JPH09232964A - Variable block length converting and encoding device and transient state detecting device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To highly precisely detect the transient state of an input signal through a simple configuration. SOLUTION: A transient state detection circuit 2 is constituted so that the input signal is fetched for every prescribed sections, and a sum of squares is obtained respectively, and the transient state of the above-mentioned input signal is detected on the basis of the degree of the change of the signal whose sum of squares is calculated for every section extending over, at least, two sections. Then, the transient state is made capable of being detected by only executing the calculation of the sum of squares of the input signal on a time base without executing orthogonal transformation processing or filter processing.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a variable block length transform coding device and a transient state detecting device.

[0002]

2. Description of the Related Art For high efficiency encoding of audio data,
There are band division encoding that divides audio data into multiple bands on the time axis and encodes it, and transform encoding that orthogonally transforms audio data on the frequency axis and divides it into multiple bands and encodes it. . There is also high-efficiency coding in which these coding methods are combined to divide audio data into a plurality of bands on the time axis, and each band signal is further orthogonally converted and coded on the frequency axis.

As an example of the prior art, the MDC shown in FIG.
Block length variable transform coding using T (Modified Discrete Cosine Transform) will be described. In FIG. 3, the audio data input from the audio data input terminal 31 has N / 2 samples (N is, for example, 512) as one block and combines the N / 2 samples of the previous block and the N / 2 samples of the current block. The N samples are input to the transient state detection circuit 32 and the MDCT circuit 33.

The transient state detection circuit 32 determines the orthogonal transform size of the MDCT processing performed by the MDCT circuit 33 according to the property of the input N-sample signal, and determines it by MDC.
It is supplied to the T circuit 33. The MDCT circuit 33 orthogonally transforms input N-sample audio data by MDCT according to the orthogonal transformation sample size supplied from the transient state detection circuit 32, and obtains N MDCT coefficient data obtained by the block floating point. Conversion circuit 34
Output to

The block floating point conversion circuit 34 is an MD
The N MDCT coefficient data input from the CT circuit 33 is converted into floating point data represented by a mantissa part and an exponent part, the exponent part data is output to the bit allocation circuit 35, and the mantissa part data is quantized. Output to the encoding circuit 36.

The bit allocation circuit 35 determines the quantization bit length of the mantissa data based on the exponent data input from the block floating point conversion circuit 34 by utilizing human auditory masking characteristics and the like. Quantization coding circuit 3
6 is output. The quantization encoding circuit 36 extracts data from the MSB (most significant bit) of the mantissa data input from the block floating point conversion circuit 34 by the amount of the quantization bit length input from the bit allocation circuit 35, and quantizes it. Performs encoding processing.

As a result of the above processing, the exponent part data is output from the first data output terminal 37, and the encoded mantissa part data (hereinafter referred to as mantissa part encoded data) is output from the second data output terminal 38. Is output. These output data are accumulated in a recording medium (not shown) or transmitted via a transmission path (not shown).

[0008]

By the way, in the coding using the orthogonal transform, the value of N, which is the orthogonal transform size, is generally set from the viewpoint of the coding efficiency of realizing high sound quality with a high compression rate. It is better to make it larger.

However, as shown in FIG. 4, when the characteristics of the input signal change abruptly within N samples (in the period of 512 samples in FIG. 4), for example, a sharp rise of the signal occurs. In such a case, if such a signal is encoded and then decoded, as shown in FIG. 5, the ratio of noise due to encoding increases in a portion where the signal level is low, and it may be detected as a pre-echo. there were.

Therefore, in the transient state detection circuit 32,
The method of analyzing the property of the input signal and adaptively switching the orthogonal transform size has been conventionally used. In this method, when the input signal is in a steady state, orthogonal transformation of N sample size is performed once to transform into N frequency data. When the input signal is in a transient state, N samples are divided into a plurality of sections, orthogonal conversion is performed on each section, and N pieces of frequency data are converted.

