JP3464371B2 - Improved method of generating comfort noise during discontinuous transmission - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method that generates comfortable noise (CN) in a mobile terminal in a non-contiguous transmission (DTX) system. SOLUTION: This method for generating comfortable noise is to make the frequency characteristic of the comfortable noise close to actual background noise by making a random exciting signal pass through a spectrum controlled filter. In the form of other execution, a sending end identifies a voice coded parameter that does not satisfactorily represent actual background noise, and when such voice coded parameters are detected, the parameters are replaced with a mean value. In this method, an balancing operation is not disturbed by a faulty parameter.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates generally to voice communication technology, and more particularly to discontinuous transmission (DTX) system, and improvement of comfort noise (CN) in discontinuous transmission.

[0002]

2. Description of the Related Art In a mobile communication system, a discontinuous transmission system is used in which a radio transmitter is disconnected while conversation voice is temporarily interrupted. By using such a discontinuous transmission method, the power consumption of the mobile station can be reduced, and thus the recharging interval of the battery can be lengthened.
Further, the use of the discontinuous transmission method generally improves the communication quality because the communication interference level is lowered.

However, if the channel is completely cut off, the background noise included in the voice signal and transmitted is also cut off during the transmission pause period, so that the receiving side has an unnatural voice signal (no Signal condition).

As one of the conventionally known methods for solving this problem, the characteristic parameter of the background noise is extracted without completely disconnecting the transmission during the transmission pause period, and this is extracted as a silent descriptor (SID) frame. There is known a method of wirelessly transmitting at a low transmission rate using. On the receiving side, this parameter is used to reproduce the background noise so as to reflect the spectral and temporal substance of the background noise actually existing on the transmitting side as faithfully as possible. The parameter that characterizes such background noise is called the comfort noise (CN) parameter. The comfort noise parameters typically include a subset of speech coding parameters, in particular synthesis filter coefficients and gain parameters.

In the comfort noise calculation method in a certain speech codec, only a part of the comfort noise coefficient is extracted from the speech coding parameters, and the other comfort noise parameters are
For example, it is extracted from the signal that is present in the speech coder but not transmitted wirelessly.

In conventional discontinuous transmission schemes, it is assumed that noise with a flat spectrum (ie white noise) can be used to approximate the excitation signal sufficiently well. That is, in the conventional discontinuous transmission method,
Comfort noise is generated by passing locally generated noise with flat spectral characteristics through a speech coder synthesis filter. However, in practice, such a method using white noise cannot generate high-quality comfort noise. The reason is that the spectrum of the ideal comfort noise is not flat and has a slightly tilted spectral characteristic or, in some cases, a spectral characteristic that deviates significantly from the flat spectral characteristic. Although depending on the type of background noise, the optimum comfort noise spectrum has, for example, a low-pass type or a high-pass type characteristic. That is, the conventionally used random excitation (random excitation) signal (white noise) deviates from the optimized desired noise excitation, so that the comfort noise generated on the receiving side is the background noise on the transmitting side. Is recognized as different from. For example, the comfort noise generated may be much more "brilliant" than the desired comfort noise, or vice versa. In discontinuous transmission, the background noise spectrum differs between during transmission (a period during which voice coding is performed) and during transmission interruption (a period during which comfort noise is generated). Such a change in background noise is perceived by the user as a deterioration in communication quality.

[0007] Full speed (FR), half speed (HR), enhanced full speed (EFR) voice channels, etc. in GSM systems,
In a speech coding system, the comfort noise parameter transmission is slow. For example, FR and FER
In the channel, the comfort noise parameter transmission rate is once every 24 frames (ie once every 480 ms), which is very slow. That is, the comfort noise parameter is updated only twice per second. Since the transmission speed is low as described above, the spectral characteristics or temporal characteristics of the background noise cannot be expressed sufficiently accurately. Therefore, in the discontinuous transmission system, the quality of the background noise is inevitable. to degrade.

Another problem that occurs in discontinuous transmissions in digital cellular systems such as GSM is related to the hangover of some voice frames before the transmission actually ends after a voice burst. If the voice burst is below a certain threshold, the voice burst is interpreted as background noise, in which case there is no hangover period following the voice burst. The hangover period is to calculate the evaluation characteristic value of the background noise of the transmitting side, which should be sent from the transmitting side to the receiving side as a comfort noise parameter message (unvoiced descriptor (SID) frame) before the transmission is finished. Used for. As described above, the receiving side uses the transmitted background noise evaluation characteristic value to generate comfort noise having characteristics similar to the background noise existing on the transmitting side at the end of transmission.

In the known discontinuous transmission schemes similar to GSM FR or HR, comfort noise quantization is performed by a non-predictive scheme. Therefore, the receiver does not need to know if there is a hangover period at the end of the voice burst. However, in the case of GSM EFR, an efficient predictive comfort noise quantization method is adopted. Therefore, when performing comfort noise inverse quantization on the receiving side, it is necessary to determine whether or not a hangover period exists. , And some computational load is spent executing the program instructions for it.

If the background noise on the transmitting side is not constant but fluctuates considerably, the following problems will occur. That is, in such a situation, it may happen that some or all of the speech coding parameters of one or some of the frames of the averaging period do not correctly represent the average background noise. . A similar problem occurs when the end of speech is unvoiced, as determined by the voice state detection (VAD) algorithm, or when the constant background noise contains a strong impulse-like noise burst. In the existing discontinuous transmission method, since the averaging period is short, if an inappropriate voice coding parameter as described above occurs,
The results of the averaging process change significantly, and the averaged comfort noise parameters no longer accurately represent the background noise features. So between the comfort noise and the actual background noise,
Differences occur in levels and / or spectra. Therefore, the period during which the call is actually being made (the period during which voice and background noise are coded in the usual way)
Then, the user feels as if the background noise has changed between the period during which the call is interrupted (the background noise during this period is generated by the comfort noise), which impairs the communication quality.

More specifically, during the discontinuous transmission hangover period, the comfort of the first unvoiced descriptor frame is improved while the frame determined by the voice state determination algorithm to be a "non-voice frame" is wirelessly transmitted. The speech coding parameters used to calculate the noise parameters are temporarily stored. The first unvoiced descriptor frame above is transmitted as soon as the discontinuous transmission hangover period expires. The length of the discontinuous transmission hangover period is determined by the length of the averaging period. Therefore, the averaging period should be fixed to a relatively short length in order to minimize system channel usage.

Prior to describing the present invention, the circuits and methods conventionally used to generate comfort noise parameters at the transmitter and the comfort noise generation at the receiver based on these parameters. The circuit and method are described below. First, FIGS. 1 (a) to 1
Reference is made to (d).

In FIG. 1A, linear predictive coding (L
A PC) analysis block 101 calculates short-term spectrum parameters 102 from the audio signal 100. Note that linear predictive coding (LPC) is a technique well known in the art. For simplicity, it is assumed here that the synthesis filter consists only of short-term synthesis filters, but in most conventional systems such as GSM FR, EFR encoders, the synthesis filter is actually short-term. It is realized by connecting an inter-synthesis filter and a long-term synthesis filter in series. It should be noted that long term synthesis filters are not required to describe the invention. Further, in the conventional discontinuous transmission method, the synthesis filter is usually cut off for a long time when the comfort noise is generated.

The LPC analysis produces a set of short-term spectral parameters 102 for each transmitted frame. The frame length at this time differs depending on the system. For example, in the GSM channel, the frame sizes are all set to 20 ms.

The audio signal is passed through an inverse filter 103, where a residual signal 104 is produced. This inverse filter has the characteristic expressed by the following mathematical formula.

[0016]

[Equation 11]

Where a (i) (i = 1, ..., M)
Is a filter coefficient, which is generated in the linear predictive coding analysis and updated for each frame. To make the change of filter parameter between frames smoother,
The interpolation process used in the conventional speech coding technique may be applied to the inverse filter. The residual signal 104 generated here by the inverse filter 103 is the optimum excitation signal, and by passing this signal through the synthesis filter 1 / A (Z) 112 (FIG. 1 (b)) at the receiving side, The original audio signal 100 is accurately restored. The excitation gain calculation block 105 measures the energy of the excitation signal sequence and calculates the scaling gain 106 for each transmission frame.

