JPS58147798A - Voice synthesizer - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention has a complex sine tI1. ℃ da k (Composite
This relates to a speech synthesis device using the e5inu-soHal Model (hereinafter abbreviated as C8M), and its purpose is to be able to synthesize scale tones with no pitch deviation, and to have a capacity as large as data. To provide a speech synthesizer that can reduce the noise and also produce a scale tone 1-* with a reduction of 1 m.Generally speaking, speech synthesis method Fi is a waveform that encodes the speech waveform of an aroused voice. Encoding method and parameter synthesis method K that utilizes audio waveforms to parameterize Oobetsu Te 14,
The former Ki! PCM, DPCM, DM, AP
CM.

There are various methods such as ADPCM and ADM. The sum is autocorrelation 1: I-J, LPC, PARCOR, LSP
By the way, the waveform encoding method simplifies the synthesis circuit, but since the degree of compression of the dark is small, it is necessary to increase the capacity due to the degree of compression.
Further, since the parameter synthesis method #i has a high degree of data compression, it is possible to reduce the data capacity, but if the synthesis circuit becomes complicated, it will increase the amount of data. Nowadays, it has become possible to integrate complex synthesis circuits into LSIs, so the parameter method has become mainstream, and among the parameter methods, the PARCOR method is the most widely put into practical use. This PARCOR method is a method that separates the human vocalization process into vocal fold vibration and vocal tract transfer characteristics, extracts feature parameters, and synthesizes speech based on these feature parameters. When extracting, the section of the audio that can be considered to be almost in the footsteps state is taken as one frame, and the audio signal flag ~ 12 of each frame is
Sampled with a KHz sampling pulse, extract the amplitude parameter representing the amplitude of the vibration, the pitch parameter representing the fundamental period, and the %K parameter consisting of the parameter representing the spectral distribution.
1-Frequency is used as an output for synthesis, and the frequency parameter is used to calculate the audio waveform. However, in this PARCOR method, the fundamental frequency of the audio to be synthesized is determined by the frequency of the sampling pulse. For example, when sampling is performed with an IOK of 7, the fundamental frequencies of the voices that can be synthesized are as shown in the table below.

Normally, the fundamental frequency of the human voice (six words) is 80 to 4.
00 Hz, and the fundamental frequency changes from moment to moment, so even if the synthesized voice is slightly out of tune with the actual voice, it is impossible to distinguish it by ear.

However, in the case of scale notes that change relatively regularly, the shift in tone phase is clearly perceived (for example, the scale note "
'U' is 880Hz, but the pitch parameter is '11'
There is a problem in that when synthesized as wL was extremely difficult mechanically. Therefore, the inventors investigated the C5M method proposed by the Musashino Masanobu Research Institute of Telecommunications Corporation, and created a 90 complex sine wave ℃ del (C5M). is a time-discrete signal (time series) (yt) as °C del, yt-2ms aim
(wet+φt)+−−−・・・+2m theory mwmt+φ
-)・tllt is something that can be crushed, where t is an integer and is expressed in the tense t-, '2ml.

wi, φ1 (i-12, . . . n) represent the amplitude, angular frequency, and phase t- of the i-th positive polar wave component, respectively, and mi, fn wealth, . . . 9n turtle > 0.0
＜w s＜w曾〈・・・・・・Wa＜π ・・・(2
) Now, the audio signal of the original audio is obviously an unsteady signal, but if it is divided into K short sections (hereinafter referred to as frames) of 10m-~g Oms 81 degrees, the signal will be divided into K in each frame. It can be regarded as a fairly stationary signal. Therefore, by finding ml, V, and tO values such that the audio signal in such a frame and the C5M signal as shown in equation (1) are in phase with each other up to the 0-2n-1 order in the 11th region, ,
The speech % mold parameters 1i (ai='2m1) and wi are extracted (the autocorrelation coefficient in the formula is independent of φ1). (Acoustical Society of Japan 1976 Fall Presentation Proceedings P, 55
7-558 1979, [Simple audio synthesis IC & J Sagayama, & Kura) using composite positive mosquito waves] sP and document 2 (Acoustical Society of Japan audio study group material 579-19. October 1979,
``Speech Synthesis Using Composite Sine Waves'' by Sagayama and Itakura [K]K. In addition, in the case of voiced sounds, KFi
, since there is a fundamental period due to vocal cord vibration, this fundamental period is also extracted as a feature as 5 meters T 〇 Below, nine feature parameters are extracted as described above @I ,,I I
Tt-Explain the speech synthesis method used 0 Now, C
The f! symbol for speech synthesis in the 3M system. As del, yt-JlxmWtt+ ...@alJllW with the phase φ1 in equation (1) above as zero
Bj...151m1 +w+'2mi e i
= 1 * 2, ......* is used.

