JPH06214585A - Voice synthesizer - Google Patents

Abstract

PURPOSE:To reduce a user's work load while he uses voice media of a computer by providing a voice rule synthesis means which controls rhythm associated with a freely varying uttered voice speed of natural voice while interacting with the computer or in other computer usages. CONSTITUTION:The synthesizer consists of a language processing section 12 which extracts language information from the content information of voices to be synthesized, a uttered voice speed deciding section 22 which decides the speed of uttered voice, a phenome processing section 14 which generates phenome information from the language information, a rhythm processing section 16 which generates rhythm information, a rhythm processing controlling section 26 which controls a rhythm processing mark generating section 24 based on the uttered voice speed outputted from the section 22 in the section 16, a synthesis parameter generating section 18 which generates synthesis parameters from the section 24, which generates rhythm marks, the phenome information and the rhythm information and a voice waveform generating section 20 which generates voice waveforms from the synthesis parameters.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice synthesizing apparatus for synthesizing voice based on a certain rule.

[0002]

2. Description of the Related Art In recent years, human-computer interaction (Human-Compute) has been performed by inputting, outputting and processing multimedia such as characters, voices, figures and images.
r Interaction) has been studied in various forms.

In the media used for human-to-human dialogue, voice includes not only linguistic information that conveys content, but also emotional information such as emotions and personality information such as men / women.
It has the feature that various information can be easily and freely transmitted. Being able to convey such information is also important for a natural dialogue in a human-computer dialogue.

On the other hand, recently, workstations and personal computers capable of handling multimedia have been developed due to dramatic improvements in memory capacity and computer power, and high quality voice and sound signals can be obtained without special hardware. I / O has become possible with the standard system. Under these circumstances, voice input / output technology has become important as an input / output medium for computers.

Among the voice input / output techniques, a technique for converting characters of an arbitrary sentence or word into voice has been developed, which is called voice synthesis technique. Currently, hardware dedicated to speech rule synthesis has been developed, and its applications such as text reading are expanding. By such a technique, it has become possible to convert characters, which are linguistic information, into audio signals as media conversion.
However, compared with the voice output of a synthesizer that edits and synthesizes the actually uttered natural voice, the current rule-synthesized voice is unnatural and the content is difficult to understand.

This is because, in the case of rule-synthesized speech, the connection and duration of phonemes, the change in pitch, etc. are expressed as rules, and the speech is synthesized based on the rules from the analysis result of the input character string. The cause of the deterioration of sound quality is due to the quality of synthesis parameters and the control of accent and intonation.

As a method of controlling prosody, a method of linearly interpolating point pitch information for each syllable as a control method of intonation such as Japanese hiragana or interrogative sentence (Hakoda, "Pitch in sentence speech synthesis" Examination of parameter control method "Acoustical Society of Japan Material, SP88-
7 (1988)). In addition, as a control method that uses a temporal change model instead of linear interpolation for the pitch change between syllables, a "fundamental frequency pattern generation process model" by a critical damping quadratic linear system (Hirose, Fujisaki, Kawai, Yamaguchi "fundamental frequency" Synthesis of Document Speech Based on Pattern Generation Process Model ", The Institute of Electronics, Information and Communication Engineers, Vol. J.
72-A, No. 1, pp. 32-40 (1989-
1)) is under study.

Further, in order to remove monotonousness and mechanical feeling of synthesized speech such as character-to-speech conversion, pitch called prosodic information is used for the purpose of synthesizing locally emphasized or weakened speech in the synthesized speech. (Fundamental frequency), amplitude,
Method to control duration such as emphasis and weakness (Takeda,
Ichikawa "Creation and Evaluation of Prominence Generation Rule for Japanese Sentences", Journal of Acoustical Society of Japan, vol 47, No. 6, pp-3
97-404 (1991)) is under consideration.

[0009]

DISCLOSURE OF THE INVENTION Prosody control in character-to-speech conversion has been studied by the above various control methods.
The quality of synthetic speech has improved. However, the utterance speed changes variously in human conversation and reading with natural voice, and the utterance speed is gradually changed by slowly or quickly uttering politely. That is, the utterance and the whole sentence are not uttered at a constant speed, and the control of the prosody in the uttered sentence is partially different. Further, even when the same sentence is uttered, control of prosody such as intonation differs depending on the speed, and the prosody changes according to the utterance speed.