From the average of the excitation gain 106 and the short-term spectrum coefficient 102 in several transmission frames, the average background noise spectrum and temporal characteristics are extracted. Normally, the above averaging process is performed for 4 frames in the case of GSM FR channel, while it is performed for 8 frames in the case of GSM EFR channel. The parameters to be averaged are blocks 107a and 108a (FIG. 1 (d)).
Buffering (temporary storage) during the averaging period. The averaging process described above is performed in blocks 107 and 108 to generate an averaging parameter that characterizes the background noise. The average parameters generated here include the average excitation gain g _mean and the average short-term spectrum coefficient. Modern speech codecs typically use 10 (M = 10) short-term spectral coefficients, which are line spectrum pair (LSP) coefficients f _mean (as is done in GSM EFR discontinuous transmission system. i) (i = 1, ..., M).
These parameters are usually quantized before being transmitted, but for simplicity, the description of the quantization is omitted here. Note that the quantization is not an essential part for understanding the operation of the present invention.

As shown in FIG. 1D, the averaging blocks 107 and 108 are respectively buffers 107.
a, 108a and through these buffers the signal 1
07b and 108b are input. Buffers 107a and 108a will be described in more detail when describing the present invention with reference to FIGS.

The calculation and averaging of comfort noise parameters is described in detail in the GSM Recommendation Criteria: GSM06.62 "Features of Comfort Noise in Enhanced Full Rate (EFR) Voice Traffic Channels". Regarding discontinuous transmission method, GSM recommended standard GSM
06.81 “Discontinuous Transmission (DTX) Scheme on Enhanced Full Rate (EFR) Voice Traffic Channels” and for voice condition detection (VAD), see G.
SM Recommended Standard GSM06.82 "Extended Full Speed (EFR)"
Voice State Detection (VAD) on Voice Channels ". Therefore, description of these various functions is omitted here.

FIG. 1 (b) is a block diagram showing a conventional decoder used to generate comfort noise at the receiving side in a conventional voice communication system. The decoder receives two comforting noise parameters of an average excitation gain g _mean and a set of short-term spectral coefficients f _mean (i) (i = 1, ..., M) and based on these parameters Comfort noise is generated. The behavior of comfort noise generation at the receiver side is that the parameters are used at a fairly low rate (eg once every 480 ms in GSM FR and GSM EFR channels) and that there is no excitation signal from the speech coder. The operation is the same as the voice decoding operation except. In the voice decoding operation, the excitation signal on the receiving side is
An index obtained from a codebook containing multiple possible excitation signal sequences and corresponding to a particular excitation vector in the codebook is transmitted along with other speech coding parameters. Speech decoding and codebook reference operations are disclosed in detail in, for example, US Pat. No. 5,327,519, "Pulse Pattern Excited Linear Predictive Speech Coder," by Jean-Hagwist, Karijavinen, Karipekka Esrolla, and Jucca Lanta. All of these contents are hereby incorporated by reference.

On the other hand, in generating comfort noise, the codebook index is not transmitted, and instead a random excitation signal (RE) generator 110 is used to obtain the excitation signal. The random excitation signal generator 110 generates an excitation signal vector 114 having a flat spectrum.
The excitation signal vector 114 is the scaling unit 11
5 is scaled by the average excitation gain g _mean so that its energy corresponds to the average gain of the excitation signal 104 at the transmitter. The scaled random excitation signal sequence 111 thus obtained
Is input to the speech synthesis filter 112, and the comfort noise output signal 113 is generated. At that time, the average short-term spectrum coefficient f _mean (i) is used in the speech synthesis filter 112.

FIG. 1 (c) is a diagram showing spectra at various portions of the prior art decoder of FIG. 1 (b). The random excitation signal generator 110 generates a random excitation signal sequence 114 (and a scaled random excitation signal sequence 111) having a flat spectrum.
This spectrum is shown by curve A. The speech synthesis filter 112 transforms this excitation signal into a signal having a non-flat spectrum as shown by curve B.

As described above, there are many problems in the conventional comfort noise generation technology. One of these problems is that the comfort noise due to random excitation generated on the receiving side is not the properly excited comfort noise as described above, and the actual background on the transmitting side is It is to be perceived as different from noise. The object of the present invention is to eliminate or reduce such problems altogether.

A first object and advantage of the present invention is to provide an improved method of generating comfort noise in discontinuous transmission, thereby minimizing the signal degradation that occurs in discontinuous transmission schemes. .

Another object and advantage of the present invention is to provide an improved method of generating comfort noise which is capable of better extracting background noise features, which further improves comfort of improved quality. Noise and improved quality communication in discontinuous transmission schemes.

Yet another object and advantage of the present invention is to provide an improved method of generating comfort noise that can better represent actual background noise within a short averaging time in discontinuous transmission. Is.

[0028]

[Means for Solving the Problems]

[0029]

[0030]

[0031]

[0032]

Further, according to the present invention, the parameters are averaged, and if there is a bad speech coding parameter within the averaging period, all or part of it is removed or replaced with the median value. In the embodiment of the present invention, the distances between the speech coding parameters between the individual frames within the averaging period are measured, these parameters are arranged in the order of the measured distances, and other parameters are averaged during the averaging period. The parameter with the largest distance from is identified, and if the maximum distance exceeds a predetermined threshold value, the parameter with the maximum distance is compared with other parameters within the averaging period. Replace with the parameter with the smallest distance (ie, median). The parameter having the median value can be regarded as having a value that most faithfully represents the characteristic of the background noise among the parameters obtained during the averaging period.
After performing the above processing, the average of the speech coding parameters is obtained by using an appropriate arbitrary method. In the embodiments of the present invention, the method of receiving and using the comfort noise parameter on the receiving side of the discontinuous transmission system is not limited.

In addition to improving the quality of comfort noise by eliminating bad CN parameters within the averaging period, embodiments of the present invention also have other advantages. For example, in the discontinuous transmission system according to the related art, it is necessary to lengthen the averaging period in order to reduce the adverse effects of the bad parameters within the averaging period. According to the present invention, since the effect of reducing the influence of the bad parameter in the averaging is applied, the averaging period that is shorter than the length of the averaging period required in the conventional technique is sufficient.
Further, in the conventional technology, since the averaging period is long,
The hangover period also becomes longer, which increases the channel usage time. On the other hand, in the present invention, since the averaging period can be short, the discontinuous transmission hangover period can be shortened, and thus the channel use time can be reduced. Furthermore, in the prior art discontinuous transmission system,
Due to the long averaging period, a fairly large amount of static memory is required to implement the CN averaging algorithm. On the other hand, in the present invention, the averaging period is short, and a small capacity static memory is sufficient to execute the CN averaging algorithm.

[0035]

DETAILED DESCRIPTION OF THE INVENTION Prior art techniques for encoding and decoding comfort noise have been described above. A circuit and method according to a first embodiment of the present invention will be described below with reference to FIGS. 2 (a) -2 (c). Note that FIG.
In (a), 2 (b), similar reference numerals are used for elements corresponding to FIGS. 1 (a), 1 (b).

"Unvoiced descriptor averaging period" is a term used in GSM, and is defined in IS-641 Rev. A
The term "comfort noise averaging period" or "C
"N averaging period" is used. In the following description of the invention, these two terms are used synonymously. Similarly, "SID frame" and "comfort noise parameter message" or "CN parameter message"
Are used synonymously.

