JPH0198000A - Method and apparatus for processing voice signal - Google Patents

Abstract

PURPOSE: To inexpensively attain improved voice signal processing by using a voice envelope as a normalized function and normalizing each point of a voice signal. CONSTITUTION: An input signal 10 is processed by an absolute value circuit 12 for outputting a signal always indicating a positive value at first to detect the peak value of the input signal 10. A low pass filter(LPF) 16 smoothes an output from the circuit 12 and outputs a signal 18 similar to the positive half of an envelope of a voice signal and the signal 18 is sampled by a subsampling block 20 to be driven by a rate lower than a practical sampling rate. An output from the block 20 is quantized by a quantizing circuit 22 and the quantized output is sent to an interpolation block 24, which executes interpolating processing in order to prepare an amplitude function. In the interpolating processing, an amplitude function point is prepared in each sample of the input voice signal and the amplitude function is divided into voice signals in each point in order to normalize the voice signals. Consequently improved voice processing can be inexpensively executed.

Description

[Detailed description of the invention] <Industrial application field> TECHNICAL FIELD This invention relates generally to the field of audio processing.

More particularly, the present invention relates to a method and apparatus for processing speech using a point-by-point biasing technique to improve speech processing during short-term and very short-term amplitude changes.

These amplitude changes are not normally reproduced accurately by audio processing systems that introduce frames or subframes. The invention has wide application in the field of audio processing, including encoding and pitch detection areas.

<Conventional technology> Conventionally, the following documents have been published.

[1 Co. B. Nis Attar and M. R. Schroeder, “Adaptive Predictive Coding of Audio Signals”, Bell System Technical Journal, VOL, 49°pp, 1
973-1986. October 1970.

(B, S, Atal and M, R, 5chroed
er, “p, dapt tve Predictiv
e Coding of 5peech Signa
ls@.

Be1l 5yst, Tech, J,, Vol, 4
9, pp. l 973-1986. Oct, 197
0. ) [2] M. Farr Sabin and R. M. Gray, “Product Code Vector Quantizer for Waveform and Speech Coding,” IEEE Bulletin on Acoustics, Speech and Signal Processing, Vol. ASSP-32, pp. 474-488.1
June 984.

(M, J, 5abin and R, M, Gray,
“Product Code Vector Quan
tizers for W, aveform and
V.

ice Coding", IEEE Trans,
on Acoust,, 5peech, and sig
nal processing, Vol, A S S
P-32, pp, 474-488. June 198
4. ) [3] Tay Tremaine, “Government Standard Adaptive Predictive Coding Algorithm”, Speech Technology, pp, 52-62
.

February 1985.

(T, 'f remain, “The Gover
nment 5standardAdaptive Pr
editive coding algorithm
”, 5peech Technology, pP, 5
2-62, February 1985,) [4] M.M. Sondhi, 'New Pitch Extraction Method'
, IEEE Bulletin, Speech and Electroacoustics, Vol, AU-16
, pp. 262-266. June 1968.

(M, M, 5ondhi, 'New Methods
of Pitch EXtraetiOn', IE
EE Trans, Audio and Electr
oacoustics, Vol, AU-16, pp,
262-266, June 196B,) [5] L. R. Rabiner and R. Double.
Schafer “Digital processing of audio signals”, pp, 15
0-157. Prentice-Hall, Englewood
Cliffs, New Jersey, 1978.

(L, R, Rabiner and R, W, 5cha
rer, Digital Processing of
5peech Signals, pp, 150-
157, Prentice-Hall, Eng.
jewood cjiffs.

N.J., 1978. ) [6 Coen Nis Giant, “-Adaptive Quantization with Word Memory”, Bell System Technical Journal, Vol. 52, pI), 1119-114
4. September 1973.

(N.S. Jayant, “Adaptive Qua
ntification with One Word M
emory', Be1l 5yst, Tech, J
,.

Vol, 52. pp, 1119-1144. Sept
[7] M. Honda and F. Itakura, “Bit Allocation in the Time and Frequency Domain for Predictive Coding of Speech,” IEEE Bulletin on Audio-Acoustic Speech and Signal Processing, Vol. ASSP-32, pp. , 465-473.
June 1984.

