d is low, and increases the weight of the sub cost corresponding to the part of speech (for example, a foreign word). As a result, appropriate voice unit data is selected, and more natural synthesized voice can be output. When evaluating the degree of coincidence of features indicated by phonemes, not only perfect coincidence but also the degree of coincidence of phoneme group features may be considered. As a result, it is possible to flexibly cope with fluctuations in notation such as “ba” in Japanese and “va” in Japanese.

(Modification 3)
Here, the modification regarding the selection method of the audio | voice unit data in this embodiment is demonstrated.

  The voice unit selection unit 108 of the above embodiment selects the voice unit data in consideration of the linguistic feature and the acoustic feature at the same time. The speech unit selection unit according to the present modification first selects speech unit data giving priority to linguistic features.

FIG. 15 is a flowchart showing the operation of the audio unit selector according to the present modification.
First, the speech unit selection unit selects speech unit data listed as candidates from the feature parameter DB 106 (step S100).

  Next, the speech unit selection unit further selects speech unit data that matches the linguistic feature of the speech unit indicated by the language information 104d from the candidates selected in step S100 (step S102). Then, the sound unit selection unit calculates the cost of the selected sound unit data (step S104).

  Here, the voice unit selection unit determines whether or not the calculated cost is smaller than a threshold value (step S106). When it is determined that the value is smaller than the threshold value (Y in step S106), the speech unit selection unit notifies the speech synthesis unit 110 of the speech unit data selected in step S102 (step S108). On the other hand, when it is determined that the value is greater than or equal to the threshold value (N in step S106), the voice unit selection unit calculates the cost for each candidate selected in step S100, as in the above embodiment (step S110). . Then, the speech unit selection unit notifies the speech synthesis unit 110 of the candidate with the lowest cost, that is, speech unit data (step S112).

(Second Embodiment)
Hereinafter, a data creation apparatus for creating sound unit data used in the first embodiment will be described.

FIG. 16 is a configuration diagram illustrating the overall configuration of the data creation device according to the second embodiment.
The data creation device creates speech unit data stored in the feature parameter DB 106 of the speech synthesizer, and includes a text storage unit 701, a speech waveform storage unit 702, a speech analysis unit 703, and a language analysis unit 704. With.

  The voice waveform storage unit 702 is a database that stores a voice waveform signal indicating a recorded voice as a waveform. The text storage unit 701 stores the recorded voice transcript as text data. That is, the contents indicated by the speech waveform signal and the contents indicated by the text data are the same. The phoneme HMM storage unit 705 stores a phoneme HMM created for each phoneme.

  The language analysis unit 704 performs linguistic analysis on the text indicated by the text data held in the text storage unit 701, thereby extracting linguistic features from the text for each speech unit (for example, phonemes). Here, the linguistic features indicate phonological environment, morpheme information, syntax information, accent phrases, and the like. The language analysis unit 704 stores the linguistic feature information indicating the linguistic feature in the feature parameter DB 106 of the speech synthesizer and outputs it to the speech analysis unit 703.

  The voice analysis unit 703 acquires the linguistic feature information output from the language analysis unit 704 and acquires a voice waveform signal corresponding to the text from the voice waveform storage unit 702. Then, the speech analysis unit 703 divides the acquired speech waveform signal for each corresponding phoneme according to the phoneme notation shown in the acquired linguistic feature information. Here, the speech analysis unit 703 uses the phoneme HMM stored in the phoneme HMM storage unit 705 when dividing the speech waveform signal. Furthermore, the voice analysis unit 703 extracts an acoustic feature for each phoneme from the divided voice waveform signal. Here, the acoustic features indicate a fundamental frequency, a duration length, a cepstrum, and the like. Furthermore, the acoustic feature may indicate a speaker's emotion when speaking.

  Then, the voice analysis unit 703 generates acoustic feature information indicating the acoustic feature and stores it in the feature parameter DB 106 of the speech synthesizer.

  The operation of the data creation apparatus in this embodiment will be described. Here, a description will be given of a procedure when the data creation device adds voice unit data related to the text “Today's weather is sunny” to the feature parameter DB 106.