FIG. 2 (a) is a block diagram showing an apparatus for generating comfort noise parameters at the transmitting side according to the present invention. The novel parts specific to the present invention are shown enclosed by dashed lines 204 to distinguish them from parts of the prior art known in the art. In this embodiment of the invention, the residual signal 1 output from the inverse filter 103
04 is further analyzed (eg, LPC analysis) to produce a further set of filter coefficients. For this second analysis, the present invention uses a random excitation (RE) linear predictive coding (LPC) analysis 200, which is generally of lower order than the linear predictive coding analysis performed at block 101.
The spectral parameters 201 output from the random excitation linear predictive coding analysis block 200 are averaged by the averaging block 203 for several consecutive frames, whereby random excitation spectrum control (RE) is performed.
SC) Parameter r _mean (i) (i = 1, ..., R)
Is required. The random excitation spectrum control parameter characterizes the spectrum of the excitation signal.

The random excitation spectrum control parameter is not a subset of speech coding parameters, but is a parameter generated and used only in comfort noise generation. It should be noted that the inventors of the present invention can obtain sufficiently good results in the linear predictive coding analysis of the first or second order to generate the random excitation spectrum control parameter (R = 1 or 2). Found out. However, it is also possible to use a spectrum model other than the all-pole linear predictive coding method. Alternatively, the averaging process by the random excitation linear predictive coding / analysis block 200 may be performed by averaging the autocorrelation coefficients in the linear predictive coding parameter calculation. Other suitable averaging methods may be used in the calculation of The averaging period of the random excitation spectrum control parameter may be the same as the averaging period of the other comfort noise parameters, although it does not have to be the same. For example, it is desirable to set the averaging period of the random excitation spectrum control parameter longer than the averaging period conventionally used for the comfort noise parameter. That is, instead of using 7 frames as the averaging period, it is desirable to use a longer averaging period of, for example, 10 to 12 frames.

Prior to calculating the back-to-back signal gain, the linear predictive coding residual signal 104 is _{fed to} the second inverse filter H _RESC.
(Z) 202 is supplied and processed. This filter 2
02 transforms the linear predictive coding residual signal 104 into a residual signal 205, which generally has a flatter spectrum. Inverse random excitation spectrum control filter H
_{As RESC} (Z), it is possible to use an all-zero filter of the following form, but it is not limited to this.

[0040]

[Equation 12]

Next, the excitation signal gain is calculated from the residual signal 205 having the flattened spectrum. In other respects, the operation of FIG. 2A is the same as the operation of FIG.

FIG. 2 (b) is a block diagram illustrating a decoder used to generate comfort noise at the receiving side according to the present invention. Excitation signal 21 by this decoder
The formation of 2 is performed as follows. First, the white noise excitation signal sequence 114 is converted into the random excitation signal generator 11
Generated by 0, which is then scaled by scaling block 115 by g _mean .

This noise signal sequence 111 having a flat spectrum is then subjected to random excitation spectrum control (RE
SC) filter 211 produces an excitation signal having the desired spectrum. The random excitation spectrum control filter 211 performs a process reverse to that of the random excitation control inverse filter 202 of the encoder shown in FIG. If a random excitation spectrum control inverse filter having the characteristic of equation (2) is used on the transmitting side, then on the receiving side,
A random number control spectrum control filter 211 having the characteristics of the following equation is used.

[0044]

[Equation 13]

Filter coefficient b (i) (i = 1, ...,
The random excitation spectrum control parameter r _mean (i) (i = 1, ..., R) that defines R) is sent to the receiving side as part of the comfort noise parameter, and the receiving side uses the random excitation spectrum control filter. 211 uses this random excitation spectrum control parameter to generate an excitation signal with an appropriately weighted, and therefore generally non-flat, spectrum and supplies it to synthesis filter 112. Random excitation spectrum control parameter r
_mean (i) (i = 1, ..., R) is the filter coefficient b
(I) may be the same as (i = 1, ..., R),
Alternatively, any other suitable parametric representation, such as line spectrum pair (LSP) coefficients, that can be efficiently quantized and transmitted, may be used. 11 (a) to 11
(G) shows a random excitation spectrum control filter 211
2 shows a typical frequency response of.

As explained above, the comfort noise excitation signal generator 210 of the present invention has a characteristic which is not available in the prior art. That is, the comfort noise excitation signal generator 210
Uses a random excitation signal generator 110 to generate a random excitation signal having a flat spectrum. The excitation signal having the flat spectrum is supplied to the random excitation vector control filter 211 after being appropriately scaled by the average gain scaler 115. In order to prevent a mismatch between the comfort noise spectrum and the background noise spectrum, comfort noise having an appropriate spectrum is generated and a random excitation signal is provided via the random excitation vector control filter 211. next,
Excitation signal 21 whose spectrum is controlled as described above
Based on 2, the speech synthesis filter 112 produces comfort noise whose spectrum better matches the background noise actually present at the sender.

The random excitation spectrum control parameter is not a subset of speech coding parameters used for speech signal processing, but a parameter calculated and used only in comfort noise generation. That is, the random excitation spectrum control parameters are calculated and transmitted only for the purpose of exciting and generating improved comfort noise during the pauses of speech. Also, the random excitation spectrum control inverse filter 202 of the encoder
And the random excitation spectrum control filter 21 of the decoder
1 is used only for controlling the spectrum of the random excitation signal.

FIG. 2 (c) shows the spectrum of the signal in the decoder of FIG. 2 (b) when comfort noise is generated according to the present invention. Random excitation signal generator 11
0 produces a random number sequence with a flat spectrum as shown by curve A. This spectrum is the same as that shown by curve A in FIG. 1 (c). Both the signal 114 and the signal 111 have such a flat spectrum. That is, the scaling by block 115 does not affect the shape of the spectrum. The white noise sequence 111 is input to the random excitation spectrum control filter 211, and the output excitation signal 21
2 is supplied to the linear predictive coding synthesis filter. The excitation signal train 212 with improved properties generally has a non-flat spectrum (curve B 1 ). Therefore, the spectrum of the output signal 113 of the synthesis filter 112 becomes, for example, as shown by the curve C in FIG. The excitation signal sequence 212 may be a low-pass type or a high-pass type, or may include a more complicated frequency component (depending on the order of the random excitation spectrum control filter). The control of the spectrum is determined by the random excitation spectrum control parameters calculated at the transmitter and transmitted to the receiver as part of the comfort noise, as described above.

3 (a) and 3 (b) are views showing still another embodiment of the present invention. FIG. 3 (a) to FIG.
When compared with (a), in the case of this embodiment, the excitation signal gain is the random excitation spectrum control inverse filter 202.
The difference is that it is calculated from the linear predictive coding residual signal 104, not from the residual signal output from. Therefore, the random excitation spectrum control inverse filter 202 is not necessary in this embodiment shown in FIG. A decoder used on the receiving side corresponding to the encoder of FIG. 3 (a) is shown in FIG. 3 (b). Compared to the decoder of FIG. 2 (b), the excitation signal scaling block (block 11
5) is moved from the input side of the random excitation spectrum control filter 211 to the output side. In other respects, the encoder and the decoder of FIGS. 3A and 3B are similar to the operations of the encoder and the decoder of FIGS. 2A and 2B.

FIG. 4 is a block diagram showing a circuit for calculating comfort noise parameters at the transmitting side according to still another embodiment of the present invention. This embodiment addresses the above-mentioned problem that occurs when some or all of the speech coding parameters deviate significantly from the average background noise when one or several frames are present during the averaging period. Is. The parts specific to this embodiment of the invention are surrounded by dashed lines 300 and 310 to distinguish them from the prior art parts. In this embodiment of the invention, the speech coding parameters buffered in blocks 107a and 108a are first subjected to a threshold median replacement process and then averaged blocks 107 and 108.
And the average excitation signal gain g _mean and the average short-term spectrum coefficient f _mean (i) are calculated. By this processing, when the averaging period includes a parameter deviated from the average value of the background noise by a certain amount or more,
That parameter is replaced with the value of the parameter representing the actual background noise, ie the median.

