JP3254953B2 - Highly efficient speech coding system - Google Patents

Abstract

PURPOSE: To reduce the arithmetic quantity when the offset amount of a masking reference curve is calculated and to improve the sound quality for higher satisfaction of the acoustic psychology. CONSTITUTION: An acoustic psychology analysis part 3 calculates the power spectrum of an audio signal from an orthogonal conversion factor, calculates the auto-correlation of the power spectrum for every band that is previously decided, and then calculates the offset amount of the acoustic psychological masking effect from the ratio between the maximum and minimum auto- correlation value. Based on this offset amount, the quantization bit number is decided for every sub-band of a quantizing/coding means 4. Then a 2nd necessary SN ratio is calculated from the signal power of every sub-band by means of a root-mean-square error minimum theory including the acoustic control, and the final necessary SN ratio is calculated by giving weighting to the 1st and 2nd SN ratios. Based on the final SN ratio, the quantization bit number is decided for every sub-band of the means 4.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a high-efficiency audio encoding apparatus for dividing an audio signal into a plurality of frequency bands (sub-bands) and quantizing and encoding the divided signals for each sub-band. The present invention relates to a high-efficiency speech coding apparatus that determines the number of quantization bits for each subband based on psychoacoustic analysis.

[0002]

2. Description of the Related Art High-efficiency audio encoding in minidiscs (MD), digital compact cassettes (DCC), karaoke CDs, and the like compresses the data amount of audio signals, and is therefore called music compression. In such an encoding method, an audio signal is divided into a plurality of subbands by a digital filter or an orthogonal transform, and the number of quantization bits for each subband is determined based on psychoacoustic analysis in a frequency domain. In the following description, the term “encode” may be used in a sense of compression other than encoding.

FIGS. 22A to 22D show an example in which a frequency band is divided by orthogonal transform in such an encoding system. (A) of FIG.
This indicates that the bit PCM audio signal has been cut out by 512 samples. Here, the description will be made on the assumption that the total information amount enclosed by the rectangle in the figure is 16 bits * 512 = 8192 bits. Of course, the number of samples to be cut out and the number of bits of PCM are not limited to this value.

FIG. 22B shows a signal obtained by frequency-transforming the signal shown in FIG. 22A by orthogonal transform such as DCT (discrete cosine transform) or FFT (fast Fourier transform). The spectrum envelope is shown. Here, assuming that the information amount is preserved by the orthogonal transformation, this entire information amount can also be represented by a rectangular area in the figure. On the other hand, according to the psychoacoustic model, FIG.
(B), when there is a signal, the signal level that is masked by the signal and becomes inaudible can be defined as a curve, which is generally called a masking effect (to be described in detail later).

If a masking curve is drawn from FIG. 22 (b), it can be expressed as shown in FIG. 22 (c). Here, it is considered that the signal shown in FIG. 22 (b) is requantized. Then, if the quantization noise level generated by the requantization is equal to or less than the level defined by the masking curve,
It can be said that the noise is inaudible to human ears. Therefore, as shown in FIG. 22D, the spectrum is divided into sub-bands for each of a plurality of data, the maximum signal level for each sub-band is set to S, and the noise level allowed from FIG. If N is re-quantized with the number of bits satisfying this S / N, the quantization noise at that time is masked and cannot be heard.

The rectangles shown in FIG. 22 (d) indicate the amount of information required at the time of compression and decompression. Particularly, the deformed rectangle at the center of the figure shows main information, and the elongated rectangle at the bottom of the figure shows auxiliary information. . The auxiliary information is information indicating the maximum value (scale value) of each subband and the number of quantization bits necessary for decoding. Therefore, the total information amount shown in FIG. 22D is the sum of the main information amount and the auxiliary information amount.
(A) and a fraction of the total information amount in (b) of FIG.
It turns out that it becomes. By repeating the above process for each predetermined section (512 sample section in this example), encoding can be performed with almost no deterioration in sound quality.

FIG. 23 shows a general encoding process. First, for example, a 16-bit PCM audio signal is cut out for 512 samples, and the audio signal of each sample is orthogonally transformed by DCT, FFT, or the like, and divided into a plurality of subbands s (step S1). Then, the maximum value (scale value) of each subband s is determined by the psychoacoustic analysis (step S2), and the allowable noise level N [s] of each subband is determined (step S3). Next, the required S / N ratio for each subband is determined (step S4), and then the number of quantization bits for each subband is determined from this S / N ratio (step S5).
Each subband is quantized and output together with auxiliary information (step S6).

Next, a method of calculating a masking curve based on auditory psychology will be described. The masking effect means that when a certain frequency spectrum exists, sounds around a certain level or lower cannot be detected. FIG. 24 shows a masking curve relating to various frequency spectra. The slope of the curve is steeper in a lower frequency range and is gentler in a higher frequency range.

When the horizontal axis (frequency) in FIG. 24 is converted into a scale proportional to the critical bandwidth of hearing, these curves have substantially the same shape and slope as shown in FIG. Have been. As shown in FIG. 26, this critical bandwidth can be expressed by dividing DC to 20 kHz into 25 bands, and the auditory characteristics such as masking may behave in proportion to this critical bandwidth. Many,

A masking curve when a general signal as shown in FIG. 22B is present is a sum (superposition) of masking curves for individual frequency spectra as shown in FIG. 24 or FIG. It can be expressed by a curve as shown in FIG. 22C. However, in the actual calculation, if a masking curve is calculated as a smooth curve as shown in FIG. Have difficulty. Therefore, as an approximation, the spectrum is replaced with the power for each analysis band, and the masking curve is evaluated as a polygonal waveform for each analysis band.

Next, in FIG. 22D, a conventional example in which a masking curve is calculated and the noise level N is derived by setting the minimum value in each subband section of the masking curve as the noise level N allowed in the subband. Will be described with reference to FIG. In FIG. 27, (1)
(1) First, q (j = 0 to q−
1) m analysis bands i (i
= 0 to m−1) is calculated for each band total power P [i].

[0012]

(Equation 1)

(2) Next, a convolution operation of the masking reference curve B corresponding to the analysis band i and the band total power P [i] is performed as shown in the following equation (Equation 2) to obtain each analysis band i. Masking level M [i] in
Is calculated. Here, the masking reference curve B can be represented by B [k] (k is an integer) as shown in FIG. 28 in the case of a constant shape regardless of the analysis band i.

