JP2000010580A - Method and device for synthesizing speech - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make a synthetic speech a human and natural one by considering the physical limit of an articulatory organ thereby deciding a phoneme continuous time length. SOLUTION: A KANJI (Chiniese character)/KANA (Japanese syllabary)-mixed sentence being the object of speech synthesis is analyzed by a language processing part 101 to obtain a speech sign string in which a phoneme sign series and accent information are described and a language analytic result including the part of speech information of an independent word incorporated in respective accent phrases. A beat interval predictive processing part 108b in a speech synthetic part 102 decides synchronous timing of respective morae by a statistic method based on the number of morae and the part of speech information at every accent phrase based on the language analytic result. A phoneme continuous time length calculation processing part 107 obtains the time change of an articulation model based on the phoneme sign series in the speech sign string and the synchronous timing, and decides respective phoneme continuous time lengthes based on the time change of the articulation model.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention analyzes text data to be subjected to speech synthesis, obtains a linguistic analysis result, and, based on the linguistic analysis result, analyzes individual phonemes contained in the reading of the text data. The present invention relates to a speech synthesis method and apparatus that determines a duration time, selects a speech unit, and connects the speech units selected based on the determined phoneme duration time to synthesize speech.

[0002]

2. Description of the Related Art It is known that a typical speech synthesizer of this type is a rule synthesizer capable of subdividing and accumulating speech and synthesizing an arbitrary speech by a combination thereof. Hereinafter, an example of the prior art of the rule synthesizing apparatus will be described with reference to the drawings.

FIG. 12 is a block diagram showing a configuration of a conventional rule synthesizing apparatus. The rule synthesizer of FIG. 12 converts input text data (hereinafter, simply referred to as text) into a symbol string composed of phonemes and prosody, and generates a speech from the symbol string (Text-to-speech conversion). : Hereinafter referred to as TTS).

The TTS in the rule synthesizing apparatus shown in FIG.
The processing mechanism is roughly divided into two processing units, a language processing unit 12 and a speech synthesis unit 13, and is generally performed as follows in the case of Japanese rule synthesis.

First, the language processing section 12 performs linguistic processing such as morphological analysis and syntactic analysis on text (kanji-kana mixed sentence) input from the text file 11, decomposes into morphemes, estimates dependency relations, and the like. At the same time, the pronunciation and accent type are given to each morpheme. After that, the language processing unit 12 determines the accent type of each phrase (hereinafter referred to as an accent phrase) serving as a delimiter at the time of reading aloud, using accent movement rules such as compound words. Normally, the language processing unit 12 of the TTS can output the reading and accent type for each accent phrase thus obtained as a symbol string (hereinafter referred to as a phonetic symbol string).

Next, in the speech synthesizer 13, the duration of each phoneme included in the obtained reading is determined by the phoneme duration determination processor 14. In general, the duration of a phoneme is determined based on the isochronism of the beats specific to Japanese, as described in Shinji Kawakami, “Japanese Voice Overview (Sakura Kaede)” p.100. The duration of the consonant is fixed depending on the type of the consonant, and as shown in FIG. 13, the reference time of each mora (beat synchronization time: here, a part extending from a consonant to a vowel, indicated by the symbol △ in the figure) The vowel duration is determined so that the interval between the vowels is constant.

The interval of the beat synchronization time (hereinafter simply referred to as a beat interval) is determined as follows. That is,
Based on the empirical rule that “humans read more as the number of mora increases, the beat interval is determined by referring to the table (mora number / beat interval correspondence table) 20 based on the mora number of the accent phrase obtained by observing the actual voice. obtain.

Subsequently, in the speech synthesis unit 13, the phoneme parameter generation processing unit 16 reads necessary speech units from the speech unit memory 15 in accordance with the "reading" obtained as described above, and reads the read speech units. For `` phoneme duration ''
Are connected while expanding and contracting in the time axis direction according to the above, to generate a feature parameter sequence of the voice to be synthesized.

Here, the speech unit memory 15 stores a large number of speech units prepared in advance. The speech unit is
After analyzing the voice uttered by the announcer or the like and obtaining a predetermined characteristic parameter of the voice, the voice is included in the Japanese voice in a predetermined synthesis unit, for example, a Japanese syllable (consonant vowel: hereinafter, referred to as CV). It is created by cutting out all syllables from the above feature parameters.

Here, a low-order cepstrum coefficient is used as a parameter. The low-order cepstrum coefficient can be obtained as follows. First, a window function (here, a Hanning window) is applied to voice data uttered by an announcer or the like at a constant width and a constant cycle, and a Fourier transform is performed on a voice waveform in each window to calculate a short-time spectrum of the voice.
Next, after the obtained short-time spectrum power is logarithmically obtained to obtain a logarithmic power spectrum, the logarithmic power spectrum is subjected to inverse Fourier transform. The cepstrum coefficient is calculated in this way. In general, it is known that higher-order cepstrum coefficients hold fundamental frequency information of speech and lower-order cepstrum coefficients hold spectrum envelope information of speech.

In the voice synthesizing unit 13, the pitch pattern generation processing unit 17 further sets a point pitch at a time when the pitch changes based on the accent type, and performs linear interpolation between the plurality of set point pitches. Generates an accent component of the pitch, into which the intonation component (usually frequency-
A pitch pattern is generated by superimposing a monotonically decreasing straight line on the time axis. A periodic pulse based on the pitch pattern is used as a sound source in a voiced section, and a white noise is used as a sound source in an unvoiced section. A filter coefficient is calculated from a characteristic parameter sequence of one voice, and the calculated coefficient is provided to a synthesis filter processing unit 18 to synthesize a desired voice. I do. Here, the synthesizing filter processing unit 18 uses the cepstrum coefficient as a filter coefficient directly as L
An MA (Log Magnitude Approximation) filter (logarithmic amplitude approximation filter) is used as a synthesis filter.

The processing up to this point is generally performed by digital processing. Therefore, the synthesized voice is a discrete signal. Therefore, the voice synthesizer 13 finally converts this discrete waveform into a D / A (digital / analog) signal. ) To the converter 19 to convert the discrete signals into electrical analog signals.
By driving a speaker or the like with the analog signal thus obtained, a sound that can be perceived by hearing can be synthesized.

[0013]

In a conventional speech synthesizer represented by the rule synthesizer described above, the speech generated by the speech synthesizer has the following problems. First, in the conventional speech synthesizer, when the duration of each phoneme included in the reading is determined in the speech synthesis unit, the reference time of each mora is determined based on the isochronism of the Japanese beat as described above. Are determined so that the time interval (beat interval) of the data becomes constant. However, when a human utters a voice, it is difficult to maintain isochronism due to physical restrictions on articulatory organs such as jaws, lips, and tongue that control the pronunciation (articulation) of words. Therefore, in reality, the isochronism is disturbed by the influence of the type of the phoneme and the phonemes before and after the phoneme, but it gives the voice a humanity and the individuality of the speaker.