For example, N samples are divided into N / 2 samples in the first half and N / 2 samples in the second half, and orthogonal transform of N / 2 sample size is performed twice in total in the first half and the second half to obtain N frequency data. Convert to. In another example, N samples are divided at equal intervals of N / 4 samples, and a total of four orthogonal transforms are performed to transform into N frequency data. As a result, as shown in FIG. 6, the section in the transient state is shortened (in this FIG. 6, by 256 samples), the range covered by the coding noise is narrowed so that it is not detected as a pre-echo. .

As a method for detecting a transient state, FIG.
Feedback type detection method shown in (Technical report: EA90-65, P
P.23-30, 1990. "Adaptive block length adaptive transform coding (ATC-AB
The 256 Kbps stereo high-fidelity audio encoder / decoder by S) ") is considered to have the best detection accuracy.

The feedback type detection method shown in FIG. 7 performs several orthogonal transforms of N sample size and a shorter sample size, for example, N / 2, N / 4 shorter sample size at the same time. , Encoding / decoding processing is performed. Then, a method of finally selecting a sample size having a low noise level from those sample sizes. Thus, in the feedback type detection method,
Since the orthogonal transform size with the lowest noise level is selected, it is considered that almost the best sound quality can be achieved in terms of hearing.

In addition to this, N samples are divided into a plurality of sections and orthogonal transformation is performed for each section (for example,
The N samples are divided into the first half N / 2 samples and the second half N / 2 samples, and the orthogonal transformation of the N / 2 sample size is performed twice in the first half and the second half), and the frequency data for the first half N / 2 samples is obtained. There is also a method of detecting the transient state based on the degree of change with the frequency data for the N / 2 sample in the latter half.

Further, without performing the orthogonal transformation, the low frequency component of the input signal on the time axis is cut off by the high pass filter,
There is also a method of detecting a transient state based on the degree of change of the high frequency component of the input signal obtained as a result.

However, the above method of detecting a transient state has a drawback that the processing becomes complicated if the detection accuracy is improved. Further, since the above-mentioned detection method performs orthogonal transform processing and filter processing, it has a drawback that the calculation load is large and it is not suitable for real-time processing of encoding.

The present invention has been made in order to eliminate such drawbacks, and an object of the present invention is to make it possible to detect a transient state of an input signal with high accuracy with a simple structure.

[0018]

According to the present invention, an input signal is taken in for each predetermined section to obtain a sum of squares, and 2 is calculated for each section.
It has a transient state detection circuit that detects the transient state of the input signal by the degree of change of the summed signals over at least two sections. This transient state detection circuit does not perform orthogonal transform processing or filter processing, but has a certain section of the input signal on the time axis (for example, N / 2 or N when the maximum orthogonal transform size of the orthogonal transform circuit is N). The sum of squares is obtained for each (/ 4 size section), and the transient state is detected by the degree of change in the value of the sum of squares. According to the present invention having such a configuration, it is possible to detect the transient state only by performing the square sum calculation of the input signal on the time axis.

[0019]

DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram showing a configuration example of a variable block length transform coding apparatus according to the present embodiment. In the present embodiment shown in FIG. 1, the MDCT circuit 3 has a maximum N
The orthogonal transform of the sample size is possible, and the orthogonal transform of the N sample size or the N / 4 sample size is performed according to the control signal from the transient state detection circuit 2.

In FIG. 1, the audio data input from the audio data input terminal 1 has N samples (N
Is, for example, 512) as one block, and the buffer circuit B
It is temporarily stored in the UF. The audio data stored in the buffer circuit BUF is then input to the transient state detection circuit 2 and the MDCT circuit 3. In the following, for simplicity, N = 512 will be described.

The transient state detection circuit 2 receives 512 as an input.
Sample signal is N / 4 samples (= 128 samples)
Each of the small blocks is divided into four small blocks, the sum of squares of samples is calculated for each small block, and 2 between adjacent small blocks is calculated.
Calculate the ratio of sums of products. When the ratios of the sums of squares of the four small blocks are all smaller than the predetermined value, the MDCT circuit 3 is instructed to perform the orthogonal transform of 512 sample size. Further, when the ratio of the sum of squares of the four small blocks becomes larger than a predetermined value between at least one adjacent small blocks, it is instructed to perform the orthogonal transform of 128 sample size.