[0014]

(Equation 2)

(3) -1: Next, when the analysis band i is different from the sub-band s, the minimum masking level M [i] in the section of the sub-band s is changed to the allowable noise level N [s (The number of subbands is n).

[0016]

N [s] = min [M [i]]

Here, i indicates each band included in the sub-band s [S], and s = 0 to n-1.

(3) -2: analysis band i and sub-band s
Are the same

[0019]

N [s] = M [s] where s = 0... N-1

(4) Signal level S of each subband s
[S] is obtained.

[0021]

(Equation 5)

(5) The required S / N ratio (S) of each subband s is determined based on the signal level S [s] and the allowable noise level N [s].
Nreq [s]) (average S / N ratio).

[0023]

## EQU6 ## SNreq [s] = 10.0 · log 10 (S
[S] / N [s])

The permissible noise level N [s] of each subband s is obtained by the above processes (1) to (3), and the required S / N ratio of each subband s is obtained by the processes (4) and (5). The number of quantization bits (and the number of inverse quantization bits) of each subband s is determined based on the required S / N ratio.

Here, the setting of the masking reference curve B [k] explained in FIG. 28 plays an important role in such a series of processing. Generally, it is said that the masking effect behaves differently depending on the properties of the signal on the masking side and the signal on the masking side, and specifically, as shown in FIG. 28, the difference between the peak value and “0.0”. The offset F "is affected by the nature of the signal.

In the high-efficiency encoding method, since the signal on the masked side is "noise", the offset amount F changes depending on the signal on the masked side. According to experiments, when the signal on the masking side is a “sine wave”, F ≒ 25 dB,
It is reported that F ≒ 5 dB in the case of “noise”. Actual music / speech signals input to the high-efficiency coding have an offset amount F with these values as upper and lower limits, and it is often necessary to appropriately measure this offset amount F and use it for psychoacoustic analysis. It can be said that it is necessary to achieve sound quality.

It is desirable that the offset amount F be measured for each section of the processing and for each frequency band. As a conventional method of measuring the offset amount F, it is general to determine tonality. Tonality is an index indicating the purity of a signal, and is 1.0 (sine wave) to
It takes a value in the range of 0.0 (noise). This tonality is represented by the F
It is calculated from the linear prediction of the FT spectra A, B, C.
Note that there may be a gap between the sections, or there may be overlapping portions. To obtain a q-point spectrum, a 2q-point FFT spectrum is required.

FIG. 30 shows a conventional method for calculating the offset amount F by obtaining the tonality. The amplitudes R ₁ [j] and R ₂ of the FFT coefficients of the three sections
[J], R ₃ [j] (j = 0 to q−1) and phase Φ ₁
[J], Φ ₂ [j], Φ ₃ [j] are obtained. Here, generally, (R ₃ , Φ ₃ ) is the spectrum of the current section,
Also, (R ₂ , Φ ₂ ) is the spectrum of the previous section, and (R ₁ , Φ ₂ )
In many cases, Φ ₁ ) is the spectrum of the previous section.
The amplitude R [j] and the phase Φ [j] are obtained from the real part (Real [j]) and the imaginary part (Imag [j]) of the FFT coefficient as follows.

[0029]

(Equation 7)

The spectra R _X [j] and φ _X [j] of the third section predicted from R ₁ , R ₂ , Φ ₁ , and Φ ₂ are obtained by linear prediction as follows.

[0031]

R _X [j] = 2 · R ₂ [j] −R ₁ [j] Φ _X [j] = 2 · Φ ₂ [j] −Φ ₁ [j]

The predicted value (R) on the (R, Φ) plane
_X , Φ _X ) and the distance c [j] between the actually measured value (R ₃ , Φ ₃ ) are evaluated. This distance is unpredictable (unpredictabi
lity).

[0033]

(Equation 9)

The unpredictability c [j] is weighted by the power spectrum for each analysis band i, averaged, and the unpredictability c
2 [i] is obtained.

[0035]

(Equation 10)

Unpredictability c2 [i] after weighting processing
To the tonality t [i].

[0037]

## EQU11 ## t [i] = a + b.ln (c2 [i])

However, a and b are 0.0 ≦ t [i] ≦ 1.0
A constant that is determined to be The offset amount F [i] is calculated from the tonality t [i].

[0039]

F [i] = α · t [i] + β · ｛1.0−t
[I]｝ [dB] where α = 25.0, β = 5.0, etc.

[0040]

However, the method of calculating the tonality t [i] has the following problems. Problem (1) The amount of calculation is large. In the processing shown in FIG. 30, since the calculation of the square root and the arc tangent is performed for each sample, the calculation amount is considerably large. The square root is also used in the distance calculation in the process 3.

Here, when the system is realized by a DSP (Digital Signal Processor) or the like, if a general product-sum operation is performed by one instruction, the function operation becomes 100
It is considered more than instruction. In processing, the square root is calculated twice and the arc tangent is calculated once.
= 512 (1024 point FFT), so at least 100 · 512 · 3 = 153600
This consumes the number of operations.

For example, if the DSP has a capacity of 20 MIPS (Mill
If the sample frequency fs = 44.1 kHz, the calculation amount per section is 20 · 10 ⁶ · 512 / 44100.0 ≒ 2322
Since it is 00 times, this DSP consumes about 66% of the calculation amount.

In some high-efficiency coding systems, orthogonal transforms that use transform coefficients that cannot be expressed as amplitude and phase, such as M (Modified) DCT, may be used instead of FFT. In this case, it is necessary to perform an FFT operation separately in order to perform the tonality calculation, and the amount of operation increases accordingly.

Problem (2) There is a problem in the tonality calculation itself when the audio signal is vibrato. For example, when the input signal is a vocal or a single instrument and vibrato is applied, as shown in FIG.
It drifts at a period of about 10 Hz. For example, when the section length is 512 samples and the sections are close,
The movement amount of the center in three sections is 1024 samples → 23
msec, which is a 1/4 cycle of vibrato of 10 Hz (2
5 msec).

