JPH01219899A - Speech synthesizing device - Google Patents

Abstract

PURPOSE:To obtain a smooth voice by specifying phoneme parameters of the sectional area, etc., of acoustic tubes by time zones, processing the phoneme parameters while superposing the energy of a sound source upon a pattern prescribed by an exponential function, and assigning a value less than the function value of the pattern for a vowel to be made voiceless. CONSTITUTION:The voice channel of a human is regarded as an acoustic tube group and made to correspond to a circuit element group of surge impedance components, and a voice is simulated according to the current wave at the output terminal. Parameters of the pitch of a current source and the sectional area of acoustic tubes are prescribed by phonemes, and parameters are interpolated as to connections between phonemes and the connections of sectioned time zones in phonemes. As for energy, energy patterns corresponding to accent prescribed previously by an exponential function by words are assigned and the interpolated values of the energy are superposed on the energy patterns. Further, the value less than the function value of the energy patterns is assigned to a vowel to be made voiceless to generate no voice, so the voice which is closer to the human voice is obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION A. Field of Industrial Application The present invention relates to a speech synthesis device using an acoustic tube model.

B9 Summary of the Invention The present invention regards the human vocal tract as a group of acoustic tubes, and by associating this with a group of circuit elements for surge impedance components, it is possible to simulate speech based on the current wave at the output end of the group of circuit elements. The generation device divides the generation time of each phoneme into multiple time periods for each phoneme that makes up a syllable, and specifies phoneme parameters such as the cross-sectional area of the acoustic tube and the energy of the sound source wave for each time period. In addition to interpolating the parameters, the interpolated value group of the energy of the sound source wave is superimposed on a pattern defined by an exponential function depending on the accent of the word, and for vowels to be devoiced, the function value of that pattern is used. By assigning a value lower than , it is possible to create a voice that is smooth and approximates human speech.

C1 Conventional technology Electronic devices that artificially synthesize and output so-called sounds, such as voice synthesis and music synthesizers (electronic musical instruments), have recently become LSs for voice recognition and voice synthesis using one or several chips.
I can now be realized at low cost through audio information processing and semiconductor large-scale integrated circuit technology, and various systems have been proposed depending on the purpose of use and constraints. This speech synthesis method involves two methods: recording and editing raw human speech, which combines them appropriately and editing them into sentences; There is a parameter method that extracts only the parameters and controls them during the speech synthesis process to artificially create a speech signal.

In the parameter method, the audio waveform is sampled at certain intervals, the audio signal value at each sampling point is converted from analog to digital, and the values are displayed as O and l signs. In order to record faithfully, it is necessary to increase the number of bits, which requires a large memory capacity.

Therefore, various highly efficient encoding methods are being researched and developed in order to reduce the amount of information as much as possible.

One such method is a delta modulation method in which at least one bit corresponds to the information of one audio signal. In this method, one bit is used to determine whether the next audio signal value is higher or lower than the current value, and if it is higher, the code "l" is used.
”, if it is low, the code “0” is given and the audio signal is encoded.In the actual system configuration, a constant amplitude step amount (delta) is determined, and the previous code is changed to prevent errors from accumulating. The residual signal between the audio value obtained by the encoding and the input audio signal is encoded.

This type of predictive coding is known as the linear prediction method (predicting from several previous sample values) and the Percoll method (using a partial autocorrelation function called the linear Percoll coefficient instead of the prediction coefficient of the linear prediction method). There is.

Problems to be Solved by the 2009 Invention Of the conventional speech synthesis methods, the recording and editing method has a problem in that the amount and types of texts that can be synthesized are limited.

In addition, in methods using predictive coding, it is difficult to create articulatory combinations that correspond to the joints between sounds, and a method for combining synthesis units has not been established. In the process of going from a steady state to a consonant through a transition, and then through a vowel transition to a steady vowel sound, the sound at the joint between the vowels becomes disconnected. Therefore, there is a problem that the sound lacks smoothness and does not give a natural feeling when heard by humans.

An object of the present invention is to provide a speech synthesis device that can synthesize sentences of any amount and provide a smooth sound that is close to actual human speech and can give a natural feeling to the listener. It's about doing.