When the MDCT circuit 3 is instructed by the transient state detection circuit 2 to perform a 512-sample size orthogonal transform, the 512-sample size is obtained based on the 512-sample data fetched from the buffer circuit BUF. Orthogonal transform of. At this time, in the present embodiment, the following processing is performed in order to ensure the continuity of the data by the windowing processing.

That is, first, the data of 512 sample size held in the buffer circuit BUF is set to 5
512 sample size MDCT processing is performed once and 512
Obtain frequency data. After that, the buffer circuit BUF
The 512 sample size data stored in
By shifting by 6 samples, 256 samples in the latter half of the current block and 256 samples in the first half of the next block to be newly input become data of a new 512 sample size,
Prepare for the next MDCT process.

As described above, in the present embodiment, the data is overlapped by N / 2 samples (= 256 samples) and MD is added.
CT processing is performed. This situation is shown in FIG.

When the 512-sample size MDCT process is performed as described above, the inverse MDCT process performed on the decoder side (not shown) is as follows. It should be noted that here also, for simplification, description will be made with N = 512.

1) The 512 frequency data restored from the bit stream of the encoded data are subjected to the inverse MDCT processing of 512 sample size once,
The 512 frequency data are inversely transformed into 512 time data. 2) The second half 256 time data of the previous block remains in the buffer, and the second half 256 time data of the previous block and the first half 256 time data of the 512 time data of the current block are added. And output from the decoder.

3) The second half 256 time data of the previous block held in the above buffer are 51 times of the current block.
Of the two pieces of time data, the latter half 256 pieces of time data are replaced. The above processes 1) 2) 3) are repeated.

Further, when the MDCT circuit 3 is instructed by the transient state detection circuit 2 to perform the orthogonal transform of N / 4 sample size (= 128 sample size), 512 fetched from the buffer circuit BUF. Orthogonal transformation is performed on the sample data every four 128 sample sizes. Also at this time, because of the windowing process,
As shown in FIG. 2B, the following processing is performed.

1) The 512 pieces of sample data held in the buffer circuit BUF are divided into four equal parts, and each of them is divided into 12 parts.
MDCT processing of 128 sample size is performed on 8 sample data. By a total of four MDCT processes of 128 sample size, 512 sample data are orthogonally transformed into 512 frequency data.

2) After the 512 pieces of sample data held in the buffer circuit BUF have been orthogonally transformed, 51
The two samples are shifted by 256 samples to form new 512 samples of data in the latter half 256 samples of the current block and the first half 256 samples of the newly input next block, and prepare for the next MDCT processing.

As shown in FIG. 2B, even when MDCT processing of 128 sample size is performed, the MDCT processing is performed by duplicating data every 256 samples. 2 in this example
This is because the data acquisition process on the decoder side can be performed irrespective of the orthogonal transformation size by setting 56 samples).

Further, the inverse MD on the decoder side when such MDCT processing of 128 sample size is performed
The CT process is shown below.

1) The 512 frequency data restored from the bit stream of the encoded data are subjected to inverse MDCT processing of 128 sample size four times,
The 512 frequency data are inversely transformed into 512 time data. 2) The second half 256 time data of the previous block remains in the buffer, and the second half 256 time data of the previous block and the first half 256 time data of the 512 time data of the current block are added. And output from the decoder.

3) The latter half 256 time data of the previous block held in the above buffer are 51 times of the current block.
Of the two pieces of time data, the latter half 256 pieces of time data are replaced. The above processes 1) 2) 3) are repeated.

As described above, the difference in processing on the decoder side between the case where the 512-sample size MDCT processing is performed on the encoding side and the case where the 128-sample size MDCT processing is performed on the encoding side is only the processing of 1).

