JP3138574B2 - Linear prediction coefficient interpolator - Google Patents

Abstract

PURPOSE:To provide a linear predictive coefficient interpolating device which can extremely reduce its calculation quantity even in a digital signal processor. CONSTITUTION:A linear predictive coefficient interpolating device is provided with a hyperbolic conversion part 61 which applies the hyperbolic conversion to a partial autocorrelation coefficient of a prescribed degree acquired from an input voice signal of a fixed time length, a linear interpolation part 62 which is connected to the part 61 and applies the linear interpolation to the hyperbolic conversion result of the part 61, and a hyperbolic reverse conversion part 63 which connected to the part 62 and applies the reverse conversion to the linear interpolation result of the part 62.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech information processing system using linear prediction coefficients, and more particularly to a linear prediction coefficient interpolation device which can be used in a speech information processing system.

[0002]

2. Description of the Related Art In a conventional technique, an input speech signal is subjected to linear prediction analysis, spectrum information is obtained from linear prediction coefficients, and the obtained spectrum information is transmitted. A system for performing voice communication at a low bit rate has been realized.

[0003] As a speech encoding / decoding method for transmitting at a low bit rate using the above-described linear prediction analysis, CELP (Code Excited Linea) is used.
r Prediction) method is known.

[0004] The CELP system is a model of voice generation, in which codebooks have candidates for signals corresponding to air currents generated in the glottis, and one of these signals corresponds to a period of opening and closing a vocal cord. The synthesized speech is generated by passing through a pitch prediction (or long-term prediction) filter for adding a pitch and a (short-term) prediction filter corresponding to articulation in the oral cavity. At this time, an optimal signal is selected from AbS (Analysis) from a code book including signal candidates corresponding to the airflow.
The signal number, gain, pitch prediction filter coefficient, lag (corresponding to pitch length), and linear prediction filter coefficient are quantized and coded by a synthesis analysis method, that is, a signal of this signal.

In the CELP system, an input audio signal is reduced to about 20
The frame is divided into ４０40 ms frames, and each frame is divided into 4 to 5 subframes for processing. In the CELP method, a linear prediction coefficient is obtained for each frame, and a linear prediction coefficient is obtained for each subframe by interpolating them. Essentially, the linear prediction coefficient can be obtained for each subframe, so that encoding with good synthesized sound quality can be performed. However, since the amount of calculation increases and the amount of codes to be transmitted increases, the above method is generally used.

Conventionally, in the CELP system, LAR (Log Area Ratio) is used as a method of interpolating linear prediction coefficients.
o: Interpolation using a logarithmic conversion type (logarithmic cross-sectional ratio) of a PARCOR coefficient is used. That is, the linear prediction coefficient is converted into the PARCOR coefficient, and the LAR is obtained from the PARCOR coefficient by the following equation (1).

[0007]

(Equation 1)

[0008] Then, linear interpolation is performed on the LAR, and the interpolation result is converted into a PARCOR coefficient by the inverse conversion of equation (1) to obtain a linear prediction coefficient in subframe units.

Here, the LAR will be described.

The PARCOR coefficient represents a correlation value, and its value ranges from -1 to +1. However, since there is a bias in the distribution due to the order of the PARCOR coefficient, nonlinear transformation is performed on the PARCOR coefficient up to the second order having a large bias. To perform interpolation. Here, twice the gm shown in the equation (1), that is, 2 gm
Is called LAR.

FIG. 5 shows a conversion result of PAR to LAR from a PARCOR coefficient. As shown in FIG. 5, since the LAR has a sensitive conversion characteristic with respect to the PARCOR coefficient near ± 1, the interpolation error is reduced by performing the LAR conversion in advance. Further, the middle and high order PARCOR coefficients are less ubiquitous and are dispersed around zero. LAR is PARCO near 0
As for the R coefficient, a nearly linear conversion is performed,
It also matches the characteristics of the ARCOR coefficient.

[0012]

However, in the above-mentioned conventional technology, the LAR generally used in CELP in the interpolation of the linear prediction coefficient at present has a small interpolation error but needs to use a logarithm. For this reason, there is a problem in that the amount of calculation increases when implemented on a digital signal processor (DSP).

SUMMARY OF THE INVENTION An object of the present invention is to provide a linear prediction coefficient interpolation apparatus which can greatly reduce the amount of calculation even when implemented on a DSP in view of the above-mentioned problems in the conventional technology.

[0014]

SUMMARY OF THE INVENTION An object of the present invention is to provide a PARCO of a predetermined order obtained from an input audio signal of a fixed time length.
Conversion means for hyperbolically converting the R coefficient, interpolation means connected to the conversion means for linearly interpolating the result of the hyperbolic conversion, and inverse means connected to the interpolation means for inversely converting the linearly interpolated linear interpolation result This is achieved by a linear prediction coefficient interpolation device including a conversion unit.