Therefore, in the conventional method of determining the phoneme duration based only on the isochronism of the Japanese mora in the speech synthesizer, such physical restrictions on articulators are not taken into account. There is a drawback that the temporal arrangement of the mora is too constant and the humanity of the synthesized voice is impaired.

The present invention has been made in view of the above circumstances, and its object is to determine a phoneme duration in consideration of physical constraints of articulatory organs, thereby making synthesized speech more human-like and natural. Another object of the present invention is to provide a voice synthesizing method and apparatus capable of synthesizing a voice that is easy to hear and that does not get tired even after long hours of listening.

[0016]

SUMMARY OF THE INVENTION The present invention provides an articulatory method in which text data to be subjected to speech synthesis is analyzed to obtain a linguistic analysis result, and based on the linguistic analysis result, a movement of an articulator is modeled. A time change of the model is obtained, and a duration of each phoneme included in the reading of the text data is determined based on the obtained time change of the articulation model, and a speech unit is selected based on the result of the language analysis. Then, the speech is synthesized by connecting the speech units selected based on the determined duration of the phoneme.

In the present invention, the articulatory model is controlled using the result of language analysis of text data to be subjected to speech synthesis, and the duration of the phoneme is obtained based on the control result of the articulatory model. Can reflect the physical constraints of the articulatory organs when uttering the voice to the phonological duration, and at the same time give subtle utterance speed in the sentence to the synthesized speech, and synthesize a human-like natural and easy-to-hear speech it can.

Here, the synchronization timing of each mora included in the reading of the text data is determined based on the result of the language analysis, and based on the phonetic information included in the result of the language analysis and the determined synchronization timing, the sound timing of the articulator is determined. If the temporal change of an articulatory model that models movement is to be obtained, the physical constraints of articulatory organs when a human utters a voice are accurately reflected in the phonological duration, and at the same time, the subtlety in the sentence A high utterance speed can be given to the synthesized speech with high accuracy, and a more human-like, natural and easy-to-hear speech can be synthesized.

In particular, when determining the time change of the articulatory model, the articulatory organ operation command time to the succeeding phoneme of each phoneme included in the phoneme information acquired by the language analysis is determined based on the previously determined synchronization timing. If so, it becomes possible to reflect the physical restriction of the articulatory organ when a human utters a voice to the phonological duration with higher accuracy.

Further, when determining the articulatory organ operation command time for the succeeding phoneme of each phoneme, it is determined whether the phoneme is a consonant or a vowel, and at the time given by the determination result and the synchronization timing. Based on this, it is preferable to determine the articulator organ operation command time for the subsequent phoneme of the phoneme. Here, if the phoneme is a consonant, the articulatory organ operation command time to the succeeding phoneme of the consonant is determined based on the time given by the synchronization timing, and if the phoneme is a vowel, the articulation of the vowel is determined. Is compared with the time given by the synchronization timing, and based on the comparison result, the articulation to the succeeding phoneme of the vowel based on the time given by the synchronization timing or the time at which the articulation of the vowel is reached The organ operation command time may be determined. Specifically, if the time to reach the articulation of the vowel is smaller than the time to reach the articulation of the vowel if the time to reach the articulation of the vowel is greater, On the basis of this, the articulator organ operation command time for the succeeding phoneme of the vowel is determined. In this way, occurrence of an unnatural articulation motion can be prevented.

In order to determine the synchronization timing of each mora, it is preferable to estimate the synchronization timing by a statistical method based on the result of language analysis. In this case, a subtle change in utterance speed in the text is estimated by a statistical method, so that a human-like voice that is easy to hear can be synthesized. Here, for example, quantification class 1 can be applied as the statistical method. In the linguistic analysis, for each utterance segment (for example, an accent phrase), the part of speech information of the content word (independent word) included in the utterance segment is obtained. It is preferable to estimate the synchronization timing of each mora by a statistical method on a per speech segment basis based on the number of mora and the part of speech information for each speech segment.

The articulatory model preferably employs an articulatory model that models the movements of the articulators of the jaw, lips, and tongue.
It may be represented by a step response function of a critical damping quadratic linear system.
In such an articulation model, the model is simplified, so that the amount of calculation is small.

Further, as the articulatory model parameters, an allowable range, which is the state of the articulatory organ in which the phoneme is recognized as being uttered, is assigned for each phoneme, and a boundary between the phonemes is determined based on the allowable range. Determining the duration of a phoneme reflects the relatively ambiguous movements of the articulators of the jaw, lips, and tongue when a human utters normally, making it more human-like, natural, and easier to hear. It is possible to synthesize a voice that does not get tired even when listening. The time in the allowable range is the time to reach the articulation of the corresponding phoneme, and the time out of the allowable range is the time to exit from the articulation of the corresponding phoneme. As a method of determining the boundary between phonemes based on the permissible range, for example, the time (tot) at which the state of any articulator first falls out of the corresponding permissible range of the phoneme (the relevant phoneme) and the state of all articulators Is an intermediate time point in a section sandwiched by a time point within a corresponding allowable range of a subsequent phoneme (subsequent phoneme) (when both the phoneme and the subsequent phoneme are vowels). In addition,
The time (tout) at which the state of any of the articulators first goes out of the corresponding permissible range of the corresponding phoneme is defined as the boundary between the phonemes (when the phoneme is a consonant), or the state of all the articulators is the succeeding phoneme. Can be used as a boundary between phonemes (when the phoneme is a vowel and the subsequent phoneme is a consonant).

Further, an articulation model parameter set composed of articulation model parameters for each phoneme for controlling the articulation model, which is created based on the actual speech, is held. If the articulatory model is controlled based on the set, the synthesized speech can be made more human, and it is possible to imitate the tone of the speaker who uttered the speech used to create the parameter.

In particular, a plurality of articulation model parameter sets created based on voices of different speakers are held, and one set of articulation model parameters is selected from the plurality of articulation model parameters during speech synthesis. Is selected and the tone model is controlled based on the selected set of tone model parameters, the tone of the synthesized speech can be variously changed.

Further, when the above-mentioned statistical method of quantification type 1 is applied, a plurality of category quantity sets created based on the voices of different speakers are held, and when estimating the synchronization timing of each mora, If one category quantity set is selected from the plurality of category quantity sets and used, subtle changes in utterance speed in the text can be variously changed.