For example, "he drives that big car"
When the sentence is uttered, the phrase voice section is as follows. The phrase voice section is a section in which one breath is spoken. The solid line in FIG. 2 shows the phrase component of the fundamental frequency pattern of the voice. This is an extraction of a phrase voice section of a voice uttered by a male subject. FIG. 2 (b) shows the case of uttering at 7 mora / sec. Here, "/" represents the boundary of the phrase voice section.

[0011]

[Equation 1] Next, when the vocalization speed is increased, for example, 10 mora / s
At the speech production speed of ec, the phrase speech section as shown in FIG.

[0012]

[Equation 2] Also, when the speaking rate is slowed down, for example, the speaking rate is set to 4
In mora / sec, as shown in (c) of FIG.

[Equation 3] Becomes

As described above, the segment of the phrasal voice section changes depending on the utterance speed, and natural and smooth utterance is performed.

However, the conventional speech synthesizer is intended for a field such as reading out a document for the purpose of character-to-speech conversion, and the prosody control method is determined by the sentence content to be read out regardless of the set utterance speed. Therefore, the prosody control that is read aloud does not change depending on the set utterance speed.In response to changes in the utterance speed of synthetic speech, the control of phonological duration and pause length is changed for phonology and linear for prosody. It only changed over time. For this reason, it is unavoidable that unnatural synthetic speech occurs when the speaking rate to be set is set high or low. In addition, it was not possible to synthesize a natural voice accompanied by a change in utterance speed by voice rule synthesis in dialogue.

In actual natural speech, prosody is controlled according to the above-mentioned vocalization speed, and a phrase may be changed by inserting a pause. Realization of a speech synthesizer capable of controlling prosodic information associated with changes in vocalization rate during dialogue with a computer and various other uses makes it possible to synthesize natural and easy-to-understand speech, reducing the burden on the user. To be done.

The present invention has been made in view of the above circumstances, and it is an object of the present invention to provide a means for realizing voice rule synthesis that enables control of prosody information associated with the vocalization speed of a natural voice that is slow and fast. And

[0017]

According to a first aspect of the present invention, there is provided a speech synthesizing apparatus for extracting linguistic information from content information of speech to be synthesized, and a phonological unit for generating phonological information from the linguistic information. And a prosody processing unit that generates prosody information,
A synthesizing parameter generating unit that generates a synthesizing parameter from the generated phonological information and prosody information, and a voice waveform generating unit that generates a voice waveform from the synthesizing parameter, and a utterance speed determining unit that determines a utterance speed. A prosody processing control unit that controls the prosody processing unit based on the vocalization speed information output from the vocalization speed determination unit.

A speech synthesizing apparatus according to a second aspect of the present invention is a speech synthesizing apparatus according to the first aspect.
The prosody processing control unit for controlling the segmentation of the speech tone component based on the speaking rate information output from the speaking rate determining unit.

[0019]

[Operation] With the above device, when synthesizing voices with different utterance speeds based on rules, in the control of prosody such as intonation which cannot be realized by the conventional voice rule synthesizer, the utterance speed of the synthesized voice is changed. Dependent control is possible, and speaking speed changes variously in human conversation and reading by natural voice, and it becomes possible to synthesize speech with slow and fast speaking speed, such as slowly and carefully speaking. .

That is, even if the sentences to be synthesized are the same, in controlling the prosody such as intonation according to the speaking speed,
It is possible to change the segmentation of the tone unit having one tone component.

[0021] Therefore, not only is it intended for the fields such as character-to-speech conversion and reading of a document which the conventional speech synthesizer has aimed at, but also in a dialogue between a human and a computer in which the speaking speed becomes fast or slow. It becomes possible to regularly synthesize a natural voice accompanied by a change in utterance speed by voice synthesis.

As described above, by realizing a speech synthesizer capable of controlling prosody information associated with a change in utterance speed in dialogue with a computer and various other uses, it is possible to synthesize a natural and easily understandable speech by a computer. This makes it possible to reduce the burden on the user when using the computer voice media.

