JPH03263100A - Audio encoding and decoding device - Google Patents

Abstract

PURPOSE:To improve the quantization characteristic of a high frequency band by dividing a rediaul spectrum and a spectrum envelope to a low frequency band and the high frequency band and constituting a residual vector and a spectrum envelope vector in each band. CONSTITUTION:A band-classified residual vector constituting means 10 and a band-classified spectrum envelope vector constituting means 11 constitute vectors in each band. That is, the spectrum envelope and the residual spectrum are divided to two high and low bands, and the spectrum envelope vector and the residual vector are constituted in each band, and thereby, weighting of only the low frequency component due to the spectrum envelope vector is prevented by vector quantization of the residual vector. Thus, the quantization distortion of the high frequency component is improved and the quality in the high frequency area of a reproduced audio signal is improved.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to an improvement in a method for quantizing a linear prediction residual signal in a speech encoding/decoding device used when digitally transmitting or storing speech signals.

[Prior art] An input audio signal is subjected to linear predictive analysis for each analysis frame to extract a linear predictive residual signal (hereinafter referred to as a residual signal), and after converting this residual signal into frequency domain components, multiple The method of dividing into vectors and vector quantizing each vector is as follows.
This is one of the effective methods for high-efficiency compression encoding of speech, and a conventional speech encoding/decoding apparatus using this method is shown in FIG.

Figure 3 is by Suda, “Application of 9tani r8kbit/s TCWVQ coding to land mobile communication” IEICE technical research report vol. 1.88, no. 453, SP8g-138
It is similar to that shown in FIG. 1 of (February 1989).

The operation of the conventional device will be explained below.

First, the encoding section shown in FIG. 3 (3a) will be explained.

Input audio signal expressed as a time series of discrete value samples (1)
A linear predictive analysis means (3) performs a linear predictive analysis on each fixed-length analysis frame (hereinafter referred to as a frame) to obtain a linear predictive coefficient. The linear prediction analysis means (3) outputs the obtained linear prediction coefficients (5) to the linear prediction inverse filter means (2), the spectral envelope extraction means (7), and the multiplexing means (18). The linear predictive inverse filter means (2) receives the input audio signal (1
) is subjected to linear prediction inverse filtering using the linear prediction coefficient (5) for each frame to obtain a residual signal (4). The conversion means (6) to the frequency domain is DCT (Discrete
Co51ne Transf.

rm), the residual signal (4) is converted for each frame into a sample sequence indicating the frequency domain components, and this is output as a residual spectrum (8). The spectral envelope extraction means (7) extracts the input audio signal (1) from the linear prediction coefficient (5).
Find the frequency spectrum envelope of and output it as the spectrum envelope (9). Note that the spectral envelope (9) is also expressed by a sample sequence in the frequency domain. The residual vector constructing means (30) constructs a plurality of vectors having each component of the residual spectrum (8) as an element from the residual spectrum (8), and outputs this as a residual vector (12). The spectral envelope vector construction means (31) constructs a plurality of vectors using each component of the spectral envelope (9) from the spectrum envelope (9) in the same manner as the residual vector construction means (29), and Spectral envelope vector (
13).

A specific example of vector construction performed by the residual vector construction means (29) and the spectral envelope vector construction means (30) is shown in FIG. In FIG. 4, (4a) is the spectral envelope (9), and (4C) is the spectral envelope vector (13) constructed by the spectral envelope vector construction means (31). (4b) is the residual spectrum (8
) and (4d) is the residual vector (12) constructed by the residual vector constructing means (30).

Figure 4 shows the spectral envelope (4a) and the residual spectrum (4a).
b) consists of 15 frequency domain samples each, and 5 spectral envelope vectors (4C) from each.
The case where the residual vector (4d) is constructed is shown.

As shown in Figure 4, the spectral envelope vector (4
c) and the construction of the residual vector (4d) are performed by extracting frequency domain samples at equal intervals. The first vector of the spectral envelope vector (4C) and residual vector (4d) uses the lowest frequency sample as the starting point of extraction, and from then on, the starting point of extraction is shifted one point at a time in the direction of higher frequencies until the fifth vector is extracted. Configuration takes place.

The vector quantization means (14) is a residual vector construction means (
In step 31), a plurality of residual vectors (12) are vector quantized using registered vectors (16) registered in advance that are output from a vector codebook (15). Specifically, the registered vector (16) with the smallest distance value Dm from the residual vector (12) given by equation (1) is searched for, and the number of the registered vector is determined by the residual quantization result (17). )
Output as .

