JPS5849878B2 - Speech analysis/synthesis method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention aims to improve effective compression methods by analyzing (recording) and synthesizing audio information, and in particular, by analyzing and synthesizing linear prediction coefficients or equivalent partial autocorrelation coefficients as main analysis/synthesis harameters. It concerns the improvement of synthetic systems.

Effective compression of audio information, so-called low bitrat
Recently, e-coding has been attracting a lot of attention and research experiments are being conducted for the purpose of applications such as explanatory voices for voice responses and image responses.

One of the effective means for compressing information is an analysis/synthesis system whose main parameter is a linear prediction coefficient or an equivalent partial autocorrelation coefficient.

The principle of this system is shown in Figure 1.

The speech waveform is parametrically analyzed and linear prediction coefficients or partial autocorrelation coefficients are extracted.

The output after parameter extraction (the difference between the measured value and the predicted value) is a so-called residual waveform, an example of which is shown in FIG.

This residual waveform is an equivalent sound source waveform for speech generation, and in the case of speech, it is a series of impulse waveforms as shown in Figure 2, and in the case of unvoiced sound, it is a random white noise waveform. Become.

Therefore, from the correlation of this residual waveform, the voiced and unvoiced V/U of the sound source, its strength A, and, in the case of voiced, its pitch period τ are extracted.

In normal parameter analysis/synthesis methods, this sound source information (V
/U, A, τ) is transmitted or recorded, and during playback, the output of the impulse sequence with period τ or white random noise is switched by V/U, and its intensity is controlled by A to reproduce the sound source waveform. .

Although this modeling of the sound source enables significant information compression, it also degrades the quality of the synthesized speech.

Therefore, in order to construct a high-quality parameter analysis/synthesis system, an effective means for compressing information on the sound source (residual) waveform is required.

The present invention relates to improvements in this respect.

Since voiced sounds, which occupy the majority of the speech waveform in terms of time and energy, mainly contribute to sound quality, we will consider voiced sounds here.

As shown in FIG. 2, the residual waveform of the voiced part consists of a main peak waveform represented by an impulse sequence and a random-looking background waveform at its base.

The contribution of the residual waveform to sound quality can be considered by dividing it into three factors: its waveform, period, and timing.

(1) The waveform is mainly an impulse-like peak waveform, and has an almost uniform energy distribution in terms of amplitude spectrum.

(2) Periodicity appears as a remarkable line spectral structure on its amplitude spectrum.

(3) Timing and background random-appearing waveform information is believed to be preserved on its phase spectrum.

FIG. 3 shows an example of the spectrum of the residual waveform.

It can be clearly seen that the features of the upper edicts 1) and (2) appear on the amplitude spectrum of FIG.

Now, the Fourier transform of a waveform consisting of N sample values is
It consists of N amplitude spectrum values and N phase spectrum values.

However, if the input waveform is a real number waveform, the amplitude and phase spectra are even and odd symmetric functions centered around the -+1st2 value.

〉※N Therefore, the required information is 2X-=N.

2 Furthermore, assuming that the amplitude spectrum is a constant value, N Then the necessary information will be -ko.

FIG. 4 shows an example of a residual waveform reproduced by inverse Fourier transform from the arc information (phase information) thus derived.

This constant amplitude spectral residual waveform (hereinafter referred to as phase approximation residual waveform) has the same phase spectrum as the original residual waveform and has a constant amplitude spectral intensity.

In order to give this phase approximation residual waveform a line spectrum structure with a period τ, it is sufficient to calculate the force after the period τ as shown in FIG. 5, and to control the chain time using the pitch period τ.

The extracted period τ is the average value of the analysis interval, and the coefficient r is used to take into account the effect of the slight change in τ in that interval and to reduce sudden discontinuous changes in pitch. It is better to perform attenuation chain addition of (r<1).

FIG. 5 shows an example of the structure of a damping chain addition circuit with a chain time τ and a damping r.