First, to explain the case of synthesizing voiced sound frames, we now have a voiced sound with vocal cord vibration #J! (“a”, C””
There is a fundamental period in the frame of ``m'', etc.).Therefore, it is necessary to assign the % parameter ml to the wit131 formula and give the fundamental period to the signal that is calculated and synthesized based on the characteristic parameter TK, and the characteristic parameter T The time tl is reset at the nine-pitch station period tp based on .The signal waveform formed by K in this way is as shown in the IR1 diagram, and the power spectrum has a harmonic structure of the fundamental frequency (1/lp). By the way, the combination F obtained in this way
When the ILf voice is reset at time t (t = OL voice waveform), it becomes discontinuous9, and the quality of the synthesized single voice deteriorates due to the occurrence of this discontinuous part. In order to alleviate the waveform discontinuity, a pitch synchronization envelope using an exponential decay function cicada as shown in Fig. 2 is given.That is, the %descriptive parameters @l, wi,
Speech waveform synthesized by TLC41/1 (single-dot chain 41) t
- Pitch synchronization envelope curve (dotted line) K & 9 waveform (solid line) K conversion is performed to alleviate waveform discontinuity. In addition, general Kf voice is composed of multiple frames, and parameters @i, wi, T f
Update and audio combination F! It is a trap to do L, but %
When the characteristic parameters aI, wI, T are updated, if the characteristic parameters ai, wI, 'l'' change suddenly, distortion will occur at the connection point of the frame.For this reason, one of the characteristic parameters 1', w+, T is gradually changed. Parameter interpolation is performed to vary the parameters.

Next, to explain the case of synthesizing 1 voiceless frame and silent frame, the fundamental period tp does not exist in the frame of voiceless voice without vocal fold vibration ("s"IIP"r-b"e etc). I want to talk and it's time t
is set to tMonO at the beginning of the frame, and the pressure is monotonically increased thereafter. - 10,000, in silent frame trap always Ft as t”o
- All you have to do is OK.The voiced frame, unvoiced frame, and silent frame are each synthesized as shown above.&?
, each of the parameters Hatakei, Wi&, and T used for synthesis is quantized to reduce the amount of memory required to store data for speech synthesis. The quantization parameter of T of the meter is 6 pits, wi
The quantization parameter F14-5 pits, aiij
It is quantized as the product (ai"AXai) of 5 meters ml', which indicates the relative amplitude, and A, which indicates the absolute amplitude, and each quantum parameter is 4 to 5 pits. Therefore, normal voice In addition, when extracting the feature parameter II IWI eT, PARCOR
As with the method, the audio signal is set to an appropriate frequency (for example, 1
It performs arithmetic processing using a sampling pulse (OKHz) as a supplicator, and each % parameter a1. Octopus, Tij is extracted as a digital value, and the frequency of the sashi printer pulse is 8K) When IZ, the order is n 4.10KH
When z, n-5 is appropriate.
I) is a voice synthesis Flt device, and (1) is a digital filter (Fi parameter code section, +31FiD/A).
):/Bark, (4) is an analog 0 demultiplexer,
ll#i Voltage controlled sine wave oscillation with phase reset (P
hasresettable VOltage Co
ntro-11ed sin Qsilator),
(・) and 171 ri voltage-controlled amplifier (Volt
age Controlled Amp +), t!
IIFi adder, 1i11 is voltage controlled fiem pulse oscillator (Voltage Controlled Pu1
se Generator ), (lla analog 0 demultiplexer (AM) output (D/A converted time series signal of %Kin parameter ml IWI , T)) A bias control section that performs frame interpolation based on K and generates bias voltages for each part. 16) gives a bias voltage Vwi k based on the characteristic parameter W1, %l @) Ic gives a bias voltage ya i l t- based on the characteristic parameter 1', 7) K gives a % characteristic bar 5 Bias voltage V based on meter AK
A is given, and (-)K is designed to give a nine bias voltage Vt based on the amnesty parameter TK.

