JP3365908B2 - Data converter - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a data conversion device for extending the bit length of digital audio data.

[0002]

2. Description of the Related Art Currently, in digital audio that digitizes an analog signal and performs recording and reproduction, 16 to 2 is currently used.
Data with a high precision bit number such as 0 bit is used. However, during the introduction of these technologies,
13-bit or 14-bit precision data is used. Therefore, when using a music source recorded by these devices (for example, 13-bit system) as a sound source for compact discs (for example, 16-bit system), The data of the 13-bit system is used as the upper 13 bits of the 16-bit system, 0 is assigned to the lower 3 bits of the 16-bit system, or the lower bit is a meaningless data area.
These music sources have a small number of effective bits, so distortion occurs when the sound of the instrument decays and disappears,
There is a feeling of noise. That is, quantization noise occurs, and as a countermeasure against this, for example, in Japanese Patent Laid-Open No. 7-193502, a data string sampled on the time axis is used as a frequency axis component (frequency coefficient).
There is shown a data conversion device which obtains effective data in lower bits by converting the data into 0 to data having a specific level or less, and inversely converting the data with bit precision larger than the input data.

That is, in this device, digital data of a predetermined number of bits N (data string sampled on the time axis)
Are divided into several blocks and converted into frequency axis components, and among the converted frequency axis components (frequency coefficients), frequency axis components (frequency coefficients) of a level below a predetermined level are 0 or attenuated Coefficient), and the frequency coefficient is inversely converted into a time axis component with a bit number M larger than the effective bit number N at the time of input and is output as digital data (sample data). This gives, for example, 1
It is possible to obtain data in which a noise component such as a level equal to or less than the rounding error ± LSB of 3 bits or the like is removed and the effective bit is expanded to 16 bits.

[0004]

However, in the above-mentioned conventional apparatus, the frequency axis component (frequency coefficient) of the quantization noise having a low level is removed and the minute music component is removed.
(Frequency axis component of music) is also removed, and the digital data reproduced by inversely converting it to the time axis component may have a deteriorated sound quality because the attenuated sound of the music signal disappears on the way. That is, the quantization noise can be removed, but at the same time, a minute musical component is also removed, and the attenuated sound disappears midway, which makes it easier to perceive deterioration in sound quality.

The present invention eliminates quantization noise when expanding the bit length of digital data, and restores the quantization level of a music signal which is apt to be buried in the quantization noise while restoring the effective bit number. It is an object of the present invention to provide a data conversion device capable of obtaining large digital data.

[0006]

In order to achieve the above object, in the present invention, digital data having an effective number of bits N is divided into several blocks, converted into frequency axis components, and converted from frequency axis components. , The frequency axis component below a predetermined level is replaced with the frequency axis component which is inferred from the frequency axis component over several blocks, and then converted into the time axis component with the effective bit number M (N <M) and output. . As a result, when expanding the bit length of digital data, the quantization noise is removed, while for a small level music signal that is apt to be buried in the quantization noise, while restoring this, digital data with a large effective bit number Can be obtained. The minute music signal buried in the quantization noise can be restored, and the sound quality of old recording sources can be improved.

[0007]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram showing a configuration example of a data conversion device according to the present invention. Referring to FIG. 1, the data conversion device includes a conversion unit 2 that divides digital data having an effective bit number N into specific blocks and converts the data into frequency axis components (frequency coefficients), and frequency axis components converted over several blocks. From the estimating means 4 for estimating the frequency axis component of the block to be converted next, and of the frequency axis components converted by the converting means 2, the frequency axis components below a predetermined level are estimated by the estimating means 4. A replacement means 7 for replacing the frequency axis component with a frequency axis component, and an inverse conversion means 8 for converting the frequency axis component from the replacement means 7 into a time axis component (digital data string) having an effective bit number M (N <M) and outputting the time axis component. ing.