1=0 (m=0.M
-1> In equation (1), Wl is the spectral envelope vector (13), El is the residual vector (12), c i "ゝ is the m-th registered vector (16), and I is the sample of the residual spectrum (8). The number of vector dimensions is determined by the number of component residual vectors (12), and m is the number of registration vectors (16) determined depending on the amount of quantization bits allocated to vector quantization. (1 ) is understood as a calculation formula for the distance between vectors (El and CI) weighted by the spectral envelope vector (13).The quantization bits are equally allocated to each residual vector.

When considering the frequency components of the input audio signal (1), samples at frequency points with large spectral envelope values are more important, and the residual spectrum (8) at those points is more important than frequency points with small spectral envelope values. It is necessary to quantize with a small quantization error. Weighting by the spectral envelope vector (13) shown in equation (1) corresponds to the necessary conditions of using inter-vector distance calculation for vector quantization of the residual vector (12). Furthermore, if the spectral envelope vector (4c) and residual vector (4d) are extracted using the method shown in Figure 4, the average power obtained by adding and averaging the samples within each spectral envelope vector will be approximately equal between the vectors, so Vector quantization means (1
4) has the advantage that when vector quantizing the residual vector (4d), equal quantization values +- can always be given to each residual vector.

The residual quantization result (17) output from the vector quantization means (14) just described and the linear prediction coefficient (5) output from the linear prediction analysis means (3) are multiplexed by the multiplexing means (18).
The signals are multiplexed and output as encoded audio information to a transmission path (19).

Next, the decoding section shown in FIG. 3 (3b) will be explained.

The separating means (20) converts the encoded speech information input from the transmission path (19) into linear prediction coefficients (5) and residual quantization results (1).
7) Separate. The vector dequantization means (21) converts the residual quantization result (17) into a vector using the registered vector (30) output from the same vector codebook (29) as the vector codebook (15) of the encoding section. A residual vector (22) is obtained by inverse quantization. The residual reproduction means (33) converts each sample of the plurality of residual vectors (22) outputted from the vector dequantization means (21) into the residual vector composition means (3a) of the encoding section (3a) in the frequency domain. 31), and obtain a reproduced residual spectrum (24). A time domain conversion means (25) converts the reproduced residual spectrum (24) into the time domain by inverse DCT to obtain a reproduced residual signal (26). The linear prediction synthesis filter means (27) synthesizes and outputs a reproduction signal (28) by performing linear prediction synthesis filtering processing using the reproduction residual signal (26) and the linear prediction coefficient (5).

[Problem to be Solved by the Invention] As explained above, according to the conventional speech encoding/decoding device, the configuration of the spectral envelope vector is such that the samples on each frequency domain of the spectral envelope are arranged at regular intervals over the entire frequency domain. Since this is done by extraction, for phonemes such as vowels that generally have a characteristic in which the amount of low frequency components is considerably larger than that of high frequency components, each spectral envelope vector is all as shown in Figure 4 (4C). Similarly, it has strong power only on the lower frequency side. Therefore, when the input audio signal has such intersonic characteristics, only the low frequency components are weighted and emphasized by the spectral envelope vector during vector quantization of the residual vector, resulting in large quantization distortion of the high frequency components. As a result, there was a problem in that the quality of the reproduced audio signal in the high frequency range deteriorated.

This invention was made to solve the above problems, and it divides the spectral envelope and residual spectrum into two bands, a low band and a high band, and creates a spectral envelope vector and a residual vector for each band. By configuring this, only low frequency components are not weighted by the spectral envelope vector in vector quantization of the residual vector.

[Means for Solving the Problems] A speech encoding/decoding device according to the present invention divides a linear prediction residual signal converted into a frequency domain sample sequence into a plurality of frequency bands, and performs processing for each frequency band. A band-specific residual vector constructing means for constructing a preset number of residual vectors, and a frequency spectrum envelope of an input audio signal represented by a frequency domain sample sequence using the same method as the band-specific residual vector constructing means. A band-specific spectral envelope vector configuring means for configuring a plurality of spectral envelope vectors, and a residual vector configured by the band-specific residual vector configuring device using a spectral envelope vector configured by the band-specific spectral envelope vector configuring device. The encoding section includes a vector quantization means for vector quantizing the residual vector quantized by the vector quantization means, and a vector dequantization means for vector dequantizing the residual vector vector quantized by the vector quantization means; It is equipped with a residual reproduction means for arranging each sample of the residual vector that has been inverse vector quantized by the childization means at a frequency position in the frequency domain before residual vector composition by the band-based residual vector composition means. be.

