JPH0969782A - Audio data encoding device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce the degradation of the tone quality at the time of decoding by not performing the clipping processing at the time of difference encoding of exponent part data. SOLUTION: With respect to mantissa part data and exponent part data calculated by a block floating-point conversion circuit 13, a smoothing circuit 14 which performs such correction that the difference of exponent part data between at least adjacent bands is put in a preliminarily determined range is provided. Exponent part data smoothed by the smoothing circuit 14 is subjected to difference encoding by a difference encoding circuit 15, and thereby, exponent part data is not clipped at the time of difference encoding of exponent part data, and error does not occur when an MDCT coefficient is reconstituted from mantissa part data and exponent part data on the decoder side.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an audio data compression apparatus, and more particularly to an audio data encoding apparatus for highly efficiently encoding digital audio data.

[0002]

2. Description of the Related Art For high efficiency encoding of audio data,
There are band division coding in which audio data is divided into a plurality of bands on the time axis for coding, and transform coding in which audio data is orthogonally transformed on the frequency axis and divided into a plurality of bands for coding. There is also high-efficiency coding in which these are combined and divided into a plurality of bands on the time axis, and each band signal is further subjected to orthogonal transform on the frequency axis for coding.

As an example of the prior art, the MDC shown in FIG.
Transform coding using T (Modified Discrete Cosine Transform) will be described. The audio data input from the audio data input terminal 1 is subjected to windowing by the windowing circuit 2 in block units, for example, 512 samples as one block. In the windowing circuit 2, 50% of the current block is overlap-added to the previous block,
The addition result is input to the MDCT circuit 3.

In the MDCT circuit 3, orthogonal transformation is performed by MDCT, and the resulting MDCT coefficient data is input to the block floating point transformation circuit 4. In the block floating point conversion circuit 4, of the 512 MDCT coefficient data input from the MDCT circuit 3, for example, 240 from the zero frequency are in some bands, for example, 64.
It is divided into bands and converted into block floating point data consisting of one exponent part data and a plurality of mantissa part data per band. At this time, the coefficient in one band is, for example,
A band division method is adopted in which the number of coefficients in one band increases from low frequency to high frequency with 1 to 8.

The exponent part data differential encoding circuit 5 differentially encodes the exponent part data input from the block floating point conversion circuit 4. Further, in the bit allocation circuit 6,
Based on the exponent part data input from the block floating point conversion circuit 4, the quantization bit length of the mantissa part data is determined by utilizing human auditory masking characteristics and the like, and is input to the quantization coding circuit 7. To do. In the quantization coding circuit 7, the quantization bit length input from the bit allocation circuit 6 is
Data is extracted from the MSB of the mantissa data input from the block floating point conversion circuit 4 and quantized and encoded.

By the above processing, the exponent part differential encoded data is output from the first data output terminal 8 and the mantissa encoded data is output from the second data output terminal 9. These output data are stored in a recording medium,
It is transmitted via a transmission line.

[0007]

By the way, in the differential encoding of the exponent part data, in order to compress the information amount, the exponent part data is not encoded as it is, but the difference value is encoded from the low frequency to the high frequency. It has become. This compresses the information amount of the exponent part data.

However, in the differential encoding of exponent part data, in general, a process of clipping a differential value of a certain value or more, for example, a differential value of 4 or more is set to 4, and a differential value of -4 or less is set to -4. Is being processed. Due to this clipping process, an error occurs when reconstructing the MDCT coefficient from the mantissa data and exponent data on the decoder side, and there is a problem that the sound quality is deteriorated.

Therefore, the present invention prevents clipping processing from being performed when differentially encoding exponent data,
It is an object of the present invention to provide an audio data encoding device that reduces deterioration of sound quality during decoding.

[0010]

According to the audio data encoding apparatus of the present invention, before the differential encoding of exponent data, the mantissa part is set so that the difference between exponent data of adjacent bands falls within a predetermined range. The data and the exponent part data are smoothed.