FIG. 2 shows a more specific example of the configuration of the data conversion apparatus shown in FIG. Referring to FIG. 2, the data conversion apparatus divides digital data (for example, PCM data used in digital audio or the like) having an effective number of bits N sampled on a time axis into specific blocks, that is, a frequency axis component, that is, a frequency axis component. FF to convert to coefficient (spectrum)
A T (Fast Fourier Transform) unit 102, a threshold value generation unit 105 for generating a threshold value TH about the quantization noise level, and an FFT
A spectrum comparing unit 106 that compares the frequency coefficient (spectrum) from the unit 102 with a threshold value TH, and a spectrum estimating unit that estimates the frequency coefficient (spectrum) of the block to be transformed next from the frequency coefficient (spectrum) transformed over several blocks. 104 and the frequency coefficient (spectrum) from the FFT unit 102 according to the comparison result of the spectrum comparison unit 106.
And spectrum selecting section 10 for selecting either the frequency coefficient (spectrum) from spectrum estimating section 104.
7 and a temporary storage buffer 1 for temporarily storing the frequency coefficient (spectrum) selected by the spectrum selection unit 107.
03, and an inverse FFT unit 108 that inversely FFT-converts the frequency coefficient (spectrum) selected by the spectrum selection unit 107 into a time axis component with an effective bit number M larger than the input effective bit number N.

Here, the FFT section 102 and the inverse FFT section 10
1 corresponds to the transforming means 2 and the inverse transforming means 8 in FIG. 1, the spectrum estimating section 104 corresponds to the estimating means 4 in FIG. 1, and the threshold value generating section 105, the spectrum comparing section 106, and the spectrum selecting section 107 respectively. 1 corresponds to the replacement means 7 in FIG. In addition, the temporary storage buffer 103 has a plurality of blocks.
It has a capacity to store (at least two blocks) of spectrum, and is used for performing estimation processing in spectrum estimation section 104.

Next, the processing operation of such a data converter will be described. Input terminal 1 of this data converter
01, for example, when PCM data received and demodulated by a digital audio interface is input, F
The FT unit 102 converts the input PCM data into frequency coefficients (spectrum) by N-point FFT processing. Note that FFT is one of the calculation methods for executing DFT (discrete Fourier transform) at high speed, and the solution of DFT shown in the following equation can be obtained with a small number of multiplications.

[0011]

[Equation 1]

In the equation (1), x (n) is N-point input data subjected to the windowing process, and a 50% overlapped Hamming window is used for this window.

X (k) obtained by the equation 1 represents the input data x (n) as a combination of a sine wave having N samples as one cycle and a sine wave having a cycle of an integer (k) times that of the sine wave. Represents the information of each sine wave (spectrum) in the case. Specifically, since X (k) is a complex number, its absolute value indicates the spectrum level and its argument indicates the phase.

FIG. 3 shows an example of the spectrum analyzed by the FFT unit 102. This spectrum consists of a music component C ₁ and a quantization noise component C ₂ . It is known that when the data level is large, the quantization noise becomes white and the noise power is about 1/12 of the LSB level. In FFT, this noise is also spectrally decomposed. The power is about -27 dB.

In the threshold generator 105, for example, this level is generated as a threshold TH for judging quantization noise. When the input data is small, the data becomes tone distortion and does not become white, so the data level of a specific spectrum rises. In order to deal with this, when the input data level is low, the noise can be removed well by slightly raising the threshold value of the band that is considered to be the noise spectrum.

The spectrum comparison unit 106 compares the FFT-analyzed spectrum in each band with the threshold value TH and outputs the result to the spectrum selection unit 107.

In each band, spectrum selecting section 107 determines which spectrum of the output (spectrum) from FFT section 102 and the output (spectrum) from spectrum estimating section 104, which will be described later, is output to inverse FFT section 108. , According to the comparison result of the spectrum comparison unit 106. That is, the spectrum selection unit 107 includes the FFT unit 1
When the output (spectrum) from 02 is larger than the threshold TH, the output (spectrum) from the FFT unit 102 is selected, and when the threshold TH is larger than the output (spectrum) from the FFT unit 102, , The spectrum from the spectrum estimation unit 104 is selected.

The spectrum selected and output from spectrum selecting section 107 is applied to inverse FFT section 108. Reverse F
The FT unit 108 converts the spectrum output from the spectrum selection unit 107 into a time axis component (digital data string) by inverse FFT. The inverse FFT is one of the calculation methods for executing the inverse DFT at high speed, and the solution of the inverse DFT shown in the following equation can be obtained with a small number of multiplications.

[0019]

[Equation 2]

In this case, it is assumed that the inverse FFT section 108 has sufficient calculation accuracy in order to obtain the output of the effective bit number M (M> N). The waveform of the time axis component (digital data string) converted by the inverse FFT unit 108 is output from the output terminal 109 while being overlapped with the preceding and following waveforms by the same ratio (for example, 50%) as the FFT unit 102.

