JPH02150900A - Method and device for linear predictive analysis - Google Patents

Abstract

PURPOSE:To obtain a linear predictive analytic result close to an actual spectrum envelope over the entire frequency band by taking a 2nd linear predictive analysis for an input speech signal which is filtered by a reverse filter when normalized predicted residual electric power in a 1st linear predictive analysis becomes extremely small. CONSTITUTION:A 1st linear predictive analysis part 1 takes a 1st linear predictive analysis of the input speech signal and a normalized predicted residual electric power calculating circuit 2 calculates the normalized predicted residual electric power in the 1st linear predictive analysis to determine an attenuation coefficient (r) by a predetermined corresponding relation with the normalized predicted residual electric power; and the attenuation coefficient (r) is applied to linear prediction coefficients alpha1 and alpha2 obtained by the 1st linear predictive analysis and the reverse filter 5 which uses the linear predictive coefficients after the attenuation coefficient is applied is constituted. Then the input speech signal is inputted to this reverse filter 5, whose output is processed by a 2nd linear predictive analysis. Consequently, the linear predictive analytic result which is close to the actual spectrum envelope over the entire frequency band is obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a linear predictive analysis method and apparatus thereof, and particularly to a linear predictive analysis method and apparatus thereof in speech analysis.

[Conventional technology]

A phenomenon of underestimation of the bandwidth of polar frequencies is known as a drawback of linear predictive analysis, which is one of the methods of speech analysis. This phenomenon occurs particularly markedly in the case of nasal sounds, and is caused by the first formant. When linear predictive analysis is performed on a nasal sound or the like in which this phenomenon is noticeable, the normalized predictive residual power becomes extremely small, for example, 1O-3 or less, when the analysis order is about second order. As a result of the underestimation, the bandwidth of the first formant becomes extremely narrow, leading to a false impression that most of the energy of the voice is concentrated in the first formant.

FIG. 2 is an explanatory diagram for explaining this phenomenon. When this phenomenon occurs, if the actual spectral envelope is like the dotted curve a, the spectral envelope shown by the result of the linear predictive analysis will be like the solid curve.

Lag windows are known as a solution to this drawback.

The lag window provides attenuation in the time delay direction to the autocorrelation coefficient sequence, and has a spectrum spreading effect. Spectral spreading of the lag window can compensate for the first formant bandwidth underestimation.

[Problem to be solved by the invention]

However, since the spectral spread effect of the lag window extends over the entire frequency band, the lag window is applied to the conventional linear prediction analysis.
If the underestimation of the formant bandwidth is compensated for, the results of the linear predictive analysis will be greatly distorted in frequency bands other than the first formant.

An object of the present invention is to provide a linear predictive analysis method and apparatus that can prevent underestimation of the bandwidth of polar frequencies and obtain linear predictive analysis results that are close to the actual spectral envelope over the entire frequency band.

[Means to solve the problem]

The linear prediction analysis method of the present invention performs a first linear prediction analysis on an input audio signal, calculates the normalized prediction residual power in the performed first linear prediction analysis, and calculates the normalized prediction residual power in the first linear prediction analysis performed. An attenuation coefficient is determined based on a predetermined correspondence relationship with respect to the residual power, and the determined attenuation coefficient is applied to the first
is applied to the linear prediction coefficient obtained by the linear prediction analysis of A second linear predictive analysis is performed on the output.

Further, in the linear prediction analysis method of the present invention, a first linear prediction analysis is performed on an input audio signal, a normalized prediction residual power in the first linear prediction analysis performed is calculated, and the calculated normal and attenuation means for determining an attenuation coefficient according to a predetermined correspondence relationship with respect to the predicted residual power, and applying attenuation determined by the determined attenuation coefficient to each of nine inputs or outputs of the unit delay element. configuring an inverse filter whose coefficients are the linear prediction coefficients obtained in the first linear prediction analysis, inputting the input audio signal to this inverse filter, and performing a second linear prediction analysis on the output of the inverse filter. You can also do it.

The linear prediction analysis device of the present invention includes a first linear prediction analysis means that performs a first linear prediction analysis on an input audio signal, and a first linear prediction analysis performed by the first linear prediction analysis means. a calculation means for calculating the normalized predicted residual power in the calculation means, a determining means for determining an attenuation coefficient in a predetermined correspondence relationship with respect to the normalized predicted residual power input from the calculating means, and an input from the determining means. applying means for applying the attenuation coefficient obtained in the first linear prediction analysis to the linear prediction coefficient obtained in the first linear prediction analysis; The inverse filter is configured to include a filter and a second linear predictive analysis means that performs a second linear predictive analysis on the output of the inverse filter.