In FIG. 5, if the input signal from the right is f (t) and the output signal from the left is g (t), then (3) the phase approximation residual waveform is obtained from the constant amplitude complex spectrum by inverse Fourier transform.

(4) The residual waveform is approximated by attenuated chain addition.

The parameter synthesis system is excited using this approximate residual waveform as a sound source.

According to this approximate residual waveform, not only the spectrum and periodicity but also the timing of excitation and the random waveform part in the background are well approximated, and it is expected that the sound quality will be improved.

A specific embodiment of the present invention will be described below with reference to the block diagram of FIG.

The audio waveform 1 to be analyzed and synthesized is amplified to an appropriate level and band-limited by a preamplifier 2 and a low-frequency p-wave generator 3, then converted to a digital signal by an A/D converter 4, and then stored in a buffer memory. 5 is stored.

When the A/D conversion is completed, data of a predetermined analysis interval length (usually taken from 10 ms to 20 ms) is extracted, and a 6-parameter analyzer is used to calculate the linear prediction coefficient or One of the partial ego relation coefficients is extracted.

Since this extraction circuit is already known, the description thereof will be omitted.

The extracted parameters 7 are stored in the audio information memory 8.

A residual correlation of the residual waveform 9 output from the parameter analyzer 6 is calculated by a correlation analyzer 10, and a sound source intensity 11(A) is extracted as a normalization factor.

Furthermore, the pitch extractor 12 extracts the peak value of the correlation from the voiced/
The pitch period 14 (τ) is determined from the unvoiced pitch 13 (V/U) and the time that gives the peak value.

Information 11, 13, and 14 is also stored in the memory 8.

On the other hand, the residual waveform 9 is converted into a spectrum by the Fourier transform device 15, its amplitude spectrum component is made constant by the amplitude constant circuit 16, and only its phase component 17 is stored.

In this case, use the symmetry N Then, only the value of - is memorized.

2. When the analysis and information extraction for one analysis section is completed, the next section is read out from the buffer memory, and the above analysis, extraction, and recording processes are repeated until all data has been analyzed.

The audio waveform is reproduced from the stored information as follows.

First, complete phase information is recovered by the phase information reproducing circuit 18 from the stored phase spectrum information.

The function of this circuit is the Fourier N of real functions. Using the property of transformation, we can find the odd function symmetric 2 with respect to the first one. N Phase component from data from 1 to 1 to 2 from 1 to N1 It is to restore to data.

An example of the circuit configuration is shown in FIG.

The phase information stored in the recording device 8 is N of the residual waveform.
N phase values obtained by Fourier transformation of sample values {θ.

}n-1~N's most N Only the first -ko.

This information {θ. }·n=1 to N2 are read out from the storage device and given to the phase information reproducing circuit 18 as a signal 30.

The details are shown in FIG.

N The registers in this circuit are registers 31 and 2. and 34, and 31 is 7 from θ1 to θ. , and 34 is the buffer from θN to θ. It's a sofa.

The input signal 30 is divided into two parts, one N 2) Enter directly into register 31 starting from the first one. are sequentially transferred by the clock from 1 to N. is stored in

One side is N by the sign inversion circuit 32. -+1st 2 as signal 33 after sign inversion N to the Nth register 34, the first one 2 are input sequentially from 2 to 2 and transferred by the clock. Stored from +1 to N.

By this, the phase information {0.}, n-1 to N is restored.

θ.

The inverse Fourier transform circuit 19 performs inverse Fourier transform on the phase spectrum restored by 1-θN-H, thereby reproducing the phase approximation residual waveform 20.

The amplitude spectrum of this waveform is approximately reproduced by an attenuation chain addition circuit 21 whose time is limited by the pitch period, and then controlled to have an appropriate amplitude by an amplitude limiting circuit 22.

The attenuation time of the attenuation force calculation circuit is controlled every 10 to 20 milliseconds based on pitch period information, and the attenuation amount r
In the case of unvoiced sounds, r = 0, making the force calculation effect zero, and in the case of voiced sounds, O < r < 1, giving an additive effect.

Also, the amplitude control is not the actual amplitude A, but is multiplied by A(1-r) in order to eliminate the change in amplitude due to the attenuation chain addition.

The approximate residual waveform 23 thus reproduced is applied to the parameter synthesis circuit 24, and a speech waveform is synthesized.

The circuit configuration for parameter synthesis using linear prediction coefficients or partial autocorrelation coefficients is well known and will not be described here.

The synthesized audio waveform is sent to the D/A converter 25 and the low frequency waveform generator 2.
6 and is reproduced as an audio waveform 27.

The audio information to be stored or transmitted using the method according to the present invention includes (1) a linear prediction coefficient or a partial autocorrelation coefficient p (p is the order of analysis and is usually set between 10 and 12);
If the analysis interval length is 20 milliseconds and the sampling period is 50 microseconds (sampling frequency is 20kHz), is the original speech waveform? Bit X20X10 X20X10 -3-32k
Bits/20ms Parameter 8 bits x 12 = 96 bits/20ms Sound source (1+
8+8) bit/20 ms phase spectrum 8 bits x 20 x 103 x 20 x 1. O-3X0.5=
1.6kbits/20ms Therefore, original audio: 150kbits/second Original method: 80kbits (96+17) x 50bits: 8
This results in information compression of 6 bits/second, or approximately 1/2.

As described above, according to the present invention, the amount of information can be compressed to about 1/2 while maintaining the quality of synthesized speech at a sufficiently high level (as compared to the normal parameter synthesis method).

Furthermore, in the explanation of the embodiments of the present invention, the explanation was given in the form of using all the phase information of the residual waveform, but when it is acceptable to make the quality of synthesized speech a little too harsh, we can use additional information as described below. Compression is possible.

Now, if the phase spectrum of the residual waveform is {θ}n=1~, then the average settling time Z is defined as follows.

Therefore, the phase information is correctly maintained up to the mth precept, and θ for higher frequency components.

One ω.・Z (mod 2π) is considered a good approximation.

This information is N by adding one piece of information called Z. Yotte θ.

: It is possible to omit the information where m<nζ1 can.

Most of the information that is recorded or transmitted is {θ.

} (5) Since the information is about 9 10 2 N, if we take m1--, we can compress the information to 32 1 1 N 1 K, and if we take m=--1, we can compress the information to 1. At 3
2 2 4 ru.

[Brief explanation of drawings]

Figure 1 is a basic diagram of the speech parameter analysis/synthesis method, Figure 2 is an example of a residual waveform, Figure 3 is an example of the amplitude spectrum of the residual waveform, and Figure 4 is a constant amplitude phase. An example of approximate residual waveform, Fig. 5 is a principle diagram of the attenuation chain addition circuit, Fig. 6
This figure is a principle diagram of approximate residual waveform reproduction according to the present invention, FIG. 7 is a block diagram of an implementation of the present invention, and FIG. 8 is a configuration diagram of a phase information restoration circuit. Steam I Rin's X direction Bono 0 Yanagi (↓voice 狽 ふ i, <) U. −、Ima Kō Hari 廁 ← Ay L (Ima leg) Mine 2 Kozue 5 Rindai Soku O -10 120 -3 (1) -251-

Claims (1)

[Claims] 1 In an analysis/synthesis method that analyzes and extracts parameters such as linear prediction coefficients or partial autocorrelation coefficients, records or transmits them, and synthesizes a speech waveform from the sound source and parameter information on the playback side, prediction errors are generated during analysis. Determine the residual waveform, residual waveform strength, pitch period, and whether it is voiced or unvoiced.
Then, extract and record only the phase spectrum component, reproduce the constant amplitude phase approximation residual waveform by inverse Fourier transform from the phase spectrum recorded or transmitted at the time of admonition, The speech analysis/speech analysis method further comprises means for approximating the residual waveform by emphasizing the periodicity of the residual waveform by using attenuated chain addition having a chien time controlled by the pitch period.
Synthesis method.