The spectrum selected and output from the spectrum selection unit 107 is temporarily stored in the temporary storage buffer 103 in parallel with the above processing. Note that, as described above, the temporary storage buffer 103 has a capacity for a plurality of blocks (for example, one block before the current block, two blocks before the current block). Thus, based on the spectrum temporarily stored in the temporary storage buffer 103, the spectrum estimating unit 104
Estimate the spectrum in (current block). As a specific example, the spectrum absolute value over the immediately preceding block, 2
The spectral absolute values over the immediately preceding block are estimated spectral absolute values X when there are X _B-1 and X _B-2 , respectively.
_B is obtained by a linear function such as the following equation.

[0022]

[Formula 3] X _B = X _B-1 + {X _B-1 −X _B-2 }

The phase is adjusted so that the waveform is continuous with the previous block. 1 for waveforms with 50% overlap
In the next term X (1), + 180 °, in the second term X (2),
Since the angle of advance is 360 ° (same below), the phase of the angle of advance is added to the phase of each term of the immediately preceding block to match this, and the phase of the spectrum of the estimation block is determined.

In the above example, the spectrum estimation unit 1
In 04, the estimated spectrum is uniquely obtained by Equation 3, but it is mainly the attenuation part of the musical sound that can be well estimated, so the size of the estimated spectrum is limited in the band where the spectrum is increasing. By setting it to zero or reducing it to zero, it is possible to reduce the generation of noise due to incorrect estimation.

As described above, according to the spectrum estimating section 104, even if the spectrum in the current block is equal to or less than the threshold value TH, this is calculated from the spectra of the block immediately before and the block two blocks before, for example. It can be calculated by analogy. Therefore, when the process in the spectrum selection unit 107 (that is, when the output (spectrum) from the FFT unit 102 is larger than the threshold value TH, the FFT unit 10
2 is selected, and when the threshold TH is larger than the output (spectrum) from the FFT unit 102, the process of selecting the spectrum from the spectrum estimation unit 104 is performed by the FFT unit 102. In the spectrum, the spectrum estimation unit 104 regards the music signal spectrum of a minute level that is removed together with the quantization noise component C ₂ equal to or less than the threshold TH.
It means that it is reconstructed by the spectrum estimated in.

As described above, according to the present invention, when the digital audio data recorded with a low bit number is reproduced and output with a bit number higher than the low bit number at the time of recording (that is, when the bit length expansion is performed). ), Even when the quantization noise component C ₂ equal to or less than the threshold value TH is removed, it is possible to estimate and restore a minute level of the music signal spectrum that is removed along with it.

Next, as a concrete processing example of the data conversion apparatus shown in FIGS. 1 and 2, of the data removed below the threshold value of the current block, a process of inferring (estimating) minute music data from the previous block. Will be described.

Now, when music data is inputted from the input terminal 101, the inputted music data is temporally divided into blocks and a time axis component (digital data string) to a frequency axis component (spectrum). ). The conversion from the time axis component to the frequency axis component can be performed by using the fast Fourier transform (FFT) as described above.
That is, it can be performed according to the equation 1.

FIG. 4A shows the spectrum of the current block when continuous music data is converted into frequency axis components (spectrum) in block units. In FIG. 4A, the spectrum of the current block is
A relatively high level music data spectrum and a relatively low level noise component (quantization noise) spectrum are mixed.

The spectrum selection unit 107 substantially determines that the spectrum of the current block in FIG. 4A is equal to or less than the threshold value TH and removes it. This state is shown in FIG. As can be seen from FIG. 4 (b), at this time, even in the spectrum that is originally music data,
Spectra below the threshold TH are removed.

In FIG. 4B, the spectrum estimation unit 104 performs spectrum estimation processing in order to restore the spectrum of the music data that is equal to or less than the threshold value TH that has been removed.
Specifically, for example, the spectrum of the block two blocks before the current block as shown in FIG. 4C and the spectrum of the block one block before the current block as shown in FIG. 4D. Then, the spectrum in the current block is inferred. To analogize the spectrum in the current block from the previous block, the spectrum of the current block is predicted from the spectra of the two blocks before the current block by using, for example, a linear function of Equation 3. Specifically, as shown in FIGS. 4 (c) and 4 (d), a temporal spectral level gradient is obtained from the spectral level of the same frequency in each block, and the gradient corresponds to the current block. Estimate the spectrum. FIG. 4 (e) shows the spectrum of the analogy block.