Now, the quantization parameters of the feature parameters mI, wi*T sent from the chitame (M) to the speech synthesis section (S) are converted into the parameters wLK necessary for synthesis by the parameters The respective parameter values (digital') are determined by the tD/A converter (3) and the analog oj diful multiplexer (4).
Ana □j ) K conversion, divided b This parameter value is interpolated by the interpolation voltage set in advance in the bias control unit, and the bias voltage (Vw 1) corresponding to the interpolated parameter value is Val)' (V
A)(VT) is generated. Each sine wave oscillator +il
generates a sine wave with an angular frequency that is thrown at the bias voltage (V, 1) [, and in the case of a sound frame, it is reset by a reset pulse output from a pulse oscillator (VCPG) every pitch period tp. 〇Each sine wave oscillation l5I
I) The outputs are each amplified by the amplifier 11 whose amplifier is set by the bias voltage (Val), and added by l511.
The output of this adder +81 is further added to the amplification f! where an amplifier is set using a bias voltage (Va). i
The signal is amplified and a synthesized speech signal (V, ) is output.

Figure 4 shows an example of the configuration of a digital speech synthesis device. tps is an interpolation circuit, (l
T-latch that latches the amnesty parameter T, (14
a Characteristic parameters wit 5 t, 01 t
i 4I So parameter Otori Latch i'・A latched to 1 ~ When α, amnesty parameter wi & al, AK Based on AK, sine wave calculation circuit that generates a sine wave with rounded corner frequency and amplitude, 0'/) is an adder that adds the outputs of each sine wave calculation (b) path, ■ is a multiplication 1', 1lFiDZA cosciverter, - is a 0-bus filter, and a 0-ts pulse for synthesis that is approximately equal to the control Fi supple:, dibars (10KH2). Pitch counter that counts t and outputs sine wave calculation circuit (l- calculation side 89), fourth is T latch field data and pitch force output @υ detects coincidence and resets pitch counter (2) is a nine-pitch synchronization envelope coefficient table that stores the envelope coefficients that give the pitch synchronization envelope #I!, @ is a voiced/unvoiced switching circuit ◎ Below, the speech synthesis section (S) The operation of the other parts is approximately the same as that of the above-mentioned analog voice synthesizer, so the explanation will be omitted. When the parameter T is O, a signal to switch the voiced/unvoiced switching circuit to the unvoiced sound IIK is output from the voiced/unvoiced judgment screen wI, and the count is sent to the multiplication FM sea bream, and the voiced/unvoiced judgment screen is output. A signal to reset the pitch counter υ at the beginning of the frame is output from the rhinoceros circuit 11,
An unvoiced sound is synthesized. However, if the previous frame is a voiced sound, the beginning of the KFi frame will be detected in the middle of one pitch period tp, but in the case of KFi, the pitch counter will be reset until the end of that pitch period tp. Ru. On the other hand, when synthesizing voiced sounds, the decoded feature parameters'
rFi'r5T111IK is latched, and when the latched data of the T latch DI and the output of the pitch counter match, a reset signal of the pitch counter @υ is output from between the -number output circuits.