(M, Honda and F, 1 takura
, “Bit A11ocation in Time
and Frequency Domain for
Predictive Coding of 5pee
ch”, I EBE Trans, onAcoust
,,5peech,and Signal Proce
ssing, V.

1, As5P-32, pp, 465-473. June
(1984, ) Conventionally, block-by-block normalization of audio waveforms has been widely used in audio processing systems. There are two examples, one is the APC method with a block (i.e. frame) size of 10-30 ms (ref. [11]), and this
In the PC method, the residual signal is normalized by a gain.

The other is waveform gain vector quantization (S
GVQ) (Reference [2]). In this method, the vector is normalized by its effective value.

In the above APC method, normalization is performed over the entire frame by a single gain value.

This clearly creates problems when the signal amplitude changes rapidly across frames. To partially solve the above problem, it is possible to divide one block into several sub-blocks by dividing several gain values (
scale factor) (Reference [3]
). A similar problem is encountered in autocorrelation methods of pitch detection, where intermediate clipping is used to avoid correlation peaks due to harmonic structures (References [4] [5]).

As shown later, this method of intermediate clipping is
In some cases, the intended purpose may not be achieved.

Jayant's backward adaptive quantization (ref. [6]) performs an unconditional point-by-point normalization of the signal fed to the quantizer by predicting the signal amplitude. However, although this form of normalization is particularly suitable for quantization methods, the normalization function (step size) also makes the normalized signal useless for other purposes.

The present invention uses a point-wise amplitude normalization method that has very broad applicability in many areas of audio processing, including encoding.

Any of the above prior art techniques, as shown by the present invention, utilizes interpolation and normalization as well as interpolation and normalization between subrate samples of the positive half of the envelope of the audio signal in order to more accurately reproduce or encode digitized audio. not doing what is to be done.

<Objects, Structure, and Effects of the Invention> Accordingly, an object of the present invention is to provide a method and apparatus for processing audio that achieves improved audio signal processing at a relatively low cost.

Another object of the invention is to provide a method for processing speech that provides improved speech processing during significant amplitude changes in short terms or very short terms. .

Furthermore, it is an object of the invention to provide a method and apparatus for processing speech that provides point-wise normalization of the speech signal by using the speech envelope as the normalization function.

The object of the invention is achieved by finding a smoothing function that varies with the amplitude of the signal. In the example described below, this smoothing function is in the form of the upper half envelope of the audio signal, but other suitable functions may be used. When this function performs a point-by-point division of the signal, the resulting waveform has a fairly constant amplitude throughout.

In the following description, the amplitude change that occurs from pitch to pitch, that is, from frame to frame, is called a short term (ST), and the amplitude change that occurs within a pitch period is called a very short term (VST) or intra-pitch change. . Its application and expected effects will determine whether short-term normalization or very short-term normalization is appropriate. Therefore, the method of calculating the amplitude function should be the parameter chosen to yield the expected normalized tones in the examples below.

The above objects of the invention will become apparent from the following description of the invention.

In one embodiment of the present invention, a method for processing an audio signal includes the steps of: obtaining the absolute value of the audio signal; and low-pass filtering the absolute value of the audio signal to obtain a positive half envelope function (amplitude function). including steps to This positive half envelope function is then subrate sampled and the values between subrate samples are determined by interpolation to create simulated amplitudes. The audio signal is then normalized by the amplitude function and wideband encoded.

In another embodiment of the invention, an apparatus for processing audio signals includes an input circuit for receiving audio signals.

This circuit converts the audio input signal to the positive (or negative) of the audio signal.
It is coupled to the audio input to convert it into an amplitude function with half-envelope characteristics. A normalization circuit is coupled to the input circuit and divides the audio signal by an amplitude function to perform normalization of the audio signal.

<Example> Hereinafter, the present invention will be explained in detail with reference to illustrated embodiments.