Therefore, in the conventional tonality calculation, since the linear prediction is performed for each spectrum, the prediction accuracy is deteriorated by vibrato, and the calculated tonality is extremely high even though the signal is originally a signal with high tonality in terms of hearing. There is a problem that the measurement becomes lower and the measurement is deviated from the hearing.

Problem (3) The calculation of the required S / N ratio (SNreq [s]) based on the psychoacoustic analysis as in the processing (1) to (5) shown in FIG. However, if the data compression ratio is high and the S / N ratio of each subband s after quantization / inverse quantization is lower than the required S / N ratio, a problem occurs. That is, in the conventional method, if the required S / N ratio by the psychoacoustic analysis is not satisfied, the S / N ratio of all subbands s is deteriorated on average. When the S / N ratio is deteriorated, noise is gradually detected in accordance with the amount, and at this time, the deterioration of a band having a large signal power tends to be more audible.
Therefore, the conventional method cannot be said to be optimal in sound quality in a situation where deterioration of the S / N ratio can be detected.

Here, in order to alleviate the above-mentioned problems, in the conventional method, when the required S / N ratio is not satisfied, information in a band with a small power is reduced and information is allocated to a larger band. Be taken. However, in this method, when information of one band and one bit is moved, for example, the S / N ratio of the movement source is deteriorated by about 6 dB, and the S / N ratio of the movement destination is improved by about 6 dB. Become. In addition, since the correction is performed using the band power itself, a new problem occurs in that a band having a large power (for example, a middle and low frequency range) is given too much importance.

Problem (4) By the way, in the above description, it is considered that an independent audio signal is coded with high efficiency.
Further, depending on the system, it is conceivable to transmit a high-efficiency coded signal and a signal without high-efficiency coding, mix these signals on the reproducing side, and reproduce them as one audio signal.

As the simplest example, for example, as shown in FIG. 32, the audio signal of channel (CH) -A is encoded by the audio encoder 20 with high efficiency, and
Are multiplexed by the multiplex unit 21 and transmitted without high efficiency coding. On the reproduction side, the channel is separated by the demultiplex unit 22 and the signal C decoded by the audio decoder 23 is output.
The audio signals of HA ′ and CH−B are mixed by the mixer 24.

As another example, as shown in FIG. 33A, a CH-A audio signal is encoded with high efficiency, and MIDI (Musical Instrument Digital) which is an international standard of a digital signal for controlling an electronic musical instrument is used. Int
(Erface) The CH-B audio signal is converted into MIDI code by the sequencer 25 and multiplexed by the multiplex unit 21 for transmission. Then, on the reproduction side, the channel is separated by the demultiplex unit 22 and the signal CH-A ′ decoded by the audio decoder 23 and the MID
The signal CH-B ′ played by the MIDI sound source 26 is mixed by the mixer 24 based on the I code.

As a modification, as shown in FIG. 33 (b), a CH-A audio signal is encoded with high efficiency, and two channels of CH-B1 and CH-B2 are MIDI-coded.
The two channels CH-B1 'and CH-B are coded and reproduced on the reproduction side from the signal CH-A' and the MIDI code.
2 ′ are mixed by mixers 24-1 and 24-2, respectively, and output on two channels. Among them, the system shown in FIG. 33B has recently been used for communication karaoke using a MIDI code, and the signal to be efficiently encoded is often a real voice chorus.

However, CH-A for high efficiency coding
Is assigned a bit by psychoacoustic analysis using only the audio encoder 10,
The effect of CH-B on the mixing side on the reproduction side is not considered. In other words, when mixing is performed on the playback side,
The CH-A audio signal will be affected by the masking effect from the CH-B signal, and thus will be optimally encoded when listening to only the CH-A audio signal, but will mix other signals. If so, the sound quality is not optimal.

FIG. 34 shows the signal spectrum of CH-A and C
Masking level M1 obtained by psychological analysis of HA
[I] and an example of a masking level M2 [i] from another channel to be mixed, where M1 <M2 in the low band and the high band, and M1> M2 in the middle band.
In this case, the optimum masking level M [i] for the CH-A signal after mixing is determined by processing (3) shown in FIG.
At

[0054]

M [i] = max (M1 [i], M2 [i]) where i = 0 to m−1

It is considered that: (A) of FIGS. 32 and 33
As shown in (b), when mixing is performed, the masking level deviates from this optimum value M [i], and there is a problem in that the masking level is not optimal in terms of hearing. In particular, when the compression ratio is high such that the actual noise level is equal to or higher than the masking level, noise is also reduced in the region where M1 [i]> M2 [i] in FIG. A phenomenon that sounds emphasized occurs.

In view of the above-mentioned problems (1) and (2), the present invention can reduce the amount of calculation when calculating the offset amount of the masking reference curve, and can improve the sound quality by further satisfying the psychological sense of hearing. It is an object of the present invention to provide a high-efficiency audio coding device. In view of the above-mentioned problem (3), the present invention also provides a high-efficiency audio coding apparatus capable of improving the sound quality when the data compression ratio is high and the required S / N ratio by psychoacoustic analysis is not satisfied. The purpose is to provide. In view of the above-mentioned problem (4), the present invention also provides a psychoacoustic system that takes into account the influence of a signal that is not efficiently coded when mixing a signal that has been efficiently coded and a signal that is not to be highly efficient. It is an object of the present invention to provide a high-efficiency speech coding apparatus capable of performing analysis to more satisfy psychoacoustics and improve sound quality.

[0057]

In order to achieve the above object, the present invention calculates a power spectrum of an audio signal from orthogonal transform coefficients and calculates an autocorrelation of the power spectrum for each predetermined band. An offset amount of the masking effect on psychoacoustics is calculated from a ratio between the maximum value and the minimum value of the autocorrelation, and the number of quantization bits of each subband is determined based on the offset amount.

That is, according to the present invention, the dividing means for dividing the audio signal into sub-bands of a plurality of frequency bands, and the audio signal of each sub-band divided by the dividing means is quantized with a variable number of quantization bits. And a quantizing / encoding means for encoding, and a power spectrum of the audio signal calculated from the orthogonal transform coefficient obtained by the dividing means or the separate orthogonal transform means, and the autocorrelation of the power spectrum is determined for each predetermined band. The offset amount of the masking effect on psychoacoustics is calculated from the ratio of the maximum value and the minimum value of the autocorrelation, and the quantization bit of each subband of the quantization / encoding means is calculated based on the offset amount. A high-efficiency speech coding device having a psychoacoustic analysis means for determining the number is provided.

