JPS5816297A - Voice synthesizing system - 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 relates to a speech synthesis method, particularly linear predictive coding (hereinafter referred to as LP).
The present invention relates to a speech synthesis method based on the method (abbreviated as C).

Normal LPC speech synthesis methods use simple pulses as the sound source signal for synthesis, but although it is possible to significantly compress information, the details of the sound source information are lost and the quality of the derived synthesized speech deteriorates. It's inconvenient.

Therefore, in order to solve this problem, a plurality of sound source signals for synthesis are used, one of them is selected for each frame, and LPC speech synthesis is performed in that frame.
A speech synthesis method as shown in the figure was previously proposed by the present inventor.

That is, FIG. 1 schematically shows its configuration. In the figure, (1) is an input terminal to which the original sound is applied, and (2) is an input terminal to which the original sound is applied.
) extracts the waveform part to be analyzed from the input original speech, and extracts characteristic parameters related to vocal tract transfer characteristics such as LPC parameters, voiced/unvoiced, etc. from the extracted waveform.
A speech analyzer extracts feature parameters related to the sound source such as pitch frequency and imaging width, (3) is a transmission path, and (4) is a speech synthesizer that resynthesizes speech from the feature parameters.

In this method, as mentioned above, transmitting the prediction residual signal itself to the speech synthesizer (4) requires a huge amount of information, so instead, the prediction residual signal in each frame is frequency-analyzed. 2. Information regarding the spectral envelope obtained by smoothing the fine components is transmitted to the speech synthesizer (4) side.

This operation will be explained according to the flowchart in FIG.

At step 0υ, for example, the original voice as shown in Figure 3 (Figure 3 represents the time waveform of the female sound "a") is applied, LPC analysis is performed at step O2, and step (t31Kr,
The pc parameter, that is, the above-mentioned linear prediction coefficient ak is extracted. In general, the parameters required for speech synthesis are voiced/unvoiced, pitch frequency, and amplitude for the sound source, and the vocal tract transfer characteristics (spectral envelope) differ depending on the method, but in the case of this LPC speech synthesis method, the above LPC parameters are handle.

In addition, step Q4) K predicted residual signal is obtained by LPC analysis, and the amplitude, which is one of the sound source parameters, is extracted from the predicted residual signal using Stella 1a. Get a signal. FIG. 4 shows a prediction residual signal corresponding to the original speech shown in FIG. Then, a pitch analysis is performed in step 06) using this prediction residual signal, and a pitch frequency (pitch period) is obtained in step (17). This pitch period is the fundamental period of sound vibration in the vocal cord sound source, and is an important parameter characterizing voiced sounds. The cycle (frame cycle) for determining each of these parameters
Usually, the time length of the waveform cutting window is about 15 to 3Q m8ec.

And these steps Ql) to αη are the speech analyzer (2)
It is done on the side and is customary.

Next, in step a, the prediction residual signal as shown in FIG. 4 is subjected to Fourier transformation using, for example, 256 sample points.
Performs time-frequency conversion. As a result, a frequency spectrum of the prediction residual signal as shown by waveform S in FIG. 5 is obtained. This frequency spectrum is substantially represented by a power spectrum with phase information removed. Then, this frequency spectrum is subjected to spectral smoothing by, for example, the cepstral method in step 11, to obtain a spectral envelope as shown by the envelope E in FIG. 5. In FIG. 5, approximately one frame is represented.