Means and operations for solving the E0 problem (1) Basic concept In order to radiate sound outward from the mouth, a sound source is required, and this sound source is produced by the vocal cords. -The vocal cords have the function of intermittent exhalation by opening and closing two folds, and these intermittent intervals generate airflow called puffs, and when the vocal cords are tensed, tension is applied to these folds, causing the folds to open and close. The frequency increases, producing a high-frequency puff sound. When the exhalation flow is increased, the sound becomes louder.

When this sound source wave passes through a cylindrical acoustic tube like the vocal tract, a resonance phenomenon causes certain components of the sound waves from the open end to be emphasized and certain components to be attenuated, creating a complex vowel waveform.

Even if the sound source waves have the same waveform, the sound emitted from the mouth is affected by the shape of the vocal tract that the sound passes through before being emitted from the lips. That is, the sounds produced by humans are determined by the length and cross-sectional area of the vocal tract from the vocal cords to the lips, and the way the vocal cords vibrate.

The present invention has been made with this in mind, and considers the vocal tract as a group of acoustic tubes with variable cross-sectional areas.
Furthermore, the starting point is to realize the traveling wave phenomenon representing the transmission of sound waves in an acoustic tube using its equivalent circuit. If the vocal tract is regarded as an acoustic tube, the propagation of sound waves in each acoustic tube can be divided into forward waves and backward waves, and can be thought of as repeated reflection and transmission phenomena at the boundary surfaces of each acoustic tube. Reflection and transmission are quantitatively defined according to the degree of mismatch of acoustic characteristic impedances at the interface, that is, the ratio of the respective cross-sectional areas of adjacent acoustic tubes.

The reflection and transmission phenomena described above are the same as the transient phenomena that occur when an innocuous current is passed through lines with different impedances in an electric circuit.

(2) Equivalent circuit Based on the above, a sound tube model consisting of n sound tubes 81 to S0 is shown in Fig. 1 (A).
It can be expressed as an electric circuit in which a group of circuit elements (TI-T-) each consisting of a lossless surge impedance component with no resistance are connected in series as shown in FIG. A,
−A is the cross-sectional area of the acoustic tubes 81 to 81 to S, respectively. Here, in the present invention, basically, the above electric circuit is applied,
Impulse current supplied to this and each circuit element T1 to Tn
By changing the surge impedance of the acoustic tube model, the current output from the final stage circuit element Tn is supplied to the vocal part such as a speaker, in response to changing the sound source wave of the acoustic tube model and the cross-sectional area of each acoustic tube. By doing this, the sound obtained from the acoustic tube model is simulated.

Specifically, we assume a circuit equivalent to the above electric circuit as shown in Figure 1 (2), change the current of the current source in this equivalent circuit with respect to time, and create the calculation formula as described later. Since the cross-sectional area ratio of the acoustic tube is introduced, each cross-sectional area A + ""' A- is changed with respect to time, and the current value of each part is calculated by this. In the figure, P is the current source, Zo is the impedance of the current source,
Z1 to Z0 are surge impedances and Zt of circuit elements Ti to Tn, respectively.

is the radiation impedance, i 0A-1(n-11A,
i IB~i nB, a OA-a (h-1)A
+ a IB-a nB is the current in the current path corresponding to each symbol, W OA-W +n-1)^, Wla to Wna are the current sources, I OA-1tn-nA is the backward wave current, I, b
~InB indicates forward wave current. In this equivalent circuit, for example, circuit element T1. Focusing on the connecting part of Tt,
A current source 'N+Ah corresponds to the current 111B flowing through the circuit element T toward Tt, and a current source W I corresponds to the current 11A flowing through the circuit element T2 toward T.
Assuming A h, the current 1+a is the circuit element Tl.

The reflected wave current f+a reflected to T at the boundary of T and T
, transmitted wave current a (Ah), and the current 11A is reflected to T at the boundary of circuit elements T, , T, reflected wave current i+A, and transmitted wave current a+B transmitted to T,
This is an equivalent representation of the division into Further, FIG. 1 (1) is a schematic diagram schematically showing such a situation.