According to the present invention configured as described above, since the mantissa part data and the exponent part data are smoothed,
When the differential encoding of the exponent part data is performed, the exponent part difference data is not clipped. This allows
MDCT from the mantissa data and exponent data on the decoder side
No error occurs when reconstructing the coefficients, and sound quality is not deteriorated.

[0012]

DETAILED DESCRIPTION OF THE INVENTION An embodiment according to the present invention will be described below with reference to FIG. The audio data input from the audio data input terminal 10 is, for example, 51
2 samples as one block, block by block,
Windowing is performed by the windowing circuit 11. Windowing circuit 1
In 1, 50% of the current block is overlap-added to the previous block, and the addition result is input to the MDCT circuit 12.

In the MDCT circuit 12, orthogonal transformation is performed by MDCT, and the resulting MDCT coefficient data is input to the block floating point transformation circuit 13. In the block floating-point conversion circuit 13, of the 512 MDCT coefficient data input from the MDCT circuit 12, for example, 240 from the zero frequency are divided into several bands, for example, 64 bands, and 1 per 1 band. It is converted into block floating point data composed of one exponent part data and a plurality of mantissa part data.

The block floating point conversion will be described below. First, the maximum value is found among the absolute values of MDCT coefficient data within one band. Next, the maximum value data F
Is expressed as follows. F = M × 2 exp (-N) where F: Floating point data M: Mantissa part data, 0.5 ≦ M <1, -1 ≦ M <-0.5 N: Exponent part data, positive integer

The exponent part 2exp (-of the maximum value data F is
NMD) divides the other MDCT coefficient data in one band to make them mantissa data. By this processing, one exponent part data for one band and mantissa part data of the number equal to the number of MDCT coefficient data in one band are calculated.

At this time, the number of coefficient data included in one band is, for example, 1 to 8, and a band division method is adopted in which the number of coefficients in one band increases from a low frequency to a high frequency in units of 2n. . For example, from low frequencies, 16 bands each having one coefficient in one band, 16 bands each having two coefficients in one band, 16 bands each having four coefficients in one band, 1
That is, there are 8 coefficients in the band and 16 in the band. In this case, the exponent part data is 16 × 4 = 64, and the mantissa part data is 16 × (1 + 2 + 4 + 8) = 240. The state of band division in this case is shown in FIG.

Next, the smoothing circuit 14 smoothes the mantissa part data and the exponent part data input from the block floating point conversion circuit 13. The smoothing method will be described below. As an example, it is assumed that the differential encoding circuit 15 performs a process of rounding a differential value greater than or equal to R and rounding a differential value less than or equal to −R.

Further, for the following description, the 64 exponent part data input from the block floating point conversion circuit 13 are N (x), x = 0, in order from low frequency to high frequency.
Let's say 1, ..., 63. In addition, 240 mantissa data are M (x, y), x = 0,1, ..., 63, y in order from low frequency to high frequency.
= 0,1, ..., L; However, when x = 0, .., 15, L = 0, x = 1
When 6, .., 31, L = 1, x = 32, .., 47, L = 3, x = 48, ..,
When 63, L = 7.

Here, x of the mantissa data M (x, y) is as shown in FIG.
Is a variable indicating which band is counted from the lowest frequency of, and y is a variable indicating which coefficient is in that band. For example, the 35th band from the lowest frequency in FIG. 2 can be expressed as x = 35. At this time, since one band includes four coefficient data in this band, y is 1
It can take a value of ~ 4.

As described above, when it is determined that the mantissa data is represented by the variables x and y, smoothing is performed by the following (1) to (3).
According to the formula, the band x = 0,1 ,.

In the case of N (x)> N (x + k) + k × R (1), k × R is set to the value between the exponent data separated by k bands.
There will be the above differences. That is, when k = 1, the difference in the exponent part data between adjacent bands is represented, and therefore there is a difference of R or more, which is the threshold value of clipping processing, between adjacent bands.