When an FFT is used to convert a music signal into a frequency axis component, the spectrum of the current block and the spectrum of the immediately preceding block are in phase when analogizing a small amount of music data in the subsequent stage. Since it is continuous, it is possible to easily determine the phase of the analogy spectrum of the current block by adding the advance angle component to the phase of each term of the spectrum obtained by the calculation. That is, when FFT is used when converting to the frequency axis component, by adding the phase of the advance angle to the phase of each term of the block immediately preceding the current block, to the calculated spectrum corresponding to the current block. , The phase of its spectrum can be determined. Further, by using FFT instead of DFT as a method of converting a music signal from a time axis component to a frequency axis component, it is possible to reduce the number of multiplications and perform a fast discrete Fourier transform, and easily determine the phase of the analogy spectrum. It can be calculated.

Next, in the spectrum selection unit 107, as shown in FIG.
Either the spectrum of the original current block shown in (a) or the spectrum of the analog current block shown in FIG. 4 (e) is selected according to the threshold value TH. By this selection processing, FIG.
In (a), the spectrum of the music data that is equal to or greater than the threshold TH is output as it is (that is, output as shown in FIG. 4B). On the other hand, since the spectrum level is equal to or lower than the threshold value TH, the spectrum of the minute music data determined to be noise and removed is replaced with the analog of the current block shown in FIG.

Therefore, the spectrum of the current block to be output becomes the spectrum shown in FIG. 4 (f), and it is possible to restore the spectrum of the minute music data which is judged to be noise and is removed.

FIG. 5 is a schematic diagram showing a case where the level of music data gradually decreases with time. Figure 5 (a) shows
The spectrum of the original music data is shown in Fig. 5 (b).
Shows the spectrum of music data in the prior art,
FIG. 5C shows the spectrum of music data according to the present invention.

When the level of the music data is gradually attenuated with time, the original original music data shows that the noise spectrum of the quantization noise changes with time as shown in FIG. 5 (a). The level is almost constant. On the other hand, the level of the music data spectrum is gradually attenuated as time passes, and minute music data becomes buried in the noise spectrum.

In order to remove the noise spectrum of the original original music data shown in FIG.
As shown in (b), a threshold TH is set, and the spectrum of a level below the threshold TH is removed. However, in this conventional technique, the noise spectrum below the threshold TH is removed, but the spectrum of the music data whose level is gradually attenuated and falls below the threshold TH is also removed at the same time.

In contrast to this conventional technique, in the present invention, FIG.
As shown in (c), the spectrum is analogized from the two blocks before the current block, and the minute music data below the threshold is replaced with the analogized music data spectrum and output.

Therefore, while the quantization noise is removed, the minute music data is estimated, reconstructed and output, so that the sound quality of the attenuated portion of the musical instrument or the like can be improved.

More specifically, a PCM having a coarse quantization resolution.
In the data, in particular, it is easy to perceive deterioration in sound quality because distortion occurs when the sound of the musical instrument is attenuated and disappears, or the attenuated sound disappears midway. The attenuated portions of these musical instruments and the like have a strong temporal correlation, and a relatively good waveform can be estimated by using FFT analysis. In the present invention, based on the estimation result by such FFT analysis,
It is possible to obtain PCM data in which the bit length is extended by adding a minute component that has disappeared due to quantization.
It is possible to greatly improve the sound quality of the attenuation part of a musical instrument or the like.

In the above example, the linear function of equation 3 is used to perform the spectrum of the current block, but instead of this, linear prediction according to the following equation may be used.

[0042]

[Equation 4]

Here, x _ni is i for the current block n.
It is an observation value of the previous block, α _i is a linear prediction coefficient, and x _n is a linear prediction value of the current block. This linear prediction uses the spectrum of the current block to
From (1 to p blocks before), for example, when the spectrum of the current block is inferred from the previous 10 blocks (p = 10), this is used to determine the current block from the block two blocks before the current block. It is possible to predict a spectrum with a more accurate value than when predicting a spectrum of.

In this case, the temporary storage buffer 103 needs to have a capacity capable of storing spectrum data of several blocks (p blocks) or more.
Further, the spectrum selection unit 107 may be provided with a determination processing unit that determines which block spectrum is used to analogize the spectrum of the current block among the spectrum levels of a plurality of blocks. When a spectrum of a plurality of blocks is used, for example, if a sudden pulse noise or the like is added 6 blocks before the current block to 4 blocks before, the 10 blocks before the current block including the block to which the noise is added are added. When a spectrum is inferred from a block, the inferred spectrum may be a spectrum having a discontinuous value with respect to the spectrum of the previous block. Therefore, since the determination processing unit is provided to remove the spectrum of sudden pulse noise and the like and analogize the spectrum, it is possible to obtain a spectrum of continuous values that are not affected by noise.