This spectral smoothing is performed for each frame to obtain a plurality of spectral envelopes in step (a). Since the obtained multiple spectrum envelopes differ from frame to frame,
Transmitting them frame by frame still requires a considerable amount of information. Therefore, in step (2I), the spectral envelopes are separated by the distance between them, that is, those with similar spectral envelope shapes are grouped into one group VC,
The spectral envelope of one of them is used as a representative pattern. Next, similar operations are performed on the remaining spectral envelope groups to sequentially extract representative patterns. By repeating this operation, it is possible to obtain spectrum envelopes 1 to +16 representing a plurality of representative patterns, for example 16, in step 0, and this K can represent the predicted residual signal information of all frames. . Also, these 16 spectrum envelopes 1 to +16 are 2=1
6, so there are 4 bits per frame, for example, [0001] for spectrum envelope +1 and ( for spectrum envelope +2).
If a bit code for each spectral envelope end such as 0010:1 is assigned in advance, any corresponding spectral envelope can be selected according to the corresponding bit information when desired. ' Next, in step (2), perform inverse Fourier transform under appropriate phase conditions to obtain time waveforms as shown in Figure 6 corresponding to each spectral envelope +1 to +16, that is, impulse ≠ 1 to
+16 step c! 4) Obtain. Phase information is required here; in other words, the phase information has been removed by the Fourier transform in step 08) above, so it is necessary to provide some kind of phase information during this inverse Fourier transform. However, since the sound quality of the synthesized speech is not expected to be significantly affected by the waveform of the sound source signal, this step (
The phase information during the inverse Fourier transform in the two countries may be provided in a manner convenient for subsequent signal processing. Step C! The waveform of impulse ÷ 1 to +16 obtained in (a) is, for example, shown in Fig. 7 A and B.
, CK. This figure 7A
, B, and C, when the duration of the impulse is constant, the waveform shown in FIG. 7C has the lowest signal level at the final end of the impulse waveform. Therefore, when an impulse waveform as shown in FIG. 7C is used, it is considered that the sound source waveform connection error will be minimized during LPC speech synthesis in step (25). Therefore, when performing inverse Fourier transform in step @, it is best to provide phase conditions so that the waveforms of impulses -#-1 to +16 that cannot be obtained in step (goods) become substantially the waveforms shown in Figure 7C. preferable. The impulse shown in FIG. 7C is an impulse response of a minimum phase shift system, and the phase conditions for obtaining it can be easily shown.

In this way, the spectral envelope +1 to ÷16, which is a representative pattern of prediction residual signal information, is subjected to inverse Fourier transform in steps and converted into a corresponding time waveform, which is then converted into a corresponding time waveform by a speech synthesizer (
4) It will be used as the sound source signal (Fig. 1). Note that this Fourier inverse transform requires a huge amount of hardware to perform on the speech synthesizer (4) side, so if real-time operation is not required, it can be performed using software on the speech analyzer (1) (Figure 1) side. It is preferable to perform an inverse Fourier transform to convert the spectral envelope into a time waveform, and then transmit it to the speech synthesizer (1) side.

Then, in the step (25) of LPC speech synthesis performed on the speech synthesizer (4) side, approximately 4 bits of information are allocated to each frame, and as described above, the 16 bits at step 0 are
One of the impulses≠≠1 to +16 is selected and used as the sound source for synthesis of that frame. In other words, the selected impulse substantially corresponds to the information regarding voiced sound among the sound source information necessary for speech synthesis of that frame.

Also, in the step Q mark, the step (+51
The amplitude information of step αD and the pitch information of step αD are added, and the LPC parameter of step (1 wrinkle) is added as information regarding the transfer characteristics of the vocal tract, and as a result, step (
26) K synthesized speech is extracted.

With such a speech synthesis method, the frequency spectrum of the synthesized speech in each frame becomes more similar to that of the original speech, and the quality of the synthesized speech is improved.

By the way, in the case of the above-mentioned speech synthesis method, one type of sound source waveform arranged for each pitch cycle is used as the sound source waveform for synthesis in one frame, so the waveform and signal level of the synthesized speech are determined at the connection part of each frame. tends to become discontinuous,
There is an inconvenience that the sound quality of the synthesized speech is not smooth.

In view of the above, the present invention provides a speech synthesis method that can reduce discontinuities between frames of synthesized speech as described above and make the sound quality smooth.

