JPS623440B2 - - Google Patents

Abstract

PURPOSE:To improve the signal-to-quantized noise ratio by using the scale factor and to reduce the audio memory of the telephone automatic response unit and further to automate the guide broadcast to the place having comparatively low noise level.

Description

[Detailed Description of the Invention] The present invention, for example, divides an audio signal into fixed intervals of several milliseconds or tens of milliseconds, and
It concerns a method for DPCM encoding (or decoding) (hereinafter referred to as interval DPCM method).
The purpose is to improve signal-to-quantization noise.

Such a section DPCM method is based on the principle of the ``speech processing method'' disclosed in Japanese Patent Application No. 105385-1985, which has already been filed.

FIG. 1 is a block diagram of the interval DPCM audio encoding system. rth (r=0, 1, 2, ......, R)
Discrete audio signal S _ri (i=0,
1, 2, ......, n-1) and the predicted value S^ _ri ,
That is, the prediction error signal e _ri is quantized by the quantizer Q, the output signal e ^* _ri is encoded by the encoder COD, and the encoded signal D _ri and the prediction coefficient A _rj (j=1, 2 ,……,J) and mixer MIX
The signals are mixed together and an output signal M _ri is output.

S ^* _ri is the quantized output signal e ^* _ri and the prediction signal S^ _ri
Although it contains errors due to quantization, it is approximately equal to the audio signal S _ri .

FIG. 2 shows the configuration when restoring an audio signal.
First, a separator SEP separates the encoded signal D _ri and prediction coefficient A _rj from the signal M _ri . Next, in the decoder DEC, D _ri
is decoded, the quantized signal e ^* _ri is derived, and the predicted value S^ _ri
A sum signal S ^* _ri is output. Although S ^* _ri includes a quantization error, it is approximately equal to _Sri .

Figure 3 shows the predictor P used in Figures 1 and 2.
This shows the configuration of D is a delay element, and the predicted value S^ _ri is the past value S ^* _r(i-1) , S ^* _r(i-
2) ,......
It is calculated from S ^* _r(i-J) as shown in the following equation.

Since the interval DPCM method divides the input signal into fixed lengths and updates the prediction coefficient A _rj for each interval, it is not considered to be a stationary stochastic process over a long period of time like an audio signal. The conventional DPCM method (prediction coefficient A _rj
is constant for all sections), a signal S ^* _ri with extremely high approximation is restored. Proceedings of the Acoustical Society of Japan (October 1975), 1-1-
According to what is described in the 6PDPCM audio encoding method, if the signal to ambient noise ratio in the sound field is about 30 dB or less, audio quality that is almost the same as the 8-bit linear PCM method can be obtained. Moreover, it has the advantage that the audio information can be compressed to 1/2.

However, places where the S/N ratio in the sound field is about 30 dB or less are places where the noise level is relatively high, such as stations and airports, and there is a drawback that the practical range is limited to such places.

The present invention eliminates these drawbacks, and makes it possible to automate guidance announcements to places with relatively low noise levels, such as hotels and hospitals, and to reduce the voice memory of automatic telephone answering devices that use public telephone lines. It is something that makes it possible.

The details of the present invention will be explained below using the drawings.
FIG. 4 is a block diagram of a speech encoding device using the present invention. γ _r is a constant coefficient (called a scale factor) that is multiplied by the input signal S _ri in order to improve the signal-to-quantization-noise ratio in divided intervals where the input signal level is low, and the maximum level does not overscale in each interval. It is decided by the degree.

A _rj (j=1, 2, 3, . . . , J) is a prediction coefficient for the r-th interval, and P is a predictor, which has the same configuration as that shown in FIG. 3.

S^ _ri is the predicted value for γ _r · S _ri in the r-th interval,
e _ri is a prediction error signal. S ^* _ri is S^ _ri and e ^* _r
It is a sum signal with _i and is approximately equal to γ _r ·S _ri . Q
_P is a selection adaptive quantizer, and the prediction error signal e _ri
The optimum quantization characteristic is selected for each interval according to the effective value σ _er . p is a selection code for selecting the optimal quantization characteristic that minimizes quantization distortion for each section. COD is an encoder that outputs a coded signal D _ri from the selection code p and the quantized signal e ^* _ri . MIX is a mixer that inserts a prediction coefficient A _rj , a scale factor γ _r , and a code p in advance of the encoded signal D _ri , and outputs a mixed output M _ri .