[0023]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

As shown in FIG. 1, the sentence-speech rule synthesizing device 10 of this embodiment has a language processing unit 12, a phoneme processing unit 14, a prosody processing unit 16, a synthesis parameter generating unit 18, a speech waveform generating unit 20, and a utterance. The prosody processing unit 16 includes a speed determination unit 22, and the prosody processing unit 16 includes a prosody processing symbol generation unit 24 and a prosody processing control unit 26.

That is, the speech rule synthesizing device 10 extracts the utterance content information of the speech to be synthesized and output from the language processing section 12.
Processing is performed to determine information necessary for the synthesis processing, and the phoneme processing unit 14 and the prosody processing unit 16 generate a phoneme symbol string to which a prosody component is added. From this symbol string, the synthesis parameter generation unit 18
Generates a synthesis parameter sequence, and the voice waveform generation unit 20 generates an output waveform. The synthesizing device 10 has a utterance speed determining unit 22 that determines the utterance speed from the utterance speed and utterance content information as input information, the information from the language processing unit 12, and the like.
The phonological processing unit 14 is provided with a prosody processing control unit 26 that controls prosody processing based on the vocalization speed information output from the vocalization speed determination unit 22.
By passing the information from No. 6 to the prosody processing symbol generation unit 24, it becomes possible to regularly synthesize natural speech according to the utterance speed.

Next, each processing unit will be described in detail.

The language processing unit 12 analyzes the input Kanji / Kana mixed sentence in the character-to-speech conversion, and is information necessary for voice processing, such as words, boundary of clauses, reading of Kanji, accent of words, and dependency. ,, part of speech, inflection, etc.

For example, by performing morphological analysis, syntactic analysis, semantic analysis, etc. in the language processing unit 12, information necessary for voice processing, that is, word / segment boundaries, kanji reading, word accent, dependency, part-of-speech , Output information such as inflections.

In the morphological analysis, an input sentence can be divided into word units, and grammatical information such as the part of speech and utilization information of each word and boundaries between words and phrases can be obtained. In addition, it is possible to obtain kanji readings, word accents, etc. from the dictionary search results in the morphological analysis. Here, a dictionary to which each word is read and accented is used.

In the syntactic analysis, the words obtained by the morphological analysis,
It is possible to create a syntax tree or the like for a bunsetsu and analyze it to obtain information on each word, the dependency between bunsetsus, and the degree of connection between bunsetsus. For example, if the previously used example sentence is syntactically analyzed, the syntactic information as shown in FIG. 3 is obtained.

In the semantic analysis, the semantic structure of each sentence can be output from the analysis result containing the semantic attributes of words obtained by the syntactic analysis.

On the other hand, in the speech rule synthesis which synthesizes from the concept, the language processing unit 12 analyzes the content to be synthesized by a conceptual expression and outputs information such as a syntactic boundary, a part of speech, a reading, and an accent. Specifically, word dependency processing, relational processing centering on predicates, grammar processing of terms, and connection processing between sentences are performed to obtain the above-mentioned information.

The utterance rate determining section 22 determines utterance rate information for the prosody processing section 16 to perform prosody control according to the utterance rate.

As the input to the utterance rate determining unit 22, information output from the language processing unit 12 can be input, such as inputting the utterance rate from the outside or including specification of the utterance rate in the utterance content information.

Then, the utterance speed determination unit 22 determines the utterance speed by inputting the utterance speed from the outside or by specifying the utterance speed in the utterance content information. Determines the speaking rate. Further, in order to determine the utterance speed from the information obtained from the language processing unit 12, for example, from the syntactic information, the embedded sentence portion is uttered earlier than other portions, or the semantically insignificant portion is uttered earlier. To set the speed.

By thus determining the utterance speed, the prosody processing section 16 outputs information for performing control according to the utterance speed.

The phonological processing unit 14 generates phonological attributes of words from the information such as the reading of Kanji and dictionary items output from the language processing unit 12, and the phrase symbols output from the prosody processing unit 16 and the like. According to the well-known phonological rules such as nasalization, devoicing, and rendaku using the prosodic information indicating pause marks and semantic breaks, etc., the reading of each word in the synthesized sentence is determined and a monophonic symbol string is output. . Also,
The prosodic processing unit 16 outputs monophonic phoneme attribute information necessary for the accent type determination in the prosodic processing unit 16.