Since the pitch counter @υ is reset with the pitch period jp in this way, the pitch period tp is given to the signal generated by each sine wave arithmetic unit (l-).

At this time, the voiced/unvoiced switching circuit lights up as voiced fllKll and I? Then, the damping coefficient values that provide pitch synchronization meridians are sequentially read out from the envelope coefficient table using the output direction of the pitch counter as an address, and the multiplier at the time of multiplication -0 is sequentially set based on the damping coefficient values. Output FiD/A when using multiplier
It is converted into an analog 0-signal by a converter, and then passed through a 0-bus filter to obtain a synthesized audio signal (V,).

By the way, if we now consider the case where scale tones are synthesized using a C5M type speech synthesizer, a single chord has one fundamental period, ν9, and an m-duplex has a fundamental period of Fim. K, all of which are voiced sounds. Therefore, such a scale f is treated as a voiced sound. Similar to the problems encountered in this technique, the fundamental frequency of synthesized speech takes on discrete values, resulting in pitch deviations, which is a problem, and it is difficult to assign multiple fundamental periods, making it impossible to synthesize chords. On the other hand, unlike the human voice, scale sounds have regularity, and in general, the scale t of l is the characteristic parameter a! , -1T are approximately the same.
It is often composed of multiple frames. Therefore, the ninth IK that reduces the capacity of the data meter (1)
- is the characteristic parameter ii of each frame of the scale note, W
t eT t- Since all of them are stored in the Etame ℃ tll, there is no need to change the feature parameters, which are almost the same as the previous frame.
ew1. A repeat code is stored as the audio dark of the frame from which T is extracted, and when the repeat code is read out from the frame (1) and the characteristic parameters a1, wI, and T of the previous frame are used, the audio of the frame is reproduced. However, when using K in this way, the synthesized speech has the same amplitude over several frames, resulting in what is called a weak synthesized speech. It's so easy to play the D special piano! There was a problem with φ when synthesizing scale tones. The present invention has been made in view of the above points.

Hereinafter, an embodiment will be explained with reference to the drawings. Figure 5 shows an embodiment of the present invention, and the general structure and operation of the speech synthesis section (S) are approximately the same as the conventional example in Figure 4 described above. Therefore, the differences from the conventional example will be explained in detail below. In the present invention, if scale tones are synthesized as voiced sounds, pitch deviations occur and chords cannot be synthesized. , this audio dark is summed so that it can be stored in the data frame (1), and in order to reduce the memory capacity of the data frame (1), the characteristic parameters, l, which are almost the same as those of the previous frame, are added. The repeat code (even 1 pit) is stored in the data memory as the audio dark of the extracted frame (1), and when the repeat code is read out from the data memory (1), the % mold parameter of the previous frame is A repeat means (X) is provided for synthesizing the voice over the frame using ilewi, and further a scale envelope imparting function is provided for imparting a scale envelope consisting of an attenuation curve to the audio signal over a plurality of frames synthesized by the repeat means. Hereinafter, the operation will be explained based on a specific configuration. ◎- detects a synthesized sound switching code indicating whether to synthesize a human voice or a scale tone. gate circuit cmtl
It is a code discrimination circuit that performs tk when a code for synthesizing a human voice is input.