[0027]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a schematic configuration of a speech rule synthesizing apparatus according to an embodiment of the present invention. The speech rule synthesizer (hereinafter referred to as a speech synthesizer) is supplied on a recording medium such as a CD-ROM, a floppy disk, a memory card, or a communication medium such as a network on an information processing apparatus such as a personal computer. This is realized by executing dedicated software (sentence-to-speech conversion software), and has a sentence-to-speech conversion (TTS) processing function, that is, a sentence-to-speech conversion processing (sentence-to-speech synthesis processing) function of generating speech from text. The functional configuration is roughly divided into the language processing unit 101,
It is divided into a speech synthesis unit 102.

The language processing unit 101 analyzes input sentences such as kanji and kana mixed sentences to generate reading information and accent information, and converts a phonetic symbol sequence and accent information based on these information into a phonetic symbol sequence. It controls the generation process.

The speech synthesizing unit 102 controls a process of generating a speech based on a speech symbol string output from the language processing unit 101. Now, in the speech synthesizer of FIG. 1, a text (here, a Japanese document) to be subjected to sentence-to-speech conversion (reading) is stored as a text file 103.
In this apparatus, according to the sentence-to-speech conversion software, a sentence mixed with kanji and kana is read out one by one from the file 103, and the sentence-to-speech conversion process described below is performed by the language processing unit 101 and the speech synthesis unit 102 to synthesize speech.

First, a sentence (input sentence) mixed with kanji or kana read from the text file 103 is input to the language processing unit 10.
1 is input to the language analysis processing unit 104. The linguistic analysis processing unit 104 performs a morphological analysis on the input kanji-kana mixed sentence to generate reading information and accent information. Morphological analysis is an operation of analyzing which character string forms a phrase in a given sentence, and what the structure of the word is.

For this purpose, the linguistic analysis processing unit 104 uses a morpheme dictionary 105 having "morpheme", which is the minimum component of a sentence, as a headword and a connection rule file 106 in which connection rules between morphemes are registered. That is, the linguistic analysis processing unit 104 obtains all morpheme sequence candidates obtained by collating the input sentence with the morpheme dictionary 105, and from among them,
A combination that can be grammatically connected back and forth with reference to the connection rule file 106 is output. In the morphological dictionary 105,
Along with grammatical information used at the time of analysis, morpheme readings and accent types are registered. For this reason, if a morpheme is determined by morphological analysis, reading and accent type can be given at the same time.

For example, if a morphological analysis is performed on the sentence “go to the park and read the book”, the following is obtained: / park / go / go / book / read / read /. And morphemes.

Each morpheme is given a reading and an accent type, and becomes / Kouen / D / Itte / Phon / ヲ / Yomi / Mas /. Here, the morpheme containing “＾” has a high pitch in the mora immediately before it, and implies that the pitch is dropped in the mora immediately after that. Also "＾"
If there is no, it means that it is a flat accent.

By the way, when a human reads a sentence, he does not read with such an accent in morpheme units, but reads several morphemes together and adds an accent for each unit.

In view of the above, the linguistic analysis processing unit 104 further summarizes the morphemes with one accent phrase (accent giving unit) and also estimates the movement of the accent due to the summary. In addition to this, the linguistic analysis processing unit 104 also adds information such as devoicing of vowels and pauses when breathing out. Thus, in the above example, the following phonetic symbol sequence is finally generated.

/ Cohenue / Itte. Here, a period “.” Indicates a pause, and “()” indicates a syllable in which a vowel is unvoiced.

The linguistic analysis processing unit 104 outputs such a speech symbol string, and at the same time, outputs part-of-speech information of a content word (independent word) included in each accent phrase and an inflected form according to the type of part-of-speech. . In the above example, / common noun (park) / verb (go) / common noun (book) /
The verb (read) / is output.

When the speech symbol string and the part of speech information are output by the language analysis processing unit 104 in the language processing unit 101, the phoneme duration calculation processing unit 107 in the speech synthesis unit 102 is activated.

The phoneme duration calculation processing unit 107 determines the duration of the consonant part and vowel part of each mora included in the input sentence according to the phoneme information in the speech symbol string generated by the language analysis processing unit 104. . The outline of the processing for determining the duration in the phoneme duration processing unit 107 is as follows.

As described above, in the process of generating human speech, physical restrictions on the movement of articulators affect the duration of phonemes. In Japanese speech, this restriction of articulatory organs disturbs the isochronicity of the beat, which is a characteristic of the time structure unique to Japanese. However, in practice, the isochronism is disturbed, but on the contrary, it gives the voice a humanity.

Therefore, one articulatory model is used in which the state of a plurality of articulators is used as a parameter, the model is controlled in accordance with the phoneme sequence to be synthesized, and the phoneme duration is determined based on the control result.

As for the articulatory model, Fujimura-Coke
Various models have been proposed, including the articulation model for r.
However, many of these models in recent years aim to correlate articulatory movements with the acoustic properties of speech, and simulate articulatory control mechanisms to approximate the acoustic properties of the vocal tract. , The structure and control of the model are complicated.

A simple model is sufficient as a model required to determine the duration of the phoneme, as long as the effect of the physical constraints of the articulator on the duration of the phoneme can be expressed.

Therefore, in the present embodiment, four articulatory organs that are likely to be physically constrained in their movements in an actual utterance are selected, and an articulatory model for phoneme duration control is formed by these. The selected articulators are the jaw opening (J), the lip rounding (L), the front tongue position (FT), and the rear tongue position (BT) shown in FIG.

Then, in order to simulate the movement of the articulatory organs, phonemes pronounced in different articulation styles, that is, all abnormal sounds are distinguished. For example, as shown in FIG. 4, the sound-repelling “n” has several different articulation styles depending on the subsequent phonemes.

Based on such phonetic classification, Japanese voices are classified into unvoiced vowels and nasal vowels for vowels and palate consonants for consonants. According to the input example of the sentence “go to the park and read the book”, each phoneme of the phoneme sequence included in the phonetic symbol sequence input from the language analysis processing unit 104 in the language processing unit 101 is: First, it is represented by a sequence (first phoneme information) as shown in FIG. In FIG. 5A, /: / indicates a long sound, / N / indicates a sound repellent, and / Q / indicates a prompting sound.

Further, each phoneme is converted from its phoneme environment by the phoneme duration calculation processing section 107 (in the articulation model time change determination processing section 107b), ie, the phoneme sequence of the above detailed classification, that is, the abnormal sound level. It is converted into a phoneme sequence (second phoneme information) as shown in FIG. Note that the conversion of the abnormal sound level into the phoneme sequence may be performed by the language processing unit 101 (for example, the language analysis processing unit 104) instead of the phoneme duration calculation processing unit 107.

In this embodiment, each phoneme ph has a unique state Ainh (k, ph) for each articulator k (k is J, L, FT, BT) and a range of the articulator k (hereinafter, referred to as “k”). Upper limit Amax (k, ph) and lower limit Ami
3 × 4 (= 12) n (k, ph) and its phoneme ph
Are assigned to the parameters of a total of thirteen articulation models having the minimum duration Dmin (ph).