As shown in FIG. 4, the prosody processing section 16 comprises a prosody processing control section 26 and a prosody processing symbol generation section 24.

The prosody processing control unit 26 controls the vocalization rate information output from the vocalization rate determination unit 22 to cause the prosody processing symbol generation unit 24 to perform processing according to the vocalization rate. As a result, the prosody processing unit 16 is selected according to the speaking rate.
It is possible to change the classification of speech tone components obtained in.

In the prosody processing symbol generation unit 24, phonological information such as the monophonic symbol string and word phonological attribute generated in the phonological processing unit 14 and the boundary of phrases, word accents and grammatical information output in the language processing unit 12. Language information such as, and further, control information relating to the speaking rate from the prosody processing control unit 26, and the like, and input the fundamental frequency pattern, power, duration,
Determine prosodic components such as pose position.

For example, regarding the generation of prosodic symbols, parameters that describe prosodic features, that is, the length and size of a pose, the size of a phrase command / accent command, are defined as a prosody symbol, and a pose symbol / phrase symbol is defined. Can be used to determine the prosody component. Further, it is also possible to classify sentence types such as a plain text, an interrogative sentence, and an imperative sentence, and create accent and phrase rules for each sentence type.

In many cases, the classification of speech tone components, that is, the phrase voice section is semantically united as a unit. Here, a chain of morphemes that always have one accent component regardless of context, such as a combination of independent words and adjuncts, is called a "prosodic word", and a parse tree in which a prosodic word is a "leaf" is shown in FIG. Like In the boundary of prosodic words, the boundary part in which the left word directly modifies the right word is called "right-dependent boundary", and the other boundary is called "non-right-dependent boundary". Further, a word chain that is separated by non-right dependency boundaries and includes only right dependency boundaries is called a "maximum right dependency chain". The maximal right dependence chain is semantically one unit, and is often vocalized in a single breath. In this case, the boundary of the phrase speech section can be regarded as corresponding to the boundary of the maximum right dependence chain. In this way, it is possible to generate a phrase speech section by using information such as the boundary of clauses in the language information and the relationship of dependency.

The prosody symbol generation unit 24 generates a prosody symbol using the above prosody symbol generation rule,
A phrase speech section is generated using the phrase speech section generation rule.

Next, the prosody processing control unit 26 controls the prosody processing symbol generation unit 24 to perform prosody processing according to the utterance speed based on the utterance speed information determined by the utterance speed determination unit 22. .

For example, when changing the division of the phrase speech section according to the utterance speed, the prosodic processing symbol generator 24 is controlled by giving a threshold value or weighting relating to the utterance speed in the phrase speech section generation rule. The following shows the generation rules of phrase speech sections using thresholds and weights according to the speaking speed.

Here, the strength of the tone connection by the relation of maximal right dependence chain and the dependency showing the semantic unity, the threshold of the number of mora in the phrasal speech section by the utterance speed, and the weighting by those utterance speeds. To generate a phrase voice section.

As for the weighting, the faster the utterance speed, the larger the unit of utterance in one breath, and the weight is often higher than the semantic information. On the contrary, if the speaking speed is slow,
The unit of utterance is shortened, and the utterance is often carefully to make it easy to convey the meaning. Therefore, the generation of the phrase speech section has the following rules.

(1) The boundary between sentences, joints, and clauses is the boundary of the maximal right dependence chain as the boundary of the phrase speech section.

(2) The boundary I of the maximum right dependence chain is obtained. Proportion I (n) at the boundary of the phrase speech section at boundary I
Is set to 1. The boundary of the clauses other than I is I (n) = 0. n indicates the boundary position between the phrases from the beginning of the sentence.

(3) Tone coupling strength K (n) due to the dependency relation between the phrases is determined from the table of FIG.

(4) Thresholds A and B based on vocalization speed information
To decide. Each threshold is referred to from the table of FIG. 8 according to the speaking rate.

(5) At the boundary I, if the number of mora of the immediately preceding maximum right dependence chain is shorter than the threshold value A, the boundary S is not determined by the threshold value. That is, S (n) = 0.
Similarly, the bunsetsu boundary is S (n) = 0.

(6) In each boundary I, when the number of mora of the maximal right dependence chain is longer than the threshold value B, the boundary is divided into a length equal to or smaller than the threshold value B and the boundary is divided. The threshold value boundary S (n) = 1 is set at a position where the number of moras in the created boundary becomes even.

