KR0138868B1 - Lsp frequency quantizer - Google Patents

Abstract

The present invention relates to an LSP frequency quantizer, and proposes an efficient LSP quantization method using a relatively simple structure, a small amount of computation, and a small memory space using a DPCM AQBW quantization method in an LSP speech coding system. It is an object of the present invention to provide an LPS frequency quantizer that provides further improved speech quality.

Description

LSP Frequency Quantizer

1 is a diagram illustrating an encoding structure of a QCELP speech coder to which the present invention is applied.

2 is a diagram illustrating a decoding structure of a QCELP speech coder to which the present invention is applied.

3 is a quantization structure diagram of an LSP frequency into an LSP transmission code according to the present invention.

4 is an inverse quantizer structure of the LSP frequency of the LSP transmission code according to the present invention.

The present invention applies a mixed quantization method of DPCM (Differential Pulse Code Modulation) and AQBW (Adaptive Quantization Backward) to a conventional LSP frequency quantization method in a speech coder using an LSP (Line Spectrum Pair) to improve LSP-specific sequential ordering characteristics. In addition to the advantages of DPCM quantization, LSP frequency quantizers can be used to improve the quality of speech by enabling more efficient quantization than conventional quantization schemes. Instead of extracting the parameters from the voice and transmitting the voice directly, these parameters are sent instead.

The receiver uses this parameter to synthesize the voice again, ultimately reducing the amount of data transmitted under a limited transmission environment, which is bandwidth.

In particular, in speech coding, efficient extraction of representative parameters and optimal quantization of extracted parameters are the key.

The QCELP (Qualcomm Code Exited Linear Prediction) speech coder adopted by CDMA digital mobile communication performs speech coding using LSP, pitch parameter, and codebook parameter. Especially, short term correlation of speech Quantization of LPC (Linear Prediction Coefficient) coefficients is an important factor in determining overall speech quality.

Since the quantization of LPC coefficients itself has a relatively large active area, it is very rare to quantize the actual LPC itself, which transforms the LPC coefficients into coefficients of other characteristics that are mathematically identical and advantageous to quantization, thereby quantizing the transformed coefficients. Methods are widely used.

However, quantization of LSPs that are good in the conversion to LSP frequency, which is one of the representative LPC conversion methods, has been suggested so far, but due to the enormous amount of storage memory and a large amount of computation, it is impossible to adopt them in real time systems.

The present invention devised to solve all the problems of the prior art, proposes an efficient LSP quantization method using a relatively simple structure, a small amount of calculation, a small memory space by using the DPCM AQBW quantization method in the speech coding system using the LSP The purpose is to provide an LSP frequency quantizer that provides improved speech quality.

Hereinafter, with reference to the accompanying drawings will be described an embodiment of the present invention;

First, a problem to be first performed for digital speech encoding according to the present invention is a problem of encoding and reducing speech information on how to efficiently encode and convert a continuous speech signal into a digital form.

The demand for voice reduction has arisen from the beginning of digital transmission of voice and is essential for the efficient use of bandwidth-limited communication channels.

In general, there are three digital speech coding schemes.

First, Waveform Coding, which attempts to faithfully reproduce speech waveforms with a relatively high transmission rate, and Waveform Protection, which encodes coefficients that assume an active model of speech with a low transmission rate and characterize the active model. Source Coding, Hybrid Coding, which takes both of the above advantages and encodes them.

The present invention relates to a speech coder that quantizes LPC coefficients used in most wave coding and mixed coding schemes to an LSP frequency, which is advantageous for quantization, and quantizes the converted LSP values into digital values. can do.

1 is an encoding structure diagram of a QCELP speech coder used in a digital CDMA mobile communication system to which the present invention is applied, and Hamming function W _H (N) in the figure is as follows.

In addition, Rate denotes a transmission rate by a variable rate algorithm, ∑ () ² denotes an average square error, and PF (z) denotes a rear filter.

The speech coder features a hybrid coding scheme and includes five parameters for speech coding: LSP frequency, pitch gain, pitch lag, codebook gain, and codebook index. Used for

The pitch parameter is for modeling long term intervals (frame to frame) of speech, long term correlation, and codebook parameters are used to vector quantize the excitation signal of speech.

The most important factor in speech coding is the LPC coefficient, which measures the short-term similarity of speech.

This is an important coefficient that is adopted in most low-speed speech coders by measuring similarity between each element of speech in a short period of 10-20ms (hereinafter referred to as a frame).

These coefficients are converted into other parameters that have a relatively large active area and are easy to quantize.

Among them, the QCELP speech coder converts LPC coefficients into LSP frequencies and then quantizes them.

This conversion to LSP is used in many speech coders because it involves effective quantization, interpolation and proper stability checking.

FIG. 2 is a structural diagram of synthesizing speech using five parameters by inverse quantization of five parameters encoded and quantized by a QCELP speech encoder.

Next, the quantization of the present invention will be described.

First, the LSP quantization method according to the present invention adopts the DPCM-AQBW method.