In the case of k = 2, the difference in the exponent part data from the band further apart from the adjacent bands on both sides is shown. Therefore, the difference of 2 × R or more means that the difference is 2 × R. This means that a difference equal to or greater than the threshold R of the clipping process always occurs between adjacent bands.

Similarly, in the case of k = i, the difference in the exponent part data from the band i−1 further apart than the bands on both sides is expressed.
The fact that this difference is i × R or more indicates that there is a portion in which a difference of R or more, which is the threshold value of the clipping process, occurs between two bands that are adjacent to the band that is ahead by i. become.

As shown in FIG. 3, even if the difference Δ1 from the band 1 to the band 2 is R or less as viewed from the band 1, if the difference Δ2 from the next band 2 to the band 3 is R or more, Is more effective than band 1 and band 3 rather than band 2 and band 3 only.
The envelope can be made smoother by performing the correction between the band 4 and the band 4 further ahead.

Therefore, in the present embodiment, the clipping process is predicted not only on both sides, but also between bands further ahead,
The exponent part data is corrected in advance as follows so that the difference between the exponent part data becomes equal to or less than the threshold value R of the clipping process.

M (x, y) = M (x, y) × (2 × exp (N (x + k) + k × R-N (x))), y = 0, .., L (2 ) N (x) = N (x + k) + k × R (3)

The above formulas (1) to (3) will be described in detail. First, according to Eq. (1), the exponent data N (x) to be smoothed and the exponent data N (x + k) separated by k bands are k × R.
Is compared with the value N (x + k) + k × R. As a result, the exponent part data to be clipped is detected in advance when the differential encoding of the exponent part data is performed.

When the above equation (1) is satisfied, (3)
As shown in the equation, the exponent data N (x) to be smoothed is the value N (x + k) obtained by adding k × R to the exponent data N (x + k) separated by k bands.
Replace with + k × R. As a result, it is possible to prevent clipping processing from being the target of differential encoding of exponent part data.

At this time, since the exponent part data is replaced, the mantissa part data also needs to be corrected accordingly. This is done by the equation (2). In equation (2), the value obtained by subtracting N (x) from the value N (x + k) + k × R obtained by adding k × R to the exponent data N (x + k) separated by k bands (2 × exp (N (x + k) +
k × R-N (x))) multiplied by the mantissa of all x bands.

This means that the mantissa data of the x band absorbs a portion exceeding the difference allowed for the exponent data separated by the k band. As an actual calculation order, the mantissa part M (x, y) must be corrected by the expression (2) before the exponent part N (x) is replaced by the expression (3).

In the above smoothing processing, the exponent data N (x + k) apart from the exponent data N (x) in the x band by k bands
The problem is how much is to be compared, that is, the value of k is changed from 1 to how many.

If the smoothing process is performed by comparing the exponent part data N (x) of the x band with all other exponent part data, there will be completely no clipping process target in the differential encoding. However, the amount of calculation in that case is enormous, which is unsuitable for actual encoding processing. Therefore, in the present embodiment, only k = -2, -1,1,2 is set as the target range for smoothing to reduce the calculation amount.

When a voice signal is input, the exponent data are highly correlated and do not change abruptly. Therefore, in most cases, the smoothing process of k = 1,2 is the target of the clipping process. Can be eliminated.

Next, the exponent part data difference encoding circuit 15 calculates the difference value of the exponent part smoothed data input from the smoothing circuit 14 and calculates the exponent part having a value range from -R to R. The difference data is encoded by replacing it with an appropriate code. Further, the bit allocation circuit 16 determines the quantized bit length of the mantissa data based on the exponential part smoothed data input from the smoothing circuit 14 by utilizing human auditory masking characteristics and the like. It is input to the quantization coding circuit 17.

In the quantization coding circuit 17, data corresponding to the quantization bit length inputted from the bit allocation circuit 16 is taken out from the MSB of the mantissa smoothed data inputted from the smoothing circuit 14 and is quantization coded. I do. Then, the first data output terminal 18 outputs the exponential part difference encoded data, and the second data output terminal 19 outputs the mantissa part encoded data. These output data are stored in a recording medium or transmitted via a transmission path.