Code discrimination circuit - gate circuit controlled by output tmF
The voiced/unvoiced determination circuit DI) is controlled so that its output is directly input to the pitch counter 1)K, and when unvoiced sounds are synthesized, the pitch counter key υ is reset at the beginning of each frame as in the conventional example. Therefore, when synthesizing the unvoiced sound of a human voice, the operation is exactly the same as in the conventional example. On the other hand, when a chord consisting of the scale note tn- is input,
The gate circuit is a voiced/unvoiced switching circuit 111 for several frames Kl [corresponding to IN'c after the first frame of the scale note]
Only the reset signal of the pitch counter output from l! of each subsequent frame is passed through. The reset signal t corresponding to lK is cut off. In this case, the first frame of the scale note can be detected by inserting a silent frame at the change point of the scale note and detecting the first frame based on the silent frame, or by detecting all 7L/- frames of the Fi scale. II # Wi parameter Feng i, Wi same & constant (repeat hand * (X
)) and set the feature parameter Otori i,
A possible method is to detect a change in wl in the first frame t. As described above, when synthesizing a pitch tone as an unvoiced tone, if we trap the pitch counter Il to be reset only at the tip of the first frame of a scale tone that has multiple frames KTh, as shown in Fig. 4, the conventional example The tIL-shaped discontinuity at the frame connection point, which occurs when synthesized as an unvoiced sound, does not occur, and distortion of the synthesized scale tone can be reduced.

- is a repeat code detection unit that constitutes the repeat hand K becomes S?
The audio of the frame in which the repeat code is read as audio data is the characteristic parameter 25 of the previous frame, wif
It will be synthesized using In addition, when applying the repeat means (X) when synthesizing human voices, it is sufficient to also hold the characteristic parameter T in the parameter decoding section (2) when a repeat code is detected. This is a scale envelope coefficient table constituting a scale envelope imparting means (Y) that imparts a scale envelope consisting of an attenuation curve as shown in Figure @6 to the signal output from the instrument Fm) ... The scale envelope coefficient that gives a scale envelope consisting of an attenuation curve over the entire range of the scale note consisting of a river is stored.0 From this fWI envelope coefficient table (c), pitch h ushita @υ Synthesized speech signal (V6) from K by setting the multiplier α4 multiplier a based on the scale envelope coefficient read out using the output as an address.
In this case, the pitch h counter @ u
#i The number of hits is 16 pits, making it possible to synthesize long scale notes spanning multiple frames. In other words, the pitch counter! 1 does not enter the reset state 114 until it counts 65,536 0-tsku pulses for synthesis. For example, if the 0-tsku pulse for synthesis is 10 KHz,
It is possible to apply the above-mentioned scale envelope to a long LnliF#fK of approximately 6.5 seconds, and also to eliminate waveform discontinuity at the frame connection point. After the pre-set time, the pitch counter (!υ ratio output exceeds a certain value), the scale envelope coefficient is continuously output from the same address in the w# envelope coefficient table area. The voiced/unvoiced switching circuit that sets the multiplier is also controlled by the code discrimination circuit output. In the case of KFi, as in the conventional example, the envelope coefficient table for pitch synchronization is set for voiced sounds.
For unvoiced sounds, set the constant as a multiplier (1 multiplier), and when synthesizing a scale sound as an unvoiced sound, set the data read from KVi and the pitch counter body υ output from the scale envelope coefficient table (to). The scale envelope coefficients sequentially read out are set as multipliers of the multipliers.

FIG. 7 shows another embodiment, each structure 1 of the embodiment in FIG. 5! ! This is achieved analogously, and the code discrimination in the conventional example shown in Fig. 3 and the code discrimination in the embodiment shown in Fig. 5 - 1 (up to)
and a repeat means (X), and a gain control circuit t7 for controlling the amplifier of power saving 7) as an envelope imparting means for the VT floor. An envelope coefficient K is generated in a travel time constant circuit using a resistor, a capacitor, etc., and the amplifier of the amplifier 171 is controlled based on this scale envelope coefficient and characteristic parameter AK. The synthesized speech signal (Vs)l- of scale tones over multiple frames outputted from )) becomes a signal with a scale envelope added, and a scale tone with reduced tension is synthesized.) The invention is configured as described above, and in a C5M type speech synthesis device, feature parameters extracted by regarding scale tones as non-F'll tones are transferred to a data file as voice data.
Since the system is designed to synthesize scale tones using the audio data read out from the data media, it is possible to synthesize scale tones without any pitch deviation, and also to synthesize chords with multiple period periods. In addition, the repeat code can be easily synthesized as frame 6 audio data, which has almost the same feature parameters as the previous frame.
When the repeat code is read out from the data meter, the repeating means for synthesizing audio based on the % & J ~ parameter of the previous frame is provided, so it is almost the same.
It is possible to reduce the amount of audio data for scale tones made up of multiple frames having the characteristic parameters of Since the signal is provided with a scale envelope imparting means consisting of an attenuation curve, there is no problem! It has the advantage of being able to synthesize scale tones of 11g1kv. In addition, Figure 5 1! In this example, when a scale note is synthesized as an unvoiced sound, the pip pre-counter is reset at the beginning of the first frame of the scale note, and is not reset at the beginning of subsequent frames. Therefore, the distortion of the waveform in the connection of each frame, which is seen in the conventional example shown in Fig. 4, does not occur, and the scale tone consisting of multiple frames can be synthesized without distortion.