FIG. 1 shows a block diagram of the normalization method of the present invention. In this embodiment, the invention is implemented using programmed processing equipment known to those skilled in the art. Therefore, the block diagram of FIG. 1 may be considered both a functional block diagram and an operational flow diagram.

The input signal applied to node 10 is the absoluteness block 12
and is input to delay block 14. The output of the absoluteness block 12 is input to a low pass filter 16. The low-pass filter 16 removes high frequency components and outputs a positive half envelope signal 18.

The positive half envelope signal 18 is sampled in a sub-sampling block 20. The output of the sub-sampling block 20 is input to a quantization circuit 22. The signal sampled by the sub-sampling block 20 is quantized by the quantization circuit 22 to one of predetermined quantization levels. In this example, quantization of only 4 or 5 bits is used, but the invention is of course not limited to this.
Much depends on the individual application in the future.

The output of the quantization circuit 22 is input to an interpolation block 24. The output of the interpolation block 24 is input to a divider 26. The divider 26 is used to create a normalized signal 28 by dividing the delayed output from the delay block 14 by the output of the interpolation block 24. Essentially, the output of the interpolation block 24 is a sampled envelope signal, shown as AF in the figure, referenced as a function of amplitude. In post-processing, this AP uses the normalized signal 28
It may be used by the post-processing block 30 in the same way as the fortune-telling delay signal 32.

The post-processing shown in post-processing block 30 may take many forms as known to those skilled in the art. for example,
Audio data may be encoded into streams or frames of binary data by circuitry including, for example, an adaptive differential pulse code change FJ (ADPCM) code. This digital signal can then be transmitted for reception at a remote location, for example by a modem.

The remote receiver uses this AP (the quantized and subrate sampled signal from the quantization circuit 22 is easily and inexpensively encoded) to simply denormalize the input signal. can be played. Of course, mock A
Interpolation is also used in the receiver if the signal from the quantizer circuit 22 rather than F is transmitted. In another example, encoding and/or pitch detection can be part of post-processing. Thus, many variations are possible in the application and practice of the present invention.

To explain the operation, the human input signal lO is first processed by the absolute value circuit 12 which outputs signals having all positive values. This process may be replaced by other equivalent processes. For example, a half-wave rectifier may be used to cut off the lower half or the upper half of the audio signal. Alternatively, the square root of the voice signal may be obtained after squaring the audio signal and removing the negative component. In any case, the purpose is to detect the peak value of the input signal in preparation for the low-pass filter 16. The low pass filter 16 smoothes the signal and outputs a signal that resembles the exact half of the envelope of the audio signal (assuming the audio signal is symmetrical). This signal 18 is sampled by a sub-sampling block 20 which operates at a rate well below the actual sampling rate required to sample the audio signal itself. For example, approximately 100 Hz for short-term amplitude normalization (STAN), very short-term amplitude sampling (V
STAN) is approximately 1000 Hz. For most audio applications, subrate sampling rates between about 100 Hz and about 1000 Hz are suitable, but are not limited thereto. The output of subsampling block 20 is quantized by quantization circuit 22 and sent to interpolation block 24 .

In the interpolation block 24, interpolation processing is performed to create an amplitude function as described later. The interpolation process creates a point on the amplitude function for each sample of the input audio signal. This amplitude function is divided point by point into the audio signal in order to normalize the audio signal.

In the above embodiments, a low pass filter is used for interpolation. Note that points on the amplitude function may be generated between subrate samples using simple linear interpolation in order to improve calculation efficiency. Other known interpolations and curve interpolations may prove advantageous in other applications.

Such normalization offers many advantages in some forms of post-processing. The quantized amplitude signal is sampled at a low frequency and compared to the sampling rate for transmitting the audio signal to the receiver, resulting in low cost and high accuracy in reproducing transient frames of audio. Furthermore, the normalization process has significant advantages when applied to speech coding.

Since the audio signal is normalized, its signal variation and the number of bits required to accurately encode the signal are reduced;
This will reduce the bandwidth required for transmission.

The amplitude function (AF) is preferably obtained by reducing the absolute value of the input digital signal.

The cutoff frequency of this filter determines the normalization rate.