First, the calculation processing of the scalar excitation signal gain parameter g by the block 300, which is performed before the averaging processing by the block 107, will be described.
A series of excitation signal gain values 107b during the averaging period
After being temporarily stored in the buffer block 107a, these values are input to the block 301 and rearranged in the order of the magnitude of the values. These series of pump gain values are
Each is given an index. The permuted series of gain parameters 302 is the median permutation block 3
Of these excitation signal gain values, L of which the difference from the median exceeds a predetermined threshold value are replaced with the median value of the series of parameters. The difference between the value of the individual parameter and the median is calculated by block 304, and if the difference exceeds a threshold value, the index corresponding to the value of the excitation signal gain is provided as signal 305 to median replacement block 303. To be done.

An odd value is preferably selected for the length N of the averaging period. In this case, the ((N + 1) / 2) th value of the rearranged series of pump gains is the central value. The variable L that determines the number of parameters to be replaced is 0
To N-1 and is determined in advance (for example, a constant).

For the value of the excitation gain such that the difference from the central value exceeds a predetermined threshold value, the selector 3
07 is switched, and the output signal 308 of the center value substitution block 303 is supplied to the averaging block 107 as the value 309 of the excitation gain. On the other hand, for the individual excitation signal gain values whose difference from the median value is predetermined and does not exceed the threshold value, the output of the buffer block 107a is used as the input parameter value 309 to directly average the averaging block 1.
The selector 307 is switched so as to be supplied to 07.

The selection state of the selector 307 is controlled by the signal 306 output from the threshold value block 304.

Next, the process for the line spectrum pair (LSP) coefficient f (k) (k = 1, ..., M) performed by the block 310 before the averaging process by the block 108 will be described. During the averaging period, a series of LSP coefficients 108b are buffered in block 108a and then provided to block 311. LSP coefficient f _i (k) of the i-th frame in the averaging period LSP coefficient f _j of the j-th frame in the same averaging period
The spectral distance from (k) can be approximately represented by the following formula.

[0056]

[Equation 14]

Here, M is the order of the linear predictive coding model, and f _i (k) is the k-th LSP parameter of the i-th frame in the averaging period.

In the averaging period of length N, the LSP coefficient f _i (k) of the i-th frame is compared with the LSP coefficients of all other frames j = 1, ..., N (i ≠ j). In order to obtain the spectral distance ΔS _i , the sum of the spectral distances ΔR _ij is used for all i = 1, ...
・, Calculate for N.

[0059]

[Equation 15]

However, ΔR _ij = 0 (that is, the distance from itself is 0). The processing represented by the above equations (4) and (5) is performed in block 311.

Spectral distances are determined by various other L
It can be approximated by a PC filter expression, and specific examples thereof are, for example, AH Gray Jr., JD Markel, “Distance Measure in Speech Processing” (, IEEE Acoustics, Speech, and Signal Processing). Transaction, Vol. 24, p.
380-391, 1976). Also, the immittance spectrum pair (ISP) can be used similarly to the line spectrum pair. For this, see, for example, Wybistriz, Esperell, "Imittance Spectrum Pair (ISP) in Speech Coding".
(Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing, Minneapolis, Minnesota, Vol. 2, p. 9-1.
2, April 27-30, 1993).

The spectral distance ΔS _i of each LSP vector f _i during the averaging period, determined by block 311, is supplied to block 313. In this alignment block 313, the spectral distances are sorted in order of their magnitude. Note that each of these spectral distances is associated with the LPS vector of the averaging period by the index. In the averaging period, vector f _i (i with the smallest distance [Delta] S _i =
, ..., N) is considered as the median vector f _med of its averaging period and its distance is represented by ΔS _med .

A series of LSP coefficient vectors f _i in the averaging period are stored in block 3 according to the order of the spectral distance.
Aligned by 13. This series of LSP vectors 314 aligned by block 313 is provided to median replacement block 315. In block 315, LSP vector f _i of P (0 ≦ P ≦ N-1 ) is,
It is replaced with the median f _med . The indices of these P vectors are ΔS _i (i
, 1, 2, ..., N) and the median ΔS _med . The index of f _i , where ΔS _i −ΔS _med is greater than a predetermined threshold, is conveyed by signal 317 to median replacement block 315.

For some i = 1, 2, ..., N, the output signal 320 from the median replacement block 315 is averaged as the parameter 321 for the difference ΔS _i −ΔS _med larger than the threshold value. The selector 319 is switched to be supplied to the block 108. On the other hand, ΔS _i −
If ΔS _med is less than the threshold for all i = 1, 2, ..., N, the input signal 321 to the averaging block 108 is the signal 1 directly from the buffer block 108a.
The position of the selector 319 is switched, as provided as 08b.

The selection state of the selector 319 is controlled by the signal 318 output from the threshold value block 316.

FIG. 5 is a diagram showing another embodiment of the present invention. The parts specific to this embodiment of the invention are surrounded by the dashed line 400 to distinguish them from the prior art parts. Figure 4 above
In the embodiment of FIG. 5, the median values of the excitation signal gain g and the LSP vector f _i are processed independently, but FIG.
In this embodiment, these parameters are processed simultaneously as described below.

When it is determined that the parameters of a frame should be replaced by the median, both the excitation signal gain g and the LSP vector f _i for that frame are replaced with the respective parameters of the frame having the center. .

The i-th frame in the averaging period and j used to align the frames for median replacement
Instead of equation (4) representing the approximate distance ΔR _ij between the parameters of the th frame, the excitation signal gain g and LSP
The following equation (6) in which both of the vectors f _i are considered is used here.

[0069]

[Equation 16]

Here, M is the order of the linear predictive coding model, and f _i (k) is the k-th LPS parameter of the i-th frame in the averaging period. Also, q
_i is the excitation signal gain parameter of the i-th frame.

ΔT _ij is first calculated, and then using equation (5), other frames j = 1 ... N (i) of the parameters of frame i within the averaging period of length N Find the distance ΔS _i for ≠ j) for all i = 1, ..., N. ΔT _ij obtained by using the equation (6) is given by the equation (5)
When applied to, the distance ΔR _ij is used instead of the distance ΔR _{ij in} the equation (5). The processing according to equations (5) and (6) above is performed by block 401. The weighting factor w in the above equation (6) is appropriately selected depending on which of the excitation signal gain and the spectral distance is to be applied to the median value replacement process. This weighting factor is set to an appropriate value empirically obtained by the user.

The distance ΔS _i of each frame in the averaging period obtained by the block 401 is supplied to the alignment block 403 as 402. The alignment block 403 rearranges the distances according to these values. Each of these distances is associated by index with one frame within the averaging period. Within the averaging period, the minimum distance ΔS _i (i = 1, 2, ...
, N) is taken as the median frame for the parameters g _med and f _med of its averaging period and its distance is denoted ΔS _med .

The excitation signal gain to be aligned by block 403 is provided from buffer 107a to block 403 as signal 107b, while the LSP coefficients are signal 10.
8b is supplied from the buffer 108a to the block 403. As already mentioned, the series of parameters of the averaging period are ordered by their block 403 in their spectral distance ΔS _i . The set of parameters aligned by block 403 are provided as signals 404 and 405 to median replacement block 406 and block 4
At 06, the parameters g _i and f _i of the L (0 ≦ L ≦ N−1) frames are replaced with the parameters g _med and f _{med of the} median frame. For the indices of these L vectors, the block 407 determines ΔS _i (i =
, 2, ..., N) is compared to the median ΔS _med, and the determined index is the signal 40
8 is transmitted to the median substitution block 406. ΔS
_{If i} −ΔS _med is larger than a predetermined threshold, block 407.
If so, the median replacement block 406 replaces the corresponding parameters g _i and f _i with g _med and f _med . Note that the minimum value and the maximum value of the value of L may be determined in advance.