The frequency data (512 MDCT coefficient data) orthogonally transformed by the MDCT circuit 3 as described above is supplied to the block floating point conversion circuit 4 as in the case of FIG. Block floating point conversion circuit 4
Is the 512 MDCTs input from the MDCT circuit 3.
The coefficient data is converted into floating point data represented by the mantissa part and the exponent part, and the exponent part data is converted into the bit allocation circuit 5.
And the mantissa part data to the quantization coding circuit 6.

The bit allocation circuit 5 is based on the exponent data input from the block floating point conversion circuit 4.
The quantization bit length of the mantissa data is determined by utilizing the human auditory masking characteristic and the like, and the determined quantization bit length is output to the quantization encoding circuit 6. The quantization coding circuit 6 includes a bit allocation circuit 5
MSB of the mantissa data input from the block floating point conversion circuit 4 by the quantization bit length input by
Then, the data is extracted from and the process of the quantization coding is performed.

As a result of the above processing, the first data output terminal 7 outputs the exponent part data, and the second data output terminal 8 outputs the mantissa coded data. These output data are accumulated in a recording medium (not shown),
It may be transmitted via a transmission path (not shown).

Next, the operation of the transient state detection circuit 2 described above will be described in more detail. N samples (= 512) input to the transient state detection circuit 2 from the buffer circuit BUF of FIG.
The data of (sample) is first calculated by the sum of squares calculation unit 21 as N /
The sum of squares is obtained for each small block of 4 samples (= 128 samples).

Assuming that the data of N samples (= 512 samples) is x (n), n = 0,1,2, ..., N-1, the sum of squares of the sample data of each small block is as follows. It is calculated by the equation (1). In this equation (1), m is an index of a small block composed of N / 4 samples.

[0041]

[Equation 1]

Next, in the sum of squares comparison unit 22, the above equation (1)
The sum of squares for each small block made up of N / 4 samples (= 128 samples) calculated in step 1 is compared with each other to detect a transient state. The detection of this transient state is performed by the following procedure. That is, the square sums between adjacent small blocks are compared by seeing whether the following expressions (2) to (7) hold. However, T shown in these equations
h is a threshold value for adjusting the detection sensitivity.

S (m)> Th × s (m + 1) (2) s (m + 1)> Th × s (m) (3) s (m + 1)> Th × s (m + 2) (4) s (m + 2)> Th × s (m + 1) (5) s (m + 2)> Th × s (m + 3) (6) s (m + 3)> Th × s ( m + 2) (7)

Next, the block size determining unit 23 determines the orthogonal transform size based on the result of the comparison in the square sum comparing unit 22. That is, the above equations (2) to
If at least one condition of (7) is satisfied, the orthogonal transform size is determined to be N / 4 sample size (= 128 sample size). If none of the conditions of the above equations (2) to (7) is satisfied, the orthogonal transform size is determined to be N sample size (= 512 sample size).

In the above equations (2) to (7), a rapid change in the property of the input signal, such as a sharp rise of the signal, is detected by comparing the square sums of adjacent small blocks. doing. The parameter that adjusts the detection sensitivity is the threshold Th. Threshold Th
The actual encoding / decoding process may be performed while varying the value, and the audible sound quality and noise level may be measured to obtain an appropriate value.

In this embodiment, good results are obtained when Th = 20. The numerical value of 20 is √20≈4.47 when converted into signal amplitude. That is, in the present embodiment, the input signal is in a transient state when the average amplitude between adjacent small blocks is larger than 4.47 times.

In the above embodiment, consideration is given to use in variable block length transform coding capable of orthogonal transform of N sample size (= 512 sample size) and N / 4 sample size (= 128 sample size). And realized the transient state detection method. Besides this, for example, N
It is also applicable to variable block length transform coding capable of orthogonal transform of sample size and N / 2 sample size (= 256 sample size).