In the linear predictive coefficient interpolating apparatus according to the present invention, the hyperbolic transform means may be configured to perform the hyperbolic transform such that the closer the PARCOR coefficient of a predetermined order approaches the value of ± 1, the steeper the slope.

[0016]

In the linear predictive coefficient interpolating apparatus according to the present invention, the converting means includes a predetermined-order P of a predetermined order obtained from an input audio signal having a fixed time length.
The ARCOR coefficient is hyperbolically transformed, the interpolation means linearly interpolates the hyperbolically transformed result, and the inverse transformation means inversely transforms the linearly interpolated linear interpolation result.

In the linear predictive coefficient interpolating apparatus according to the present invention, the hyperbolic conversion means performs hyperbolic conversion so as to have a steep slope as the predetermined order PARCOR coefficient approaches a value of ± 1.

[0018]

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram showing an embodiment of a linear prediction coefficient interpolation apparatus according to the present invention.

FIG. 1 is a block diagram showing an embodiment of a prediction coefficient interpolation section which is a linear prediction coefficient interpolation apparatus according to the present invention.

FIG. 2 is a block diagram showing the configuration of an embodiment of a CELP encoding / decoding apparatus having the main parts of FIG.

Before explaining FIG. 1, the role of the prediction coefficient interpolation unit in FIG. 1 will be described with reference to the encoding / decoding apparatus in FIG.

First, the encoding unit shown in FIG. 2 will be described.

In the encoding unit shown in FIG. 2, the linear prediction analysis (short-term prediction) unit 1 performs linear prediction analysis of the digital input signal s (t) sampled at the sampling frequency fs at a certain frame period, and performs quantization. The obtained PARCOR coefficient is determined for each frame. Here, t indicates the sampling time.

Thereafter, in the prediction coefficient interpolator 6 (see FIG. 1), a hyperbolic converter 61, which is a conversion means, performs a conversion for alleviating the uneven distribution of low-order PARCOR coefficients, and converts the conversion results by the interpolation means. A certain linear interpolation unit 62 performs linear interpolation. Thereafter, a hyperbolic inverse transform unit 6 as an inverse transform means
In step 3, the inverse conversion of the interpolation result is performed, and an interpolation PARCOR coefficient for each subframe is obtained.

The auditory weighting filter 2 has a transfer function W (z) represented by equation (2).

[0026]

(Equation 2)

Here, α ^(m) _i (0 ≦ m <M, 0 <i ≦
P) represents the ith-order interpolated linear prediction coefficient of the m-th subframe, and M represents the number of subframes included in one frame.

That is, the auditory weighting filter 2 comprises: an inverse filter 201 for obtaining a residual signal of the input signal based on the equation (3);

[0029]

(Equation 3)

A weighted linear prediction filter 202 based on equation (4)

[0031]

(Equation 4)

The signal u (t) is formed by emphasizing the valley portion of the spectrum of the input signal s (t).

Here, λ is a parameter for determining how much the valley portion of the spectrum is emphasized, and the valley portion is emphasized as it approaches 0. A signal as similar as possible to this signal u (t) (a composite signal weighted by hearing) u ′
Encoding is performed to synthesize (t). This processing is performed in subframe units. The number of samples per frame is F.

With the auditory weighting filter 2, the error of the synthesized signal s' (t), which will be described later, from the input signal s (t) decreases as the power of the spectrum decreases, and these spectra are hard to be masked by the masking characteristics of the auditory sense. Since the noise of the component can be reduced, the synthesized sound quality is improved audibly.

The codebook 3 contains N excitation signal vectors (where N is a positive integer), and one of the excitation signal vectors cj (t) (0 ≦ j <N) is Multiplied by γ by the multiplier 4, the pitch component is added by the pitch prediction filter 5 and passed through the weighting synthesis filter 7 (the transfer function is the same as that of the filter 202), so that the perceptually weighted synthesized signal u′j (t) is obtained. can get. Here, in order to select the optimal excitation signal vector cj (t), AbS (Analysis by Synthesis) is used.
s) Method is used.

The AbS method will be described with reference to FIG. The multiplier 4 multiplies the excitation signal vector cj (t) in the codebook 3 by the gain γ, and the pitch prediction filter 5 adds a pitch component. After that, the weighted synthesis filter 7 generates a synthesized speech signal u′j (t). The difference between the synthesized voice signal and the input voice signal u (t) of the target is calculated by the subtracter 8, and an error signal ej (t) is generated. The power Pj of the error signal ej (t) is calculated by the power calculator 9 based on equation (5).

[0037]

(Equation 5)

Then, the error minimizing unit 10 searches for the excitation signal vectors cj, γ of the codebook 3 and the coefficients of the pitch prediction filter in which the power Pj is minimized.

The linear prediction coefficient, the index j of the excitation signal, the gain γ, the coefficient of the pitch prediction filter, and the pitch length are encoded and multiplexed by the encoding / multiplexing section 11 and sent to the transmission line 12. The transmission path 12 includes a wireless system, a wired system, and a storage system.

Next, the decoding unit shown in FIG. 2 will be described.

In the decoding unit, the transmitted code string is demultiplexed and decoded in the demultiplexer / decoding unit 13, and the linear prediction coefficient, the index j of the excitation signal, the gain γ, the lag (pitch length), The pitch prediction filter coefficients are obtained. An excitation signal vector cj (t) indicated by j in a codebook 14 having the same excitation signal vector as the codebook 3 of the encoding unit is converted by the multiplier 15 into γ.
The pitch component is added by a pitch prediction filter 16 having the same structure as that of the pitch prediction filter 5 of the encoding unit, and further, a transfer function F (z) represented by Expression (6) is added.

[0042]

(Equation 6)

Linear prediction filter (synthesis filter) having
17, the synthesized signal s' (t) is obtained. The linear prediction coefficient α ^(m) _i used in the equation (6) is obtained by calculating the linear prediction coefficient for each transmitted frame by the prediction coefficient interpolation unit 1
In FIG. 8, interpolation is performed for each subframe.

The prediction coefficient interpolation unit 18 has the same structure as that of the prediction coefficient interpolation unit 6, and performs a hyperbolic transformation, performs a linear interpolation, and then performs a hyperbolic inverse transformation to obtain a PARCOR coefficient for each subframe. .

Next, the prediction coefficient interpolation unit 6 of FIG. 1 will be described in detail.

The short-term prediction unit 1 in FIG. 2 obtains a P-order (P is a positive integer) PARCOR coefficient for each voice section of one frame. Thereafter, the PARCOR coefficient kj (j = 1 to P) is quantized. Hereinafter, this quantized PARCOR coefficient is used.

The prediction coefficient interpolating unit 6 calculates the PAROR coefficient obtained in the immediately preceding frame and the PAR coefficient obtained in the current frame.
By performing linear interpolation with COR coefficients, M (M is a positive integer) PARCOR coefficients in subframe units are obtained. This linearly interpolated PARCO coefficient is calculated using the low-order PAR
The COR coefficient is converted by the hyperbolic conversion unit 61, the result is interpolated by the linear interpolation unit 62, and then inversely converted by the hyperbolic inverse conversion unit 63. In addition, only the linear interpolation is performed by the linear interpolation unit 62 for the middle and higher order PARCOR coefficients.

The PARCOR coefficient determined by the above method is converted into a linear prediction coefficient αi (i = 0 to P), and the linear prediction coefficient obtained in each subframe is used. However,
Interpolation of the first-order PARCOR coefficient k1 is performed by calculating k for each frame.
1 is converted by the following equation (7), linearly interpolated, and inversely converted by the equation (8).

[0049]

(Equation 7)

[0050]

(Equation 8)

According to the above equations (7) and (8), k1 unevenly distributed to +1 is distributed as shown in FIG. In this conversion formula, as the value of the constant a increases, the inclination around ± 1 increases. The secondary PARCOR coefficient can be similarly converted and interpolated. In addition, a P value very close to ± 1
When the ARCOR coefficient is expressed by the LAR, the value becomes a very large value, and a large dynamic range is required. According to this conversion formula, such a PARCOR coefficient can be converted and expressed with a small dynamic range.

Next, referring to the flowchart of FIG.
The operation of the PARCOR coefficient interpolation processing according to the present embodiment will be described.

First, the PARCOR coefficient ki (order i = 1 to P) of the n-th frame is input to the prediction coefficient interpolator 6 (FIG. 1) in the input processing (401), and the transform i representing the order of the PARCOR coefficient is calculated. Initialized to 1 (402). PAR
The order of the COR coefficient is determined (403) and the lower order PAR is determined.
The COR coefficient is converted to yi (n) according to equation (7) (404), and the middle-high order PARCOR coefficient is yi (n).
Is performed (405). Then, the variable i representing the order is incremented (406). Find yi (n) in the PARCOR coefficients of all orders. The end of the process is determined (407). Then yi
(N) is interpolated with yi (n-1) of the previous frame to obtain y
i '(n) is obtained (408). Equation (9)
Shown in

[0054]

(Equation 9)

Referring back to FIG. 4, the PARC of the previous frame
The OR coefficient yi (n-1) is calculated using the PARCO of the current frame.
It is updated with the R coefficient yi (n) (409), and the initial value 1 is substituted for the variable i representing the degree (410), as in the above step 402.

After the processing, the low-order yi '(n) is subjected to hyperbolic transformation (Equation (8)) and transformed into PARCOR coefficients (411, 412). The middle and high order yi '(n) is directly substituted for the PARCOR coefficient (413). These conversion processes are performed for all the orders yi '(n), and thereafter, the variable i representing the order is incremented (414), and the end of the process is determined (415). The PARCOR coefficient thus obtained is applied to the processing on a subframe basis (416).

In the flowchart of FIG. 4, steps 402 to 407 are processed by the hyperbolic converter 61 of FIG. 1, steps 408 are processed by the linear interpolation unit 62 of FIG. 1, and steps 409 to 415 are processed by the inverse hyperbolic converter 63 of FIG. Is done.

In this embodiment, the hyperbolic functions shown in equations (7) and (8) are used as the conversion equations in the hyperbolic transformation step 404 and the hyperbolic inverse transformation step 412.
If the function has a characteristic of distributing and transforming the PARCOR coefficient unevenly distributed near ± 1 as shown in (1), the interpolation result is improved, and a higher-order curve having such a characteristic may be used. A hyperbolic function is appropriate from the viewpoint of the computational complexity of. The reason is that if the hyperbolic function is used, its inverse transformation is simple and it can be determined that it is suitable from the viewpoint of the amount of calculation.

As described above, according to the present invention, a linear interpolation taking into account the distribution of the linear prediction coefficients before interpolation is performed, thereby exhibiting good interpolation characteristics without conversion to LAR. The amount can be greatly reduced. That is, as an effect of reducing the influence of the interpolation error, as described in the embodiment, the error due to the interpolation of the presented PARCOR coefficient unevenly distributed around ± 1 is reduced by performing the conversion in consideration of the distribution. Further, as an effect of reducing the amount of calculation, in the interpolation of the linear prediction coefficient according to the present method, the calculation of the logarithm required for the calculation of the LAR becomes unnecessary. .

[0060]

The linear predictive coefficient interpolating apparatus according to the present invention provides a PARC of a predetermined order obtained from an input speech signal of a fixed time length.
Conversion means for hyperbolically transforming the OR coefficient, interpolation means connected to the conversion means for linearly interpolating the result of the hyperbolic transformation, and inverse means for inversely converting the linearly interpolated linear interpolation result connected to the interpolation means Conversion means,
The calculation of the logarithm required for the calculation of the AR becomes unnecessary, and the amount of calculation can be reduced when this method is actually realized on a computer.

In the linear predictive coefficient interpolating apparatus according to the present invention, the hyperbolic conversion means performs the hyperbolic conversion so as to have a steep slope as the predetermined-order PARCOR coefficient approaches the value of ± 1, so that it is close to ± 1. An error due to interpolation of unevenly distributed presentation PARCOR coefficients can be reduced by performing conversion in consideration of the distribution.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of an embodiment of a linear prediction coefficient interpolation device according to the present invention.

FIG. 2 is a block diagram illustrating a configuration example of a speech encoding / decoding device including the linear prediction coefficient interpolation unit illustrated in FIG. 1;

FIG. 3 is a diagram showing a PARCOR by a linear prediction coefficient interpolation unit shown in FIG. 1;
FIG. 9 is an explanatory diagram illustrating a distribution of coefficients.

FIG. 4 is a diagram showing a PARCOR by the linear prediction coefficient interpolation unit shown in FIG. 1;
It is a flow chart for explaining operation of coefficient interpolation processing.

FIG. 5 is an explanatory diagram of a conversion result from PARCOR to LAR.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Short-term prediction part 2 Perceptual weighting filter 3,14 Codebook 4,15 Multiplier 5,16 Pitch prediction filter 6,18 Prediction coefficient interpolation part 7,202 Weighting synthesis filter 8 Subtractor 9 Power calculation part 10 Error minimization part 11 Encoding and multiplexer unit 12 Transmission path 13 Demultiplexer, decoding unit 17 Synthesis filter 61 Hyperbolic transformation unit 62 Linear interpolation unit 63 Hyperbolic inverse transformation unit 201 Linear prediction filter

Continuation of front page (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 13/00 G10L 19/00-19/14 H03M 7/30 H04B 14/04

Claims (2)

(57) [Claims] 1. A conversion means for performing hyperbolic conversion of a predetermined order Percoll coefficient obtained from an input audio signal having a predetermined time length, and an interpolation means connected to the conversion means for linearly interpolating the hyperbolically converted result. A linear prediction coefficient interpolation apparatus, comprising: an inverse conversion means connected to the interpolation means for inversely converting the linearly interpolated linear interpolation result. 2. The linear prediction coefficient according to claim 1, wherein the hyperbolic transformation means performs the hyperbolic transformation so as to have a steep slope as the Percoll coefficient of the predetermined order approaches a value of ± 1. Interpolator.