The 9-bit linear representation of the above AP is adequate, and the division is done very efficiently by table lookup. For those speech coding applications where normalization function information is required at the receiver side, the amplitude function is sampled at the Nyquist rate and quantized using 5) log-PCM. Regeneration of the quantized AP is done at both transmitter and receiver by interpolation, and the same AF is used at both ends.

If explicit information of the AP is not required at the receiving machine, the blocks within the dotted line in FIG. 1 may be bypassed.

The amplitude normalized signal, delayed input signal and AP are then provided to a post-processing block 30. As shown later,
Certain parameters may be estimated from the normalized signal, certain other parameters may be estimated from the input signal, and certain other parameters may be renormalized from the normalized signal. It is best to estimate it from the conversion.

For short-term amplitude normalization (S T A N), the cutoff frequency of the low-pass filter is set to 25-30 Hz in an embodiment with a 50 Hz stopband rejection and a 100 Hz sampling rate.

For Very Short Term Amplitude Normalization (VSTAN) lζ, it is desirable to set all frequencies to 10 times the above value (approximately 250 to 300H for the cutoff frequency).
z, stop band rejection is 500H2, sampling rate is 100
0Hz).

However, in other embodiments, other filter characteristics are generally desired, and the cutoff frequency typically lies somewhere within or between these ranges. Of course, these quantities are merely illustrative and not limiting, and they depend largely on the future application. If AF
'If no transmission is required, an I[R filter (infinite impulse response filter) can be used. Otherwise, a FIR filter (finite impulse response filter) is recommended for computational efficiency in the decimation and interpolation process ζε. The computation required for normalization is almost 10 operations per point, which is 2 in most DSPS.
- Consumes 5% real time.

The delay introduced for 5TAN is 10-20 ms and the delay introduced for VSTAN is 2-4 ms. These delays are determined by the delays inherent in the low-pass filter, and are mostly set equal to it.

These shifts are therefore dependent on the exact characteristics of the low-pass filter used for the particular application. The delay properly normalizes the input signal by the portion of the envelope. The parts of the envelope are made by identical parts of the input signal so that a more accurate normalization occurs.

FIG. 2 shows three typical examples of changes in short-term amplitude obtained through the computer simulation of the present invention. It should be noted that the normalized signal has a fairly constant shape of amplitude throughout and does not change the mid-pitch characteristics for voiced sounds. The normalized signal is A
It is clear that it has more desirable amplitude characteristics for PC pitch detection and quantization, but not for calculation of prediction parameters. For example, a segment containing an unvoiced sound preceding a voiced sound with amplitude variation is shown.

This voiced sound is followed by silence.

This 125 millisecond segment contains several frames worth of data. In frames containing voice changes (consisting of the part of the frame containing voiced sounds and the remaining part containing unvoiced sounds/silence), the autocorrelation function calculated from the normalized signal, and therefore the prediction parameter, is calculated from the input signal. will be significantly influenced by unvoiced/silence compared to the predicted parameters. This results in poorer predictions. Therefore, it is generally better to use the input signal (not normalized) for calculating the linear prediction parameters. For this reason, the selection of the normalized signal is made according to the actual example. The usefulness of 5TAN in the autocorrelation method of pitch detection is illustrated in FIG.

The audio segment with amplitude changes is shown in FIG. 3a. The AP and normalized signals are shown in FIG. Intermediate clipping the signal before calculating the autocorrelation function helps to attenuate extraneous peaks caused by strong harmonics. At least three known methods of intermediate clipping may be used in accordance with the present invention.

(1) Divide the frame into 4 ms mini-frames, find the maximum absolute value for each mini-frame, and set the intermediate clipping threshold to 30% of this maximum absolute value.

(2) Find the maximum absolute value in the first 1/3 and the last 1/3 of the frame and set the intermediate clipping threshold to 64% of these two minimum values.

(3) Use the amplitude normalized signal for intermediate clipping with the intermediate clipping threshold set to 50% of the maximum absolute value.