The present invention also calculates a first required S / N ratio for each subband based on psychoacoustic analysis of the frequency domain of the audio signal, and includes a root-mean-square method including auditory control from the signal power for each subband. The second required S / N ratio is calculated by the minimum error theory, the first and second required S / N ratios are weighted to calculate the final required S / N ratio, and this final required S / N ratio is calculated. , The number of quantization bits of each subband is determined.

That is, according to the present invention, the dividing means for dividing the audio signal into sub-bands of a plurality of frequency bands, and the audio signal of each sub-band divided by the dividing means is quantized with a variable number of quantization bits. And a quantization / encoding means for encoding, and a first required S / N ratio for each subband based on psychoacoustic analysis of a frequency domain of the audio signal, and auditory control based on signal power for each subband. The second required S / N ratio is calculated according to the root mean square error theory including the following, and the first and second required S / N ratios are weighted to calculate the final required S / N ratio. A high-efficiency speech coding apparatus is provided which includes psychoacoustic analysis means for determining the number of quantization bits of each subband of the quantization / coding means based on a required S / N ratio.

The present invention also relates to a psychoacoustic system in which a first audio signal to be encoded with high efficiency and a second audio signal which is not encoded with high efficiency and mixed with the first audio signal on the reproduction side are respectively in the frequency domain. Analyzing to calculate first and second masking levels, calculating a final masking level based on the first and second masking levels,
The number of quantization bits of each subband is determined based on the final masking level.

That is, according to the present invention, the dividing means for dividing the first audio signal to be encoded with high efficiency into sub-bands of a plurality of frequency bands, and the audio signal of each sub-band divided by the dividing means is variable. Quantization / encoding means for quantizing and encoding with the number of quantization bits, the first audio signal, and a second audio signal which is mixed with the first audio signal on the reproduction side without being encoded with high efficiency. Each of the audio signals is subjected to psychoacoustic analysis in the frequency domain to calculate first and second masking levels. A final masking level is calculated based on the first and second masking levels. And a psychoacoustic analysis means for determining the number of quantization bits of each subband of the quantization / coding means based on the speech / voice efficiency coding apparatus. .

[0063]

In the present invention, the power spectrum of the audio signal is calculated from the orthogonal transform coefficients, the autocorrelation of the power spectrum is calculated for each predetermined band, and the psychoacoustic is calculated from the ratio of the maximum value and the minimum value of the autocorrelation. Since the offset amount of the above masking effect is calculated, and the number of quantization bits of each subband is determined based on the offset amount, the amount of calculation when calculating the offset amount of the masking reference curve is reduced, and Even when vibrato is applied to the signal, the auditory psychology can be further satisfied and the sound quality can be improved.

Further, in the present invention, the first necessary S / N ratio for each subband is calculated based on the psychoacoustic analysis of the frequency domain of the audio signal, and the square including the auditory control is calculated from the signal power for each subband. The second required S / N ratio is calculated by the average error minimum theory, the first and second required S / N ratios are weighted to calculate the final required S / N ratio, and the final required S / N ratio is calculated. Since the number of quantization bits of each subband is determined based on the ratio, the data compression ratio is high, and the sound quality can be improved when the required S / N ratio by psychoacoustic analysis is not satisfied.

Further, according to the present invention, the first efficient coding is performed.
Audio signal and the first signal on the playback side without high efficiency encoding
The audio signal and the second audio signal to be mixed are each subjected to psychoacoustic analysis in the frequency domain to calculate first and second masking levels, and based on the first and second masking levels, a final masking level is calculated. Is calculated, and the number of quantization bits of each subband is determined based on the final masking level. Therefore, when the high-efficiency coded signal and the non-high-efficiency coded signal are mixed on the reproduction side, high-efficiency coding is performed. The psychoacoustic analysis is performed in consideration of the influence of the no-signal, so that the psychoacoustic can be more satisfied and the sound quality can be improved.

[0066]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a first embodiment of a high-efficiency audio coding apparatus according to the present invention, FIG. 2 is a block diagram showing a modification of FIG. 1, and FIG. 3 is a block diagram showing another modification of FIG. FIG. 4 is an explanatory diagram showing a case of calculating the autocorrelation of the power spectrum with the preceding and following subbands, and FIG. 5 is a flowchart for explaining a process of calculating the offset amount.
FIG. 6 is an explanatory diagram showing an example of a spectrum of an audio signal in which vibrato exists, and FIG. 7 is an explanatory diagram comparing an offset amount obtained by a conventional tonality calculation method and an autocorrelation method of the first embodiment.

The first embodiment shown in FIG. 1 shows a case where the band division of an audio signal is performed by orthogonal transformation. FIG.
In, for example, a 16-bit PCM audio signal is cut out by the windowing / cutout unit 1 for 512 samples,
The audio signal of each sample is converted into a DC signal by the orthogonal transform unit 2.
Orthogonal transform by T, FFT, etc., and a plurality of subbands s
Is divided into

Then, the psychoacoustic analyzer 3 calculates the offset amount F of the masking reference curve, determines the number of quantization bits, and the quantization / encoding unit 4 uses the number of quantization bits to Each divided subband s
Is quantized and encoded. The data quantized and encoded by the quantization / encoding unit 4 and compressed, and the number of quantization bits determined by the psychoacoustic analysis unit 3 are multiplexed by the multiplex unit 5 to obtain M
Output to D, DCC, etc. At the time of decompression, the compressed data is inversely quantized and decoded based on the number of quantization bits for each subband s.

In the modification shown in FIG. 2, the input audio signal is divided into sub-bands s by the digital filter 6, quantized and encoded by the quantizing / encoding unit 4 and compressed, and the psychoacoustic analysis is performed. The number of quantization bits determined by the unit 3 is configured to be multiplexed by the multiplex unit 5. Here, in the sub-band division method using the filter bank, the low band resolution required by the present invention cannot be obtained. Therefore, as in the case shown in FIG. The signal is divided into a plurality of sub-bands s by the orthogonal transformation unit 2, and the offset amount F of the masking reference curve is calculated by the psychoacoustic analysis unit 3, and is quantized.
The number of quantization bits of the encoding unit 4 is determined.