In the present invention, when two voiced frames (or two unvoiced frames) are consecutive, by interpolating the corresponding sample values of the sound source signal waveform in each frame between those frames, the sound source is Make the signal waveform change smoothly and little by little.

An embodiment of the present invention will be described in detail below with reference to FIGS. 8 to 10.

FIG. 8 shows the configuration of this embodiment, and in the figure,
C31) is a clock generator, 0 is an address counter, (3→ is a frame counter, and 04) is an interpolation counter. Different types of timing signals are generated.

Reference numeral 051 denotes a sound source signal waveform memory, and as explained in connection with FIG. Inverse Fourier transform is performed to convert it into a time waveform (impulse response), which is stored in advance as sound source signal waveform data for each frame in order to be used as a sound source for LPC speech synthesis. From among the sound source signal waveform data stored in the memory c3, one frame of sound source signal waveform is selected by the output of the address counter.

C(6) is a buffer memory that temporarily stores the sound source signal waveform of the frame one frame before the current frame, and its contents are updated every frame by the output of the frame counter (to). Ru. Both (37) and (38) are coefficient units, and coefficient unit 07) is a memo! JC35
It works to add coefficients as described below to the sound source signal waveform of the current frame output from 1, It works to add coefficients different from those mentioned above to the sound source signal waveform of the frame.The coefficients added by these coefficient multipliers 07) and (to) C37
) and (to) are updated each time they are supplied. Further, the frequency of the interpolation clock varies depending on how many equal parts one frame is divided into for interpolation. For example, when one frame is divided into four equal parts and interpolation is performed, the frequency is set to four times the frame frequency.

0! is an adder for adding the respective outputs of the coefficient unit 037) and □□□, and (40) is an output terminal from which the interpolated sound source signal is taken out.

Next, the operation of this embodiment will be explained. Now Memo IJ G5!
Among the sound source signal waveforms stored in and corresponding to each frame, for example, the sound source signal waveform at frame -#-n is en (
m), and the sound source signal waveform at the subsequent frame +n+1 is assumed to be en+1(m). m is the number of sample points of the sound source signal waveform stored in the memory c351, m = 1°2
, ... is represented by y. For example, at a sampling frequency of 10kHz (sampling period of 100μs), t=
30, the length of the sound source signal waveform is Q, 1m8
30=3ms.

As shown in FIG. 9, consider the case where one frame section is divided into a plurality of parts, for example, divided into four equal parts. 3.4
and number it.

And the sound source signal waveform enJ(
m) is determined by linear interpolation as follows.

In the above equation (1), the number of divisions J is J = 1.2, 3.4, and the number of sample points m is m = l, 2,
..., Q.

From the above equation (1), the sound source signal waveform e in the divided section of J=1
nl(m) is enl(m)=en(m)...
...(2), and it can be seen that it matches the sound source signal waveform before interpolation at frame number n.

Further, the sound source signal waveform en2(m) in the divided section of J=2 is based on the above (1). FIG. 10 shows a case in which numerical values are actually entered and interpolation is performed for J=2.

That is, from the above equation (3), en(
m).

If the levels of en+ 1('') are respectively 1.0 and 0.9, the level of e112(m) after interpolation is 0.975, and similarly en(m) and en+t when Km=2.
If each level of (m) is -0, 8, -0, and 7, the level of enz(m) after interpolation is -0,775, and m
= en(m) when 3.

If each level of en+x(m) is 0.5 and 0.7, the level of en2(m) after interpolation will be 0.55, and m
When = 4 (7) en(m), en+t(m)
If the respective levels of are -〇, 2゜-〇, and 3 respectively, the level of en2(m) after interpolation becomes -0,225, and as a result, the sound source signal waveform en(m) shown by the solid line in Fig. 10A is obtained. The sound source signal waveform en+s ('') shown by the solid line in Figure 10B
As a result, an interpolated sound source signal waveform en 2 (m) as shown by the broken line in FIG. 10A is obtained.