Figures 3d and 3e show the intermediate clipped waveform and the autocorrelation function derived therefrom. The time coordinates for the autocorrelation functions have been expanded by a factor of three to improve analysis. For the method shown at 3 in FIG. 3, a renormalized intermediate clipped signal obtained by multiplying the intermediate clipped signal by AP is used for the autocorrelation calculation. From these figures, it is clear that the method shown in 3 is most effective in eliminating most of the harmonic information and preserving the pitch peak of the low level signal after intermediate clipping.

This success is reflected in the autocorrelation function for the method presented in 3. This is independent of the extraneous peak shown for 1 and 2 in FIG.

FIG. 4 shows two examples of VSTAN and its application to wideband coding. Figures a, b, c, and d show the input waveform, step size adaptation, Niheli short-term amplitude function, and normalized waveform in Giant's 4-bit APCM, respectively. It is clear that the main difference between the normalized signals of FIGS. 3 and 4 is that the latter normalizes intermediate pitch amplitude changes.

While the normalized signal of FIG. 4 is highly undesirable as an input to a pitch detector, it has amplitude characteristics that are highly desirable for fixed-step quantizer circuits. 1 rubel fixed form quantization circuit (3,46 bits) is 4 bit AP
Gives a sound quality very similar to commercials. If it includes the bitrate required to transmit the amplitude function,
This nigga method has almost the same rate. However, if the adaptive bit allocation in the time domain is based on AF information and keeps the total number of bits in a frame (10-20 ms) constant, the sound quality is better than APCM. It is noteworthy that APCM is relatively simple and does not involve coding delays. FIG. 4e shows a histogram of the normalized signal.

Thus, point-wise normalization of audio signals is disclosed as a pre-processing technique to improve post-processing;

ing. Furthermore, the amplitude function is used for normalization. This amplitude function is obtained by passing the above signal through a low band y to produce a signal having a shape similar to half the envelope of the audio signal. The cut-off frequency of the low-pass filter plays a very important role in the characteristics of the normalized signal. Two types of normalization rates have been found to provide suitable amplitude characteristics for different applications. The present invention is VS
Improvements can also be made when combining vector quantization of the T amplitude function with a specific normalized signal (Reference [2]).

The invention is advantageously implemented using a programmed processing device such as a microcomputer.

The list of computer programs at the end represents a typical Fortran implementation of the audio processing of the present invention. Number 1 in this list is a software implementation of the magnitude circuit, low-pass filter and envelope oscillator according to the invention.

The list l accepts an input file, takes the absolute value of each point in the file, and also uses the routine DO with the file coefficients to perform low-pass filtering of the input file.
Use TPR (dot creation routine). Then the above input file is a routine A M P L that creates an amplitude function.
TD to file the amplitude function in the output file.

Listing 2 uses an input file and an amplitude function file to create a normalized signal. This normalized signal has two pairs!
It is input to a two-pair ITDH9 compressor to produce a compressed output. This 2:1 output is input to the APC-AB quantization circuit and then to the TDHS decompressor. The output of the TDHS decompressor is denormalized to produce reproduced audio. This Fortran implementation can be easily adapted to suit individual needs.

The computer program listed at the end is AP-120B
Fortran was used to operate the VAXI+-750 minicomputer using an array computer.
) was written in 77.

It is thus clear that what has been described above is a device of the invention which fully satisfies the above objects and advantages.

Although the invention has been described with respect to specific embodiments, it will be appreciated that many substitutions, modifications, and variations will be apparent in light of the above description. Accordingly, the invention includes all substitutions, modifications and changes falling within the spirit and scope of the claims.

- Below margin 12 AMPF I L (K) = f N
FI L (K) AMPF I L (LEN-2) ='
A'AMPF I L(LEN-1) = I NF I
L(LEN-1)AMPP I L(LEN)=I
NF I L (LEN) AMPF I L (LEN+
1)=OAMQF IL(LEN-2)='A'AM
QF I L (LEN-1)=”M'AMQF I L
(LEN)='Q'AMQF I L(LEN+1
)=0OPEN(IN I T=LUN I, F I
LE= l NI> I LOPEN (UNIT)
N2. FILE=AMPF 11
RECORDS I ZE=90)CO,
PEN(tJNIT=LUN3.FILE≠AMQFI
CALL VCLR(0,1,3000)CALL
INITLZ Do 8000 1FRAME=1,10000C
ALL AMPLTD IP (IEND, NE, O) GOTo 999.
5TATUS='OLD', rtEcORDs I Z
E=90)L, TYPE='NEW', FOIIM='
UNFORMATTED'.