(7) The weighting Ig, Kg, Sg of each boundary according to the utterance speed is calculated from the table of FIG. 7, and the probability P (n) of the phrase speech section at the boundary between the phrases is calculated by the following formula. .

[0055]

[Equation 4] (8) When P (n) is larger than the specified value T, it is set as the boundary of the phrase voice section. Here, T = 0.65.

Here, the threshold values A and B and the weighting values are changed according to the vocalization speed information, so that it is possible to change the segment of the phrase voice section corresponding to the vocalization speed. As a result, it is possible to generate natural prosody information according to the speaking rate.

In this embodiment, the prosody corresponding to the utterance speed is provided by providing the above-mentioned prosody symbol generation rule with the prosody processing control unit 26 for controlling the prosody process from the utterance speed information output from the utterance speed determining unit 22. Allows control.

The prosodic component determined in this way divides the monophonic symbol string output by the phonological processing unit 14 into phrase speech sections, and prosody such as the rise and fall of the pause component and phrase component in each phrase speech section. A phonological symbol string to which a prosody symbol string indicating information is added is generated and output to the synthesis parameter generation unit 18.

As shown in FIG. 5, the synthesis parameter generator 18 synthesizes the speech synthesizer 10 on the basis of the phonetic symbol which is the processing result of the phoneme processing unit 14 and the pause symbol which is the processing result of the prosody processing unit 16. At the same time as generating the transfer characteristic parameter time series for controlling the characteristics of the filter, at the same time as generating the fundamental frequency time series for controlling the sound source frequency of the speech synthesizer 10 based on the speech tone and accent symbols which are the processing results of the prosody processing unit 16. To do.

That is, a time series of transfer characteristic parameters is generated by connecting syllable parameter sequences corresponding to a single phonetic symbol and a pause symbol in accordance with the sequence of symbols, while a fundamental frequency pattern is generated according to an accent symbol and a tone key symbol. By generating commands to the model and inputting commands to the fundamental frequency pattern generation model, the fundamental frequency time series is generated.

The syllable parameter series and the transfer characteristic parameter series are composed of a series of synthesis filter parameters of the speech synthesizer 10.
Changes in frequency characteristics of each syllable are represented by a series of synthesis filter parameters at fixed time intervals. In the pause section, since there is no voice output, the sound source gain of the synthesis parameter is set to zero. The transfer characteristic parameter sequence connects the syllable parameter sequence according to the set vocalization rate and phoneme duration. In connection, interpolation processing between parameters is performed so that each parameter changes smoothly like natural speech.

A fundamental frequency time series for controlling the sound source frequency of the speech synthesizer 10 based on the speech tone symbol and the accent symbol for controlling the speech tone component and the accent component which are the processing results of the prosody processing unit 16 is generated. To generate. Here, a case will be described in which a tone symbol and an accent symbol are input to a fundamental frequency pattern generation process model to generate a fundamental frequency time series.

The fundamental frequency pattern generation process model includes (1) a gradual downward curve (speech component) from the beginning of a sentence to the end of the generated fundamental frequency, and (2) a curve with a strong local undulation (accent component). Two components of critical braking 2
It is modeled as a response of a second-order linear system. The time change of the logarithmic fundamental frequency is expressed as the sum of these two components. The fundamental frequency F0 (t) (t is time) is formulated as the following equation.

[0064]

[Equation 5] Here, Ln () represents a logarithm, and SUM [i =
[n, m] () represents the sum of n to m for the parameter i, F0 (t) represents the fundamental frequency, Fmin represents the minimum fundamental frequency, and Gpi (t) represents the impulse of the speech component control mechanism. Represents a response function, Gaj (t) represents a step response function of the accent component control mechanism, and Api,
Aaj indicates the tone component instruction, the magnitude of the accent component instruction, T0i indicates the time point of the tone component instruction, and T1j and T2j indicate the start point and the end point of the accent instruction.

Here, when t> 0, Gpi (t), Gaj
(T) is

[Equation 6] And when t <0, Gpi (t) = Gai (t)
= 0.

It should be noted that min [n, m] indicates that n or m is smaller, and it does not take a value larger than TH (<= 1.0). exp () is an exponential function, and Xi and Yj are natural angular frequencies of the speech tone component control mechanism and the accent component control mechanism, respectively.

In order to generate the fundamental frequency time series based on the fundamental frequency pattern generation process model, it is necessary to set each parameter, but the parameters are changed so as to minimize the error between the natural speech and this model. The parameter value is initialized by the best approximation estimation of the parameter value.

The speech waveform generating section 20 includes a transfer characteristic parameter time series for controlling the characteristics of the synthesis filter of the speech synthesizer 10 which is the processing result of the synthesis parameter section and a fundamental frequency for controlling the sound source frequency of the speech synthesizer 10. The synthesis filter of the speech synthesizer 10 is driven based on the sequence to generate a speech waveform.

For example, an example of voice waveform generation by the formant type synthesizer 100 as shown in FIG. 6 will be described.

The source of the voiced sound source is a pulse generator, and the source of noise is a noise generator. RG in Figure 6
P indicates a glottal resonance circuit, RGZ indicates a glottal anti-resonance circuit, A1 to A5 indicate amplitude control circuits, AV indicates an amplitude control circuit for voiced sound sources, and AH and AF indicate amplitude control circuits for unvoiced sound sources. , AB are amplitude control circuits for bypass. In addition, F0 indicates the fundamental frequency, and F1 to F5 are
A resonance circuit controlled by the resonance frequency and the bandwidth is shown.

This formant synthesizer 100 generates a sound source waveform while controlling the frequency of the voiced sound source and the gains of the voiced sound source and the unvoiced sound source based on the fundamental frequency time series and the gain parameter of the sound source of the transfer characteristic parameter time series. . Further, the sound source waveform is input while controlling the synthesis filter based on the transfer frequency parameter time series synthesis filter center frequency parameter and bandwidth parameter, and the output waveform of each filter is obtained, and the sum of the output waveforms of the respective synthesis filters is calculated. Are output from the synthesizer 100. For example, a synthesis parameter can be updated every 8 msec at a sampling frequency of 12 KHz, a low-pass filter for impulses can be used for voiced sound sources, and a random number generator can be used for unvoiced sound sources.

Although the waveform generation by the formant type synthesizer 100 has been described here, a generally known one can be used as the configuration of the synthesizer, the type of the sound source, the sampling frequency and the like.

The present synthesis device 10 is characterized in that it is possible to generate a synthesized voice that is easy to hear at any utterance speed by changing the segment of the phrase voice section according to the utterance speed.

In natural speech, prosody is controlled according to the utterance speed, and, for example, the segmentation of the phrase speech section changes as shown in FIG. In order to convert the utterance content into a natural and smooth synthetic speech according to the utterance speed, it is necessary to change the segment of the phrase speech section according to the utterance speed.

When the utterance speed is high, a plurality of phrase sound sections are connected to form one phrase sound section, or when the utterance speed is low. Within one phrase voice section, the phrase voice section is divided into a plurality of phrase voice sections at a structurally and semantically cut portion of a sentence.

The change of the segment of the phrase speech section according to the utterance speed will be described with a specific example.

Here, a description will be specifically given by using a sentence example in which the above-described processing is applied by applying the above-described rule for changing the division of the phrase speech section according to the utterance speed.

From the result of the syntactic analysis, the syntax tree of the sentence example is as shown in FIG. 3. From this, the boundary of the maximal right dependency chain is as follows. Here, "|" represents the boundary of the maximum right dependence chain. It also shows the ratio that can be the boundary between each clause.

[0079]

[Equation 7] In addition, the strength of the tone combination due to the dependency relation between each bunsetsu is as follows. However, the larger the value, the weaker the degree of binding.

[0080]

[Equation 8] The boundary S obtained from the threshold according to the speaking rate is as follows. The speech rate here is 7 mora / sec
From the table of FIG. 8, A is 3
Mora, B is 13 Mora. The threshold boundary is represented by "|".

[Equation 9] In this case, the boundary of the maximal right dependency chain after "he is" is the boundary S according to the generation rule of the phrase speech section, but "that" is 2
Since it is a mora, it is not a boundary after that due to the production rules. Since the boundary of the last maximum right dependency chain satisfies the production rule, it becomes a phrase-speech section boundary, which is a section like the above sentence.