[Operation] The band-based residual vector composition means of the present invention divides the linear prediction residual signal converted into a frequency domain sample sequence into a plurality of frequency bands, and divides the linear prediction residual signal converted into a frequency domain sample sequence into a plurality of frequency bands, and divides the linear prediction residual signal converted into a frequency domain sample sequence into a plurality of frequency bands. Samples are sequentially extracted at regular intervals preset for each frequency band using the sample as a starting point, and the extracted samples are used as elements to construct the first residual vector in each frequency band, and then the sample that becomes the starting point is By shifting one point at a time in the direction of higher frequencies and extracting samples at the constant intervals, a number of residual vectors set in advance for each frequency band after the second one is constructed, and the band-specific spectral envelope vector construction means , the vector quantization means constructs a plurality of spectral envelope vectors from the frequency spectrum envelope of the input audio signal represented by a frequency domain sample sequence in the same manner as the band-wise residual vector constructing means; Each residual vector constructed by the vector construction means is loaded with each element sample value of the spectral envelope vector constructed by the band-based spectral envelope vector construction means, which has the same positional relationship in the frequency domain. The vector quantization means vector quantizes the residual vector quantized by the vector quantization means, and the residual reproduction means vector quantizes the residual vector by vector-to-vector distance calculation. Element samples of the dequantized residual vector are arranged at positions on the frequency domain before vector construction by the band-specific residual vector construction means.

[Example] An embodiment of the present invention will be described below with reference to FIG.

In FIG. 1, the same parts as in FIG. 3 are designated by the same reference numerals, and their explanation will be omitted.

The operation of one embodiment of the present invention will be described below.

First, the encoding section shown in FIG. 1(1a) will be explained.

The band-specific residual vector construction means (10) divides the residual spectrum (8) into two frequency bands, low and wide, and extracts the sample of the first residual spectrum having the lowest frequency within each frequency band. Samples are extracted at regular intervals set in advance for each frequency band using the
The second residual vector (12) is constructed, and the starting sample is shifted one point at a time in the direction of higher frequencies, and the second and subsequent residual vectors (12) are constructed in a preset number for each frequency. .

The band-based spectral envelope vector construction means (11) constructs a spectral envelope vector (13) from the spectrum envelope (9) in the same manner as the band-based residual vector construction means.

FIG. 2 shows a specific example of vector construction performed by the band-based residual vector construction means (10) and the band-based spectral envelope vector construction means (11). In FIG. 2, (2nd) is the spectrum envelope (9), and (2C) is the spectrum envelope vector (13) constructed by the band-based spectrum envelope vector construction means (11). In addition, (2b) is the residual spectrum (8), and (2d) is the residual vector (12) configured by the band-specific residual vector configuration means (10) to calculate the distance during vector quantization (formula (1) ), only the low frequency components are no longer weighted with emphasis by the spectral envelope vector.

Next, the decoding section shown in FIG. 1(1b) will be explained.

The residual reproduction means (23) is a vector inverse quantization means (21).
Each sample of the plurality of residual vectors (22) output from the band-specific residual vector constructing means (1
0) is placed at the original position before the vector configuration performed, and the reproduced residual spectrum (24) is obtained.

In the above embodiment, the band-specific residual spectrum configuration means (
lO) and band-specific spectral envelope vector construction means (11
) divided the residual spectrum (8) and spectral envelope (9) into two bands, a low frequency band and a high frequency band, but
These may be divided into three or more bands. Further, in the vector quantization means (14), more quantization bits are always fixedly assigned to the residual vector formed from the low frequency band than to the residual vector formed from the high frequency band. Figure 2 shows the spectral envelope (2a) and the residual spectrum (
2b) consists of 15 frequency domain samples each, which is divided into 9 samples for the low frequency band and 6 samples for the high frequency band, and 3 and 2 spectral envelope vectors (2C ) and residual vector (
2d) is shown.

The vector quantization means (14) converts the residual vector (12) output from the band-specific residual vector composition means (10) into a spectral envelope vector (13) output from the band-specific spectrum envelope vector composition means (11). Vector quantization is performed using a distance calculation formula (formula (1)) using .

Since the band-specific residual vector configuration means (10) and the band-specific spectral envelope vector configuration means (11) perform vector configuration for each band, even if the spectral envelope has a large component only in the low frequency band, the configured spectral envelope All of the vectors do not have strong power only on the low frequency side as explained in the example of the conventional device (Figure 2 (2C:)
), or the quantization bit amount may be set for each band by controlling the quantization bit amount according to the average power value of the spectral envelope of each frequency band.