If any of i = 1, 2, ..., N is greater than the distance ΔS _i −ΔS _med , the signal 411 output from the median replacement block 406 is the averaging block 108. Signal 41 input to the median substitution block 406 as a parameter 321 to
The position of the selector 410 is switched so that 2 is input to the averaging block 107 as the parameter 309. On the other hand, ΔS _i −ΔS _med is the same for all i = 1, 2, ...
.., N is smaller than the threshold value, the input signal 321 of the averaging block 108 is the buffer block 108a
The position of the selector 410 is switched so that the signal 108b is directly supplied from the buffer block 107a and the input signal 309 of the averaging block 107 is directly supplied as the signal 107b from the buffer block 107a. The position of the selector 410 corresponds to the signal 4 from the threshold block 407.
Controlled by 09.

As described above, the median distance is subtracted from the individual distance values (that is, ΔS _i −ΔS _med is obtained).
Instead, use the quotient (ΔS _i / ΔS _med ) of the individual distances divided by the median distance, for example, in blocks 316 and 407 to represent the difference between the individual distances and the median distance. In many cases, it is more preferable to use the quotient in this way. When the quotient is used, the relative displacement or normalized displacement of the individual distance values from the median distance is obtained, which is independent of the absolute value of ΔS _i or ΔS _med. It is possible to express the gap.

Next, another embodiment of the present invention will be described. Before that, first, the voice encoder discontinuous transmission system on the transmission side will be described with reference to the block diagram of FIG. The signal 601 from the A / D converter 600 is processed by the speech encoder 602 on a frame-by-frame basis. As in the above case, the frame length is typically 20 ms. The sampling frequency of the audio signal 601 is generally 8 kHz. Speech encoder 602 encodes the input speech frame by frame into a set of parameters 603. The parameters thus obtained are transmitted from the radio subsystem 611 of the digital mobile radio unit to the reception (RX) side.

The discontinuous transmission operation is indirectly controlled by the voice state detecting device on the transmitting side. Voice state detection device 6
The basic function of 04 is the noise present in the voice signal,
It is to distinguish the noise when there is no voice but only noise. The voice state detection device 604 continuously monitors and determines whether the input signal contains voice or does not contain voice at all. The operation of the voice state detecting device 604 is performed based on the voice encoder 602 and its internal variable 605. The voice state detection device 604 outputs a binary voice state flag 606 that takes a value of 1 when voice is present and takes a value of 0 when voice is not present. The operation of the voice state detecting device 604 is based on the standard GSM0.
The frame is used as a unit as defined in 6.82.

Traffic frames, each associated with a binary SP flag 607, are continuously sent from the voice encoder discontinuous transmission handler 612 to the radio subsystem 611. The wireless subsystem 611 determines whether the traffic frame sent from the discontinuous transmission handler 612 is a voice frame (SP flag = "1") or a so-called unvoiced identifier (SID) frame (that is, a comfort noise parameter message) ( Whether the SP flag = “0”) can be known by the SP flag 607. The wireless subsystem 611 transmits the frame as a radio wave to the air at a timing based on the state of the SP flag 607.

The basic problem in the above discontinuous transmission is as follows. That is, the audible background noise that was transmitted with the voice disappears when the transmission is paused, so that the background noise becomes discontinuous on the receiving side. If the transmission and the transmission pause are frequently repeated within a short period of time, the discontinuity of the background noise described above causes an unpleasant feeling to the receiving user. This unnaturalness becomes particularly noticeable in an environment where background noise is large, such as an automobile. In the worst case, this discontinuity may make speech indistinguishable.

As one of the known solutions to this problem, when the transmission is interrupted, the receiving side can synthesize noise similar to the background noise present at the transmitting side (ie comfort noise). It is being appreciated. As already mentioned, the parameters needed to generate the comfort noise are determined by the sender's speech coder (block 608 in FIG. 6) and transmitted to the receiver by the SID frame before the transmission is interrupted. To be done. After that, the parameters are repeatedly transmitted at low speed. By doing so, it becomes possible to make the comfort noise generated on the receiving side correspond to the actual change of the background noise on the transmitting side.

If the comfort noise parameter determined at the transmitter properly represents the level of actual audible background noise and the envelope of the spectrum, it is possible to generate high quality comfort noise at the receiver. Is known to be. These characteristics of background noise change slightly over time. Therefore, in order to generate good comfort noise, the parameters representing the level of background noise and the envelope of the spectrum are generated by the speech coder and these parameters are averaged over several speech frames. is necessary. In the discontinuous transmission scheme in the GSM full-rate speech encoder and the extended full-rate encoder (see GSM06.31 and GSM06.81), the length of the SID averaging period is 4 frames and 8 frames, respectively, and 1 The frame length is 20 milliseconds.

Before the transmission is interrupted at the end of the voice burst, the optimized noise parameter is obtained and included in the first SID frame to be transmitted to the receiving side as described above. , A hangover period is used. The hangover is a period (that is, VAD flag 606 = “0”) in which it is determined that the voice is not present by the voice state detection device 604, and the transmission of the voice frame is not stopped yet (that is, the SP flag 607 =). "1") period. In this regard, FIG.
See also. During the hangover period, the voice state detection device 604 has already determined that there is no voice, so it is guaranteed that the voice frame contains only noise (not voice), and thus these The comfort noise parameter can be determined by averaging the speech coding parameters of the hangover frame of.

The length of the hangover period is determined by the length of the SID averaging period. That is, the hangover period should have a sufficient length needed to average the obtained comfort noise parameters before transmitting them in the SID frame. In the discontinuous transmission method of the GSM full-speed speech encoder, the length of the hangover period (the length of the SID averaging period) is used to determine the comfort noise by the method of updating the SID frame using the parameters of the previous frame. ) Is
4 frames is enough. On the other hand, in the discontinuous transmission method of the GSM extended full speed speech encoder, the length of the hangover period is 7 frames (the length of the SID averaging period-1), and the length of the 8th frame of the SID averaging period is The parameters are obtained while the speech encoder is processing the first SID frame. FIG. 7 shows a hangover period and S in the GSM extended full-speed voice coding discontinuous transmission system.
The concept of the ID averaging period is shown.

At the end of the hangover period, the first SID
The frame is transmitted. The background noise characteristics are subsequently evaluated using the comfort noise calculation algorithm, and the updated SID frame is sent to the wireless subsystem 611 on a frame-by-frame basis. The SID frame is continuously transmitted while the voice state detecting device 604 determines that there is no conversation voice. When the SID averaging period ends, the discontinuous transmission handler 61 on the transmitting side
2 is the SID averaging period comfort noise calculation algorithm 6
08 is notified by a flag 609 that the SID averaging period has ended. The flag 609 is normally in a reset state of "0", but is set to "1" when the updated SID frame is sent to the wireless subsystem 611. When the flag 609 is set, the comfort noise calculation algorithm 608 averages the parameters to create an updated SID frame used by the wireless subsystem 611.
The updated SID frame is transmitted to the wireless subsystem 611.
And written to the SID memory block 610 as well. The latest SID frame is always stored in the SID memory block 610 and is used in the subsequent processing.

When the voice burst ends, the last SI
If the time elapsed since the D frame was sent to the wireless subsystem is 24 frames or less, the last S
The ID frame is repeatedly read from the SID memory 610 and sent to the wireless subsystem 611. The above operation is continued until an updated new SID frame is prepared, that is, until the next SID averaging period ends. This improves transmission efficiency because it is not necessary to insert a hangover period for calculating a new SID frame at the end of a voice burst when a short spiked background noise is mistaken for voice. be able to.

FIG. 8 shows, by way of example, a voice burst when it lasts as long as possible without a hangover. The binary flag 613 is the SID memory 61.
0 is used to indicate when to store an updated new SID frame in the SID memory 610 and when to send the latest updated SID frame from the SID memory 610 to the wireless subsystem 611. S
If the P flag 607 is "0", the SID memory 6
10 determines in each frame whether the SID frame should be stored or transmitted.

In the GSM enhanced full rate speech coder discontinuous transmission system, the binary flag 614 is also necessary to inform the noise estimation algorithm of the end of the hangover period. Flag 614, which is normally reset to "0", is raised to "1" for one frame period during which the first SID frame after the hangover period after the voice burst is transmitted.