L, TYPE='NEW', RECOr(DS I Z
E=90) FLTM(91) = FLTM(91
)*, 001DO+0 1=1.90 FLTM(1) = FLTM(1)*, 0011
0 FLTM (182-1) = FL
TM(1) CPUUT FILTERINTOARrt
A) CALL APPUT (FLTM, IFLTM,
181.2) CALL APWAIT ETURN ND 5U13ROUTINE AMPLTDBYTE
A (9G), C INTEGer (AL2LMR EQUIVALENCE (C, IC) COMMON
/5RCBLK/FLTM (181), DAT (l
fI LUNI, LUN2. LUN3. L.R.E.
C,IFLTM,IDAT,.

DIMENSION DATABS (180) CRE
AD TWORE CORDS ANDIBASE
= 0 Doj00 Summer REC=1.2 READ(LUNI,l 01.END=999)AD
o20J=1. LREC C= A(J) DAT(IBASE+J) = AL2LN11(
IC) 20 DATABS (IBASE+J)
= ABS(DAT(IBASItoo)
IBASE = rBAsE + LRECl
ol FORMAT(90AI)I PRO
C, MEMORY +O), AMP (l sO), AMQ (18G), AQ
NT(6).

IDAT, IAMP, JAMP, LFRMCONVER
T FROM LOG TOLINE/'1(+
J)) IRCADR= ICMP @[AADR=
InCADR IALADR= IAADR IRADR= IALADrt IERADR= IRADrt INEXT = IERADR CONST(1) = 1°C0N5T(2) = , 5 CONST(3) = 256°Do 2 1=1, 256 2 1HISTO(1) = OCALL
APIIIIT(0,0゜TYPE *,' INPUT ACCEPT 105.LEN,: TYPE *,' ACCEPT*, ITDH9,I( 105FORMAT(Q, 32A1) I NF I L(LEN+ 1)=0+ 8 + 9 + 9 + 9 10 1 ISTAT) FILE (*, D6K)' +NFIL ITD Summer (S, Summer QNT, INS, NRBITS=';
INT, INS, NRBITS Do 12 K=1. LEN=30UTF I
L(K)=I NF I L(K)+2
AMPPIL(K)=INFIL(K)AMPP I
L(LEN-2)='A'AMPF' I L(LEN
-1) = INF I L (LEAMP I L (LE
N)=I NP [L(LEN)AMPP I L(
LEN+1)=00UTF I L(LEN-2)=
"E.

1F (IQNT, EQ, 1) 0UTFIL
(LBjIP(IQNT, EQ, 2) 0U
TFIL(LEIIF(IQNT, EQ, O)
0UTFIL(LEfIF(I QNT, EQ,
O, AND, I TDH8, EO, UTF I L
(LEN-1) = I NP I L (LEN-OUT
PIL(LEN)=INFIL(LEN)OUTF I
L(LEN+1)=0OPEN(UNIT=LUN
1. FILE=INF1'J=2)='Q'N-2)='R' q-2) 1,000°T.

Q, 0) 0UTF [L(LEN-2)='O.

