JP2001100777A - Method and device for voice synthesis - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method and a device for voice synthesis which synthesizes a voice of natural intonation which is easy to hear and does not fatigue a listener when being heard for a long time. SOLUTION: A pitch pattern generation means generates a first pitch pattern or its feature quantity from analysis information of an input text in accordance with a prescribed rule and generates a second pitch pattern, which is obtained by correction based on a pitch pattern generation process model, from the first pitch pattern or its feature quantity and outputs this second pitch pattern.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizing method and apparatus for generating a pitch pattern of speech to be synthesized from an analysis result of an input text.

[0002]

2. Description of the Related Art A typical typical conventional speech synthesizer is a rule synthesizer capable of subdividing and accumulating speech, and synthesizing an arbitrary speech by a combination thereof. This rule synthesizing device converts input text data into a symbol string composed of phoneme information and prosody information, and performs a sentence-to-speech conversion process for generating a speech from the symbol string. The sentence-to-speech conversion processing mechanism in the rule synthesizing apparatus is roughly divided into a language processing unit and a speech synthesis unit.

The linguistic processing unit performs linguistic processing such as morphological analysis / syntax analysis on text data composed of kanji and kana mixed sentences input from a text file, performs decomposition into morphemes, and estimates dependency relations. At the same time, give each morpheme a reading and accent type.

[0004] Thereafter, the language processing unit determines the accent type of each phrase (hereinafter referred to as an accent phrase) serving as a delimiter of the accent at the time of reading aloud using the accent movement rules such as compound words with respect to the accent.

[0005] The morpheme, dependency, accent type, and reading information thus obtained by the language processing unit are hereinafter collectively referred to as linguistic information. Next, the speech synthesis unit determines the duration of each phoneme included in the obtained “reading”.

In general, the phoneme duration is determined based on the isochronism of beats peculiar to Japanese. For example, the duration of the consonant is fixed depending on the type of the consonant, and the duration of the vowel is determined so that the interval between the consonant and the vowel, which is the reference time of each mora, is constant.

Subsequently, in accordance with the "reading" obtained as described above, the phoneme parameter generation processing unit in the speech synthesizing unit reads necessary speech units from the speech unit memory, and reads the read speech units. The connection is made while expanding and contracting in the time axis direction in accordance with the phoneme duration determined by the method, to generate a feature parameter sequence of the speech to be synthesized.

[0008] A large number of speech units prepared in advance are stored in the speech unit memory in the speech synthesis unit. The speech unit analyzes a voice uttered by an announcer or the like, obtains a predetermined voice characteristic parameter expressing the envelope characteristic of the spectrum, and then obtains a predetermined synthesis unit, in this example, a Japanese syllable unit (consonant + Vowels), and is created by cutting out all syllables included in Japanese speech from the above-mentioned feature parameters.

As a feature parameter generated by the phoneme parameter generation processing unit, a low-order coefficient of a cepstrum is used. The low-order cepstrum coefficient can be obtained as follows.

First, voice data uttered by an announcer or the like is cut out by a window function (for example, a Hanning window) 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. I do. Next, the power of the obtained short-time spectrum is logarithmically obtained to obtain a logarithmic power spectrum, and the logarithmic power spectrum is subjected to inverse Fourier transform. The cepstrum coefficient is calculated in this way.

It is well known that cepstrum characteristics include that higher-order coefficients hold voice fundamental frequency information and lower-order coefficients hold voice spectrum envelope information. On the other hand, a unit pattern obtained by extracting a pitch pattern extracted from a human uttered voice using an autocorrelation method or a cepstrum method in units of accent phrases is stored in a unit pattern storage unit in the voice synthesis unit.

Each unit pattern is associated with linguistic information based on a statistical method such as quantification type 1, and unit pattern selection / selection in the speech synthesizer is performed.
The connection unit determines a unit pattern to be used from the language information output from the language processing unit according to the association rule.

Further, the unit pattern selection / connection unit smoothly connects the unit patterns determined for each accent phrase constituting the sentence while performing interpolation processing, and finally generates a sentence pitch pattern.

In a synthesis filter processing section provided in the voice synthesis section, a periodic pulse based on a pitch pattern is used as a sound source in a voiced section and white noise is used as a sound source in an unvoiced section, and filtering is performed as a filter coefficient calculated from a feature parameter sequence of the voice. Then, a desired voice is synthesized. As a synthesis filter of the synthesis filter processing unit, a logarithmic amplitude approximation filter that directly uses cepstrum coefficients as filter coefficients is used.