FIG. 9 shows the reception (R
It is a block diagram which shows the audio | voice decoder in the (X) side. A series of speech coding parameters 701 input from the radio subsystem 700 of the digital mobile radio unit are processed frame by frame by the speech decoder 702 to synthesize a speech signal 703. The obtained audio signal 703 is D /
It is supplied to the A (digital-analog) converter 704. The D / A converter 704 generates an audible voice signal for the receiving user.

The receiving side discontinuous transmission system receives the binary SP flag 705 from the radio subsystem. This flag operates in contrast to the sending SP flag. That is, if an audio frame is received, S
When the P flag is "1" and the SID frame is received or the transmission is interrupted, the SP flag is "0".
Becomes Also, a binary flag 706 from the radio subsystem 700 informs the comfort noise generation algorithm 707 of the presence or absence of a newly received SID frame. That is, this flag is normally reset to "0", but is set to "1" when a new SID frame is received while the SP flag 705 is "0".

The SP flag 705 = "0", so when the transmission is dormant, the comfort noise generation block 707 of the speech decoder 702 represents the characteristic of the background noise on the transmitting side contained in the SID frame. Based on the data
Generate comfort noise. During the transmission suspension period,
Although the updated SID frame is sent at a slow repetition rate, it is possible to smoothly change the characteristics of the comfort noise by interpolating the decoded comfort noise parameter between these updated SID frames. Can be

In the GSM full-speed voice coding discontinuous transmission system, when a new updated SID frame is newly calculated and sent to the radio subsystem 611 (FIG. 6), the parameters indicating the characteristics (level and spectrum) of background noise are set. , SID is averaged during the SID averaging period and then scalar quantized using the same quantization method as in the normal speech coding mode. Similarly, if the SID frame is G
Upon arrival at the SM full speed speech decoder 702, the unvoiced descriptor parameters are decoded using the same dequantization method used in the normal speech decoding mode (see eg GSM06.12).

In the GSM extended full-speed voice coding discontinuous transmission system, when calculating a new SID frame, a parameter (LSP) representing the spectrum of background noise is used.
The average of (parameters) in the SID averaging period is calculated, and the vector is quantized using the same quantization table that is used to quantize the parameters in the normal speech coding mode. These spectral parameters are dequantized by the decoder 702 using the same table as the predictive dequantization table used in the normal speech coding mode. When calculating a new SID frame, the parameter (fixed codebook gain) representing the level of background noise is set to SI.
After being averaged in the D averaging period, quantization is performed using a scalar prediction quantization table used for quantization in a normal speech coding mode. At the decoder, these parameters are dequantized using the same table as the predictive dequantization table used in the normal speech decoding mode.

However, in the case of adaptively controlling the predictive quantizer, it is difficult to use the above quantization method for the quantization of the comfort noise parameter to be transmitted by the SID frame. The reason is that it is not possible to maintain the frame-by-frame synchronization of the quantizer and dequantizer predictors of the encoder and decoder, respectively, because the transmission is interrupted when the speech is interrupted. Because it is. However, it is possible to locally estimate the quantizer predictor value in the same way in the encoder and decoder, as described below. Ie the newest 7
The quantized LSP and fixed codebook gain parameters of the speech frame are stored locally in both encoder 602 and decoder 702. At the end of the hangover period at the end of the voice burst, the average value of these stored parameters is determined. The average parameters thus obtained, that is, the reference LSP parameter vector f ^ref , the reference fixed codebook gain g
_c ^ref has the same value in the encoder 602 and the decoder 702. The reason is that in quantization, both the quantized LSP and the fixed codebook gain have the same value in the encoder 602 and the decoder 702 in the normal speech encoding mode. The average value of the above reference LSP parameter vector f ^ref and the reference fixed codebook gain g _c ^ref is kept frozen until the hangover period after the next voice burst. This frozen average value is
It is used to quantize the comfort noise parameters instead of the normal predictor of the quantization algorithm.

Returning to FIG. 9, the discontinuous transmission handler 708 on the receiving side receives the SP flag 705 as an input,
1 if a hangover period occurs after a voice burst
The binary flag 709 is set to "1" and output only during the frame period. The flag 709 is reset to "0" in other periods. The flag 709 is used for the comfort noise generation algorithm 7 in the discontinuous transmission system of the GSM extended full speed speech decoder 702.
07, the reference LSP parameter vector f ^ref and the reference fixed codebook gain g _c ^ref by the averaging process.
Used to signal when to update the
See SM06.62). For a method of determining the value of flag 709, see Finnish Patent Application No. 95325.
No. 2 and corresponding US patent application Ser. No. 08 /
No. 672,932 (filed June 28, 1996), International Application No. PCT / FI96 / 00369, all of which are incorporated herein by reference.

As described above, in recent speech encoders, the speech coding parameters are quantized using the prediction method. That is, the quantizer attempts to predict the quantized value as accurately as possible. In such a predictive quantizer, the difference or quotient between the actual parameter value and the predicted parameter value is usually quantized and transmitted to the receiving side. The inverse quantizer on the receiving side, which corresponds to the above quantization, has a predictor similar to that of the quantizer. Therefore, the value of the parameter quantized on the transmission side can be reproduced by adding or multiplying the received difference or quotient value to the predicted value.

In such a predictive quantizer, the adaptation is usually performed by using the result of the quantization and updating the predictor after each quantization. The predictors of both the quantizer and dequantizer are updated with the values of the reconstructed quantization parameters so that they remain synchronized with each other.

However, there are difficulties in using such a quantization scheme that adapts the predictive quantizer to quantize the comfort noise parameters sent by the SID frame. That is, since the transmission is interrupted while the conversational voice is interrupted, the quantizers of the encoder 602 and the decoder 702 and the predictors of the dequantizer cannot be synchronized on a frame-by-frame basis.

The same quantization table used by the predictive quantizer to encode normal speech can be used to quantize comfort noise parameters, if possible:
That is preferable. For that purpose, it is necessary to perform prediction during discontinuous transmission without performing adaptation. To enable the quantizer to encode the value of the parameter in response to changes in the background noise characteristic,
It is necessary that the predictor has a value as close as possible to the value of the current background noise average parameter. Further, it is desirable that the same predicted value can be used in both the quantizer and the dequantizer.

As already mentioned, in quantizing the comfort noise sent by the SID frame, one way to obtain a good predictive value is to hang over the value of the parameter quantized by the normal speech coding mode. During the period, store and average the value of this stored quantized parameter at the end of the hangover period and freeze the average value of the predictor thus obtained until the next hangover period. It is to hold. However, this method has the problem that in GSM or similar discontinuous transmission schemes, the speech decoder 702 cannot know when the last hangover period of a speech burst has occurred.

In view of the above problems, it is an object of the present invention to provide a method for informing the speech decoder 702 of the occurrence of the last hangover period of a speech burst. The above purpose is preferably a speech coder 60.
2 to the speech decoder 702 by sending the hangover period information as side information in a SID frame (ie a comfort noise parameter message).

This method of the present invention will be further described with reference to FIG. In the scheme in FIG. 10, the binary flag 709 does not use the value generated by the receiving side discontinuous transmission handler, but uses the SI via the transmission channel.
The value sent from the encoder 602 by the D frame is used. Therefore, in order to perform the inverse quantization by the prediction method of the present invention, the flag 709 is locally generated in the decoder 702.
Need not be generated, and therefore the discontinuous transmit handler block 708 on the receiving side is no longer needed. In this scheme of the invention, flag 709 is raised to "1" in the first SID frame after the hangover period. However, if the hangover period preceding the first SID frame does not exist, the flag 709 of the first SID frame is reset to "0".
In the second and subsequent SID frames during the comfort noise insertion period, the flag 709 is always reset to "0".

An advantage of this method of the invention is that the speech decoder does not need the discontinuous transmit handler 708 to locally determine if there is a hangover period at the end of the speech burst. Therefore, the program instruction used by the discontinuous transmission handler 708 on the receiving side becomes unnecessary, and the calculation load of the speech decoder 702 is reduced.