When one phoneme ph is considered, the state of each articulatory organ k of a typical articulatory model for uttering the phoneme is the eigenstate Ainh (k, ph). On the other hand, the state of the articulatory organ in which this phoneme is recognized as being uttered is not one point in the eigenstate but has a certain allowable range.
Therefore, the range of each articulator k that can be accepted as the articulation of the phoneme is defined as Amax (k, ph) and Amin as described above.
Expressed as (k, ph). In this embodiment, Ainh (k,
ph), Amax (k, ph) and Amin (k, ph) are normalized with the movable range of the articulator being 0 to 1. For example, parameter values for phoneme [i] are as shown in FIG.

On the other hand, a time series M (k, t) representing the movement of each articulator k is calculated by the following equation (1) based on the phoneme series to be synthesized. Here, ΣRi (k, t) represents the number of phonemes in the phoneme series as i = 1.
If i = N, i = 1 of Ri (k, t)
Ｉi = N−1.

Also, Ri (k, t) is defined as Ri (k, t) in the range of t <ti, assuming that the start time of transition of the model from the ith phoneme phi to the succeeding phoneme phi + 1 (i + 1th phoneme) is ti. k, t) = 0, and Ri (k, t) = ｛Ainh (k, phi + 1) −Ainh (k, ph) in the range of t ≧ ti.
i) It is represented by｝ S (t-ti).

S (t) is approximated using the step response of the critical damping quadratic linear system, ie, S (t) = 1- (1 + at) e- ^at (2). Here, a represents the natural angular frequency αk of the articulator k. The natural angular frequency differs depending on the articulator, and the higher the speed of the articulator, the greater the value.

As described above, the beat synchronization time is a plurality of times at equal time intervals T given based on Japanese isochronism. By adjusting the time interval T, that is, the beat interval T, the speech speed of the synthesized voice can be changed. In the description of the related art, the beat synchronization time is a portion extending from a consonant to a vowel, and is a position indicated by a symbol に お い て in FIG. However, Watanabe et al. Of Utsunomiya University, "Examination of time structure of speech considering speech generation process" (Abstracts of the 6th Autumn Meeting of the Acoustical Society of Japan, 2-5-
19, pp. 313-314), the reference to the start time of the articulation phenomena, that is, the transition start time from the preceding articulation movement to the following articulation movement, is based on the conventional section from the consonant to the vowel. Rather than saying, it can better explain the isochronism of the beat. Therefore, based on the observation result of the voice uttered by the human, it is appropriate to adjust the transition start time ti from the corresponding phoneme of each articulatory organ to the subsequent phoneme to the beat synchronization time in the articulatory model of the present embodiment.

Therefore, the above ti is obtained as follows. First, when the articulatory model is shifted to the i-th phoneme phi by moving each articulator from the preceding (i-1) th phoneme phi-1 based on the above equation, all articulators (J, L, FT) are used. , BT) fall within the respective allowable ranges (articulation allowable ranges) of the phoneme phi, and further, a time tready advanced (added) by the minimum duration Dmin (phi) of the phoneme phi. The phoneme p
If hi is a consonant, this time tready is
The transfer start time ti of the model to hi + 1 is defined as the phoneme p
When hi is a vowel, the time tready is compared with the beat synchronization time, and the larger one is defined as ti. That is, although the beat synchronization time is usually set to ti, if the time ready exceeds the beat synchronization time, that is, if the articulator does not reach the articulation allowable range of the phoneme (vowel) by the beat synchronization time, the articulation is adjusted. By delaying the start of the transition of the model to the subsequent phoneme until the organ reaches the articulation tolerance of the phoneme (vowel),
Care is taken to prevent unnatural articulation.

FIG. 7 shows an example of a temporal change of the articulatory organs J, L, FT, and BT (the state of the articulatory model in which the movement of the articulators J, L, FT, and BT) is controlled based on the above rules. in this way,
The movement of the articulator is represented as a continuous quantity with respect to the time axis.

On the other hand, the beat interval T is obtained as follows. In the conventional example, based on an empirical rule that "humans read the more the number of mora, the faster the number of mora", there is prepared in advance a table describing the correspondence between the number of mora in the accent phrase obtained by observing the actual voice and the beat interval T. In addition, the beat interval (interval of the beat synchronization time as the reference time) is obtained by referring to the table from the number of mora of the accent phrase to be read out. However, in actual human utterances, not only the number of mora included in utterance segments such as accent phrases,
The fact that various factors such as the type of part-of-speech affect the duration of phonemes is described in, for example, "Setting the duration of phonemes in sentence speech" by Miki et al. It has been reported in academic conference reports SP90-2, p9-16).

Based on such a report, in the present embodiment, in the beat interval estimation processing unit 108b, the part-of-speech information of the content word (independent word) for each accent phrase generated by the language analysis processing unit 104 together with the number of mora in the utterance segment. , And the beat interval T is adjusted to control the phoneme duration time with higher accuracy.

The beat interval T is approximated as being constant within the accent phrase, and is estimated by the following equation: beat interval T = a0 x0 + a1 x1 + ... + aN xN + b0 y0 + b1 y1 + ... + bM yM (3)

Here, when the number of mora in the accent phrase is n (where n ≦ N + 1), xn−1 = 1 and xi = 0 (i ≠
n-1), and when exceeding N + 1, xN = 1,
xi = 0 (i ≠ N). There are M + 1 types of parts of speech of the accent phrase content word, and the m-th (m = 0, 1, 2,
.., M), ym = 1, yi = 0 (i ≠
m). This is an estimation method based on quantification class 1 well known in the field of statistics, where x and y are called dummy variables, and a and b are called categorical quantities. The set of category quantities a0 to aN and b0 to bM in the above equation (3) is stored in the beat quantity estimation category quantity memory 108a, 1.
08a '. This memory 108a, 10
8a 'is a specific area secured in, for example, a main memory (not shown). Note that the category quantities a0 to aN, b
How to obtain 0 to bM will be described later.

The beat interval estimation processing unit 108b performs the processing in (3) above.
According to the formula, the beat interval T is obtained for each accent phrase from the number of mora in the accent phrase and the part of speech information of the accent phrase content word, and the beat synchronization time is set and output based on the obtained beat interval T.

As described above, the articulation model time change determination processing unit 107b in the phoneme duration calculation processing unit 107 determines the following phoneme determined based on the beat synchronization time output from the beat interval estimation processing unit 108b. Then, a time-series pattern of each articulator is generated.

Subsequently, the phoneme boundary determination processing unit 107c
The phoneme duration is determined based on the time series pattern of each articulator generated by the articulatory model time change determination processing unit 107b.