(3) Calculation First, the block including the current source P that plays the first role shown in FIG. 1 can be thought of as a superposition of two circuits as shown in FIG. Therefore, the voltage of current source P is
Current a1 in the same figure. at is expressed by equations (1) and (2), respectively, and as a result, the current a. A is expressed by equation (3).

a I=v/Z, +Z+ −(1) at =
zo/zo+zt・Iot −(2)aO^= a
H0a1-1/Zo±Z1(v+Zo-IOA)...(
3) If we now supply current into the equivalent circuit for the first time, aoA can be found by setting IOA to zero.

Then, calculations are performed sequentially based on this value.

Taking as an example the calculation formula for the current value of the first stage block and the second stage block located at the left end of the figure, the following (4) ~
It is expressed as equation (12).

aoA'= 1/Zo+Zl(V'+Zo・I OA)
...(4) ioAI=aO^'-Io^
...(5) 1 o^"= i lB'+
a +e'”' (6) a+a
'-s+i+(I IB+ I IA)
...(7) i+a'=a+a' I IB
-(8)1+B'=jo^+ao^'
...(9) a lA'= S I
A (I IB+ 11 =-(10)j
lA'= a +A' I +8
”' (11) 11A'= i ta'+ a t
B' - (12) As we proceed with this calculation, the arithmetic expression for the final stage block becomes (1
3) to (15).

a nB': zL, / Zn 10ZL-1nB
+・-(13) l nB”” a nB' I n
BI na'='L tn-++A+ a (n-11
A'...(14) In this way, the current jnB corresponding to the sound wave emitted from the sound tube Sn at the final stage is determined.

However, S18゜SIA is a coefficient expressed by the cross-sectional area ratio of adjacent sound tubes, and is (15) and (1
6) Formula % Formula % A series of calculations for the current values of the blocks from the first stage to the final stage are executed instantaneously, and these calculations are performed one after another at a predetermined timing. Here, the above (4) to (1)
In formula 4), the value with a dash is the calculated value at the time, and the value without the dash is the calculated value obtained by the calculation one time before the calculation at the time. i is the digital value obtained in this way. is converted from digital to analog to create an analog current, and this current is supplied to a speaker or the like to produce sound. The timing of the calculation is determined in consideration of the speed of sound, and for example, by setting the propagation time of one acoustic tube as the time interval of the calculation, the backward wave current I is determined. A to I (□=, , A and forward wave current 1
+a-1 nil is the same speed as the speed of sound, and each circuit element T, -
A state equivalent to that flowing through 1 nA is created, thereby matching the acoustic tube model and the electric circuit model.

The present invention is based on the realization of the equivalent model and calculations described above, and specifically, the utterance time of each phoneme is divided into one or more time periods for each phoneme that makes up a syllable, and each phoneme is divided into one or more time periods. For each time period, each initial value of the pitch, which is the repetition frequency of the sound source wave, the energy of this sound source wave, and the cross-sectional area of the acoustic tube, and each initial value X of the next time period is determined from the above-mentioned initial value Xo of the relevant time period.
a phoneme parameter storage unit that stores a constant that defines how to change to r and a sound source wave pattern; a parameter interpolation processing unit that performs interpolation processing of the pitch, energy, and cross-sectional area corresponding to the input phoneme data; , the energy patterns of the sound source waves corresponding to the accents of each word are classified into multiple types, and the corresponding energy pattern type is assigned to each word, and the vowel in the word that should be devoiced is a dictionary section for registering each word with a code; and a function for defining the energy pattern for each type; and referring to the dictionary section, an energy pattern corresponding to the type of word to be uttered is determined a pattern processing unit that assigns a value lower than the function value of the energy pattern to the selected vowel and a devoicing code; and an output of the circuit element group based on the parameters interpolated by the parameter interpolation processing unit. a calculation unit that calculates the current value output from the end, and uses a value obtained by superimposing a function corresponding to the pattern selected by the pattern processing unit on the interpolation-processed interpolation value group for the calculation value of the energy; , and a voice generation section that generates a sound based on the calculation result of the calculation section, and the parameter interpolation processing section calculates the initial value Xo, the X1 corresponding to the target value, and the constant during each time period. The method is characterized in that interpolation calculations are performed multiple times using , and the energy is performed based on an exponential function for interpolation.