By performing interpolation in the same manner, as the VC progresses from 2 to 3 to 4, the sound source signal waveform en,1(m) gradually becomes the sound source signal waveform before interpolation en+ at the next frame +n+1.
It approaches x(”).

, (I9 The above is a case where one frame section is divided into four equal parts, but in general, when one frame section is divided into four equal parts, the sound source signal waveform en,1(m) in each divided section is The following formula % Formula % In other words, the sound source signal interpolated by this formula (4) is a linear combination of the corresponding sample values of the sound source signal waveform in these two consecutive frames, so the coefficient multiplier C37)
, (to) and the adder 0.

In addition, in the above equation (4), J=l, J...km=
12,...e. The number of divisions is 2, 4.
It is convenient to select a power of 2, such as 8, because the interpolation calculation of the above equation (4) can be easily performed by bit shifting of binary data.

Such an interpolation operation is performed using the circuit shown in FIG. 8. First, the clock from the clock generator 01) is counted by the address counter 02,
Based on the address information, the waveform data of each corresponding frame in the memory 4351, for example, the frame number n sound source signal waveform en(m) is selected.Then, this frame 4n sound source signal waveform en(m) is The buffer memory (to) is reached by the output of counter 03).
is accumulated in

Subsequently, the sound source signal waveform en(m) of the next frame +n+1 of frame number n is similarly selected from within the memory G by the address information of the address counter C32, and is stored in the buffer memory (36) by the output of the frame counter O. be done. The sound source signal waveform en(m) of frame 4n is stored in the buffer memory 06) and supplied to the coefficient multiplier.In other words, the contents of the buffer memory (to) are stored in the buffer memory 06) and the frame 4n is supplied to the coefficient multiplier. Updated every time.

Also, at the time when the sound source signal waveform en +1('') of frame +n+1 is supplied from the memory (35) to the buffer memory (g), it is also supplied to the coefficient unit (37).Then, the output of the interpolation clock counter 04) is Vessel G′?) and (
(to), the respective coefficients are added by these coefficient units. That is, from the above equation (4), between the coefficient units -J+1, there is a coefficient for the sound source signal waveform en(m).

k-J+1 is added, and the signal en(m)・-Y- is taken out on the output side, while the coefficient unit 07) receives the sound source signal −
1 A negative signal is taken out at the output side of en+1(m). Then, these extracted signals are supplied to the adder 0 and added, and the output terminal (4 (IIK is the interpolated sound source signal waveform en, 1 (m) as expressed by the above equation (4) is output.

As described above, according to the present invention, when two voiced frames (or two unvoiced frames) are consecutive, interpolation is performed between the frames to the corresponding sample value of the sound source signal waveform in each frame, Since the sound source signal waveform changes smoothly and little by little between each frame, discontinuities in the synthesized speech waveform and signal level at the connections between each frame are removed, resulting in smooth sound quality and excellent quality. You can get synthesized speech.

[Brief explanation of drawings]

FIG. 1 is a block diagram schematically showing an example of the prior art of the present invention, FIGS. 2 to 7 are line diagrams for providing the operating station 1 of FIG. 1, and FIG. A configuration diagram showing one embodiment, FIGS. 9 and 10 are diagrams for explaining the operation of FIG. 8. Gel is a clock generator, C321 is an address counter, 03) is a frame counter, (goods) is an interpolation clock counter, C35i is a sound source signal waveform memory, (to) is a buffer memory, (37), (to) is a coefficient unit, 0kon is an adder. Figure 1 Figure 7

Claims (1)

[Claims] In a speech synthesis method that changes the sound source signal waveform for each frame, interpolation is performed between consecutive frames to the corresponding sample value of the sound source signal waveform in each frame, so that the sound source signal waveform changes smoothly between frames. A speech synthesis method characterized by the following.