riL, 5TATUS='OLD', rLEcORDs I
ZE≠90) CALL XMINV (MSE, l,
90.64) CALL APGET(RMSMIN,
90.1.2) CALL APGET (RJPIT,
91, 1.2) JPIT=IFIX(rLJPIT) CALL APWD IPIT = IPVAL(JPIT)CALL
XMAXY(MSE,!,90.64)CALL
APGET (RMSMAX, 90, 1.2) TYPE
*,' MIN, MSE, MAX, MSE
,FIF(ITDHS,EQ,0) Go
To 160CPERFORM TDHS
COMFIPIT2 = IPIT/2 NCMP = 0 NEXP' = 0 NPIT = 0 MTCH =', RMSMIN, RMsMAX, IPIT
νrtEssION 15 GAIN=SQRT(AL(1)/AL(
MCI)) Do2G summer = NC1, N+NC 205PQ (1) = SP (1) IP (IQNT, EQ, 0) Go TO21GDO5
01=l, N TEMP earthquake 0°Do 40 J=1. NC1 40TEMP=TEMP+A(J)*5P(1+NC1
-J) CRES(1)=GAIN*TEMP rLES(1)=TEMP 50 ABSRES(1)=ABS(RES(1
))LB IN=(IPIT+NRE/2)/NR
EDo 60 1=1. DBIN=1 60 ABSRES(+ +N)=AnSrtE
S(1)rtEsMAX=0゜MAXLOC=1 Do 70 1=l, IPIT Do 65 J-1, LBIN-165ABSRE
S(1)=・ABSRES(1)+ABSI'1tES
(1+J)IP(ABSRES(1),LT,RESM
AX) GOTo 70RESMAX=ABSRES(
1) MAILOC=1 70 C0NTINUE Do T5 Summer= l, NRE75
NBIN (1) -LBINNPEM Snow I P I T-
NRE) kLB I NIP (NREM) 80.9
0.8580 NREM=-NREM Do 82 1=l, NREM 82 NBIN(1)=NBIN(+)-1G
o To 90 85 Do 88 I=1. NREM8s
NBIN(1)=NBIN(1)+1100
AQErtR=AQ-ACIF'(IFRAME,
GE, 11B? TYPEllol FORMA
T(IOX, 14.4F15.6)ETURN ND COPTIMUM UNIFORM 5UBROUTINE QNTEC(A, 5TPS+Summer CLIP=1 AK=ABS(A)/5TPS I ZI P(AK,
GT, l 5. ) AK= 15゜K=AK+, 5 IP (K, LE, ICLIP) K=O IP (K, EQ: O) co To 10AQ
=STPS I ZI (K+ I CL I P) I
P (A, LT, 0.) AQ Tomi-AQGo TO15 10AQ 100.

[11T, 5=1 Go To 100 15 1F, (AQ) 30,30.2020
1BITS=2)kK Go To 100 30 1BITS =+1 to 2* 100 AQErtR=AQ-ACTYPE
101. IBI TS, A, 5TPSIZ, Nolot
FORMAT(IOX, 14.4F15.6)r
lETURN ND 01 , NB I T, A, ASD, AQ, AQERR
GAUSSIAN OυANT I ZER[Z,A
Q,AQERR,IBITS)\Q,AQERR 20 DAT(1-MINPTI)=1. /FLO
AT(3*DO301=LPIT2+l, LPIT!I
PVAL(1-MrNPTl)=1 1FLEN(1-M!NPT+)=2*IIPLEN(
+-MINPTI)=r 30 DAT(1-M(NPTi)=1./FLO
AT(1)Do 40 r=LPITI-MIN
PTl + I, 6IPVAL (summer) = IPVAL (1-1
)+2I PLEN(+)=I PVAL(1)I F
LEN(1)=2*[PVAL(1), io DA
T(D=1./pLoAT(rpt, EN(D)CAL
L APPUT(DAT, IPNOrLM, 64,2
CALL APWAIT Do 100 1=l, NG SP(1)=0°100 ' 5PQ(1)=0°102 FORtMAT(5X, 1614) rt
ETUrtN ND CGET IMPUT DATA AND N
, O5UBROUTINE GETNOR(LDAT
, LABYTE A (90), C INTEGERAL2LNrt EQUIVALENCE (C,IC)COMMON
/APMEM/AP(10000)COMMON
/BLK/DAT (180), AMP (1RMAL I
ZE B'/AMPL I TUDE I? L
JNCT I ONMP, I END) 80), LUNI, LUN2. LUN3. LREC,I
DAT, JDAT.