In the modification shown in FIG. 3, a sequence for an audio signal and a windowing / cutout unit 1 for determining the number of quantization bits are provided.
a, 1b, orthogonal transform units 2a, 2b (and an offset calculation amount calculation unit 7) are provided. In the case where the orthogonal transform units 2a and 2b are configured in such a manner, the orthogonal transform units 2a and 2b may be configured to have different numbers of points. .

Next, the processing for calculating the autocorrelation of the power spectrum with the preceding and succeeding subbands will be described with reference to FIG. For example, when the power spectrum of the predetermined sub-band s and the sub-bands s−1 and s + 1 before and after the sub-band s are as shown in FIG. , S + 1 is calculated. If the result is as shown in FIG. 4B, the ratio of the maximum value and the minimum value of the autocorrelation value is logarithmically converted to calculate the offset amount F. As a result, the offset F increases in the case of a tone-like signal with clear harmonic components, and decreases in the case of a noise-like signal. Note that, as shown in FIG. 4B, the slide amount = 0 and its peripheral position are excluded from the maximum value search.

Next, the processing for calculating the offset amount F of the masking reference curve will be described with reference to FIG. FIG. 5 shows, as an example, 2q points F as an orthogonal transformation.
This shows a case where FT is used, and the number of points of this orthogonal transformation is 2
It is desirable that q is a value of about 1024 to 2048. In FIG. 5, first, the real part Real [j] and the imaginary part Imag [j] of the orthogonal transform coefficient
From the power spectrum p [j].

[0073]

## EQU14 ## p [j] = Real [j] ² + Imag [j] ² where j = 0 to q-1

Next, the autocorrelation Sc [s] [i] is determined for each predetermined band. If q samples are divided into n bands,

[0075]

(Equation 15)

Finally, the autocorrelation Sc for each band
[S] Offset amount F from maximum and minimum values of [i]
[S] is calculated.

[0077]

(Equation 16)

FIGS. 6A and 6B show that the number of points of the orthogonal transform units 2 and 2a shown in FIGS.
The spectrum of the audio signal with vibrato is 1
This shows a case where the peak is shifted by 024 points (23 msec), and the peak is offset as is clear from the figure. FIG. 7 shows the offset amount F obtained for each of the 32 bands by the conventional tonality calculation method and the autocorrelation method of the present embodiment. As is apparent from FIG. There is a lot of information,
The offset amount F according to the autocorrelation method of this embodiment matches the audibility.

The calculation amount for obtaining the offset amount F is, for example, 2048 of the orthogonal transformation unit 2b shown in FIG.
Assuming that the FFT point is included, the processing shown in FIG. 4 is approximately 90,000 times assuming that the function operation is performed 100 times, the multiplication is performed once, and the division is performed 20 times. Can be halved compared to about 180,000 times.

Next, a second embodiment of the present invention will be described with reference to FIGS. FIG. 8 is a block diagram showing a high-efficiency audio coding apparatus according to a second embodiment, FIG. 9 is an explanatory diagram showing a relationship between a noise shaping factor and quantization noise, and FIG. 10 is a diagram showing first and second necessary S / S. From N to the final required S
A flowchart for explaining a process of calculating / N;
11 is an explanatory diagram showing a weighting function for calculating the final required S / N, FIG. 12 is an explanatory diagram showing a spectrum of a source in which deterioration of the S / N ratio is easily detected, and FIG. 13 is a source diagram shown in FIG. FIG. 14 is an explanatory diagram showing a comparison result of sound quality between the conventional example and the second embodiment.

In the second embodiment shown in FIG. 8, the first necessary S /
Psychoacoustic analyzer 3 for calculating the N ratio, and second required S / N calculation for calculating the second required S / N ratio based on the root mean square error theory based on the signal power of each subband s ( And final required S / N calculation) section 8 and bit allocation section 9
Having. The first necessary S / N is obtained by a psychoacoustic model centering on the masking effect purely as in the conventional example, and the second required S / N ratio is determined based on the signal power of each subband s. It is obtained based on the root mean square error theory to which a parameter for controlling the quantization noise aurally is added.

Here, in the latter, the required S / N ratio of a band having a higher power tends to be slightly emphasized as compared with the former. Therefore, first, the first required S / N ratio is normalized such that the total average value of the second required S / N ratio of each subband matches that of the second required S / N ratio. The reason is that the first required S / N ratio is an amount consistent with the psychoacoustic psychology, and the second required S / N ratio is used for the assistance, otherwise the first and second required S / N ratios are used. This is because a malfunction occurs if there is a difference between the average values of the required S / N ratios.

Finally, the first required S / N ratio and the normalized second required S / N ratio are weighted and added to obtain the final required S / N ratio to obtain the quantum of each subband s. Determine the number of coded bits. In this case, the weighting ratio is added with emphasis on the first necessary S / N ratio, for example, first required S / N ratio: second required S / N ratio = 0.7: 0.3. I do. According to the above method, even when the compression ratio is high and the S / N ratio is detected, auditory deterioration can be minimized, and the problem that a band with a large power is overemphasized does not occur.

Next, a method of allocating bits according to the root mean square error theory will be described. In general, it is said that a speech waveform can be approximated by a Gaussian process. In this case, the bit allocation (the number of bits of each band s) bit [s] that minimizes the root mean square error after quantization is represented by the transmission rate-distortion. From the theory, it is expressed as in the following Expression 1 (Equation 17).

[0085]

[Equation 17]

In practice, the coefficients a and b are adjusted so that the sum of bit [s] becomes the number of usable bits. Here, Equation 1 (Equation 17) shows a case where no auditory control is performed, and the obtained bit [s] strongly reflects the band power, and the resulting quantization noise is white noise similar to PCM coding. Becomes Therefore, in this embodiment, when performing auditory control, a weighting factor w [s] is added to Expression 1 (Equation 17), and the following Expression 2 (Equation 18) is obtained.
Get.

