JPS6031198A - Encoding of forecast residual signal - 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 Field of Industrial Application The present invention relates to a predictive residual signal encoding method used in a residual drive type speech synthesis device used for automatic guidance broadcasting equipment and voice communication such as telephones. Example configuration and its problems FIG. 1 shows the configuration of a conventional residual-driven speech synthesis method.

The configuration of this conventional example will be explained below. In FIG. 1, 1 is an input terminal to which a voice input is applied, 2 is a prediction residual signal, 3 is a drive signal, and 4 is a synthesized voice output. The audio signal l is subjected to linear predictive analysis by the predictive analysis filter 10, and spectral parameters (eg, PAROOR coefficients, etc.) expressing the short-time (eg, 20 m5) spectrum of the audio signal are extracted. Further, the audio signal is passed through an inverse filter formed by the extracted parameters and converted into a prediction residual signal 2. If this prediction residual signal 2 is input as a drive signal to the synthesizer 22, the original voice will be reproduced, but there will be no effect from the viewpoint of information compression. In the conventional example, a prediction residual signal 2 is input to a 3-page low-pass filter (LPF') to obtain a residual baseband signal, which is resampled to 1 by a down sampler 12 and then resampled by an encoder 13. The data is encoded and sent to the transmission channel 14.

When the original audio signal is sampled at 10KHz, the LPF
This has the effect of slowing down the cutoff frequency of II from, for example, 600 Hz to 1. As the encoder 3, an adaptive POM encoding method, a pitch adaptive predictive encoding method, or the like is used.

When restoring the original audio signal from the residual baseband signal encoded as described above, first, the decoder 15
After restoring the residual baseband signal resampled with
The up-sampler 16 converts it to the original sampling frequency.

This is passed through a low-pass filter (LPF) 7 having the same characteristics as the low-pass filter 1 to reproduce the original residual baseband signal. In order to restore the original audio signal, a drive signal corresponding to the predicted residual signal 3 is required, but the reproduced residual baseband signal has a band (for example, JP-A-0RGO).
-31198 (2) below 600 Hz), it is necessary to generate and add high frequency components. The full-wave rectifier 18 performs full-wave rectification on the reproduced residual baseband signal to generate high-frequency components. This is input to a spectrum smoother 19 to obtain a flat spectrum.

The flattening of the spectrum originates from the fact that the prediction residual signal 3 has substantially flat spectral characteristics.

Bypass filter (HPF) 20 is LPFII but still L
This filter has the opposite characteristics to PF17 and removes spectral components that fall within the baseband. Next, an adder 21 adds the output of the HPF 20 and the output of the LPF 17 to obtain a drive signal 3. This is input to the synthesizer 22 to obtain a synthesized speech output.

As described above, in the conventional example, since nonlinear processing called full-wave rectification is performed, spectral distortion remains even after spectral flattening, and noise is mixed into the synthesized speech.

OBJECT OF THE INVENTION The present invention eliminates the drawbacks of the above-mentioned conventional example.
The present invention provides a prediction residual signal encoding method for synthesizing speech with less noise.

Structure of the Invention In order to achieve the above object, the present invention reduces redundancy due to the harmonic structure of the short-time spectrum of the prediction residual signal, and can synthesize high-quality speech with a small amount of information. It is effective.

DESCRIPTION OF EMBODIMENTS The configuration of an embodiment of the present invention will be described below with reference to the drawings. Figure 2 shows the short time period (
2oms) shows the Fourier spectrum amplitude component.

The voiced part has a harmonic structure in which similar peaks appear for each pitch period, and there is a lot of redundancy on the frequency axis. In the example shown in Figure 2, 512 points of fast Fourier transform are applied to 200 data frames of voiced parts from the prediction residual signal of a voice signal sampled at 1 OKHz, and the sample points on the frequency axis are is every 19.53Hz, totaling 256
Consists of points. On the other hand, since the pitch frequency fluctuates independently of the sampling points on the frequency axis mentioned above, it is necessary to resample each frequency point proportional to the pitch frequency in order to reduce the redundancy of the harmonic structure on page 6. 0 Therefore, as shown in FIG. 2, in the present invention, first, the pitch frequency f. f in the frequency band, starting from the mouth point of . After each partial spectrum is divided into partial spectra, the band of each partial spectrum is resampled to 1 (rL is a positive integer). FIG. 3 shows an example of a partial spectrum (amplitude component), where the (0) mark indicates the original sampled value, and the (•) mark indicates the value obtained by resampling the area within the band. When resampling, it is necessary to interpolate between the original sample values, and the present invention uses a method of interpolation using an interpolation function such as linear interpolation or a spline function.

Next, redundancy is reduced on the frequency axis using the spectral distance between adjacent bands of the partial spectra resampled as described above as a measure.

Now, the band number of the partial spectrum is i (i=Q, l.

2.3. ......), the resampled spectrum of the band is S, (no 0.1.2.-", n: 5io
= 5(z-+)yL), and on page 7, the spectral distance between the first band and the (?knee 1) band is defined by the following equation.

However, W is a weighting function that cannot be evaluated by weighting the vicinity of the peak of the partial spectrum. Further, αZ is a coefficient for correcting the level difference between both bands and focusing only on the spectral outline to determine the distance scale. αt is determined so that the following formula is minimized.