[0015]

The conventional speech synthesizer represented by the rule synthesizer as described above has the following problems with the generated pitch pattern. For the control of pitch pattern, see the Journal of the Acoustical Society of Japan, 1971.
vol. 27, pages 445 to 453, a pitch pattern generation process model that describes its operation characteristics by superimposing a global descent component and a local undulation component is widely used. It has been shown to approximate the pitch pattern of human uttered speech very well.

However, since it is difficult to automatically separate the two components, a global descent component and a local undulation component, a generation rule is automatically constructed by a statistical method from a large amount of voice data based on a voice corpus to control pitch. Is performed without considering the superimposed generation process model as described above.

However, since the modeling that is not conscious of the superimposed generation process model focuses on the pattern shape and the rate of change of the pitch, compared to the pitch pattern generation process model, the synthesized voice of the human voice is compared with the synthesized voice. There is a problem that the restriction of the human vocalization mechanism and the linguistic restriction appearing in the pitch pattern are not reflected, resulting in unnatural inflection of voice.

In view of the above problems, the present invention provides a voice synthesizing method and apparatus for performing pitch control based on a voice corpus, by performing pitch control capable of taking into account restrictions on a vocal mechanism and linguistic restrictions when generating a pitch pattern. The object of the present invention is to enable generation of a natural human-like pitch pattern.

[0019]

SUMMARY OF THE INVENTION In order to solve the above-mentioned problems and achieve the object, a voice synthesizing method and apparatus of the present invention are configured as follows. The speech synthesis method of the present invention comprises:
A first pitch pattern or a characteristic amount thereof is generated from analysis information of the input text according to a predetermined rule, and a second pitch pattern corrected based on the pitch pattern generation process model from the first pitch pattern or the characteristic amount Is generated, and a voice is synthesized based on the second pitch pattern.

Further, the speech synthesizing apparatus according to the present invention comprises a text analyzing means for analyzing input text to generate text analysis information, and a pitch pattern generating a pitch pattern based on the text analysis information output by the text analyzing means. Generating means for generating a first pitch pattern or its characteristic amount from the text analysis information in accordance with a predetermined rule, and correcting the first pitch pattern or its characteristic amount based on a pitch pattern generation process model. Pitch pattern generation means for generating a second pitch pattern and outputting the second pitch pattern, and voice synthesis means for synthesizing voice based on the second pitch pattern output from the pitch pattern generation means Have.

[0021]

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 synthesis device according to one embodiment of the present invention. This voice rule synthesizing apparatus is realized by executing dedicated software supplied by 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. Is done.

This speech rule synthesizing apparatus has a sentence-to-speech conversion function for generating speech from text data, and is composed of a text analysis unit 101 and a speech synthesis unit 102.
The text analysis unit 101 analyzes the input sentence, which is a mixture of kanji and kana characters, identifies words (morphological analysis), estimates the composition of the sentence based on the obtained part-of-speech information and the like (dependency analysis),
Output the result.

The speech synthesizer 102 includes the text analyzer 10
A voice is generated based on the text analysis result which is the output of 1. Text data (in this case, a Japanese document) to be subjected to sentence-to-speech conversion (speech) is stored as a text file 103.

In this apparatus, in accordance with the sentence-to-speech conversion software, a sentence mixed with kanji or kana is read from the text file 103, and is read by the text analysis unit 101 and the speech synthesis unit 10.
2, a sentence-to-speech conversion process described below is performed to synthesize speech.

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

The morphological analysis unit 104 checks the input sentence against the Japanese analysis dictionary 105 to obtain all morphological sequence candidates, and outputs a grammatically connectable combination from among them. In the Japanese analysis dictionary 105, readings of individual morphemes and accent types are registered together with information used at the time of morphological analysis. Therefore, if a morpheme is determined by morphological analysis, a reading and an accent type are given at the same time.

Next, the grammatical attributes of the individual words included in the sentence determined by the morphological analysis unit 104 are stored in the dependency analysis unit 10.
Then, the sentence structure is analyzed to estimate the dependency relation of each word.