  First, the language analysis unit 704 reads text data from the text storage unit 701, performs morphological analysis and syntax analysis on the text indicated by the text data, and analyzes fields, readings, emotions, and the like. For example, the language analysis unit 704 generates linguistic feature information indicating the same content as the linguistic information 104d shown in FIG. 3 as an analysis result, and stores the linguistic feature information in the feature parameter DB 106.

  Next, the speech analysis unit 703 acquires a speech waveform signal corresponding to the text “Today's weather is sunny” from the speech waveform storage unit 702 and also acquires linguistic feature information from the language analysis unit 704. The speech analysis unit 703 segments the speech waveform signal for each phoneme by using the phoneme HMM of the phoneme HMM storage unit 705 based on the phoneme notation indicated by the linguistic feature information. In this example, the speech unit is a phoneme, but it is not particularly limited to a phoneme.

  When the speech analysis unit 703 segments the speech waveform signal for each phoneme, the speech analysis unit 703 analyzes the fundamental frequency, the duration, and the power for each phoneme. The analysis method is not particularly limited, and any analysis method may be used. The voice analysis unit 703 stores the analysis result in the feature parameter DB 106 as acoustic feature information.

  Note that the speech analysis unit 703 may analyze the emotion and add the analysis result to the acoustic feature information without the language analysis unit 104 analyzing the emotion. In addition, when the speech waveform signal includes information indicating emotion in advance, the information may be added to the linguistic feature information or the acoustic feature information.

  By operating as described above, the data creation device creates vector unit speech unit data as shown in FIG. 6 in the feature parameter DB 106 for each phoneme. In the case of the text “Today's weather is sunny”, 21 vector-displayed voice unit data are created.

  As described above, according to the present embodiment, speech unit data including both linguistic feature information and acoustic feature information for each phoneme can be easily constructed in the feature parameter DB 106.

  As mentioned above, although this invention was demonstrated using 1st and 2nd embodiment and a modification, this invention is not limited to these.

  For example, in the first and second embodiments, Japanese text is converted into speech, but the present invention may convert text in other languages into speech, particularly for languages having foreign words and final particles. Has an effect on text.

  In the first and second embodiments, phonemes are handled as speech units, but other units may be handled as speech units.

  The speech synthesizer according to the present invention produces an effect that a natural synthesized speech can be output as indicated by text data, and is useful as, for example, a car navigation system or a reading device in the entertainment field.

It is a block diagram which shows the structure of the speech synthesizer in the 1st Embodiment of this invention. It is a block diagram which shows an example of an internal structure of a language analysis part same as the above. It is a figure which shows an example of the content of the language information same as the above. It is a figure which shows the content of the acoustic feature information same as the above. It is a figure which shows the content of the linguistic feature information same as the above. It is a figure which shows the content of the one audio | voice unit data stored in feature parameter DB same as the above. It is a block diagram which shows the specific structural example of an audio | voice unit selection part same as the above. It is a figure which shows a target vector same as the above, a candidate, and a target cost vector. It is a figure which shows the content of the linguistic information produced | generated from the text data which shows a foreign word same as the above. (A) is explanatory drawing for demonstrating the phoneme used as the selection object of audio | voice unit data same as the above, (b) is a figure which shows the target vector and candidate corresponding to phoneme "u" same as the above. (A) is explanatory drawing for demonstrating the phoneme used as the selection object of audio | voice unit data same as the above, (b) is a figure which shows the target vector and candidate corresponding to phoneme "e". (A) is a figure which shows the analysis result with respect to the phoneme "yo" of adverb "well" (fully), (b) is a figure which shows the analysis result with respect to the phoneme "yo" of common noun "yoru" (night) (C) is a diagram showing an analysis result for a phoneme “yo” that is a final particle included in one sentence, and (d) is an analysis for a phoneme “yo” that is a final particle included in another sentence. It is a figure which shows a result. It is a block diagram which shows the structure of the speech synthesizer which concerns on the modification 1 of 1st Embodiment. It is a block diagram which shows the internal structure of the cost calculation part which concerns on the modification 2 of 1st Embodiment. It is a flowchart which shows operation | movement of the audio | voice unit selection part which concerns on the modification 3 of 1st Embodiment. It is a block diagram which shows the whole structure of the data preparation apparatus in the 2nd Embodiment of this invention. It is a block diagram which shows the structure of the conventional speech synthesizer.