By partially differentiating both sides of the above equation with respect to αL, we obtain. 73
F,"i/aαi = O, then α2 is (3)
Determined by Eq. In the present invention, the spectral distance D
When i is less than a certain threshold, only αt is stored as the amplitude parameter, and Si is rejected. Also,
The restored value S' of Sz is determined from the following equation.

S's=αtS(i-+)ノ °Old 11° mark...°...
...(4) If S is rejected, calculate the spectral distance Di+ from Si for S(i+1), and use the same procedure to determine the restored value S'(z-+-+) do. As a result of the above processing, the partial spectrum Si
) or the amplitude parameter αk (ink) is saved.

Although only the amplitude component has been explained above, in the present invention,
Exactly the same processing is applied to the phase component. That is, the phase component θ or the phase parameter βk is stored for each band.

After reducing redundancy on the frequency axis as described above, θ, and βk are encoded in the stored S, α.

Next, a method for restoring the original speech from Si)', α, θ, and βk will be explained. Now, in the example of Figure 2 (
So, ), S+), α2. α3. α4. α5*Sn)
, α7. α8. α9+5IOJ°.

Sn), 512j, S, s), α14 r 8157
It is assumed that the amplitude part of the partial spectrum and the amplitude parameter are stored as in '+516)'y517J'). In this case, the amplitude of the rejected subspectrum is 1↑
'' component is determined by equation (4) as follows:
': ノ -11 2131 """l rL)S ノ - α3817', (//)S'Iノ - α+S
+7', (ri)j91. No - α5sa7', ('
/ ) another no - α5sa7', (Iku) S'8no - α786.7', (//)8%ノ =
α5sa7', ('') is restored as S'+4no α148I3.7', (// ). Therefore, the amplitude component of the entire spectrum is (
So)', Sl)', S's)', 8
'+), S'l), S's)', S', I).

S'8), "'+1)'+ S'111J'+ 810
．． 7'+ 5IIJ'+ 512)'+ 5IB)'+
S'IQ'+SIM).

S+6), S+q)'). The phase component is also restored in the same manner. Next, the sample values 10 of the amplitude and phase components of the restored entire spectrum are
After interpolating between pages, resampling is performed to restore the original sampled values. This is inversely Fourier transformed to reproduce the time waveform. 20
By connecting the time waveforms reproduced every m5 and inputting them to a synthesizer, the original voice is restored.

As described above, in the present invention, since encoding is performed focusing on the redundancy of the harmonic structure of the short-time spectrum of the prediction residual signal, the prediction residual signal can be encoded at a low bit rate.

Furthermore, if the resampling rate 1 is set to a constant value over all analysis frames, it is possible to perform spectrum matching between different partial spectra (amplitude and phase) of frames, which has the advantage of further reducing redundancy. .

Effects of the Invention The present invention has the above-described configuration, and provides the following effects.

(a) The short-time spectrum of the prediction residual signal is divided into partial spectra in a band equal to the pitch frequency, and the divided bands are resampled every - (tL is a positive integer) of the pitch frequency. Therefore, the redundancy of the harmonic structure of the spectrum can be reduced.

Page 11 (1)) By setting the resampling rate to the same value over the entire prediction residual signal, it is possible to perform spectral matching between different partial spectra of the frame, which allows encoding at a significantly lower bit rate. It has the advantage of being scalable.

(C) Since both the amplitude component and the phase component of the spectrum are encoded, there is an advantage that the reproducibility of the temporal waveform is good and high quality speech can be synthesized.

[Brief explanation of the drawing]

Figure 1 is a block diagram of a device that implements a conventional residual-driven speech synthesis method, Figure 2 is a diagram showing a short-time spectrum of a predicted residual signal, and Figure 3 is a partial spectrum (wave + 11 components).
FIG. l...Input terminal, 2...Prediction residual signal, 3...Drive signal, 4...Output terminal, 10...Prediction analysis filter, 11...Low pass filter (LPF), 12...・
Down sampler, 13...Transmission channel, 15...
- Decoder, 16... Up sampler, 17... Low pass filter, 18... Full wave rectifier, 19... Spectrum smoother, 20 Bypass filter (HPF), 21...
...Adder, JP-A-Sho GO-31198 (4) 22...Synthesizer. Name of agent: Patent attorney Toshio Nakao and 1 other person No. 3
Figure-695-

Claims (3)

[Claims] (1) Divide the short-time spectrum of the prediction residual signal into partial spectra in a band equal to the pitch frequency, and set φ in the divided band to one of the pitch frequencies! - (rL is a positive integer), a predictive residual signal code characterized in that, after resampling for each signal, redundancy due to the harmonic structure of the pitch frequency is removed using the spectral distance between adjacent partial spectra as a measure; method. (2) A claim characterized in that the resampling rate is set to the same value over the entire prediction residual signal, allowing matching with partial spectra belonging to different frames. 2. The predictive residual signal encoding method according to item 1. (3) The predictive residual signal encoding method according to claim 1 or 2, wherein both the amplitude component and the phase component of the short-time spectrum are encoded. 2 pages