As described above, the text analysis unit 101
Then, the part-of-speech and the dependency relation are output to the speech synthesis unit 102 together with the information of the word reading and the accent. Voice synthesis unit 102
Then, the phoneme duration determination processing unit 107 is activated first.
In the present embodiment, as in the case of the prior art, a method of determining the phoneme duration based on the isochronism of the beat unique to Japanese is used, and therefore the description is omitted.

Next, the pitch pattern generation processing unit 108 is activated. The pitch pattern generation processing unit 108 includes a pitch or turn approximate shape estimation processing unit 109, a model parameter calculation processing unit 110, and a pitch control processing unit 111.

The pitch pattern rough shape estimation processing unit 109
Two point pitches representing a pitch outline (hereinafter, referred to as representative pitches) using a pitch outline estimation rule created in advance.
Is estimated. The rule for estimating the pitch outline is created as follows.

First, voices of a predetermined text read out by a human (for example, an announcer) are recorded. The recorded voice is taken into a computer and analyzed to extract a pitch (or fundamental frequency) pattern representing a voice pitch pattern.

Various methods have been proposed for extracting a pitch pattern, such as a method using autocorrelation and a method using cepstrum, and an appropriate method may be used. As shown in FIG. 2, two points of the maximum value and the pitch value of the accent phrase end mora, that is, the representative pitch described above, are extracted from the pitch pattern thus obtained, as shown in FIG.

Next, the correspondence between the result obtained by text analysis of the text prepared at the time of recording and the representative point pitch is regularized. If the recorded contents have a sufficient amount, a rule for estimating the representative point pitch can be automatically constructed by using a statistical method.

In this embodiment, the accent type, mora number, part of speech, and dependency relation of the accent phrase constituting the sentence which is the result of the text analysis are used as explanatory functions, and the individual representative point pitches are used as an external reference. Theory and Data Processing, "Quantification I" described by Tomio Hayashi et al., Asakura Shoten (1982)
By applying "class", a rule for estimating the representative point pitch is automatically constructed.

Using the representative point pitch estimation rule constructed as described above, from the accent information, part-of-speech information, and dependency information passed by the text analysis processing unit 101, the pitch pattern rough shape estimation processing unit 109 performs an accent phrase unit unit. Generate a representative pitch.

Next, the model parameter calculation processing unit 110
In, the model parameters of the generation process model are calculated from the representative point pitch. A model of the process of generating a pitch pattern is described in Journal of the Acoustical Society of Japan, 1971, vol. 27, 44
Although there is a model of the critical damping quadratic linear system described on page 5 to page 453, the model used in this embodiment uses the following phrase components so that model parameters can be estimated from two representative pitches per accent phrase. Simplify.

The phrase component is represented by using a gentle descending component represented by an exponential function and a vocal component at the beginning of the accent phrase. Let Gpi and Zpi denote the descending component and the vocal component of the i-th accent phrase, respectively, as functions of time t: G _pi (t) = 0 (outside the i-th accent phrase) e ^{- t} (i-th (1) Z _pi (t) = 0 (t <0, t> 1 / γ) γte ^{1− t} −1 (0 ≦ t ≦ 1 / γ) (2)

On the other hand, the accent component is approximated by a critical damping quadratic linear system step response function. When the step response of the i-th accent phrase and G _ai, as a function of time _{t, G ai (t) =} 0 (t <0) mim [1- (1 + βt) e - t, θ] (t ≧ 0) (3) is represented.

In the equations (1)-(3), α, β, and γ are the natural angular frequencies of the respective controls, and correspond to the fact that the accent component of the actual voice reaches the upper limit in a finite time, θ ≦ 1.

The F ₀ pattern superimposes these, Becomes

Here, A _pi and A _ai correspond to the size of the phrase command and the accent command, respectively, T _0i is the time of the phrase command, and T _1i and T _2i are the start time and end time of the accent command.

If F _min , α, β, and γ are fixed, A _pi and A _ai can be uniquely obtained by applying the point pitches (t _F0mx , F _0mx ) and (t _F0e , F _0e ) to the above equation. it can. A _pi If the accent type of head height type or relief ^{_{type, A pi = e (tF0e-}} T0i) ln (F 0e / F min) , and the when the accent type of the plate type, A _pi = ln (F _{0mx / F 0e) / (e} - (tF0mx-T0i) -e
^{- (tF0e-T0i)} ). A _ai becomes _{A ai = ln (F 0mx /} F min) -A pi e (tF0mx T0i) using A _pi.