This quantization method is divided into a quantization algorithm of a transmitter and an inverse quantization algorithm of a receiver. First, a quantization algorithm of a transmitter is as follows.

The formant filter coefficient LPC values for the current frame are calculated by Durbin's recursive algorithm.

A (z) = 1-a ₁ z ^-1- . … a ₁₀ z ^-10 (1)

These LPC coefficients are converted to LSP, which is the same coefficient in mathematical terms.

These LSP frequencies have a mathematical characteristic that, if the formant filter composed of LPC coefficients is stable, their magnitude is between 0 and 0.5 and can be arranged sequentially according to the index value of the coefficients as shown in Equation 2.

W ₁ W ₂ W ₃ W ₄ W ₅ W ₆ W ₇ W ₈ W ₉ W ₁₀ (2)

Once the LSP frequencies have been calculated and the data rate selected, each LSP frequency is quantized for transmission.

The structure of the LSP quantizer is shown in FIG.

Ten LSP frequencies exist around each bias value (these frequencies are the same as when the input speech has flat spectral characteristics and formant prediction cannot be performed). The bias value used for each LSP frequency is shown in Equation 3.

Where P is 10.

The predictor P _W is equal to Equation 4.

There is one predictor for each LSP frequency.

The quantizer Q _Wi for the i th LSP frequency is a linear quantizer whose change area and step size change according to the transmission speed, and are quantized as follows.

In FIG.

As the input value of the LSP quantizer, the bias value Bias _i is subtracted from the i th LSP frequency value W _i , and the error value e _i (n) is obtained by subtracting the predictor value from this value.

Dequantized (n) is added to the predicted value to obtain quantized W _i (n).

Quantized (n) is used as shown in Equation 9 to predict the LSP frequency for the next frame.

Here, e _i (n) is used as the quantizer input value, and its active area is determined by the values shown in Table 2.

For example, at transmission rate 1, the active region of e _i (n) is between -0.025 and +0.025.

Linear quantization in this region is DPCM LSP quantization.

However, if we use the ordering property of LSP of Equation 2, it has the following characteristics.

That is, e _i (n) must be smaller than x _i (n) due to the sequential heat characteristics.

Therefore, the active region of e _i (n) is changed as follows.

if Or if i = 10 the original quantization region Quantize within

if, If the quantization region And then quantize within this region.

Quantizer

Here, Q _ti (x) is as follows.

if Or if i = 10

if Back side

Where n represents the nth frame and N is the number of bits in the quantization Is the quantization level for the i th coefficient, and round (x) is a function rounded to the nearest integer of x.

The number of LSP quantization bits N according to the transmission rate is shown in Table 1, and the maximum quantization level according to the transmission rate. Is shown in Table 2.

Note that Eq. 11 is a limiting function to ensure that Q _wi (x) is between 0 and 2 ^N −1.

Therefore, after the quantization process described above, the transmitter transmits LSP _i , which is a transmission code of the quantized LSP frequency.

The dequantizer structure at the receiving end is shown in FIG.

Where P (z) is equal to Eq. 4, this predictor is updated every frame, and the bias is equal to Eq. Is expressed as

if Or if i = 10

if Back side

When the LSP quantization and inverse quantization are performed as described above, the present invention is capable of more efficient quantization than conventional LSP quantization using a relatively simple structure, a small amount of computation, and a small memory space. It is effective to construct a speech encoding system having a speech quality of.

Claims (2)

In a Line Spectrum Pair (LSP) frequency quantizer using a mixed quantization of DPCM (Differential Pulse Code Modulation) and AQBW (Adaptive Quantization Backward), a first addition means for subtracting an LSP frequency and a frequency bias value corresponding to each LSP frequency ( 31); Second adding means (32) for predicting an LSP frequency for the next frame to an output value of the first adding means (31); When determining the quantized transmission LSP parameter by varying the size of the change region and the step according to the transmission speed of the voice according to the transmission speed of the voice, the LSP using the following equation (10) A linear quantizer 33 which quantizes the quantization range by using Equation (12) and Equation (13); An inverse quantizer (34) for inversely quantizing the LSP parameter of the linear quantizer (33); Third adding means (35) for adding the output value of the inverse quantizer (34) to the predicted value; And predicting means (36) for predicting predicted values input to the second and third adding means (32,35) through the addition result of the third adding means (35). (Equation (10) is Equation 12 is Equation (13) is In a Line Spectrum Pair (LSP) frequency dequantizer using mixed quantization of DPCM (Differential Pulse Code Modulation) and AQBW (Adaptive Quantization Backward), the quantized transmission LSP parameters are expressed using Equations (14) and (15). Inverse quantization means 41 for inverse quantization; First adding means (42) for adding a predicted value to an output value of said inverse quantization means (41); Predicting means (43) for predicting the predicted value through the added value of the first adding means (42); And a second adding means (44) for outputting the dequantized LSP frequency by adding a bias value to the added value of the first adding means (42). (Equation 14 is