Further, the linear prediction analysis device of the present invention includes a first linear prediction analysis means for performing a first linear prediction analysis on an input audio signal, and a first linear prediction analysis means performed by the first linear prediction analysis means. A calculating means for calculating normalized predicted residual power in predictive analysis, a determining means for determining an attenuation coefficient based on a predetermined correspondence relationship with respect to the normalized predicted residual power input from this calculating means, and this determining means. an attenuation means for applying attenuation determined by the attenuation coefficient inputted from the means to each input or output of the unit delay element, and using the linear prediction coefficient obtained in the first linear prediction analysis as a coefficient and the input audio signal It may also be configured to include an inverse filter that inputs the inverse filter, and a second linear predictive analysis means that performs a second linear predictive analysis on the output of the inverse filter.

[Example]゛ Next, the present invention will be explained with reference to the drawings.

FIG. 1 is a block diagram showing one embodiment of the present invention.

The embodiment shown in FIG. 1 has A/D converters 1 connected in sequence.
1. Window processing circuit 12. Autocorrelation calculation circuit 13 and LPG
A first linear prediction analysis section 1 having an analysis circuit 14 and performing a second-order linear prediction analysis on an input audio signal, and a normalized prediction residual power U based on the parameters input from the first linear prediction analysis section 1. A normalized predicted residual power calculation circuit 2 that calculates a normalized predicted residual power, an attenuation coefficient determination circuit 3 that determines an attenuation coefficient γ based on the normalized predicted residual power U input from the normalized predicted residual power calculation circuit 2, and an attenuation coefficient determination circuit 3 that determines an attenuation coefficient γ. A damping coefficient applying circuit 4 applies the damping coefficient γ inputted from the circuit 3 to the α parameter inputted from the first linear prediction analysis unit 1, and the α parameter to which the damping coefficient γ is applied by the damping coefficient applying circuit 4 is used as a coefficient and the first An inverse filter 5 which inputs the output of the A/D converter 11 included in the linear prediction analysis unit 1, and a window processing circuit 61 which is connected in sequence, an autocorrelation calculation circuit 62 and an LPU analysis circuit 63, and the output of the inverse filter 5 A second linear predictive analysis unit 6 performs linear predictive analysis of a required order, for example, 12th order, and outputs parameters.

The A/D converter 11 of the first linear prediction analysis unit 1 band-limits the input audio signal using a reduction filter with a cut-off frequency of 3.4 kHz, and converts the band-limited audio signal to a sampling frequency of 8 kHz.
Sample in Hz and quantize each sample to a predetermined number of bits. The window processing circuit 12 performs window cutting in which the 30 m5 quantized samples input from the A/D converter 11 are multiplied by a Hamming window function every 20 m5, which is the analysis frame period. The autocorrelation calculation circuit 13 calculates the autocorrelation of the signal input from the window processing circuit 12 for each analysis frame. LP
The C analysis circuit 14 performs a second-order linear prediction analysis using the autocorrelation input from the autocorrelation calculation circuit 13, and calculates α1. of the α parameter, which is a linear prediction coefficient. α2 and the partial autocorrelation coefficient are output as + and 2 as parameters.

The normalized predicted residual power calculation circuit 2 includes the LPG analysis circuit 14
Using the parameter kl+kt inputted from , the normalized prediction residual power U in the linear prediction analysis performed by the first linear prediction analysis unit 1 is calculated by the following equation.

u=n (1-ki2) The attenuation coefficient determination circuit 3 uses the table lookup method to calculate the normalized prediction input from the normalized prediction residual power calculation circuit 2 using a correspondence relationship as shown in FIG. Attenuation coefficient γ is determined for residual power U.

The attenuation coefficient applying circuit 4 receives the α parameter α1. input from the LPC analysis circuit 14. The damping coefficient γ input from the damping coefficient determining circuit 3 is applied to α2, and the resulting coefficient γα1
．． γ2α2 is used as a coefficient of the inverse filter 5.

FIG. 4 is a block diagram of the inverse filter 5.