Another advantage of the method of the present invention in providing information about the presence of a hangover period to the decoder 702 is that the pseudo noise excitation signal is generated at the encoder 602 and the decoder 702 each time the hangover period ends. It is possible to reinitialize the generator synchronously. Yet another advantage of the method of the present invention for providing information about the presence of a hangover period to the decoder 702 is the received comfort depending on whether the hangover period is present or absent at the end of the voice burst. It will be possible to interpolate the noise parameters using various methods, which result from the level of comfort noise or a step change in the spectrum that can occur when the speech burst is short. The perceptual unnaturalness of comfort noise can be reduced.

Prior to describing the invention in further detail, FIG.
With reference to 2 and 13, a wireless user terminal or a mobile station 10 (for example, but not limited to, a cellular phone or a mobile phone) to which the present invention is applicable will be described. The mobile station 10 has an antenna 12 for transmitting and receiving signals to and from the base station 30. The base station 30 is, for example, part of a cellular communication network including a base station / mobile station switching center function (BMI) 32 including a mobile switching center (MSC) 34. The MSC 34 connects the mobile station 10 to the ground relay line when the mobile station 10 makes a call.
In the present invention, it is assumed that the mobile station 10 is the transmitting side, while the ground station is the receiving side. It is also assumed that the base station 30 is equipped with a suitable receiver and speech decoder for receiving and processing the encoded speech parameters and the discontinuous transmission comfort noise parameters, as will be described later.

The mobile station comprises a modulator (MOD) 14A, a transmitter 14, a receiver 16, a demodulator (DEMOD) 16A, a transmitter 14, a receiver 16 and a controller 18 for transmitting and receiving signals. The signals include signal information according to the cellular system's air interface standard and the user's voice and / or user generated data. In the present invention, air interface standards are assumed to include standards for physical and logical frame structure. However, the present invention is
It is not limited to these specific physical and logical frame structures. The present invention also provides a compatible mobile station, such as IS-136 or similar, or TDM.
The application is not limited to the A-type system. Also, in the present invention, the air interface standard is assumed to support discontinuous transmission mode operation.

The controller 18 also includes circuits for implementing the voice and logic functions required in the mobile station. For example, the controller 18 includes a digital signal processing device, a microprocessor, an A / D converter, a D / D converter,
It is composed of an A converter and various other circuits. Mobile station control and signal processing functions are assigned to these devices according to their capabilities. It is assumed that the controller 18 has the speech coder and other functions required in the improved method and apparatus for generating comfort noise in discontinuous transmission according to the present invention. All of these functions can be realized by software, or all of them can be realized by hardware. Alternatively, it can be realized using both hardware and software.

As a user interface, the earphone or speaker 17, which is usually used, and A /
Microphone 1 coupled with D converter and speech coder
9, a voice transducer such as 9, a delay circuit 20, a user input device such as a keypad 22 are provided,
These are all connected to the controller 18. The keypad 22 is a standard numeral (0-9) or related key (#, *) 22a, which is necessary for operating the mobile station 10,
And other keys 22b. Other keys 22b include, for example, a "send" key, various menu scroll keys, soft keys, and a power switch key. In addition, the mobile station 10 has a battery 26 for supplying electric power to various circuits necessary for the operation of the mobile station.

Further, the mobile station 10 has various memories, and these memories are collectively represented as a memory 24. These memories store various constants and variables required by the controller 18 in the operation of the mobile station. For example, the memory 24 may include various cellular system parameters and number assignment modules (NAM).
Memorize The memory 24 also includes a controller 18
A program that controls the operation of R is also stored (usually R
Stored in OM). Further, the memory 24 is a BMI3.
It is also used to store data received from the user, such as user messages, prior to displaying it on the display. Also stored in memory 24 is a routine for performing the method described below for transmitting comfort noise parameters in a discontinuous transmission operation.

The mobile station 10 may be a device mounted on a vehicle or a portable device. The mobile station 10 may be based on various radio interface standards, modulation schemes, and access schemes. For example, the mobile station is not limited to IS-136 and may be based on any other standard, such as the GSM standard. That is, it will be apparent that the present invention is not limited to a particular type of mobile station or a particular air interface standard.

Next, as a specific example, the present invention will be described as IS-
Although an example in which the present invention is applied to 136 will be described, it is repeatedly mentioned here that the present invention is not limited to this wireless interface standard.

In the discontinuous transmission method (IS-136.1, Rev. A, section 23.11.2) on the digital traffic channel, in the discontinuous transmission high power state, the transmitter 14 moves Transmit at the power level specified by the most recent power control command received by the station 10 (initial traffic channel designation message, digital traffic channel (DTC) designation message, handoff message, dedicated DTC handoff message, or physical layer control message). To do.

On the other hand, in the discontinuous transmission low output state,
The transmitter remains off. CDVCC is not transmitted except for the Fast Associated Control Channel (FACCH) message. In the discontinuous transmission low output state, the mobile station 1
All slow associated control channels (SA
CCH) message is transmitted as a FACCH message, after which transmitter 14 discontinues transmission (DTX).
If is not prohibited, it returns to the off state.

When the mobile station 10 switches from the discontinuous transmission high output state to the discontinuous transmission low output state, all the SACCH messages in progress in the high output state are completed or terminated, or the SACCH message transmission is interrupted. , The entire interrupted SACCH message is retransmitted as a FACCH message again in the discontinuous transmission low output state.

[0114] In the transition state when the mobile station switches from the discontinuous high output state to the discontinuous transmission low output state, the transmission output is discontinuous transmission until the transmission of the FACCH message in the middle of transmission is completely completed. Maintained at high output level.

In the preferred embodiment of the present invention, the mobile station 10 terminates all transmissions of the comfort noise block (including the six DTX hangover slots and associated comfort noise parameter messages). Until then, it is in a transition state. The comfort noise block is transmitted without interruption interruption. Any other F
If the ACCH message slot occurs at the same time as the transmission of the comfort noise block, the mobile station 10
Delay either the CH message or the comfort noise block so that one is sent first and the other later. In either case, the FACCH messages are grouped or separated so that they do not interrupt or enter the slot used to transmit the comfort noise block. This allows the speech / comfort noise decoder at the base station to generate the best quality comfort noise.

In this regard, co-assigned, pending US patent application Ser. No. 08 / 936,755 “Transmission of Comfort Noise in Discontinuous Transmission” (September 2, 1997).
5th application, Sepo Aranara, and Pekka Kapanen)
checking ...

According to a specific embodiment, the comfort noise (CN) parameter message, shown in Table 1 below, is transmitted on the Return Digital Traffic Channel (RDTC), and more specifically on the FACCH logical channel. It
The comfort noise parameter message consists of 38 bits, 26 bits of which are LSF quantized using the same split vector quantization (SVQ) codebook used in the IS-641 speech codec.
Used to represent the residual vector. It should be noted that the quantization / inverse quantization algorithm of the voice codec has been revised so that the above codebook can be used. LS
The F parameter represents the spectral envelope of the background noise on the transmit side, preferably using a 10th order LPC model of the spectrum.

The next 8 bits are used as a comfort noise energy quantization index that represents the energy of the background noise on the transmit side. The remaining 4 bits of the message are used to carry the Random Excitation Spectral Control (RESC) information element.

[0119]

[Table 1]

That is, the present invention is intended to solve the above-mentioned problems in the conventional technique by causing the receiving side to generate synthetic noise similar to the background noise existing on the transmitting side. Before the transmission of conversational speech is interrupted, the comfort noise (CN) parameter is calculated at the sender and this is sent to the receiver. After that, the comfort noise parameter is transmitted at a constant slow repetition rate.
According to this method, the comfort noise can be adapted to the change in noise on the transmission side.
The discontinuous transmission method according to the present invention evaluates acoustic background noise using the voice state detection (VAD) function 21 (FIG. 12) on the transmission side and the controller 18 on the transmission side,
A function to send characteristic parameters of background noise to the receiving side,
While the transmission of conversational voice is interrupted, it is necessary to have a function of generating noise called comfort noise, which is similar to background noise on the transmitting side, on the receiving side.