FIG. 5 shows M _ri (r=0, 1, 2, ......,
FIG. 3 is a configuration diagram when restoring the original audio signal from R; i=0, 1, 2, . . . , n-1). SEP
is a separator that separates D _ri , A _rj , γ _r , and P from M _ri . The DEC uses a restorer to output a quantized signal e ^* _ri from the encoded signal Dri and _the selected code p. S ^* _ri is e
^* _ri
and the predicted value S^ _ri (the configuration of the predictor P is the same as in Fig. 3)
Although it contains a quantization error, a value approximately equal to γ _r ·S _ri ) is obtained. Therefore, S ^* _ri
By dividing γ _{r by γ r} , a signal S _ri that is almost the same as the original signal is obtained.

As described above, in the present invention, predictive coding is performed after amplifying the input signal in the section where the input signal level is low, and according to the effective value of the prediction error signal, the quantization distortion is Since the conversion characteristics are selectively adapted, a reconstructed signal with an extremely high degree of approximation can be obtained.

Next, an example will be described. Now, the audio signal
The 12-bit linear PCM encoded signal is _Sri , and the division interval is approximately 25 ms. Also, for each interval, first, the prediction coefficient A _rj , the scale factor γ _r ,
It is assumed that the selection code p is determined, the audio signal is encoded using these parameters, and the process proceeds to the next section in sequence. in this case,
Since the predicted value S^ _ri is calculated from the past value as shown in equation (1), the first predicted value of the r-th interval is the (r-th)
1) Prediction will be made from J sample values at the end of the interval. Therefore, among the J sample values at the end of the (r-1)th interval (J=2 in the example) and all the sample values (n=256) of the rth interval, the maximum absolute value | S _ri
| γ _r was determined from _nax as follows.

1/2 FS<γ _r・|S _ri | _nax ≦FS ………(2) However, FS is the full scale level. In the example, γ _r =2 ^k , and k is set so as to satisfy equation (2).
There were five values: 0, 1, 2, 3, and 4. The quantizer Q _P is configured so that three types of characteristics Q ₀ , Q ₁ , and Q ₂ can be selected depending on the effective value σ _re of the prediction error signal e _ri . As is well known, the quantization characteristic that minimizes the quantization distortion is calculated by the following equation.

x _n =(y _n +y _n-1 )/2 ……………(3) y _n =∫ ^xm+1 _xn xP _p (x)dx/∫ ^xm+1 _xn P _p
(x)dx
……………(4) (m=1, 2, ………, 7) Here, x _n is the threshold value, y _n is the representative value, P _p (x)
is the probability density function of the prediction error signal.

In the embodiment, in order to compress the audio signal to 4 bits per sample value, 15 stages are used, and the effective value σ _e of the entire prediction error signal, i.e. Based on , the effective value σ _re for each interval, that is, was evaluated, and the quantization characteristic Q _P was determined. Note(5)(6)
n in the formula is the number of data in the interval (number of samples), R
is the total number of intervals for a sufficiently long known signal of finite length.

In the case of this example, for a sufficiently long known signal of finite length (for example, an audio signal of about 10 seconds), from the distribution of prediction error signals in all sections where σ _e < σ _re (when P = 0), Estimate the probability density function P ₀ (X). Similarly, the probability density function P ₁ (X) is estimated from the distribution of prediction error signals in all sections where 1/2σ _e <σ _re ≦σ _e (in the case of P=1). Furthermore, the probability density function P ₂ (X) is estimated from the distribution of prediction error signals in all sections where σ _re ≦1/2 σ _e (in the case of P=2). As is well known, when a signal can be regarded as having a Gaussian distribution, an exponential distribution, or a gamma distribution, the respective density functions may be used. For example, an audio signal can be approximated by an exponential function (however, the variance varies depending on P).

For the above P ₀ (X), P ₁ (X), P ₂ (X), (3)
Equation (4) is the quantization characteristic Q ₀ that minimizes the quantization distortion,
Give Q ₁ and Q ₂ . Even if the function form is the same, if the dispersion is different, the quantization characteristics that minimize quantization distortion will be different. Therefore, in order to minimize the overall quantization distortion, distortion minimum quantization must be performed corresponding to each individual probability density function. In the present invention,
It is assumed that the above P ₀ (X), P ₁ (X), and P ₂ (X) are different.