The inverse filter 5 includes a unit delay element 51.52 that sequentially delays the quantized sample input from the A/D converter 11 of the first linear prediction analysis unit 1 by one sampling period, and an output of the unit delay element 51.52. Coefficient γα1. γ2α, an adder 55 that adds the outputs of the multipliers 53 and 54, and a subtracter 56 that subtracts the output of the adder 55 from the input of the unit delay element 510. ing. The inverse filter 5 has coefficients α1 and γ instead of α2.
α1. There is no difference from a normal second-order inverse filter except that γ2α2.

The window processing circuit 61 of the second linear prediction analysis section 6 performs a window cutting process on the output of the inverse filter 5 in the same manner as the window processing circuit 12 of the first linear prediction analysis section 1 performs. The autocorrelation calculation circuit 62 and the LPU analysis circuit 63 are the window processing circuit 6
A 12th order linear predictive analysis is performed on the signal input from 1, and 1□ is output to the K parameter.

Now, a normal second-order inverse filter has a coefficient α parameter α2. It has the effect of flattening the spectral envelope of the input signal near the polar frequency determined by α2. On the other hand, the inverse filter 5 has coefficients γα1. γ2α2, and as shown in FIG. 3, the damping coefficient γ is set in the range O≦γ×1, so if γ=0 (multiplier 53.
output of 54 becomes zero), outputs the input signal as is, and γ
If ≠0, it has the effect of partially flattening the spectral envelope of the input signal near the polar frequency. The degree of this flattening increases as γ approaches 1.

If the phenomenon of polar frequency bandwidth underestimation does not occur,
Since the normalized prediction residual power U does not become so small,
The damping coefficient γ outputted by the damping coefficient determining circuit 3 becomes zero.

As a result, the input audio signal is quantized by the A/D converter 11, and the input audio signal is quantized by the inverse filter 5 without undergoing any modification.
The signal is input to the linear prediction analysis section 6, and normal linear prediction analysis is performed. When the phenomenon of bandwidth underestimation occurs, the normalized predicted residual power U becomes smaller depending on the degree of the phenomenon, and the determined attenuation coefficient γ is no longer zero, but approaches 1 as the degree of underestimation becomes more significant. As a result, the inverse filter 5 controls the A/D converter 11 so as to suppress the spectral density of the first Holt mant, which is the pole frequency where the bandwidth underestimation occurs.
Filter the input signal from. If this filtering effect is expressed as an amplitude characteristic of the inverse filter 5, the amplitude characteristic will be as shown by curve C shown in FIG. However, curves a and b in FIG. 5 are reproductions of curves a and b in FIG. 2. Suppression of the spectral density of the first formant by the inverse filter 5 is performed by the second linear predictive analysis unit 6
Since the concentration of energy in the first formant is compensated for, the underestimation of the bandwidth of the first formant is eliminated.Moreover, in this case, the inverse filter 5 does not change the spectral envelope in frequency bands other than the first formant.

Therefore, the second linear prediction analysis unit 6 outputs a parameter of 1 to 1□ to faithfully represent the spectral envelope of the input audio signal over the entire frequency band.

The embodiment shown in FIG. 1 has been described above.

FIG. 6 is a block diagram of an inverse filter that can be used in place of the inverse filter 5 in the embodiment shown in FIG.

The inverse filter shown in FIG. 6 is the inverse filter 5 shown in FIG.
The multipliers 57 and 58 are inserted immediately after the unit delay elements 51 and 52, respectively. However, the attenuation coefficient γ outputted from the attenuation coefficient determining circuit 3 is input to the multipliers 57 and 58, and the coefficients of the multipliers 53 and 54 are γα1. α parameter α3. which is output by the LPC analysis circuit 14 instead of γ2α2.
α2 is used directly. Therefore, when using the inverse filter shown in FIG. 6, the attenuation coefficient applying circuit 4 is not necessary. Multipliers 57 and 58 apply attenuation determined by attenuation coefficient γ to the outputs of unit delay elements 51 and 52, and the filtering characteristics of the inverse filter shown in FIG. 6 become the same as those of inverse filter 5.

Attenuation by multipliers 57 and 58 is achieved by unit delay elements 51 and 52.
Since it may be given to the input instead of the output of the multiplier 57.5
8 may be inserted just before the unit delay elements 51 and 52 instead of immediately after.

Further, in the embodiment shown in FIG. 1, the output of the window processing circuit 12 can be input to the inverse filter 5, and the output of the inverse filter 5 can be input to the autocorrelation calculation circuit 62. In this case, the window processing circuit 61 is unnecessary.

〔Effect of the invention〕

As explained above, the present invention performs a first linear predictive analysis on an input audio signal, and when the normalized predictive residual power in this first linear predictive analysis becomes significantly small, the input signal is By configuring an inverse filter to suppress the spectral density of the first formant of the speech signal and performing a second linear predictive analysis on the input speech signal filtered by the inverse filter, the bandwidth of the polar frequency is under-estimated. It is possible to prevent this, and moreover, it is possible to obtain a linear predictive analysis result that is close to the actual spectrum envelope over the entire frequency band.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is an explanatory diagram for explaining the phenomenon of polar frequency bandwidth underestimation in linear predictive analysis, and FIG. 3 is an implementation of the embodiment shown in FIG. 1. A graph showing an example of correspondence between input and output of the attenuation coefficient determination circuit 3 in the example, FIG. 4 is a block diagram of the inverse filter 5, FIG. 5 is an explanatory diagram for explaining the action of the inverse filter 5, and FIG. The figure is a block diagram of an inverse filter that can be used in place of the inverse filter 5. ■...First linear prediction analysis unit, 2...Normalized prediction residual power calculation circuit, 3...Attenuation coefficient determination circuit, 4...Attenuation Coefficient application circuit, 5...
... Inverse filter, 6... Second linear prediction analysis section. Agent: Patent Attorney Susumu Uchihara 2nd. Mouth 1st look I and child measurement remaining barley meeting 3rd figure

Claims (4)

[Claims] (1) Perform a first linear prediction analysis on the input audio signal, calculate the normalized prediction residual power in the first linear prediction analysis, and calculate the normalized prediction residual power for the calculated normalized prediction residual power. Deciding an attenuation coefficient according to a predetermined correspondence relationship, applying the determined attenuation coefficient to the linear prediction coefficient obtained in the first linear prediction analysis, and inversely using the linear prediction coefficient to which the attenuation coefficient has been applied as a coefficient. A linear predictive analysis method comprising configuring a filter, inputting the input audio signal to the inverse filter, and performing a second linear predictive analysis on the output of the inverse filter. (2) Perform a first linear prediction analysis on the input audio signal, calculate the normalized prediction residual power in the first linear prediction analysis, and calculate the normalized prediction residual power for the calculated normalized prediction residual power. The attenuation coefficient is determined according to a predetermined correspondence relationship, and the attenuation means is configured to apply the attenuation determined by the determined attenuation coefficient to each input or output of the unit delay element obtained by the first linear predictive analysis. A linear prediction analysis method characterized by configuring an inverse filter whose coefficients are linear prediction coefficients, inputting the input audio signal to this inverse filter, and performing a second linear prediction analysis on the output of the inverse filter. . (3) A first linear predictive analysis means that performs a first linear predictive analysis on the input audio signal, and a normalized prediction residual in the first linear predictive analysis performed by the first linear predictive analysis means. a calculation means for calculating power; a determination means for determining an attenuation coefficient according to a predetermined correspondence relationship with respect to the normalized predicted residual power input from the calculation means; and a determination means for determining the attenuation coefficient input from the determination means. an application means that applies an application to the linear prediction coefficient obtained in the first linear prediction analysis; an inverse filter that uses the linear prediction coefficient to which the attenuation coefficient has been applied by the application means as a coefficient and receives the input audio signal as an input;
A linear predictive analysis device comprising: second linear predictive analysis means for performing a second linear predictive analysis on the output of the inverse filter. (4) A first linear predictive analysis means that performs a first linear predictive analysis on the input audio signal, and a normalized prediction residual in the first linear predictive analysis performed by the first linear predictive analysis means. a calculation means for calculating the power; a determination means for determining the attenuation coefficient according to a predetermined correspondence with the normalized predicted residual power input from the calculation means; and a determination means for determining the attenuation coefficient input from the determination means. an inverse filter having an attenuation means for applying attenuation to each input or output of the unit delay element, and inputting the input audio signal using the linear prediction coefficient obtained in the first linear prediction analysis as a coefficient; A linear predictive analysis device comprising: second linear predictive analysis means for performing a second linear predictive analysis on the output of the filter.