As described above, the present invention extends the bit length by digital signal processing.
It is applicable when playing an old PCM recording source recorded at about 3 or 14 bits with a reproducing device having an accuracy of 16 or 20 bits, or when playing a compact disc at a higher resolution such as 20 bits. While removing the frequency axis component (frequency coefficient) of low-level quantization noise, while removing the quantization noise when expanding the bit length of digital data, the level of the minute level that is likely to be buried in the quantization noise is reduced. For the music signal, digital data having a large effective bit number can be obtained while restoring the music signal.

That is, according to the present invention, the original waveform is divided into specific block lengths to perform frequency conversion to obtain frequency coefficients, and from these frequency coefficients, a spectrum below a level considered to be a quantization noise level is obtained. Of the frequency band over several blocks in parallel with it, by analogy with the frequency coefficient in the next block, apply the analogical coefficient to the extracted band,
The complete set of frequency coefficients thus obtained is inversely transformed by using an inverse transformation device having sufficient calculation accuracy to obtain bit-extended PCM data.

In this way, from the frequency coefficients, the coefficients below the level that is considered to be the quantization noise level are
By substituting with the spectrum inferred from the frequency analysis result and performing the inverse transform process, the quantization noise can be removed, and the minute signal that may have been buried in the quantization noise can be restored, improving the sound quality of the old recording source. , Data with a large number of effective bits can be obtained.

[0048]

As described above, according to the present invention, the digital data of the effective bit number N is divided into several blocks, converted into frequency axis components, and a predetermined level is selected from the converted frequency axis components. The following frequency axis components are replaced with frequency axis components that are analogized from the frequency axis components over several blocks, and then converted into time axis components with the number of effective bits M (N <M) for output. When expanding the bit length of data, while removing quantization noise,
With respect to a music signal of a minute level which is apt to be buried in quantization noise, it is possible to obtain digital data having a large effective bit number while restoring the music signal. The minute music signal buried in the quantization noise can be restored, and the sound quality of old recording sources can be improved.

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration example of a data conversion device according to the present invention.

FIG. 2 is a diagram showing a more specific configuration example of the data conversion device of FIG.

FIG. 3 is a diagram showing an example of a spectrum analyzed by an FFT unit.

FIG. 4 is a diagram for explaining a process of analogizing a small amount of music data in the data conversion device of the present invention.

FIG. 5 is a schematic diagram showing a case where the level of music data gradually decreases with time.

[Explanation of symbols]

2 conversion means 4 Estimating means 7 Replacement means 8 Inverse conversion means 102 FFT section 103 temporary storage buffer 104 spectrum estimation unit 105 threshold generator 106 spectrum comparison unit 107 spectrum selection unit 108 Inverse FFT section

Claims (2)

(57) [Claims] 1. Specifying digital data having an effective bit number N
Converting means for dividing into blocks and converting to frequency axis components
And from the frequency axis component transformed over several blocks,
Estimating means for estimating the frequency axis component of the block to be converted to
And among the frequency axis components converted by the conversion means,
Estimate frequency axis components below a certain level
Replacement to replace the frequency axis component estimated by the means
Means and the frequency axis component from the replacing means, the effective bit number M
Inverse conversion means for converting to a time axis component when (N <M) and outputting
In the data conversion device including: the estimating unit, the spectral absolute value over the block immediately before, the spectral absolute value over the block before two, respectively,
When X _B-1 and X _B-2 are used, the frequency of the two blocks before the current block is calculated using a linear function of X _B = X _B-1 + {X _B-1 −X _B-2 }. A data conversion device characterized in that the frequency axis component X _B of the current block is predicted from the axis components X _B-1 and X _B-2 . 2. Digital data having an effective bit number N is specified.
Converting means for dividing into blocks and converting to frequency axis components
And from the frequency axis component transformed over several blocks,
Estimating means for estimating the frequency axis component of the block to be converted to
And among the frequency axis components converted by the conversion means,
Estimate frequency axis components below a certain level
Replacement to replace the frequency axis component estimated by the means
Means and the frequency axis component from the replacing means, the effective bit number M
Inverse conversion means for converting to a time axis component when (N <M) and outputting
In the data conversion device, the estimation device estimates the frequency axis component of the current block from the frequency axis components of a plurality of previous blocks by linear prediction.