F. Embodiment FIG. 1 is a diagram showing a block configuration of an embodiment of the present invention. ■ is the Japanese language processing section, which divides the input Japanese sentences into phrases, and the dictionary section 1. Refer to , and perform reading conversion, etc. Reference numeral 2 denotes a sentence processing section which performs processing for adding intonation to sentences. 3 is a syllable processing unit, which adds accents to the syllables that make up a sentence according to the intonation. For example, the sentence "Sakura ga Saita" is broken down into syllables such as rsAJ, rKUJ, rRAJ, etc., and each syllable is accented. Since the intonation of a sound is determined by the repetition frequency of the sound source wave, its energy, and time, which will be described later, adding an accent means determining coefficients for these parameters. In particular, the determination of coefficients for energy is carried out by the pattern processing section 31 in the syllable processing section 3.
It will be executed at

Regarding the pattern processing unit 3, energy patterns of sound source waves corresponding to the accent of each word are classified and stored into three types, for example, a head height pattern, a tail height pattern, and a middle height pattern, and these three types of energy It has a function of selecting a pattern corresponding to the word to be uttered from among the patterns. Figure 4 (A) to (T) are respectively “
This is a diagram showing the results of analyzing the speech actually uttered by humans for the words "It worked,""Abnormal", and "The circuit breaker", where the solid line shows the change in the energy of each syllable, and the dotted line shows the energy of each syllable. The energy patterns corresponding to the accents of the words obtained by connecting the peak values of are shown. In this example, the patterns shown by the dotted lines in FIG. Figure 5 (A) - (T)
The pattern processing unit 3 is defined by an exponential function as shown in FIG.
It is stored in advance in the storage section in I.

In this case, each stored pattern has an exponential function defined by a different formula, for example, assigned to an ascending portion and a descending portion. Regarding the extraction of this energy pattern, for example, the dictionary section 1. A code indicating the type of energy pattern is attached to the word to be registered in advance, and the pattern processing section 31 selects the energy pattern corresponding to this code.

Also, in the word, rIJ of rslj shown in Figure 4 (a)
In some cases, devoiced vowels whose energy level is considerably lower than the level corresponding to the energy pattern indicated by the dotted line may be included, such as rNJ and rl in rKIj shown in FIG. For this reason, the dictionary section 1. A devoicing code is attached to the vowel to be devoiced in the word, and a value lower than the function value of the energy pattern is assigned to the vowel with this devoicing code in the pattern processing unit 3. ing.

4 is a phoneme processing unit, 41 is a syllable parameter storage unit,
The phoneme processing unit 4 converts manually generated syllable data such as rsAJ into phonemes by referring to data in the syllable parameter storage unit 41 that defines the correspondence between syllables and phonemes, which are units of vowels and consonants. Process of decomposition, e.g. syllable rsAj
On the other hand, the phoneme "S".

Take out rAJ.

5 is a parameter interpolation processing section, 5 is a phoneme parameter storage section, and S is a sound source parameter storage section.

As shown in FIG. 6, the phoneme parameter storage unit 5I divides the utterance time of each phoneme into a plurality of time periods, for example, three time periods 0° to 03, and stores the duration time, the repetition frequency of the sound source wave, and the repetition frequency of the sound source wave for each time period. Stores each initial value of the energy of the sound source wave and the cross-sectional area of the sound tube, a time constant that specifies the manner of change from each of the above-mentioned initial values in the relevant time period to each initial value in the next time period, and a sound source wave pattern. ing. In this example, the human vocal tract (17cx in the male case) is connected to an acoustic tube of length lcx.
The model is made up of individually connected light beams, and for this reason, 17 cross-sectional area values (Al to A, ?) are determined for one time zone light. Further, in the sound source parameter storage unit 5°, waveform components of three types of sound source wave patterns G1-03 are stored as 50 sample data, as shown in FIG. 7, for example. The parameter interpolation processing unit 5 calculates each time period (0, -
This is the part that performs interpolation processing of pitch, energy, and cross-sectional area in 03), and this processing is performed based on the pitch, energy, and cross-sectional area of the relevant time period.
If the initial value of each parameter of energy and cross-sectional area is expressed as Xo, the initial value of the next time period is expressed as Xr, the nth interpolated value is expressed as X(n), and the time constant corresponding to each parameter is expressed as D, then This is a process in which calculations are performed n times during the time period according to the recurrence formula shown in equation (17). However, initial @X(0)
is the above Xo.

X(n)-D (X,-X(n-1)) +X(n-1
)-(17) For example, time zone 0. The energy interpolation process in is calculated according to equation (18) since Xoh<El and Xr corresponds to E.

X(n)=DE+(Et-X(n-1))+X(n-1
)...(18) Here, the above equation (17) is a recurrence equation of the following equation (19).

X=X, -e-'' -(19) That is, when equation (19) is differentiated, equation (20) is established, and therefore (21) is established.

d x / d t = De −” − (20) Δ
X=X(n+1)X(n)=Δt++D e −”
"'=Δt-D(X,-X(n))-(2D Therefore, equation (22) is obtained.

X(n+1)=Δt-D(Xr-X(n))×X(n)
-(22) Here, since the time interval of the interpolation calculation is constant, Δ and D can be collectively replaced with the time constant,
It is expressed as equation (17).

Reference numeral 6 denotes a calculation unit, which calculates the digital value of the current 1na shown in FIG. 1 (2) at the same timing as the interpolation calculation, for example, at a time interval of 100 μs, based on the parameters calculated by the parameter interpolation processing unit 5. 7 is a digital/analog (D/A) converter, which generates a current wave (analog current) based on the digital value obtained by the calculation section 6. Reference numeral 8 denotes a voice generating section such as a speaker, which generates voice based on analog current.

Next, the operation of the above embodiment will be described.

A Japanese sentence inputted by a word processor or the like passes through a Japanese language processing section 1, a sentence processing section 2, and a syllable processing section 3, where it is divided into syllables with intonation added, and further, each syllable is divided into syllables by a phoneme processing section 4. It is decomposed into Next, the parameter interpolation processing unit calculates the pitch of each phoneme,
Energy and cross-sectional area are taken out from the phoneme parameter storage unit 5, and interpolation processing is performed on these parameters for each time period (0° to 0.0°). This interpolation process is (1
For example, the energy in time period O1 is performed according to equation (18). 8th
The figure shows this situation, and each interpolated value of energy E (1), E (2) obtained by interpolation calculation is shown in the figure.
・=E (n) are arranged along the curve expressed by the following equation (23).

E = Et e-Dt・-(23) Also, sample data of the sound source wave pattern specified for each time period 0 and 03 is taken out from the sound source parameter storage unit 5, and this sample data is interpolated with pitch, etc. The value is given to the calculation unit 6, and the calculation unit 6 calculates the above E, term (3) “
The calculations detailed in "Operations" are executed. In this calculation, the coefficient or function corresponding to the accent added to each syllable unit in the syllable processing unit 3 is multiplied by each parameter obtained in the parameter interpolation processing unit 5, so that the intonation of the sentence is expressed. is calculated. In particular, regarding energy, the pattern processing unit 3. The energy pattern corresponding to the accent of the word (fifth
(see figure), read out the exponential function value that defines the pattern (the value on the vertical axis in Figure 5) so that the energy pattern is drawn during the utterance time of that word, and add the energy to the read value. The interpolated values are multiplied and combined, and the combined values are used as calculation elements. For a vowel with a devoicing code, for example, l is added to the read value.
The value is multiplied by a coefficient of about /8 and the interpolated value of energy is multiplied to match the value.

In this way, the digital value of the current wave corresponding to the sound wave emitted from the final stage sound tube is obtained, and based on this value, the D/
A current wave is created by the A converter 7, and a corresponding sound is emitted from the vocalization sound 8.