[Brief explanation of the drawing]

1 and 2 are operation explanatory diagrams of a C5M type speech synthesizer, FIGS. 3 and 4 are block circuit diagrams of conventional examples, respectively, and FIG. 5 is a Boolock circuit diagram of an embodiment of the present invention, @ FIG. 6 is a -50 block circuit diagram of another embodiment. +1) ii5'-1j151J, C3) ti-speech production part, (X) #i milli-peat stage, (Y) Fi-speech scale envelope imparting means Patent attorney Ishi 1) Chief Figure 1 γ2 y1 Procedure Amendment (voluntary) 1. Indication of the case 16 Wa 57 years old KF order No. 30733 2, name of the invention 3, person making the amendment Relationship with the case Patent applicant Location 1048 Kadoma, Kadoma City, Osaka Prefecture Name (583) ) Matsushita Electric Works Co., Ltd. Representative Iku Kobayashi 4, Agent postal code 530 Telephone Osaka (06) 345 = 7777 (Representative) Procedures
Amendment (No. 9) 1. Indication of the case 1982 Special Examination No. 80788 2. Name of the invention Speech synthesis device 3. Relationship with the person making the amendment Address of the special liver applicant Kadoma City, Osaka Prefecture 1048 Oaza Kadoma Name (583) Matsushita Electric Works Co., Ltd. Representative Iku Kobayashi 4, Agent Zip code 530 Address 1-12-17 Umeda, Kita-ku, Osaka (Ume 141)
Hill 5th floor) Name (6176) Benri Doshi [11th Chief
Telephone Osaka (06) 345 7777 (Representative) 5,
Voluntary correction of the date of the amendment order.Application number: 1988-80788 No. 1, 6th line, 5th line of the dedication document, corrected to r90m8J1kF80m8J. 2. Figures 4 and 5 of the attached drawings of this application are corrected as shown in the attached sheet. Agent Patent Attorney Ishi 1) Choshichi

Claims (1)

[Claims] [Divide the 11-boo voice into multiple frames as an interval kl frame that can be almost in a steady state, convert the audio signal in each frame into a delta as a composite signal of multiple sine waves, and calculate the amplitude and angle of each sine wave. The quantized feature parameters corresponding to the frequency and the fundamental period of the threatening voice signal are extracted as appropriate and stored in the database as voice synthesis data, and the voice synthesis f! B
Then, the amplitude and angular frequency tU of multiple sine waves are added based on the amplitude and angular frequency tU read out from the data meter. 1 for the characteristic parameter corresponding to the fundamental period! In a speech synthesizer that synthesizes a speech signal by adding a pitch period to the data, the feature parameters corresponding to the amplitude and angular frequency extracted from the data file are considered to be unvoiced sounds. At the same time, the characteristic parameters that are almost the same as those of the first frame are extracted, and the repeat code is stored as the audio data of the good frame. repeating means for synthesizing the audio signals of the frame using the repeating means, and a scale envelope imparting means fr for imparting a scale tone envelope consisting of an attenuation curve to the voice signal of a plurality of frames synthesized by the repeating step #ItK. Audio stand F1. Device 9