[Effects of the Invention] As described above, according to the present invention, the residual spectrum and the spectral envelope are divided into a low frequency band and a high frequency band, and a residual vector and a spectral envelope vector are configured for each band. Therefore, even if the spectral envelope has a strong component only in the low frequency band, the weighting using the spectral envelope vector in calculating the distance between vectors when vector quantizing the residual vector may be biased only in the low frequency band. However, the quantization characteristics in the high frequency band are improved compared to conventional devices, and it has the effect of obtaining high-quality reproduced audio signals.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is an explanatory diagram explaining the operation of the embodiment shown in FIG. 1, and FIG. 3 is a block diagram showing a conventional audio encoding/decoding device. 4 are explanatory diagrams for explaining the operation. In the figure, (1) is the input audio signal, (2) is the linear prediction inverse filter means, (3) is the linear prediction analysis means, (4) is the residual signal, (5) is the linear prediction coefficient, and (6) is Conversion means to frequency domain, (7) is spectrum envelope extraction means, (8
) is the residual spectrum, (9) is the spectral envelope, (l
O) is a band-specific residual vector composition means, (11) is a band-specific spectral envelope vector composition means, (12) is a residual vector, (13) is a spectral envelope vector, (
14) is a vector quantization means, (15) is a vector codebook, (16) is a registered vector, (17) is a residual quantization result, (18) is a multiplexing means, (19) is a transmission line, (2
0) is a separation means, (21) is a vector dequantization means,
(22) is the residual vector, (23) is the residual reproduction means,
(24) is the reproduced residual spectrum, (25) is the time domain conversion means, (26) is the reproduced residual signal, (27) is the linear prediction synthesis filter means, (28) is the reproduced audio signal, (
29) is a vector codebook, (30) is a registered vector,
(31) is a residual vector composition means, (32) is a spectral envelope vector composition means, (33) is a residual reproduction means,
(2a) is the spectral envelope, (2b) is the residual spectrum, (2'c) is the spectral envelope vector, (2d) is the residual vector, (4a) is the spectral envelope, (4b
) is the residual spectrum, (4C) is the spectrum envelope vector, and (4d) is the residual vector. Note that the same reference numerals in the figures indicate the same or corresponding parts.

Claims (1)

[Claims] Linear prediction analysis is performed on the input audio signal expressed as a time series of discrete value samples for each analysis frame of a certain length, and the linear prediction residual signal of the input audio signal within the analysis frame is calculated using the obtained linear prediction coefficients. An encoding unit that calculates the linear prediction residual signal in the analysis frame and converts it into a frequency domain sample sequence, vector quantizes it, and outputs it, The linear prediction residual signal is vector quantized and then converted into a time domain sample sequence, and the reproduced audio signal is generated by linear prediction synthesis filter processing using the linear prediction residual signal converted to the time domain sample sequence and the linear prediction coefficients. In the audio encoding/decoding device, the encoding unit includes a spectral envelope extraction means for obtaining a frequency spectrum envelope of the input audio signal in the analysis frame as a frequency domain sample sequence; The linear prediction residual signal of the analysis frame converted into a sample sequence in the frequency domain is divided into a plurality of frequency bands, and the first sample with the lowest frequency in each frequency band is the starting point. Samples are extracted sequentially at regular intervals set to band-specific residual vector construction means for constructing a preset number of residual signal vectors for each frequency band after the second one by extracting samples at regular intervals; and input obtained by the spectrum envelope extraction means. A band-specific spectral envelope vector configuring means for configuring a plurality of spectral envelope vectors from a sample sequence of a frequency spectrum envelope of an audio signal in the same manner as the band-specific residual vector configuring means, and a band-specific residual vector configuring means. calculating an inter-vector distance using the residual signal vector obtained as a weight by each element sample value of the spectral envelope vector constructed by the band-based spectral envelope vector construction means having the same positional relationship in the frequency domain; The decoding section includes vector dequantization means for vector dequantization of the residual signal vector quantized by the spectrum quantization means, and vector dequantization means for vector dequantization of the residual signal vector quantized by the spectrum quantization means. The method is characterized by comprising a residual reproduction means for arranging all the element samples of the residual signal vector vector-dequantized by the means at positions in the frequency domain before vector construction by the band-based residual vector construction means. Audio encoding/decoding device.