[0087]

(Equation 18)

The noise shaping factor γ in Equation 2 (Equation 18) takes a value in the range of −1.0 to 0.0, and coincides with Equation 1 (Equation 17) when γ = 0.0. Conversely, when γ = −1.0, the bit allocation of Expression 2 (Equation 18)
bit [s] is a constant, and the number of quantization bits for each band is the same. FIG. 9 shows quantization noise in the case of γ = −1.0 to 0.0. Generally, it is considered that the audibility matches well when γ = −0.2 to −0.1.

Next, referring to FIG. 10, each processing for calculating the final required S / N from the first and second required S / N
Will be described. First, from the orthogonal transform coefficients, the band total power P
[S] is calculated. For example, when dividing q spectra into n bands,

[0090]

[Equation 19]

Next, the entire band average error power (constant) b is determined from the predetermined band average S / N ratio (SNavr).

[0092]

(Equation 20)

The bit allocation bit [s] of each band s is calculated by Equation 2 (Equation 18).

[0094]

[Equation 21] bit [s] = a + 0.5 · log 2 (w [s]
・ P [s] / b)

A temporary second required S / N ratio (= SNreq '[s]) is calculated from the bit allocation bit [s].

[0096]

[Mathematical formula-see original document] SNreq '[s] = 6.02.bit [s] [dB]

First required S / N ratio and provisional second required S
/ N ratio average value SNreq [s] _avr, SNreq '
[S] _avr is calculated.

[0098]

(Equation 23)

Temporary average value SNre of required second S / N ratio
q ′ [s] _avr is normalized to obtain a second required S / N ratio (SN
req 2 [s]).

[0100]

[Formula 24] SNreq 2 [s] = SNreq '[s] (SNreq_avr / SNreq'_avr) [dB]

The average value of the first required S / N ratio (SNreq
[S] _avr) as a parameter, the first required S / N
Ratio (SNreq [s]) and the second required S / N ratio (SNreq
2 [s]) to the final required S / N ratio (SNreq_fin)
[S]).

[0102]

[Mathematical formula-see original document] SNreq_fin [s] = f [SNreq_avr] · SNreq [s] + (1.0-f [SNreq_avr]) · SNreq2 [s] [dB]

Here, f [x] is, as shown in FIG.
A weighting function for values in the range of 0.0 to 1.0,
When the average value of the required S / N ratio (SNreq [s] _avr) is large, the second required S / N ratio (SNreq 2 [s])
Is set to increase.

Here, as shown in FIG. 12, in the case of a source spectrum in which deterioration of the S / N ratio is easily detected, FIG.
As shown in FIG. 13A, the first and second required S / N ratios take values like thick lines and thin lines, respectively, and the first required S / N ratio as shown in FIG. And the final required S
The / N ratio has a value like a thick line and a thin line, respectively. In the case of such a source, the S / N ratio near 2 to 4 kHz where the power is large is corrected, and therefore, the S / N ratio on the audibility can be improved. Further, as shown in FIG. 14, when comparing the three first required S / N ratios in the conventional example and the present embodiment, according to the present embodiment, the average value of the first required S / N ratio (SNreq [ It can be seen that the improvement effect is large when [s] _avr) is large.

Next, a third embodiment of the present invention will be described. FIG. 15 is a block diagram showing a high-efficiency audio coding apparatus according to a third embodiment, FIG. 16 is a block diagram showing a modification of the high-efficiency audio coding apparatus shown in FIG. 15, and FIG. 17 is an audio encoder shown in FIGS. Block diagram showing an example of
FIG. 18 is a flowchart for explaining a necessary S / N ratio calculation process when synchronization between two channels is guaranteed with sufficient accuracy, and FIG. 19 is a necessary S / N ratio when synchronization accuracy between two channels is poor. FIG. 20 is a flowchart for explaining a ratio calculation process, FIG. 20 is an explanatory diagram comparing MNR after mixing according to the prior art and the present embodiment, and FIG. 21 is an explanatory diagram comparing sound quality evaluation after mixing according to the prior art and the present embodiment. It is.

The third embodiment shown in FIG. 15 is a case where the third embodiment is applied to the audio encoder 20 shown in FIG.
In this case, C to be efficiently encoded by the encoder 20 is used.
The first audio signal of H-A and the CH mixed with the first audio signal on the reproduction side without being encoded with high efficiency
-A psychoacoustic analysis of the second audio signal in the frequency domain to calculate first and second masking levels, and calculate a final masking level based on the first and second masking levels; Based on the final masking level, the number of quantization bits of each subband is determined, and the first audio signal is quantized and encoded, and is output as a bit stream. The bit stream and the CH-B non-efficiently coded signal are multiplexed by the multiplex unit 21.

The third embodiment shown in FIG. 16 shows a case where the third embodiment is applied to the audio encoder 20 shown in FIG. In this case, the MIDI sequencer 25
Converts the CH-B audio signal into a MIDI code, generates a signal CH-B 'played by the MIDI sound source 26 based on the MIDI code, and outputs the CH-A first audio signal and the signal CH-B by the encoder 20. '
First and second masking levels are calculated, and the first and second masking levels are calculated.
Calculating a final masking level based on the second masking level, determining the number of quantization bits for each subband based on the final masking level, quantizing and encoding the first audio signal, Output as The bit stream and the MIDI code are multiplexed by the multiplex unit 21.

On the reproduction side, as shown in FIG. 33A, the channels are separated by the demultiplex unit 22, and the MIDI sound source 2 is separated based on the signal CH-A 'decoded by the audio decoder 23 and the MIDI code.
The signal CH-B ′ played by the mixer 6 is mixed by the mixer 24.

The encoder 20 shown in FIG. 17 can be applied to, for example, a case where the band division of an audio signal is performed by orthogonal transform, and the band division is performed by the digital filter 6 as shown in FIG. In FIG. 17, a signal of a channel CH-A for performing high-efficiency coding and a signal of CH-B to be mixed on the reproduction side without performing high-efficiency coding are windowed / cut out sections 1A and 1B and an orthogonal transform section, respectively. It is divided into subbands by 2A and 2B and applied to the psychoacoustic analyzers 3A and 3B. Note that CH-A, C
If the mixing ratio of the H-B signal on the reproduction side is not 1: 1, the level between CH-A and CH-B is adjusted in consideration of the ratio (level adjusting unit 11).