From this, the probability of becoming a phrase speech section at the boundary between each clause is obtained as follows, and the phrase speech section represents the portion indicated by "/" from the specified value of the probability.

[0082]

[Equation 10] The following shows the division of phrase speech sections when the utterance speed is increased. When the speech rate is high, the threshold value becomes large. Speaking at a speech rate of 10 mora / sec, the threshold A is 6 mora and B is 15 mora from the table of FIG. 8, and the phrase speech section of the above sentence is as follows.

[0083]

[Equation 11] When the speaking rate is slow, the threshold value is set small.
When uttered at 4 mora / sec, A is 2 mora, B is 12 mora, and the phrase speech section of the above sentence is as follows.

[0084]

[Equation 12] In this way, by generating the phrase voice section of the above example sentence, a synthesized sentence of the phrase voice section can be generated as in the case of the natural voice as shown in FIG. That is, it is possible to control the prosody according to the speaking rate.

[0085]

As described above, in the case of synthesizing voices having different utterance speeds according to the present invention according to the present invention, in the prosody control such as intonation which cannot be realized by the conventional voice rule synthesizing device, the synthetic voice The control depending on the utterance speed becomes possible, and the voice can be synthesized according to various changes in the utterance speed in the natural voice of human.
That is, even if the sentences to be synthesized are the same, it is possible to change the segmentation of the tone unit having one tone component in prosody control such as intonation according to the utterance speed.

That is, in the voice rule synthesizing device, the utterance speed determining unit is provided, so that the utterance speed can be determined, and the prosodic processing control unit for controlling the prosody corresponding to the determined utterance speed is provided. As a result, it becomes possible to change the division of the phrase speech section according to the utterance speed.

Therefore, not only is it intended for the fields such as character-to-speech conversion and reading of a document that a conventional speech synthesizer has aimed at, but also in a dialogue between a human and a computer in which the speaking speed becomes fast or slow. By using voice rule synthesis, it becomes possible to perform rule synthesis of a natural voice accompanied by a change in utterance speed, such as slowly and slowly speaking.

As described above, the realization of the speech synthesizer capable of controlling the prosody information associated with the change in the utterance speed in the dialog with the computer and various other uses makes it possible to synthesize a natural and easily understandable voice by the computer. This makes it possible to reduce the burden on the user when using the voice media of the computer.

[Brief description of drawings]

FIG. 1 is a block diagram of a sentence voice rule synthesizing device according to an embodiment of the present invention.

FIG. 2 is a phase component showing a fundamental frequency pattern of speech, in which (a) is a speech rate of 10 m / sec and (b) is 7 m / sec. Is the speaking rate,
(C) is a speech rate of 4 mora / sec.

Figure 3 is the syntax tree for the sentence "He drives that big car".

FIG. 4 is a block diagram of a prosody processing unit.

FIG. 5 is a block diagram of a synthesis parameter generation unit.

FIG. 6 is a block diagram of a formant synthesizer.

FIG. 7 is a table showing boundary weights.

FIG. 8 is a table showing threshold values of A and B based on a speaking rate.

FIG. 9 is a table showing strengths of tone combinations.

[Explanation of symbols]

 10 ... Sentence rule synthesis device 12 ... Language processing unit 14 ... Phonological processing unit 16 ... Prosodic processing unit 18 ... Synthesis parameter generation unit 20 ... Speech waveform generation unit 22 ... Speaking speed determination unit 24 ... Prosody processing symbol generation unit 26 ... Prosody processing control unit

Claims (2)

[Claims] 1. A linguistic processing section for extracting linguistic information from content information of a synthesized voice, a phonological processing section for generating phonological information from the linguistic information, a prosody processing section for generating prosody information, and the generated phonological information. A speech synthesis apparatus including a synthesis parameter generation unit that generates a synthesis parameter from prosody information and a voice waveform generation unit that generates a voice waveform from the synthesis parameter. A speech synthesis apparatus comprising a prosody processing control unit for controlling the prosody processing unit based on the speech rate information output from the speech synthesis device. 2. The speech synthesis apparatus according to claim 1, further comprising a prosody processing control unit for controlling segmentation of speech tone components based on the speech rate information output from the speech rate determination unit.