Explanation of symbols

100t text data 104 language analysis unit 104d language information 106 feature parameter DB
108 speech unit selection unit 109 prosody estimation unit 109d prosody information 110 speech synthesis unit 111 speaker

Claims (18)

A speech synthesizer that acquires text data and converts text indicated by the text data into speech,
Storage means for storing, for each voice unit, a foreign word attribute as to whether or not the voice unit belongs to a foreign word, and voice unit data indicating an acoustic feature of the voice unit;
Feature estimation means for acquiring text data and estimating the foreign word attribute and acoustic feature of each speech unit for each of a plurality of speech units representing the text of the text data;
Selection means for selecting, from the storage means, speech unit data indicating contents similar to the foreign word attributes and acoustic features of the speech units estimated by the feature estimation means for each speech unit representing the text;
A speech synthesizer comprising: speech output means for generating and outputting synthesized speech using a plurality of speech unit data selected by the selection means. The selection means includes
When the feature estimation means estimates that the speech unit belongs to a foreign word as the foreign word attribute of the speech unit, the speech unit data indicating that it belongs to the foreign word is preferentially selected as the foreign word attribute. Item 2. The speech synthesizer according to Item 1. The speech unit data further indicates a final particle attribute as to whether the speech unit belongs to a final particle;
The feature estimation means estimates, for each of a plurality of speech units representing the text of the text data, the foreign word attribute, acoustic feature, and final particle attribute of the speech unit,
The selecting means selects, from the storage means, speech unit data showing contents similar to the foreign word attribute, acoustic feature, and final particle attribute of the speech unit estimated by the feature estimating means. The speech synthesizer according to claim 1. The selection means includes
When the feature estimation means estimates that the speech unit belongs to a final particle as the final particle attribute of the speech unit, the speech unit data indicating that it belongs to the final particle is preferentially selected as the final particle attribute. Item 4. The speech synthesizer according to item 3. The speech synthesizer according to claim 3, wherein the acoustic feature indicates at least one of a duration length of a speech unit, a fundamental frequency, and power. The speech unit data further indicates phonological environment to which the speech unit belongs, syntax information about the syntax of the speech unit, and accent phrase information about the accent phrase of the speech unit,
The feature estimation means includes, for each of a plurality of speech units representing the text of the text data, the foreign word attribute, acoustic feature, final particle attribute, phonological environment, syntax information, and accent phrase information of the speech unit. Estimate
The selection means includes speech unit data indicating content similar to the foreign word attribute, acoustic feature, final particle attribute, phonological environment, syntax information, and accent phrase information of the speech unit estimated by the feature estimation means. 6. The speech synthesizer according to claim 5, wherein the speech synthesizer is selected from storage means. The selection means includes
A first sub-cost is derived by quantitatively evaluating the similarity between the foreign word attribute of the speech unit estimated by the feature estimating means and the foreign word attribute of the speech unit data stored in the storage means. 1 derivation means;
A second sub-cost is derived by quantitatively evaluating the similarity between the acoustic feature of the speech unit estimated by the feature estimation means and the acoustic feature of the speech unit data stored in the storage means. Two derivation means;
Cost deriving means for deriving a cost using each of the first and second sub-costs derived by the first and second deriving means;
The speech synthesis apparatus according to claim 1, further comprising: a data selection unit that selects speech unit data from the storage unit based on the cost derived by the cost deriving unit. The cost calculating means includes
The speech synthesizer according to claim 7, wherein the cost is derived by weighting and integrating each of the first and second sub-costs derived by the first and second derivation means. . The speech synthesizer further includes:
A weight determination unit that identifies the reliability of the acoustic feature estimated by the feature estimation unit and determines a weight for each of the first and second sub-costs according to the reliability;
The cost deriving means is
9. The speech synthesizer according to claim 8, wherein the weight determined by the weight determining means is added to the first and second sub-costs. The weight determining means includes
When the reliability of the acoustic feature is low, the similarity of the foreign word attribute contributes to the selection of the voice unit data by the data selection means rather than the similarity of the acoustic feature. The speech synthesizer according to claim 9, wherein a weight for 2 sub-costs is determined. The selecting means further includes:
A third deriving unit for quantitatively evaluating acoustic distortion when connecting a plurality of audio unit data stored in the storage unit and deriving a connection cost;