When the articulatory model transitions from the phoneme to the following phoneme, in the initial state, all articulatory organs are within the allowable articulatory range of the phoneme, but when the state of the articulatory model changes, any of the articulatory organs changes. One exits its tolerance at time tout. Then, as the state transition of the model progresses, at a certain time tin, all articulatory organs enter the articulation allowable range of the succeeding phoneme. This means that at t <tout, all articulators are within the allowable articulation range of the phoneme, and at t ≧ tin, all articulators are within the allowable articulation range of the subsequent phoneme.

Here, when the phoneme is a consonant, that is, when the phoneme is a consonant and the subsequent phoneme is a vowel, tout is defined as the boundary between the phoneme and the subsequent phoneme (phonemic boundary between the consonant and the vowel).
When the phoneme is a vowel and the subsequent phoneme is a consonant, t
Let in be the boundary between the phoneme and the subsequent phoneme (phonemic boundary between vowel and consonant). When both the phoneme and the subsequent phoneme are vowels, the time (tot + tin) / 2 is set as a boundary between the phoneme and the subsequent phoneme (a phoneme boundary between a vowel and a vowel). In other words, the boundary between a consonant and a vowel is the time when one of the articulators first escapes from the articulation allowable range of the consonant (the relevant phoneme). ) Is the time when it enters the articulation allowable range. The boundary between vowels and vowels is defined as a section between the time when any articulator first exits the articulation allowable range of the phoneme and the time when all articulators enter the allowable range of the succeeding phoneme. It is assumed to be an intermediate time point.

With the above procedure, all phoneme boundaries are determined, and the time length of all phonemes included in a given phoneme sequence, that is, the phoneme duration is determined from the time difference between adjacent boundaries.

Incidentally, in order to control the articulatory model as described above, the unique state Ainh (k, ph) of each articulatory organ k assigned to each phoneme ph and its allowable range Ama
x (k, ph) and Amin (k, ph), the minimum duration Dmin (ph), the natural angular frequency a (= αk) determined for each articulator k in the above equation (2), and the above (3) It is necessary to appropriately set the category quantities a0 to aN and b0 to bM in the equation. Therefore, in the present embodiment, these values are set in advance by optimizing (learning) using a large amount of volume data actually generated by a human.

A method of optimizing each parameter value of the individual phonemic articulation model using a large amount of voice data will be described with reference to FIG. In FIG. 8, a voice database 130 is a file obtained by digitizing a voice uttered by a human, and includes a phonological label (as phonological information) indicating the content of the voice, phonological boundary information, and linguistic analysis of the uttered content. The part-of-speech information of each accent phrase content word obtained as a result is stored together.

Real speech phoneme duration calculation processing section 131
Extracts the information of the phoneme label and the phoneme boundary position (time) from the speech database 130, and calculates the difference between adjacent phoneme boundary positions (time) to calculate the duration of each phoneme in real speech.

The phoneme duration estimation processing section 132 performs processing by the same method as that applied in the above-described phoneme duration calculation processing section 107 in FIG. As an input, the duration of the phoneme is estimated. The difference from the phoneme duration calculation processing unit 107 is that an optimal beat interval is automatically set internally and output.

The time length comparing unit 133 calculates the phoneme duration of the real speech obtained by the real speech phoneme duration calculation processing unit 131 and the phoneme duration estimated by the phoneme duration estimation processing unit 132. Are compared to calculate the estimation error of the duration. In the present embodiment, an average square error obtained by dividing the sum of square errors of all phonemes included in the voice database 130 by the number of all phonemes is used as the estimation error.

The parameter changing unit 134 controls the phoneme-specific articulation model parameter memory 135 so that the estimation error of the duration time obtained by the time length comparing unit 133 is reduced.
The value of the articulatory model parameter for each phoneme is changed.

By repeating such feedback control, an articulatory model parameter set for each phoneme that minimizes the estimation error of the duration can be obtained in the phoneme-based articulatory model parameter memory 135.

On the other hand, the category quantity calculation processing unit 136
The category quantities a0 to aN and b0 to bM are calculated from the beat interval at the time of obtaining the articulatory model parameters for each phoneme that minimizes the estimation error and the content word part of speech information of the accent phrase in the corresponding speech database 130. . In calculating the category quantity, a general method described in Tomomi Hayashi and Tsutomu Komazawa, “Quantification Theory and Data Processing” (Asakura Shoten) can be used. The calculated category quantity is stored in the beat interval estimation category quantity memory 137.

As described above, the voice database 13
From 0, the parameter values for the articulatory model control are acquired in the phoneme-based articulatory model parameter memory 135, and the category quantity is acquired in the beat interval estimation category quantity memory 137,
When used for speech synthesis in the speech synthesis unit 102 as described below, the synthesized speech is stored in the speech database 130.
It will be very close to the tone of the speaker recorded in.

In the present embodiment, two sets of parameters for articulation model control and two sets of beats are obtained from two types of speech data files (speech database 130) created from speeches of different speakers by the above-described method. The category quantity for interval estimation is obtained. That is, a first voice data file created by the voice of the first speaker as a voice data file (including information on a phoneme label and a phoneme boundary, and part-of-speech information on an accent phrase content word) recorded in the voice database 130 And two types of second voice data files created by the voice of the second speaker are prepared, the voice data file is switched, and the above-described method is applied. The obtained articulation model parameter set is obtained in the phoneme-based articulation model parameter memory 135, and the category quantity set is obtained in the beat interval estimation category quantity memory 137.

One of the articulatory model parameter sets corresponding to the first and second speakers obtained in this way is stored in the phonemic-specific articulatory model parameter memory 10 in FIG.
7a, the other is also stored and used in another phoneme-based articulation model parameter memory 107a 'in FIG. Similarly, one of the category quantity sets respectively corresponding to the first and second speakers is stored in the beat interval estimation category quantity memory 108a in FIG. 1, and the other is placed in the other beat interval estimation category quantity in FIG. It is stored and used in the memory 108a '. In the present embodiment, this memory 107a, 1
07 'and the memories 108a and 108
By switching and using one of a 'based on the internal state of the system determined by the user's designation or the like, the tone of the synthesized voice can be switched.

Next, the operation of the phoneme duration calculation unit 107 will be described in detail with reference to the flowcharts of FIGS. First, the phoneme duration calculation unit 107
Is the above-mentioned phoneme-specific articulation model parameter memory 107
a and 107a ', an articulatory model time change determination processor 107b for performing an articulatory model time change determination process, and a phoneme boundary determination processor 107c for performing a phoneme boundary determination process based on the processing result of the processor 107b. Consists of

In the present embodiment, two types of articulatory model parameter files (not shown) corresponding to different speakers obtained by the above method, that is, each articulator J, L, FT, Two types of phonetic-based articulatory model parameter files in which the parameters of the BT articulatory model are stored are prepared. At 107a, the contents of the other file are read into the phoneme-based articulation model parameter memory 107a '. The memories 107a and 107a 'are, for example, specific areas secured in the main memory.