Next, with reference to FIG. 18, a description will be given of the processing in the case where the synchronization processing (for example, within ± 1 msec) between the channels CH-A and CH-B has been performed in advance.
FIG. 18 corresponds to the processing described in FIG. 27 of the conventional example, and the processing (1) to (5) are the same, and the processing (1) ′,
(2) ′ and (x) are added. The psychoacoustic analyzer 3B calculates the masking level M2 obtained by the psychoacoustic analysis of the frequency domain of the CH-B signal in the processes (1) ′ and (2) ′.

On the other hand, the psychoacoustic analyzer 3A calculates the masking level M1 obtained by the psychoacoustic analysis of the frequency domain of the CH-A signal in the processes (1) and (2), and in the subsequent process (x) This masking level M
1 and the masking level M2 calculated by the psychoacoustic analyzer 3B, M [i] = max shown in the equation (13).
(Final masking level M taking into account the influence of a signal that is not efficiently coded based on M1 [i] and M2 [i]
Is calculated. Next, in processes (3) to (5), the required S / N ratio is calculated based on the final masking level M. The bit allocation unit 12 allocates the number of quantization bits of each subband based on the required S / N ratio, and the quantization / encoding unit 4 quantizes and encodes the CH-A side based on the number of quantization bits. Become

Next, referring to FIG. 19, CH-A, CH-
Processing when the synchronization accuracy between B is poor will be described. Synchronous deviation during mixing is permissible on the auditory perception, but when it is a problem on psychoacoustic analysis, for example, a synchronization error of ± 5 to 10 m
In the case of sec, if the psychoacoustic analysis unit 3B performs the processing (1) ′ and (2) ′ shown in FIG. 18, there is a possibility that the change of the masking level M may have an adverse effect due to the synchronization deviation during actual mixing. There is.

Therefore, the psychoacoustic analysis unit 3B on the CH-B side
Is calculated by setting the orthogonal transform length of CH-B to about twice that of CH-A and calculating the total power P2 [i] of each analysis band in the process (1) ′ shown in FIG. Is flattened and reduced, and a masking level M2 is calculated from the total power P2 [i] and the masking reference curve B (k) in the subsequent processing (2) '. 3A is obtained from M1 [i] and M2 [i] based on the equation (Equation 13).
The processing (x) is performed without determining the maximum value when determining [i].
Then, assuming that the weighting coefficient a is, for example, a = 0.6

[0114]

M [i] = M1 [i] · 0.6 + M2 [i] · 0.4

As described above, M [i] is emphasized with emphasis on M1 [i].
Is determined, the problem of psychoacoustic analysis when the synchronization accuracy between CH-A and CH-B is poor can be solved. Therefore, according to the third embodiment, when the high-efficiency coded signal and the non-high-efficiency coded signal are mixed on the reproducing side, the high-efficiency coding is performed so that the mixed sound quality is optimized. it can.

Here, in general, when objectively evaluating the reproduction quality of an audio signal, an MNR (Mask to Noise Radar) is used.
tio) is often measured. Specifically, as shown in FIG. 34, the masking level M in the frequency domain
And the noise (quantization) noise N actually occurring in the signal. In this case, the masking effect on psychoacoustics is satisfied in the region where the MNR is positive, and no noise is detected. Conversely, in the region where the MNR is negative, the psychoacoustic masking effect is not satisfied, and noise is detected. Also, M
Even if the NR is positive, it is considered preferable to show a frequency characteristic that is as flat as possible in terms of psychoacoustics. The reason for this is that if there is a difference in MNR depending on the band, a sound that is slightly unnatural will be felt in terms of balance.

FIG. 20 shows an example of measuring the MNR [dB] after mixing when a certain audio signal is processed by the prior art and this embodiment, respectively. In the present embodiment (solid line), the characteristics are flat in almost all frequency regions, whereas in the conventional technology (broken line), the characteristics have undulations, and in some regions (around 10 kHz in the figure), negative values are shown. I have. As described above, when the average MNR is close to 0 dB, the effect is particularly large.

FIG. 21 shows an example in which subjective sound quality evaluation (five-level evaluation) is performed after mixing with a large number of sources according to the prior art and this embodiment. According to this embodiment, the average value of the evaluation values is improved. In particular, it can be seen that the variation in the evaluation value is reduced.

[0119]

As described above, according to the present invention, the power spectrum of the audio signal is calculated from the orthogonal transform coefficients, the autocorrelation of the power spectrum is calculated for each predetermined band, and the maximum of the autocorrelation is calculated. Calculate the offset amount of the masking effect on psychoacoustics from the ratio of the value and the minimum value,
Since the number of quantization bits for each subband is determined based on this offset amount, the amount of calculation when calculating the offset amount of the masking reference curve is reduced, and even when the audio signal is vibrato, The sound quality can be improved by further satisfying the psychology.

Further, according to the present invention, the first necessary S / N ratio for each subband is calculated based on the psychoacoustic analysis of the frequency domain of the audio signal, and the square including the auditory control is obtained from the signal power for each subband. The second required S / N ratio is calculated by the average error minimum theory, the first and second required S / N ratios are weighted to calculate the final required S / N ratio, and the final required S / N ratio is calculated. Since the number of quantization bits of each subband is determined based on the ratio, the data compression ratio is high, and the sound quality can be improved when the required S / N ratio by psychoacoustic analysis is not satisfied.

Further, according to the present invention, the first efficient coding is performed.
Audio signal and the first on the playback side without high efficiency encoding
The audio signal and the second audio signal to be mixed are each subjected to psychoacoustic analysis in the frequency domain to calculate first and second masking levels. Based on the first and second masking levels, the final masking level is calculated. Is calculated, and the number of quantization bits of each subband is determined based on the final masking level. Therefore, when mixing the high-efficiency coded signal and the high-efficiency coded signal on the reproduction side, high-efficiency coding is performed. The psychoacoustic analysis is performed in consideration of the influence of the no-signal, so that the psychoacoustic can be more satisfied and the sound quality can be improved.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a first embodiment of a high-efficiency audio coding apparatus according to the present invention.