In this case, 4 shown in the above equation (1)
Calculate the sum of squares for each two small blocks in the first half N of N samples
/ 2 samples and latter half N / 2 samples are replaced with two square sum calculations, and the comparison of the square sum calculation results shown in the above equations (2) to (7) is performed by comparing the first half N / 2 samples and the latter half N The transient state detection method can be realized by replacing the sum of squares with the / 2 sample. How to duplicate the data for the windowing process at this time is shown in FIG.
It may be done as in (c).

[0049]

As described above, according to the present invention, the input signal is taken in every predetermined section to obtain the sum of squares, and the sum of squares is set to 2 for each section.
Since the transient state of the input signal is detected by the degree of change of the summed signals over at least two sections, the square of the input signal on the time axis can be obtained without performing orthogonal transformation processing or filter processing. The transient state can be detected only by performing the sum calculation, the calculation load is small, and it is suitable for the real-time processing of encoding. In the present invention, it is possible to realize detection accuracy that is substantially the same as that of the conventional transient state detection method using orthogonal transformation processing and filter processing. Therefore, from the viewpoint of real-time processing, both the calculation load and the detection accuracy are improved. Is excellent in.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration example of a variable block length transform coding apparatus according to an embodiment of the present invention.

FIG. 2 is a diagram for explaining how data is taken into the MDCT circuit shown in FIG.

FIG. 3 is a block diagram showing a configuration example of a conventional block length variable transform coding apparatus.

FIG. 4 is a waveform diagram showing an input signal including a steep rising portion.

5 is an M of 512 sample size for the waveform of FIG.
Obtained after performing the encoding / decoding processing by DCT,
It is a wave form diagram which shows the signal containing a pre-echo part.

6 is an M of 256 sample size for the waveform of FIG.
Obtained after performing the encoding / decoding processing by DCT,
It is a wave form diagram which shows the signal which the pre-echo part reduced.

FIG. 7 is a diagram for explaining a conventional feedback type transient state detection method.

[Explanation of symbols]

 1 Audio Data Input Terminal 2 Transient State Detection Circuit 3 MDCT Circuit 4 Block Floating Point Conversion Circuit 5 Bit Allocation Circuit 6 Quantization Coding Circuit 7 First Data Output Terminal 8 Second Data Output Terminal 21 Square Sum Calculator 22 Square sum comparison unit 23 Block size determination unit BUF buffer circuit

Claims (4)

[Claims] 1. A variable block length transform coding apparatus having a transient state detection circuit for detecting a transient state of an input signal and determining an orthogonal transform size, wherein the transient state detection circuit has an input signal for each predetermined section. A block length variable conversion code, characterized in that the transient sum of the input signals is detected based on the degree of change in the signals summed up and squared in each section over at least two sections. Device. 2. A transient state of the input signal is detected by taking in an input signal for each predetermined section and obtaining a sum of squares for each section, and detecting a degree of change of the signal summed for the square for each section over at least two sections. A transient state detection circuit that determines the orthogonal transformation size by performing the orthogonal transformation by the unit of the orthogonal transformation size determined by the transient state detection circuit, and an orthogonal transformation performed by the orthogonal transformation circuit. A block floating point conversion circuit for converting the frequency data into a block floating point, a bit allocation circuit for allocating bits by the exponent part data supplied from the block floating point conversion circuit, and a mantissa part data obtained by the bit allocation circuit. According to the quantization bit length, the mantissa data supplied from the block floating point conversion circuit is quantized and encoded. Block length variable transform coding apparatus comprising: a quantization encoding circuit. 3. The transient state detection circuit takes in an input signal for each size of an integer fraction of the maximum orthogonal transform size of the orthogonal transform circuit, obtains the sum of squares of each, and at least the sum of squared signals. The variable block length transform coding apparatus according to claim 2, wherein the transient state of the input signal is detected by a degree of change over two or more sections. 4. A sum-of-squares calculating means for fetching an input signal for each predetermined section to obtain a sum of squares, and at least two or more sections of the signals summed squares for each predetermined section by the sum-of-squares calculating means. And a transient state detecting means for detecting the transient state of the input signal according to the degree of change over the transient state detecting device.