The language analysis processing unit 10 in the language processing unit 101
4 generates the reading information and activates the phoneme duration calculation processing unit 107 in the speech synthesis unit 102, and the articulation model time change determination processing unit 107 in the processing unit 107
b is a variable i indicating a phoneme position in a phoneme sequence to be synthesized (the number of phonemes is N) included in the reading information, a variable i indicating a first phoneme, a time t being set to 0, and a beat synchronization time being set to 0. Variable tsy
nc is (for example, a value determined by the speech rate specified by the user) T
Is a variable tin indicating the time at which all articulators J, L, FT, and BT enter the respective articulatory tolerances of the i-th phoneme.
(i) Initialize (= tin (1)) to 0 (step S)
1).

Next, the articulatory model time change determination processor 107
b indicates that the time t is the minimum duration of the i-th phoneme (Dmin
(phi)) is updated to a value advanced (step S2).
The minimum duration (Dmin (phi)) can be obtained by referring to the phoneme-based articulation model parameter memory 107a or 107a 'using the i-th phoneme.

Next, the articulatory model time change determination processor 107
b checks whether the i-th phoneme is a consonant (step S3), and if it is a vowel, compares the time t with the beat synchronization time tsync (step S4).

If the time t does not exceed the beat synchronization time tsync, the time t is updated to the beat synchronization time tsync (step S5), and the beat synchronization time tsync is advanced by T (step S6). On the other hand, the time t is the beat synchronization time t
If it exceeds sync, the process proceeds to step S6 without updating the time t, and the beat synchronization time tsync is advanced by T.
After step S6, the articulation model time change determination processing unit 107b sets the value of the current time t to the shift start time t
i (that is, the start time of shifting the model from the i-th phoneme to the subsequent phoneme) is determined (step S7).

On the other hand, if the i-th phoneme is a consonant,
Proceeding directly to step S7, the value of the current time t is determined as the shift start time ti. When the articulatory model time change determination processing unit 107b executes step S7, the time t
(Movement) of each articulator J, L, FT, BT
(= M (J, t)), ML (= M (L,
t)), MFT (= M (FT, t)), MBT (= M (B
T, t)) is calculated by the above equation (1) (step S8).

Next, the articulatory model time change determination processor 107
b is the articulatory range of each i-th phoneme at the position (MJ, ML, MFT, MBT) of the articulators J, L, FT, and BT at time t, that is, Amin (J, phi) to Amax (J,
phi), Amin (L, phi) to Amax (L, phi), Amin
(FT, phi) to Amax (FT, phi), Amin (BT, p
It is checked whether or not all the values are included in hi) to Amax (BT, phi) (step S9).

If the articulators J, L, F at time t,
If the positions of T and BT (MJ, ML, MFT, MBT) are all within the allowable range of articulation of the i-th phoneme, the articulatory model time change determination processing unit 107b sets the time t
Is advanced by a predetermined minute time δ (for example, 5 ms) (step 10), and the process returns to step S8, where the new time t
Position of each articulator J, L, FT, BT at MJ, ML,
MFT and MBT are calculated, and the determination in step S9 is performed again.

The articulation model time change determination processing unit 107b
Repeats the above operation until it detects that at least one of the positions of the articulators J, L, FT, and BT deviates from the corresponding articulation allowable range of the i-th phoneme.

As described above, if any of the positions of the articulators J, L, FT, and BT at the time t deviate from the corresponding articulation allowable range of the i-th phoneme, the articulatory model time change determination processing unit 107b Determines that the time t is the time tout (i) at which at least one of the positions of the articulators J, L, FT, and BT goes out of the articulation allowable range of the i-th phoneme, and stores the time tout in a memory (not shown). (Step S11).

Next, time change determination processing section 107b performs the same processing as step S8 at time t (step S8).
12). However, immediately after step S11 is performed as in this example, the values of MJ, ML, MFT, and MBT representing the positions of the articulators J, L, FT, and BT are set to the values in step S11.
, ML, MFT, in step S8 performed immediately before
Since it matches the calculation result of MBT, step S1
Step S12 immediately after step 1 may be skipped.

Next, the time change determination processing unit 107b determines that the position of the articulators J, L, FT, and BT at the time t is the next i.
The articulation permissible range of the + 1st phoneme, ie, Amin
(J, phi + 1) to Amax (J, phi + 1), Amin (L, phi
+1) to Amax (L, phi + 1), Amin (FT, phi + 1) to Am
ax (FT, phi + 1), Amin (BT, phi + 1) to Amax (B
T, phi + 1) are checked (step S13).

If the articulators J, L, F at time t,
If any one of the positions of T and BT is out of the corresponding articulation allowable range of the (i + 1) th phoneme, the articulation model time change determination processing unit 107b advances the time t by a predetermined minute time δ (Step S14), Step S1
2 and each articulator J, L, at the new time t.
MJ, ML, MFT, and MBT representing the positions of FT and BT are calculated, and the determination in step S13 is performed again.

The articulation model time change determination processing unit 107b
Performs the above operation with all articulatory organs J, L, FT, BT
Is detected until it is detected that the position is within the allowable articulation range of the (i + 1) th phoneme.

In this way, if all of the positions of the articulators J, L, FT, and BT at the time t are within the allowable articulation range of the (i + 1) th phoneme, the articulatory model time change determination processing unit 107b performs , The time t in (i.e., the time at which the positions of all the articulators J, L, FT, and BT fall within the allowable articulation range of the (i + 1) -th phoneme (the next phoneme).
+1), and stores it in a memory (not shown) (step S15).

Next, the articulatory model time change determination processing unit 107
b indicates whether or not the processing has proceeded to the (N-1) th phoneme (the last to second phoneme in the phoneme sequence including N phonemes).
It is checked whether or not the current value of i is N-1 (step S16).

If the current value of i is not N-1, the articulation model time change determination processing unit 107b increments the value of i (+1) (step S17), that is, sets the value of i in the phoneme sequence. After updating to point to the next phoneme, the process returns to step S2.

As described above, the articulatory model time change determination processing unit 107b performs the processing from step S2 onward for i = 1 to i
= N−1, and the sequence of tin (i) (i = 1, 2, 2,
3,..., N), that is, tin (1), tin (2), tin (3),
.., Tin (N) and a column of tout (i) (i = 1, 2, 3, 3)
.., N-1), that is, tout (1), tout (2), tout (3),
.., Tout (N-1).