FIG. 3 shows an example in which model parameters of an undulating accent are obtained. The model parameter calculation processing unit 110 further imposes physiological and physical constraints and linguistic constraints on the vocalization mechanism on the model parameters obtained as described above. This operation can reduce the unnaturalness of the generated pitch pattern caused by the estimation error of the representative point pitch. Specifically, the model parameters are corrected as follows. 1. A negative voice (A _pi <A _{pi 1} G _pi-1 (T _0i −T
_{If 0i 1} )) has occurred, the magnitude of the phrase command is corrected so that the loudness becomes zero. (To remove negative voices that are not actually possible) The size of the accent command is lower limit value Aaimin (≧
Correct the size of the phrase command so as not to fall below 0). (Because the accent nucleus always exists) If the sum of the size of the phrase command and accent command of the accent phrase preceded by the flat type is smaller than the sum of the phrase component and the accent component at the end of the preceding accent phrase, the voice is set to zero and the size of the accent command is set to Correct to match accent phrase. (In order not to create an accent nucleus at the last mora of the flat accent) Finally, based on the corrected model parameters (4)
The pitch control processing unit 111 generates a pitch pattern according to the formula.

On the other hand, the speech unit selection unit 112 in the speech synthesis unit 102 selects a speech unit based on the reading of each accent phrase. In this embodiment, the sampling frequency 110
Analysis of real speech sampled at 25 Hz using the improved cepstrum method with a window length of 20 msec and a frame period of 10 msec. 0th to 25th-order low-order cepstrum coefficients required for the synthesis of Japanese speech in consonant + vowel units A speech unit file (not shown) storing a total of 137 speech units obtained by cutting out all the syllables is prepared. It is assumed that the contents of the speech unit file have been read, for example, in the speech unit area 113 secured in the main memory at the start of the sentence-to-speech conversion process according to the sentence-to-speech conversion software.

In this way, the above-described speech units in the consonant + vowel units are sequentially read out from the speech unit area 113 and passed to the unit connection unit 114. The unit connection unit 114 generates phoneme parameters (feature parameters) of the speech to be synthesized by sequentially interpolating and connecting the speech units input from the unit selection unit 112.

As described above, when the pitch pattern is generated by the pitch pattern generation processing unit 108 and the phoneme parameters are generated by the speech unit connection unit 114, the synthesis filter unit 115 in the speech synthesis unit 102 is activated. You. This synthesis filter unit 115 uses white noise in a silent section and impulse in a voiced section as a driving sound source.
Voice is output by an LMA filter that directly uses cepstrum coefficients, which are phonemic parameters, as filter coefficients.

Although one embodiment of the present invention has been described above, the present invention is not limited to this embodiment. Instead of using cepstrum as a speech characteristic parameter, other parameters such as LPC, PARCO, and formant are used. You may. In this embodiment, the analysis-synthesis method using the characteristic parameters is adopted.
Or a formant synthesis type system. Regarding the control of the phoneme duration, for example, a statistical method may be used instead of the method using the isochronism as described above.

[0048]

According to the present invention, a natural human-like pitch pattern can be generated by realizing pitch control that can take into account the restrictions of the vocal mechanism and linguistic restrictions when generating the pitch pattern. , Easy to hear,
It is possible to provide a speech synthesis method and apparatus having a natural intonation that does not become tired even after listening for a long time.

[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 diagram illustrating a method of extracting a representative point pitch.

FIG. 3 is an explanatory diagram when a model parameter of an undulating accent is obtained.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 101 Text analysis part 102 Speech synthesis part 103 Text file 104 Morphological analysis part 105 Japanese analysis dictionary 106 Dependency analysis part 107 Phoneme duration length determination processing part 108 Pitch pattern generation processing part 109 Pitch pattern rough shape estimation processing part 110 Model parameter Calculation processing unit 111 Pitch control processing unit 112 Speech unit selection unit 113 Speech unit region 114 Unit connection unit 115 Synthesis filter processing unit

Claims (20)