G3 Effects of the Invention According to the present invention, the propagation of sound waves in a sound tube model is replaced by the flow of current in an equivalent circuit, parameters such as the pitch of the current source and the cross-sectional area of the sound tube are defined for each phoneme, and the Since parameter interpolation processing is performed on the seams between 2 and 2 or between divided time periods within a phoneme, smooth speech can be obtained, giving a natural feel to the listener.

Energy is interpolated based on an exponential function, so the arrangement of interpolated values is similar to that of actual speech. Moreover, each word is assigned an energy turn corresponding to an accent predefined by an exponential function. In addition to superimposing the energy interpolation value on the energy pattern, devoicing is achieved by assigning a value lower than the function value of the energy pattern to the vowel that should be devoiced, so it is possible to obtain a voice that is even more human-like. Can be done. Furthermore, instead of storing all parameter values in the area corresponding to the joints between phonemes in the memory, it is sufficient to store data in units of phonemes or units of time, so the memory capacity can be reduced.

[Brief explanation of the drawing]

FIG. 1 is an explanatory diagram showing an equivalent model of an acoustic tube, FIG. 2 is an equivalent circuit diagram showing a block including a current source, FIG. 3 is a block diagram showing an embodiment of the present invention, and FIGS. 4 and 5 are explanatory diagrams each showing an energy pattern. FIG. 6 is a data diagram of phoneme parameters, FIG. 7 is an explanatory diagram showing a sound source wave pattern, and FIG. 8 is an explanatory diagram showing the state of parameter interpolation processing. 11... Dictionary section, 31... Pattern processing section, 4...
- Phoneme processing section, 41...Syllable parameter storage section, 5.
...Parameter interpolation processing unit, 51... Phoneme parameter storage unit, 5... Sound source wave pattern storage unit, 6... Calculation unit, 7... Digital/analog conversion unit, 8... Generation unit. isoline e decimal diagram fig. 2 SHA DA N KI WA Fig. 5 Construction route f-7-n t @ diagram order AIY tool Middle and high high school Hoon Odaka Batun Fig. 6 Hard prefecture parameter take l Fig. 7? ;Undiluted paboon ajtl! Figure time early 01

Claims (1)

[Claims] (1) By regarding the human vocal tract as a plurality of acoustic tubes connected in tandem, and by associating the acoustic tube group with the circuit element group of the surge impedance component and making the sound source correspond with the current source, the acoustic tube In a speech synthesizer that simulates a speech wave emitted from an output end of a group of circuit elements based on a current wave of an output end of a group of circuit elements, the utterance time of each phoneme is divided into one or more time periods for each phoneme constituting a syllable. For each time period, the initial values of the pitch, which is the repetition frequency of the sound source wave, the energy of this sound source wave, and the cross-sectional area of the acoustic tube, and the initial values a phoneme parameter storage unit that stores constants and sound source wave patterns that define how to change to each initial value X_r; and the pitch corresponding to the input phoneme data;
A parameter interpolation processing unit performs energy and cross-sectional area interpolation processing, and the energy pattern of the sound source wave corresponding to the accent of each word is classified into multiple types, and the corresponding energy pattern type is attached to each word. With,
A dictionary section that registers each word by assigning a devoicing code to the vowel to be devoiced in the word, and defines the energy pattern by a function for each type, and utters it with reference to the dictionary section. a pattern processing unit that selects an energy pattern corresponding to the type of word to be used, and assigns a value lower than the function value of the energy pattern to a vowel with a devoicing code; and a parameter interpolation processing unit that performs interpolation. The current value output from the output terminal of the circuit element group is calculated based on the processed parameters, and the calculated value of the energy is applied to the interpolated value group according to the pattern selected by the pattern processing section. The parameter interpolation processing section includes a calculation section that uses a value obtained by superimposing a corresponding function, and a voice generation section that generates a sound based on the calculation result of the calculation section, and the parameter interpolation processing section calculates the initial value during each time period. X_o
A speech synthesis device characterized in that interpolation calculations are performed multiple times using the X_r corresponding to the target value and the constant, and the energy is performed based on an exponential function for interpolation.