The cost deriving unit derives the cost using the first and second sub-costs derived by the first and second deriving units and the connection cost derived by the third deriving unit. The speech synthesizer according to claim 10. A speech synthesis method for acquiring text data and converting text indicated by the text data into speech using data stored in a storage means,
The storage means stores, for each voice unit, a foreign word attribute indicating whether the voice unit belongs to a foreign word, and voice unit data indicating an acoustic feature of the voice unit,
The speech synthesis method includes:
A feature estimation step of obtaining text data and estimating the foreign word attribute and acoustic feature of each speech unit for each of a plurality of speech units representing the text of the text data;
A selection step of selecting, from the storage means, speech unit data indicating contents similar to the foreign word attributes and acoustic features of the speech unit estimated in the feature estimation step for each speech unit included in the text;
A speech synthesis method comprising: a speech output step of generating and outputting synthesized speech using a plurality of speech unit data selected in the selection step. The speech unit data further indicates a final particle attribute as to whether the speech unit belongs to a final particle;
In the feature estimation step, for each of a plurality of speech units representing the text of the text data, the foreign word attribute, acoustic feature, and final particle attribute of the speech unit are estimated,
In the selection step, speech unit data showing content similar to the foreign word attribute, acoustic feature, and final particle attribute of the speech unit estimated in the feature estimation step is selected from the storage unit. The speech synthesis method according to claim 12. A program for acquiring text data and converting the text indicated by the text data into speech using data stored in a storage means,
The storage means stores, for each voice unit, a foreign word attribute indicating whether the voice unit belongs to a foreign word, and voice unit data indicating an acoustic feature of the voice unit,
The program is
A feature estimation step of obtaining text data and estimating the foreign word attribute and acoustic feature of each speech unit for each of a plurality of speech units representing the text of the text data;
A selection step of selecting, from the storage means, speech unit data indicating contents similar to the foreign word attributes and acoustic features of the speech unit estimated in the feature estimation step for each speech unit included in the text;
A program that causes a computer to execute a voice output step of generating and outputting synthesized voice using a plurality of voice unit data selected in the selection step. A data creation device for creating speech unit data used for speech synthesis,
Voice storage means for storing a voice waveform signal indicating the voice as a waveform;
Text storage means for storing text data indicating text corresponding to the voice of the voice waveform signal;
Language analysis means for obtaining text data from the text storage means, dividing the text of the text data into speech units, and analyzing foreign language attributes as to whether each speech unit belongs to a foreign language;
Obtaining an audio waveform signal from the audio storage means, dividing the audio indicated by the audio waveform signal into audio units, and analyzing acoustic characteristics for each audio unit;
Creating means for creating the speech unit data for each speech unit and storing it in the storage means so as to indicate the foreign word attribute analyzed by the language analysis means and the acoustic features analyzed by the acoustic analysis means; A data creation device comprising: The language analysis means further analyzes a final particle attribute as to whether or not each speech unit belongs to a final particle,
The creation means creates the speech unit data for each speech unit so as to indicate the foreign word attribute and final particle attribute analyzed by the language analysis means and the acoustic features analyzed by the acoustic analysis means. The data creation device according to claim 15. The data creation apparatus according to claim 16, wherein the acoustic feature indicates at least one of a duration length of a voice unit, a fundamental frequency, and power. A data creation method for creating speech unit data used for speech synthesis using data stored in a storage means,
The storage means stores a speech waveform signal indicating speech as a waveform, and text data indicating text corresponding to the speech of the speech waveform signal,
The data creation method is:
A language analysis step of obtaining text data from the storage means, dividing the text of the text data into speech units, and analyzing a foreign language attribute as to whether each speech unit belongs to a foreign language;
An acoustic analysis step of obtaining a speech waveform signal from the storage means, dividing the speech indicated by the speech waveform signal into speech units, and analyzing acoustic features for each speech unit;
A creation step of creating the speech unit data for each speech unit and storing it in a storage means so as to indicate the foreign language attribute analyzed in the language analysis step and the acoustic feature analyzed in the acoustic analysis step; A data creation method characterized by including.

Images