In this way, the quantization characteristics of each interval are determined from the relationship between σ _re and σ _e , and when χm≦e _ri <χ _n+1, the representative value is determined for each signal range, as e ^* _ri = ym. (ym).

Generally, the signal to quantization noise ratio SNR is SNR=10 logE[(S _ri ) ² ]/E[(S _ri −sr
i) ² 〕(dB)…………(7) However, it is expressed as sri=S ^* _ri / _γr . Note that E [ ] indicates the expected value.

The solid line in FIG. 6 shows the SNR in the above example, and the dotted line shows the scale factor fixed at 1,
SNR when the quantization characteristic is also fixed to one type
According to this example, the SNR is significantly improved.

Next, specific processing steps of the present invention will be explained.

First, a known signal, for example, an audio signal of about 10 seconds, is sampled at 10 KHz and quantized to 12 bits (S _ri ). After multiplying this S _ri by the scale factor γ _r , a linear prediction method is applied to obtain a prediction coefficient A _rj (r is the interval number) for each 25 ms interval.

Next, a prediction error signal without quantization is determined. This is obtained by using S _r (i-j) instead of S ^* _r (i-j) in equation (1). Note that S _r (i-j) is a signal delayed by j samples from S _ri . That is, In the case of an audio signal, J=4 is good. e in the above equation
_ri is calculated for all γ (r=0, 1, ......R-1), and based on the magnitude relationship between σ _re and σ _e , for example, three populations (P=0, 1, 2) Classify into.
P _P (X) is estimated for each population, and three quantization characteristics Q _P (P=0, 1, 2) are determined using equations (3) and (4). FIG. 7 shows this quantization characteristic. From now on, for unknown signal inputs, only σ _re exists, so σ _e and Q _P (P=0, 1, 2) are 10
The above value estimated from the second audio signal (signal population) is used.

Note that when applied to a playback-only automatic guidance broadcasting device, etc., it is sufficient to estimate σ _e and Q _P for one speaker.

When encoding an unknown signal, use equation (1)′,
e _ri without quantization, i=0, 1, 2...
n-1 is determined, and which quantization characteristic to use is determined based on the magnitude relationship between σ _re and σ _e . That is, the value of P is determined.

Next, while performing quantization, the prediction error signal e
e ^* _ri is obtained by repeating the operation for obtaining _ri .

As described above, in the present invention, since quantization is performed to minimize distortion for each section, the quantization error is reduced and the signal-to-quantization noise ratio is significantly improved, compared to conventional stations, airports, etc. The scope of application of the section DPCM method, which was previously limited to places with high noise levels, has been expanded, and it can now be applied to relatively quiet places such as hotels and hospitals, as well as public lines.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of a device using the interval DPCM speech encoding method, Fig. 2 is a block diagram of a device for restoring speech, Fig. 3 is a block diagram of a predictor, and Fig. 4 is an implementation of the present invention. A configuration diagram of a device applying the interval DPCM audio encoding method according to an example, FIG. 5 is a configuration diagram of a voice restoration device, and FIG. 6 is a characteristic diagram of the signal to quantization noise ratio.
Figure 7 shows the quantization characteristics Q ₀ , Q ₁ , Q ₂ in the present invention.
FIG. Q _P ...quantizer, P...predictor, COD...encoder, MIX...mixer.

Claims (1)

[Claims] 1. Multiplying means for dividing the audio signal into sections of a certain time and multiplying the discrete audio signal in each section by a predetermined coefficient; and subtracting means for obtaining a prediction error signal from the output value and predicted value of the multiplication means; quantization means for quantizing the prediction error signal; addition means for adding the output value of the quantization means and the predicted value; prediction means for obtaining the predicted value from the output of the addition means; and the output of the quantization means. the predetermined coefficient is set to a value that does not cause overscaling of the maximum level of each section, and the predetermined coefficient is set to a value that does not cause overscaling, and the predetermined coefficient is set to a value that does not cause overscaling, and the predetermined coefficient is set to a value that does not cause overscaling, and the predetermined coefficient is set according to the magnitude relationship of the section effective value with respect to the overall effective value of the prediction error signal. An interval DPCM method characterized in that a probability density function is selected for each interval, and a quantization characteristic that provides a minimum quantization distortion condition for each interval is selected based on the selected probability density function.