Then, the articulatory model time change determination processor 1
Control is passed from 07b to the phoneme boundary determination processing unit 107c in the same phoneme duration calculation processing unit 107. First, the phoneme boundary determination processing unit 107c sets the variable i indicating the phoneme position in the phoneme sequence to be synthesized to 1 indicating the head phoneme, the variable Bi indicating the phoneme boundary with the preceding phoneme of the i-th phoneme, that is, B1.
Is initialized to tin (i), that is, tin (1) (step S21).

Next, the phoneme boundary determination processing unit 107c checks whether the i-th phoneme is a consonant or a vowel (step S22). If it is a vowel, the next (i + 1) -th phoneme is a consonant. Is checked (Step S
23).

If the ith phoneme is a vowel and the next i + 1
If the phoneme is a consonant, the phoneme boundary determination processing unit 1
07c sets tin (i + 1) to a variable Bi + 1 indicating a phoneme boundary with the preceding phoneme of the (i + 1) th phoneme (step S2).
4) If the i-th phoneme is a vowel and the next (i + 1) -th phoneme is also a vowel, the phoneme boundary determination processing unit 107c
intermediate time between out (i) and tin (i + 1) (tout (i) + tin (i + 1)
) / 2 is set to Bi + 1 (step S25).

On the other hand, if the ith phoneme is a consonant (in this case, since there is no consonant-consonant combination, the next (i + 1) th phoneme is a vowel), the phoneme boundary determination processing unit 107c Tout (i) is set to Bi + 1 (step S26).

When determining the value of Bi + 1 in step S24, S25 or S26, the phoneme boundary determination processing unit 107c determines the difference between Bi + 1 and Bi, that is, the preceding phoneme (i-th phoneme) of the (i + 1) -th phoneme. ) And i + 1
The time difference between the preceding phoneme of the i-th phoneme (i-1st phoneme) and the phoneme boundary Bi is determined, and the duration time Di of the i-th phoneme is determined (step S27). In the first step S27, the duration D1 of the first phoneme is B2-B
It is obtained by the operation of 1.

Next, the phoneme boundary determination processing section 107c sets N-
It is checked whether or not the processing has proceeded to the first phoneme, based on whether or not the current value of i is N-1 (step S).
28).

If the current value of i is not N-1, the phoneme boundary determination processing section 107c increments the value of i (+1) (step S29), and then returns to step S22.

In this way, the phoneme boundary determination processing section 107
c indicates that the processing after step S22 is i = 1 to i = N-1.
To the column of Di (i = 1, 2, 3,..., N−
1) That is, D1, D2, D3,..., DN-1 are obtained.

Next, the phoneme boundary determination processing unit 107c calculates the duration DN of the Nth phoneme, that is, the last phoneme (= vowel) in the phoneme sequence by the following calculation: DN = tin (i + 1) -Bi + 1 + DFO (3) is obtained (step S30). Here, DFO is the vowel fade-out time.

As a result, the phoneme boundary determination processing section 107c
(The phoneme duration calculation processing unit 107 provided with) has durations D1, D2, D of N phonemes included in the phoneme sequence.
3,..., DN.

Now, as described above, the speech synthesis unit 102
When the duration of each mora (consonant part and vowel part) included in the reading of the input sentence (input text) is determined by the phoneme duration calculation processing unit 107 within the The pattern generation processing unit 109 is activated.

The pitch pattern generation processing unit 109 first determines the point pitch position based on (the sequence of) the durations determined by the phoneme duration calculation processing unit 107 and the accent information determined by the language analysis processing unit 104. Set. Next, a plurality of set point pitches are interpolated by a straight line to obtain a pitch pattern every 10 ms, for example.

On the other hand, the phoneme parameter generation processing unit 110 in the speech synthesis unit 102 performs a process of generating phoneme parameters based on the phoneme information of the speech symbol string, for example, a pitch pattern generation process by the pitch pattern generation processing unit 109. Is performed in parallel as follows.

First, in the present embodiment, the 0th to 25th order cepstrum coefficients obtained by analyzing real speech sampled at a sampling frequency of 11025 Hz by the improved cepstrum method with a window length of 20 ms and a frame period of 10 ms are used as consonants + vowels (C
A speech unit file (not shown) storing a total of 137 speech units obtained by cutting out all syllables necessary for the synthesis of Japanese speech in units of V) is prepared. The contents of the speech unit file are read into, for example, a speech unit area (hereinafter referred to as a speech unit memory) 111 secured in the main memory at the start of the sentence-to-speech conversion process according to the sentence-to-speech conversion software. And

The phoneme parameter generation processing unit 110 performs the above-described processing according to the phoneme information (here, the first phoneme information, but may be the second phoneme information) in the speech symbol string passed from the language analysis processing unit 104. The speech units in CV units are sequentially read from the speech unit memory 111, and the read speech units are connected to generate phoneme parameters (feature parameters) of the speech to be synthesized.

A pitch pattern is generated by the pitch pattern generation processing unit 109, and the phoneme parameter generation processing unit 1
When the phoneme parameters are generated by the, the synthesis filter processing unit 112 in the speech synthesis unit 102 is activated. As shown in FIG. 2, the synthesis filter processing unit 112 includes a white noise generation unit 118, an impulse generation unit 119,
Driving sound source switching unit 120 and LMA filter 121
And synthesizes a speech from the pitch pattern and the phoneme parameters generated as described below.

First, the voiced portion (V) of the voice is switched to the impulse generating portion 119 by the driving sound source switching portion 120. The impulse generation unit 119 generates impulses at intervals according to the pitch pattern generated by the pitch pattern generation processing unit 109, and drives the LMA filter 121 using the impulse as a sound source. On the other hand, in the unvoiced part (U) of the sound, the driving sound source switching unit 120 switches to the white noise generating unit 118 side. The white noise generator 118 generates white noise,
The LMA filter 121 uses this white noise as a sound source.
Drive.

The LMA filter 121 directly uses the cepstrum of the voice as a filter coefficient. In the present embodiment, since the phoneme parameter generated by the phoneme parameter generation processing unit 110 is a cepstrum as described above, this phoneme parameter is
, And is driven by a sound source switched by the driving sound source switching unit 120 to output a synthesized voice.

The sound synthesized by the synthesis filter processing unit 112 (the LMA filter 121 therein) is a discrete sound signal, which is converted into an analog signal by the D / A converter 113 and output to the speaker 115 through the amplifier 114. , Can be heard as a sound for the first time.

Although the embodiment of the present invention has been described above, the present invention is not limited to the embodiment. For example, in the above embodiment, the cepstrum is used as the feature parameter of the voice, but the LPC or PA
The present invention is applicable to other parameters such as RCOR and formant, and similar effects can be obtained. Regarding the language processing unit, even if syntax analysis etc. other than morphological analysis is inserted, there is no problem. Regarding pitch generation, it is not necessary to use a method based on point pitch. For example, the present invention is applicable even when using a Fujisaki model. is there.