[Claims] 1. A speech synthesis method for generating a pitch pattern based on information obtained by analyzing an input text, and synthesizing a speech based on the generated pitch pattern, comprising the steps of: Generating a pitch pattern or a characteristic amount thereof, generating a second pitch pattern corrected from the first pitch pattern or the characteristic amount based on a pitch pattern generation process model, and generating a voice based on the second pitch pattern. A speech synthesis method characterized by combining 2. A text analysis means for analyzing input text to generate text analysis information, and a pitch pattern generation means for generating a pitch pattern based on text analysis information output by the text analysis means, A first pitch pattern or a feature thereof is generated from the text analysis information in accordance with rules, and a second pitch pattern corrected based on the pitch pattern generation process model is generated from the first pitch pattern or the feature thereof. A pitch pattern generating means for outputting the second pitch pattern; and a voice synthesizing means for synthesizing a voice based on the second pitch pattern output from the pitch pattern generating means. Synthesizer. 3. A voice synthesizing method for generating a pitch pattern based on information obtained by analyzing an input text and synthesizing a voice based on the generated pitch pattern. Generating a pitch pattern or a characteristic amount thereof, calculating a model parameter of a pitch pattern generation process model from the first pitch pattern or the characteristic amount thereof, and calculating a second parameter from the model parameter based on the pitch pattern generation process model. A speech synthesis method comprising: generating a pitch pattern; and synthesizing speech based on the second pitch pattern. 4. A text analysis means for analyzing input text to generate text analysis information, and a pitch pattern generation means for generating a pitch pattern based on text analysis information output by the text analysis means, wherein A first pitch pattern or a feature thereof is generated according to a predetermined rule based on the analysis information of the first pitch pattern, a model parameter of a pitch pattern generation process model is calculated from the first pitch pattern or the feature thereof, and the model parameter A pitch pattern generating means for generating and outputting a second pitch pattern based on the pitch pattern generating process model from the above, and a voice synthesizing means for synthesizing voice based on the pitch pattern output by the pitch pattern generating means. A speech synthesizer characterized by the following. 5. The speech synthesis method according to claim 1, wherein said predetermined rule is created by a speech corpus based on a learning algorithm such as a statistical method. 6. The speech synthesizer according to claim 2, wherein said predetermined rule in said pitch pattern generation means is created by a speech corpus based on a learning algorithm such as a statistical method. Speech synthesizer. 7. The speech synthesis method according to claim 5, wherein the feature amount of the first pitch pattern includes at least a pitch maximum value in an accent phrase. 8. The speech synthesizer according to claim 6, wherein the feature amount of the first pitch pattern includes at least a pitch maximum value in an accent phrase. 9. The speech synthesis method according to claim 5, wherein the feature amount of the first pitch pattern includes at least a pitch maximum value in an accent phrase and a pitch of an accent phrase final mora. Speech synthesis method to be used. 10. The speech synthesizer according to claim 6, wherein the feature amount of the first pitch pattern includes at least a pitch maximum value in an accent phrase and a pitch of an accent phrase final mora. Speech synthesizer. 11. The speech synthesis method according to claim 1, wherein the pitch pattern generation process model is a superposition type model. 12. The speech synthesis apparatus according to claim 2, wherein the pitch pattern generation process model is a superposition type model. 13. The speech synthesis method according to claim 11, wherein said superimposition type model approximates a global descending component by an exponential function. 14. The speech synthesizer according to claim 12, wherein said superimposition type model approximates a global falling component by an exponential function. 15. The speech synthesis method according to claim 11, wherein said superposition type model has a constraint that prohibits generation of a negative rumbling when generating a phrase component. Method. 16. The speech synthesis apparatus according to claim 12, wherein said superposition type model has a constraint for prohibiting generation of a negative rumbling when generating a phrase component. apparatus. 17. The speech synthesis method according to claim 11, wherein said superimposition type model has a constraint that when generating an accent component, its amplitude does not fall below a predetermined lower limit. Speech synthesis method. 18. The speech synthesizer according to claim 12, wherein said superimposition type model has a constraint that when generating an accent component, its amplitude does not fall below a predetermined lower limit. Speech synthesizer. 19. The speech synthesis method according to claim 11, wherein said superimposed model includes a ending mora of said preceding accent phrase if said preceding accent phrase is a flat accent when generating a phrase component and an accent component. A speech synthesis method characterized by correcting phrase components and accent components so that accent nuclei do not occur. 20. The speech synthesizer according to claim 12, wherein the superimposition type model adds a ending mora of the preceding accent phrase if a preceding accent phrase when generating a phrase component and an accent component is a flat accent. A speech synthesizer that corrects a phrase component and an accent component so that an accent nucleus does not occur.