FIG. 2 is a block diagram showing a modification of FIG.

FIG. 3 is a block diagram showing another modification of FIG. 1;

FIG. 4 is an explanatory diagram showing a case where an autocorrelation of a power spectrum with a preceding and following subband is calculated.

FIG. 5 is a flowchart illustrating a process of calculating an offset amount.

FIG. 6 is an explanatory diagram showing an example of a spectrum of an audio signal in which vibrato exists.

FIG. 7 is an explanatory diagram comparing the offset amounts obtained by the tonality calculation method of the related art and the autocorrelation method of the first embodiment.

FIG. 8 is a block diagram showing a high-efficiency audio coding apparatus according to a second embodiment.

FIG. 9 is an explanatory diagram showing a relationship between a noise shaping factor and quantization noise.

FIG. 10 is a diagram showing a first required S / N to a final required S / N.
9 is a flowchart illustrating a process for calculating N.

FIG. 11 is an explanatory diagram showing a weighting function when calculating a final required S / N.

FIG. 12 is an explanatory diagram showing a spectrum of a source in which deterioration of the S / N ratio is easily detected.

FIG. 13 is an explanatory diagram showing the S / N ratio of the source shown in FIG.

FIG. 14 is an explanatory diagram showing comparison results of sound quality between the conventional example and the second embodiment.

FIG. 15 is a block diagram illustrating a high-efficiency audio coding apparatus according to a third embodiment.

FIG. 16 is a block diagram showing a modification of the high-efficiency audio coding apparatus of FIG.

FIG. 17 is a block diagram showing an example of the audio encoder of FIGS. 15 and 16 in detail.

FIG. 18 is a flowchart illustrating a necessary S / N ratio calculation process when synchronization between two channels is guaranteed with sufficient accuracy.

FIG. 19: Required S when synchronization accuracy between two channels is poor
It is a flowchart for explaining / N ratio calculation processing.

FIG. 20 is an explanatory diagram comparing the MNR after mixing according to the conventional technique and the third embodiment.

FIG. 21 is an explanatory diagram comparing sound quality evaluation after mixing according to the conventional technique and the third embodiment.

FIG. 22 is an explanatory diagram schematically showing a high-efficiency audio encoding method.

FIG. 23 is a flowchart illustrating the high-efficiency audio encoding process of FIG. 22;

FIG. 24 is an explanatory diagram showing an example of a masking curve in various frequency spectra.

FIG. 25 is an explanatory diagram showing a masking curve in which the frequency on the horizontal axis in FIG. 24 is replaced with a critical band.

FIG. 26 is an explanatory diagram showing a critical bandwidth of 25 bands.

FIG. 27 is a flowchart for explaining a conventional required S / N ratio calculation process.

FIG. 28 is an explanatory diagram showing an example of a masking reference curve.

FIG. 29 is an explanatory diagram showing a method of linearly predicting a spectrum of three sections.

FIG. 30 is a flowchart illustrating a conventional offset calculation process.

FIG. 31 is an explanatory diagram showing an example of a spectrum of a signal in which vibrato exists.

FIG. 32 is a block diagram showing a conventional mixing circuit.

FIG. 33 is a block diagram showing another conventional mixing circuit.

FIG. 34 is an explanatory diagram showing a signal to be encoded with high efficiency, a masking level thereof, and a masking level of a signal not to be encoded with high efficiency.

[Explanation of symbols]

 1, 1a, 1b, 1A, 1B Windowing cutout section 2, 2a, 2b, 2A, 2B Orthogonal transformation section (division means) 3, 3A, 3B Perception psychological analysis section (perception psychological analysis means) 4 Quantization / encoding Unit (quantization / encoding unit) 5 multiplex unit 6 sub-band filter unit (division unit) 7 offset amount calculation unit (psychological analysis unit) 8 second necessary S / N calculation unit (psychological analysis unit) 9 Bit allocation unit (psychological analysis means)

──────────────────────────────────────────────────続 き Continuation of front page (56) References JP-A-7-46137 (JP, A) JP-A-3-250923 (JP, A) JP-A-6-232761 (JP, A) JP-A-7-46 66733 (JP, A) (58) Field surveyed (Int. Cl. ⁷ , DB name) H03M 7/30

Claims (3)

(57) [Claims] A dividing unit that divides an audio signal into subbands of a plurality of frequency bands; and a quantizer that quantizes and encodes the audio signal of each subband divided by the dividing unit with a variable number of quantization bits. Encoding and encoding means, calculate the power spectrum of the audio signal from the orthogonal transform coefficients obtained by the dividing means or separate orthogonal transform means, calculate the autocorrelation of this power spectrum for each predetermined band, Calculate the offset amount of the masking effect on psychoacoustics from the ratio of the maximum value and the minimum value of the autocorrelation,
A high-efficiency audio coding apparatus, comprising: psychoacoustic analysis means for determining the number of quantization bits of each subband of the quantization / coding means based on the offset amount. 2. A dividing means for dividing an audio signal into sub-bands of a plurality of frequency bands, and a quantizer for quantizing and encoding the audio signal of each sub-band divided by said dividing means with a variable number of quantization bits. Means for calculating and calculating a first required S / N ratio for each subband based on psychoacoustic analysis of a frequency domain of an audio signal, and including aural control from signal power for each subband The second required S / N ratio is calculated by the minimum theory, and the first and second required S / N ratios are weighted to obtain the final required S / N ratio.
/ Acoustic analysis means for calculating the / N ratio and determining the number of quantization bits of each sub-band of the quantization / coding means based on the final required S / N ratio. apparatus. 3. Dividing means for dividing a first audio signal to be encoded with high efficiency into sub-bands of a plurality of frequency bands, and a variable number of quantization bits for the audio signal of each sub-band divided by said dividing means. Quantizing / encoding means for quantizing and encoding the first audio signal, and a second audio signal which is not efficiently encoded and mixed with the first audio signal on the reproduction side.
Of each audio signal in the frequency domain to calculate first and second masking levels, calculate a final masking level based on the first and second masking levels, and calculate the final masking level. And a psychoacoustic analysis means for determining the number of quantization bits for each sub-band of the quantization / coding means based on the above.