FUNCTION LNR2AL (IOC) ISIG
N=O IF (IOC, LT, 0) ISIGN=”20r
cc=I ABS (I 0C) KTH child 16 Do 80 K=0.6 IP (ICC, LT, KTH) Go To 660
KTH=KTH+KTH 611F (K, EO, 0) GOTo 62ICC
=(IOC-KTI-1/2)/2**(K-1)62
1CC=ICC+++6+l5IGNLNrt2
AL=IEO11(ICG,"+25)lETUI't
N ND CI”UNCTION To C0NVERtT
FrLIINTEGER FUNCTION AL2
LNI C=I EOI(I C,”+25)ISI
GN=1 1PCIAND(IC,”200),NE,0)
IIEXP=1AND(IC,”+60)/+6I
C=IAND(IC,”17) IP(IEXP,EO,0)Go To 20IC
=ISHFT(IC+16.IEXP-1)20AL2
LNR=SummerC*l5IGN nETtJRN ND DM A-LAW To LINEARR (IC
) SI GN=-1

[Brief explanation of the drawing]

FIG. 1 is a block diagram of the audio processing method and apparatus of the present invention. FIG. 2 shows three examples of signals with short-term amplitude changes and the results of normalization according to the invention, a) the input signal, b) the amplitude function AF, and C) the normalized signal. shows. FIG. 3 shows three examples of applying the present invention to the pitch detection autocorrelation method, a) for the input signal, b) for the AP, C) for the normalized signal, and d) for the intermediate clip. signal, e
) indicates the autocorrelation function derived from the intermediate clipped signal. Figure 4 shows three examples of very short-term amplitude normalization when wideband coding is applied; a) shows the human input signal, b) shows the step size from Jayant's 4-bit APCM, C) is AP,
d) shows a normalized signal, and e) shows a histogram of the normalized signal. lO...input signal,! 2...Absolute value block, 14
...Delay block, 16...Low pass filter, 18.
... Positive half envelope signal, 20... Sub-sampling block, 22... Quantization circuit,
24... Interpolation block, 26... Divider, 28
... Normalized signal, 30 ... Post-processing block, 32.
...Delayed signal. Patent Applicant: Racal Data Communications, Inc. Agent: Patent Attorney: Aoyama Aoyama and 2 others Apparatus 3, Relation to Amendr Case Patent Applicant Address Florida, United States 33323. Sunrise, 1601 North Harrison Parkway
Name: Racal Data Communications, Inc. Representative: J. Neil Robertson Nationality: United States of America 4, Agent Address: 2-1-6 Mihigashi-ku, Osaka City, Osaka Prefecture, 540 Prefecture
No. 1 No. 6, Subject of amendment: Specification: Detailed description of the invention column. 7. The following parts of the detailed statement of amendment will be corrected. 1. Pages 26 to 39 of the Detailed Description of the Invention column will be corrected as shown in the attached sheet. that's all

Claims (5)

[Claims] (1) Detecting the peak signal level of the audio signal; and calculating, from the peak signal level of the audio signal, an amplitude function signal having an amplitude similar to the envelope of the amplitude of the peak signal of the audio signal. A method for processing an audio signal, comprising: normalizing the audio signal by the amplitude function; and wideband encoding the normalized audio signal. 2. The method of claim 1, wherein said step of calculating comprises the step of low-pass filtering said signal peak level to produce a low-pass filtered signal. (3) After the low-pass filtering step, subrate sampling the low-pass filtered signal, and interpolating between points of the subrate sampling signal to create a simulated function of the amplitude function. A method for processing audio according to claim 2, comprising the steps of: (4) input means (10) for receiving an audio signal; envelope means (16) for converting the audio signal into a relatively slowly changing envelope signal that approximately follows the peak of the audio signal; , sampling means for sampling the envelope signal at a sampling rate lower than the Nyquist rate for the audio signal;
20); and normalizing means (26) for normalizing the audio signal by the sampled envelope signal. (5) interpolation means (24) for interpolating between points of the sampled envelope signal in order to create an interpolation signal; the normalization means (26) normalizes the audio signal by the interpolation signal; An apparatus for processing audio signals according to claim 4.