In the above embodiment, the case where two types of tones can be synthesized by switching the articulatory model parameters has been described. However, three or more types of parameters are prepared by creating parameters from various human voices. Then, they may be switched and used. In short, the present invention can be variously modified and implemented without departing from the gist thereof.

[0117]

As described in detail above, according to the present invention, the articulatory model is controlled by using the linguistic analysis result for the text data to be subjected to speech synthesis, and is included in the reading of the text data based on the result. Since the duration of each phoneme is determined, the physical constraints of the articulatory organs when a human utters the voice are reflected in the duration of the phoneme, and the subtle utterance speed in the text is synthesized speech. And can synthesize natural and human-friendly voice that is easy to hear. In particular, by determining the synchronization timing of each mora included in the text data reading based on the linguistic analysis result, by controlling the articulatory model based on the determined phonological information and the synchronization timing included in the linguistic analysis result, It is possible to accurately reflect the physical constraints of articulatory organs when a human utters a voice on the phoneme duration, and to give a subtle utterance speed in a sentence to a synthesized voice with high accuracy.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a schematic configuration of a speech rule synthesis device according to an embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of a synthesis filter processing unit 112 in FIG.

FIG. 3 is an exemplary view showing four articulatory organs constituting an articulatory model applied in the embodiment;

FIG. 4 is a diagram showing, by way of example, a case of a sound-repellent “n” having several different articulation modes depending on a subsequent phoneme (that is, depending on the phoneme environment) regarding phoneme segmentation.

FIG. 5 is a diagram showing an example of a phoneme sequence included in a phonetic symbol string generated by performing language processing on a sentence “Go to a park and read a book” before and after considering a phoneme environment.

FIG. 6 is a diagram showing an example of parameters of an articulatory model for phoneme [i].

FIG. 7 is a diagram showing an example of a temporal change in the state of an articulatory model that models the movement of four articulatory organs.

FIG. 8 is a diagram for explaining a method of optimizing each parameter value of the articulatory model of each phoneme and the number of beat interval estimation categories using a large amount of voice data.

FIG. 9 is a diagram showing a part of a flowchart for explaining an articulatory model time change determination process by an articulatory model time change determination processor 107b in the phoneme duration calculation processor 107;

FIG. 10 is a diagram showing the remainder of the flowchart for explaining the articulatory model time change determination processing by the articulatory model time change determination processor 107b in the phoneme duration calculation processor 107.

FIG. 11 is a flowchart illustrating a process of determining a phoneme boundary and a phoneme duration by a phoneme boundary determination processing unit in a phoneme duration calculation processing unit.

FIG. 12 is a block diagram showing a configuration of a conventional rule synthesis device.

FIG. 13 is a view for explaining a conventional phoneme duration determining method in the rule synthesizing apparatus of FIG. 12;

[Explanation of symbols]

101: Language processing unit 102: Speech synthesis unit 104: Language analysis processing unit 107: Phoneme duration calculation unit (phoneme duration determination means) 107a, 107a ', 135 ... Phoneme train articulation model parameter memory 107b ... Articulation model Time change determination processing unit 107c: Phoneme boundary determination processing unit 108a, 108a ', 137: Category quantity memory for beat interval estimation 108b ... Beat interval estimation processing unit 109: Pitch pattern generation processing unit 110 ... Phoneme parameter generation processing unit 112: Synthesis Filter processing unit 130 Voice database 131 Real phoneme duration calculation processing unit 132 Phoneme duration estimation processing unit 133 Time length comparison unit 134 Parameter changing unit 136 Category number calculation processing unit

Claims (7)

[Claims] 1. After analyzing text data to be subjected to speech synthesis and obtaining a language analysis result, a time change of an articulatory model that models a movement of an articulator is obtained based on the language analysis result. Determine the duration of each phoneme included in the reading of the text data based on the time change of the obtained articulation model, and select a speech unit based on the language analysis result. A speech synthesis method comprising: synthesizing speech by connecting the selected speech units based on a duration time. 2. After synthesizing text data to be subjected to speech synthesis to obtain a linguistic analysis result including phonological information, based on the linguistic analysis result, a synchronization timing of each mora included in the reading of the text data. Based on the phonological information included in the language analysis result and the determined synchronization timing, a time change of an articulator model that models the movement of the articulatory organ is obtained.Based on the obtained time change of the articulator model, In addition to determining the duration of each phoneme included in the phoneme information,
A speech synthesis method comprising: selecting a speech unit based on the result of the language analysis; and connecting the selected speech unit based on the determined duration of the phoneme to synthesize a speech. 3. The method according to claim 1, wherein, when determining a time change of the articulatory model, an articulatory organ operation command time to a subsequent phoneme of each phoneme included in the phoneme information is determined based on the synchronization timing. Item 3. The speech synthesis method according to Item 2. 4. When deciding an articulatory organ operation command time for a succeeding phoneme of the phoneme, it is determined whether the phoneme is a consonant or a vowel, and based on a result of the determination and a time given by the synchronization timing. 4. The speech synthesis method according to claim 3, wherein the articulator organ operation command time for the succeeding phoneme of the phoneme is determined by using the same. 5. When the phoneme is a consonant, an articulator operation command time for a succeeding phoneme of the consonant is determined based on the time given by the synchronization timing. By comparing the time at which the vowel articulation arrives with the time given by the synchronization timing,
5. The articulatory organ operation command time for a succeeding phoneme of the vowel is determined based on a time given by the synchronization timing or a time at which articulation of the vowel is reached based on the comparison result. Speech synthesis method. 6. A language analyzing means for analyzing text data to be subjected to speech synthesis to obtain a language analysis result, and obtaining a temporal change of an articulatory model which models a movement of an articulator based on the language analysis result. A phoneme duration determining means for determining the duration of each phoneme included in the reading of the text data based on the time change of the articulation model; and selecting a speech unit based on the language analysis result. And a speech generation processing unit for generating a speech by connecting the selected speech unit based on the duration of the phoneme determined by the phoneme duration determination unit. Synthesizer. 7. A language analysis means for analyzing text data to be subjected to speech synthesis to obtain a language analysis result including phoneme information, and based on the language analysis result, each of the mora included in the reading of the text data. Synchronization timing determining means for determining a synchronization timing, based on the phonetic information included in the language analysis result and the synchronization timing determined by the synchronization timing determining means, the time change of the articulatory model that models the movement of the articulatory organ. Phoneme duration determining means for determining the duration of each phoneme included in the reading of the text data based on the time change of the articulatory model, and selecting a speech unit based on the language analysis result. Then, by connecting the selected speech units based on the duration of the phoneme determined by the phoneme duration determination means, A speech synthesis device comprising: speech generation processing means for generating speech.