[0036]

EXAMPLES The above-mentioned smoothing will be described below using specific numerical values. Here, K = 1, N (x) = 8, M (x, y) = 0.8, N
(x + 1) = 2 and R = 4.

In this case, first, the exponent part data N of the band x
Since the difference between (x) and the exponent part data N (x + 1) of the band x + 1 is 6, it means that R = 4, which is the threshold value of the clipping process, is exceeded. Actually substituting each value into the equation (1) gives the following equation. 8> 2 + 1 x 4

As described above, since the comparison condition of the expression (1) is satisfied, the smoothing processing of the expressions (2) and (3) is performed.
That is, each value is substituted into the equation (2) to obtain the following equation. M (x, y) = 0.8 × (2 × exp (2 + 1 × 4-8)) = 0.8 × (2 × exp (-2)) = 0.8 × 0.25 = 0.2

Further, by substituting each value into the equation (3), the following equation is obtained. N (x) = 2 + 1 × 4 = 6

These smoothed mantissa data M (x, y)
= 0.2 and exponent part data N (x) = 6, MDCT coefficient data 0.2 × 2exp (-6) = 0.003125 is the mantissa part data before smoothing M (x, y) = 0.8 and exponent part data MDCT coefficient data reconstructed from N (x) = 8 0.8 × 2 exp (-8) = 0.003125
It is found that there is no error.

[0041]

As described above, according to the present invention, the smoothing process of the mantissa part data and the exponent part data is performed before the differential encoding of the exponent part data. It is possible to prevent the exponential part difference data from being clipped during the differential encoding. This allows
MDCT from the mantissa data and exponent data on the decoder side
It is possible to prevent an error from occurring when reconstructing the coefficient, and it is possible to prevent deterioration of sound quality. in this case,
Even if the smoothing target range is narrowed in a form suitable for actual calculation, there is almost no error when reconstructing MDCT coefficient data on the decoder side due to the correlativity of exponent data when a voice signal is input. It can be lost.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration example of an audio data encoding device using MDCT according to an embodiment of the present invention.

FIG. 2 is a diagram showing an example of band division of MDCT coefficients.

FIG. 3 is an explanatory diagram for explaining a difference in exponent part data between different bands.

FIG. 4 is a block diagram showing a configuration example of an audio data encoding device using a conventional MDCT.

[Explanation of Codes] 10 Audio Data Input Terminal 11 Windowing Circuit 12 MDCT Circuit 13 Block Floating Point Conversion Circuit 14 Smoothing Circuit 15 Differential Encoding Circuit 16 Bit Allocation Circuit 17 Quantization Encoding Circuit 18 First Data Output Terminal 19 Second data output terminal

Claims (1)

[Claims] 1. A windowing circuit for windowing audio data, and an MD for the data windowed by the windowing circuit.
MDCT circuit for orthogonal transformation by CT, block floating point conversion circuit for performing block floating point conversion on MDCT coefficient data orthogonally transformed by the MDCT circuit, and mantissa part data and exponent part data calculated by the block floating point conversion circuit And a smoothing circuit for correcting the difference of the exponent data so that the difference between the exponent data is at least between adjacent bands, and the exponent data smoothed by the smoothing circuit is subjected to a difference sign. Exponent part data differential encoding circuit for converting, bit allocation circuit for allocating bits by exponential part smoothed data calculated by the smoothing circuit, and quantization bit length of mantissa part data calculated by the bit allocation circuit And the mantissa part calculated by the smoothing circuit according to the band to which the mantissa part data belongs A smoothing data is quantized and coded, and a quantization coding circuit is provided. In the smoothing circuit, the mantissa part data and the exponent part data are smoothed, and the smoothed exponent part data is differentially coded. An audio data encoding device characterized by: