JPH06118998A - Vector quantizing device - Google Patents

Abstract

PURPOSE:To provide a vector quantizing device which decreases a quantization error of a parameter exerting large influence on a speech signal and also reduces deterioration in speech quality even when the number of quantized bits is small in a vector quantizing device for an LSP parameter used for various speech signal processes. CONSTITUTION:Weighting coefficients of LSP parameters found by an LSP parameter analyzer 1 are controlled by a weighting coefficient controller 2 according to the distribution of the LSP parameters and the LSP parameters are quantized into vectors by a vector quantizer 3 by using the weighting coefficients and a code book 4. Further, quantization errors of some of the LSP parameters whose weight is small are quantized into vectors by a vector quantizer 5 while using a code book 6.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to vector quantization for performing vector quantization of a line spectrum pair (hereinafter referred to as LSP) parameter that represents vocal tract information of a voice used in a voice codec or voice synthesis. Regarding the device.

[0002]

2. Description of the Related Art In recent years, vector quantizers for LSP parameters have been used in low bit rate speech codecs and the like because they can express vocal tract information of speech with a smaller number of bits. Regarding the above LSP parameters, "Basics of Speech Information Processing", published by Ohmsha, November 3, 1981.
Issued 0th day, pp. 137-143.
A conventional LSP parameter vector quantizer will be described below.

FIG. 2 is a diagram showing a conventional LSP parameter vector quantizer. In FIG. 2, 11 is an LSP parameter analyzer for obtaining an LSP parameter from an input voice signal, 12 is a codebook, 13 is a vector quantization for quantizing the LSP parameter obtained by the LSP parameter analyzer using the codebook 12. It is a vessel. Further, 15 is an LSP quantization error quantizer for vector-quantizing the quantization error by the vector quantizer 13, and 14 is L
This is a codebook used when vector quantization is performed by the SP quantization error quantizer 15.

Next, the operation of the above conventional example will be described.
First, the LSP parameter analyzer 11 obtains an LSP parameter from a voice signal. The obtained LSP parameter is subjected to the operation shown in (Equation 1) by the vector quantizer 13, and a code that minimizes the sum of the square distances of the plurality of code vectors accumulated in the codebook 12 and the LSP parameter. The vector and the code vector number (quantization result) in the code book 12 are obtained.

[0005]

[Equation 1]

However, i: code vector number in code book 12 j: LSP parameter order e [i]: i code vector quantization error cb [i] [j]: quantization code book x [j]: jth-order LSP parameter Further, the quantization error between the vector quantized value obtained by the vector quantizer 13 and the LSP parameter obtained by the LSP parameter analyzer 11 is obtained, and this error is vectorized using the codebook 14. The quantizer 15 uses the above (Equation 1) to perform vector quantization. For a configuration in which such vector quantization is performed in two steps, see IEA ESPS magazine (I
EEE ASSP MAGAZINE) 1984 4
Mon, pp. 12-13.

[0007]

The LSP parameters are:
There is a characteristic that peaks on the frequency axis due to two adjacent parameters are shown, and the smaller the interval between these two parameters is, the steeper the peak is, which serves as information representing the characteristics of voice.

However, in the above-described conventional vector quantizer, quantization is performed irrespective of the size of the interval between adjacent parameters, and therefore a parameter showing a steep peak on the frequency axis is also quantized with the same weight as other parameters. As a result, there is a problem in that the quantization error becomes large, and the voice has a large influence such as the voice being deteriorated. In the case of performing vector quantization in two stages, the LSP
When vector-quantizing the parameters with a small number of quantization bits, the number of code vectors in the second stage is smaller than the vector length, which causes a problem that the quantization error becomes large.

The present invention solves the above-mentioned problem, and reduces the quantization error of the LSP parameter showing a steep peak waveform on the frequency axis, and when the vector quantization is performed in two steps, the second step is performed. An object of the present invention is to provide an excellent vector quantizer which can shorten the vector length in vector quantization to reduce the quantization error.

[0010]

In order to achieve the above object, the present invention obtains the difference between adjacent LSP parameters and varies the weighting of the LSP parameters at the time of vector quantization according to the difference. Is.

Further, in the case where the vector quantization is performed in two stages, the vector quantization in the second stage is performed by using only the quantization error of some parameters which are less weighted in the vector quantization in the first stage. It is the one.

[0012]

According to the present invention, when the difference between the adjacent LSP parameters is small, the weighting at the time of vector quantization of the adjacent LSP parameters is increased by the above-mentioned configuration, so that another LSP parameter
Make the quantization error smaller than the parameter. As a result, it is possible to reduce the quantization error of the parameter showing a steep peak on the frequency axis.

In the case where the vector quantization is performed in two stages, the vector quantization in the second stage can be executed only with some LSP parameters, and the number of code vectors can be increased compared with the quantized vector length. .

[0014]

FIG. 1 is a block diagram showing an embodiment of the present invention. In FIG. 1, reference numeral 1 is an LSP parameter analyzer for obtaining an 8th-order LSP parameter from an input voice signal. 2 is the LSP obtained by the LSP parameter analyzer 1
A weighting controller that selects which of a plurality of weighting factors accumulated based on the distribution of parameters is used,
3 uses the 8-bit codebook 4 to generate L of 1st to 8th order
It is a vector quantizer that vector-quantizes SP parameters. 5 uses the 4-bit codebook 6 to perform vector quantizer 3 for the 4th to 8th order LSP parameters.
Is a vector quantizer that vector-quantizes the quantization error due to the vector quantization executed in. Also (Table 1)
Shows an example of the weighting coefficient used in the weighting controller 2.

[0015]

[Table 1]

[0016]

[Equation 2]

Here, i: code vector number in codebook 4 j: LSP parameter order e ₁ [i]: i code vector quantization error w [j]: j-th parameter weighting coefficient cb ₁ [i ] [J]: Quantization codebook x [j]: jth-order LSP parameter

[0018]

[Equation 3]

Here, i: code vector number in codebook 6 j: order of LSP parameter e ₂ [i]: quantization error of i code vector cb ₂ [i] [j]: quantization codebook y [j ]: Quantization error vector in the first-stage quantization With respect to the vector quantization device configured as described above,
The operation will be described using (Table 1), (Equation 2), and (Equation 3). With respect to the LS parameter obtained by the LSP parameter analyzer 1, the weighting controller 2 sets the primary and secondary L
When the SP parameter difference is 0.02 or less, the weighting coefficient B in (Table 1), and when the difference between the secondary and tertiary LSP parameters is 0.04 or less, the weighting coefficient C in (Table 1) and the first order And the quadratic LSP parameter difference is 0.02 or less and the quadratic and cubic LSP parameter difference is 0.04 or less, the weighting coefficient D of (Table 1), and (Table The weighting factor A of 1) is selected. Next, using the weighting coefficient selected by the weighting controller 2 and the codebook 4, the vector quantizer 3 executes the vector quantization shown in (Equation 2). Specifically, in (Equation 2), e ₁
A code vector i that minimizes [i] is obtained, and this is output as quantized data.

Next, in the weighting coefficient of the weighting controller 2,
Is the 4th to 8th order LSP parameter with less weight, and
The vector quantization performed by the vector quantizer 3
The quantization error with the code vector i is calculated, and this quantization error
In the vector quantizer 5 using the codebook 6
The vector quantization shown in 3) is executed. Specifically, e
₂Find the code vector i that minimizes [i], and
Output as quantized data.

Therefore, according to the present embodiment, the weighting coefficient at the time of vector quantization is controlled based on the distribution of the LSP parameter, so that the quantization error of the LSP parameter showing a steep peak can be reduced. Further, by performing the second-stage vector quantization only with the quantization error of a part of the LSP parameters with small weighting,
More effective second-stage vector quantization can be realized, and LSP
It is possible to realize excellent vector quantization considering the characteristics of parameters.

[0022]

As is apparent from the above embodiment, the present invention controls the weighting coefficient at the time of vector quantization based on the distribution of the LSP parameter, thereby reducing the quantization error of the LSP parameter showing a steep peak. can do.

Further, by performing the second-stage vector quantization only with the quantization error of a part of the LSP parameters with less weighting, a more effective second-stage vector quantization can be realized. It is possible to realize excellent vector quantization in consideration of the characteristics of.

[Brief description of drawings]

FIG. 1 is a block diagram of a vector quantization device according to an embodiment of the present invention.

FIG. 2 is a block diagram of a conventional vector quantizer.

[Explanation of symbols]

 1 LSP parameter analyzer 2 Weighting controller 3 5 Vector quantizer 4 6 Codebook

Claims (2)

[Claims] 1. An analyzer for obtaining a line spectrum pair (hereinafter referred to as LSP) parameter representing vocal tract information of a voice, a weighting controller for selecting a weighting coefficient based on a distribution of LSP parameters, and A vector quantizer which quantizes the LSP parameter obtained by the analyzer using the weighting coefficient. 2. An analyzer for obtaining an LSP parameter representing vocal tract information of a voice, a first vector quantizer for quantizing the LSP parameter obtained by the analyzer, and the first vector quantizer. Weighting controller for selecting a weighting coefficient to be used for vector quantization in, and second for quantizing part of the quantization error by the first vector quantizer.
Vector quantizer and a vector quantizer.