In addition to these functions, if it is determined that the parameters reaching the receiving side have a significant error, the data is replaced and voice or comfort noise is generated from the replaced data. Therefore, the user on the receiving side is not made uncomfortable.

Further, the traffic frame labeled with the flag SP is continuously sent from the wireless transmitter 14 by the discontinuous transmission function on the transmission side. Here, when the SP flag = "1", it indicates that the frame is a voice frame, and when the SP flag = "0", a set of comfort noise parameters encoded by the frame is sent. Indicates that. The timing of transmitting the frame to the wireless interface is controlled by the wireless transmitter 14 based on the SP flag.

In the preferred embodiment of the present invention, all frames are infinitely long voice frames prior to resetting the mobile station 10 to enable accurate verification of the discontinuous transmission capability of the sender. Treated as if there were. Therefore, the first 6 frames after reset are always marked with the SP flag = "1" even if the VAD flag = "0" (ie, the hangover period, see FIG. 14).

The voice state detector (VAD) 21 continuously monitors whether or not the signal input from the microphone 19 contains conversational voice, and a binary flag (VAD flag = "1" or VAD flag =
"0") is output.

The VAD flag indirectly controls the entire discontinuous transmission operation on the transmission side through the discontinuous transmission handler operation on the transmission side described below.

When the VAD flag = "1",
The speech coded output frame with the SP flag = "1" is directly supplied to the wireless transmitter 14.

At the end of the voice burst (transition from VAD flag = "1" to VAD flag = "0"), the following 7 consecutive frames will generate a new set of updated CN parameters. used. Normally, the first six voice coded output frames after the end of the voice burst are fed directly to the wireless transmitter 14 with the SP flag = "1", which constitutes a "hangover period". It Subsequently, as the seventh frame after the end of the voice burst, the first new set of CN parameters is supplied to the wireless transmitter 14 with SP flag = "0" (see FIG. 14).

However, when the time elapsed since the immediately preceding CN parameter was calculated and supplied to the radio transmitter 14 was 24 frames or less at the end of the voice burst, a newly updated set was obtained. CN parameter of (VA
The immediately preceding CN parameters are repeatedly supplied to the wireless transmitter 14 until the D flag = successive 7 frames with “0” added) are ready. By doing so, the spike-like short background noise is erroneously interpreted as a conversational voice, and a “hangover” occurs to start calculation of the CN parameter, resulting in unnecessary operation of the wireless interface. Can be reduced. FIG. 15 shows, as an example, a voice burst in the case of the longest possible duration without a hangover.

After the first set of CN parameters after the voice burst is calculated and supplied to the wireless transmitter 14, while the VAD flag is "0", the discontinuity of the transmitting side continues. The transmission handler calculates the updated CN parameter, attaches the SP flag = “0” to the wireless transmitter 1
Supply to 4.

If the SP flag = "1", the speech coder operates in the normal speech coding mode, while
When the SP flag = “0”, not all the coding functions are required to calculate the CN parameter, and therefore the simple mode is used.

The radio transmitter 14 transmits traffic frames in the following order. That is, all S
P flag = frame with "1"; SP flag = first SP flag that appears after one or more frames with "1" have been transmitted = frame with "0"; SP flag = A frame that is “0” and is used for transmitting the CN parameter update message.

As a result, when the speaker stops speaking, the CN parameter message is transmitted, and the discontinuous transmission low power state is entered. While the conversation voice is interrupted, the transmission is restarted at regular intervals, and the CN parameter message for updating the comfort noise generated at the receiving side is transmitted.

[0133]

[Outer 1]

[0134]

[Outside 2]

[0135]

[Outside 3]

[0136]

[Outside 4]

The CN parameter message is always sent before the end of the voice burst and interruption of radio transmission,
Therefore, the CN parameter message also plays a role of initializing the comfort noise generation on the receiving side.

The timing when the CN parameter message and the voice frame are transmitted on the radio transmission path has already been described with reference to FIGS. 7 and 8.

The evaluation of background noise involves the process of averaging three different types of average parameters: LSF parameters, random excitation signal gain parameters, and RESC parameters. The comfort noise parameter encoded in the comfort noise parameter message consists of consecutive N = 7 frames, in the comfort noise averaging period with VAD = "0", as described in detail below. It is calculated.

Before calculating the average value of the LSF parameters in the comfort noise averaging period, first, the median substitution process is performed on a series of LSF parameters for which the average value is to be calculated, and the characteristics of the background noise on the transmitting side are calculated. Remove parameters that do not represent First, another LSF of the LSF parameter vector f (i) in the comfort noise period
Parameter vector f (j) (i = 0 ... 6, j = 0
The vector distance for 6 (i ≠ j) is calculated and approximated using the following equation.

[0141]

[Equation 17]

Where f _i (k) is the LS of frame i
It is the kth LSF parameter of the F parameter vector f (i).

All other frames j = 0 of the LSF parameter vector f (i) within the comfort noise averaging period.
... N (j ≠ i) LSF parameter vector f
To find the spectral distance ΔS _i for (j),
The sum of the spectral distances ΔR ^ij is calculated for all i = 0 ... 6 (j ≠ i) according to the following formula.

[0144]

[Equation 18]

All LSFs within the comfort noise averaging period
Among the parameter vectors, the minimum spectral distance ΔS
^The LSF parameter vector f (i) with i is taken as the central LSF parameter vector f _med for that averaging period and its spectral distance is denoted ΔS _med . The central LSF parameter vector can be regarded as the LSF parameter vector that best represents the short-term spectral characteristics of the background noise among all the LSF parameter vectors in the averaging period. Next, the LSF parameter vector f within the comfort noise averaging period
In (j), it is determined whether or not there is one that satisfies the following expression (6).

[0146]

[Formula 19]

Here, TH _med = 2.25 is the median replacement threshold. If there is an LSF parameter vector that satisfies the above equation, a maximum of two (ie, from among the LSF vectors larger than TH _med ) among them are calculated before calculating the average of the LSF parameter vector f _mean. 2) with the central LSF parameter vector.

A set of LSF parameter vectors obtained as a result of the above median replacement process is represented by f '(ni).
Where n is the index of the current frame and i
Is an averaging period index (i = 0 ... 6).

When the median replacement process is performed at the end of the hangover period (first comfort noise updating process), the immediately preceding 6
One frame (i = 1 ... 6 within the hangover period)
Frames L) of all LSF parameter vectors f
(N-i) has quantized values, but the LSF parameter vector f (n) of the newest frame n has unquantized values. In the subsequent comfort noise update, the LSF parameter vector of the comfort noise averaging period of the frame that overlaps the hangover period has a quantized value, but the parameter vector of the newest frame of the comfort noise averaging period is It has an unquantized value. If the period of the newest 7 frames does not overlap with the hangover period, then LSF
The parameter median replacement process is performed using only the non-quantized parameter values.

The average LSF parameter vector f ^mean in frame n is calculated using the following equation.

[0151]

[Equation 20]

Here, f ′ (n−i) is the newest 7 frames (i = 0.multidot.i) after the median replacement process.
... 6) is one LSF parameter vector, i is an averaging period index, and n is a frame index.

The average LSF parameter vector f ^{mean for} frame n is preferably the same quantization table that the speech coder uses to quantize the non-averaged LSF parameter vector in the normal speech coding mode. Is quantized using. However, the quantization algorithm used is modified to be suitable for quantization of comfort noise. LSF about to quantize
The prediction residual is obtained according to the following formula.

[0154]

[Equation 21]

[0155]

[Outside 5]

[0156]

[Outside 6]

[0157]

[Equation 22]

[0158]

[Outside 7]

[0159]

[Outside 8]

Further, the random excitation signal gain for each subframe is calculated as follows based on the LP residual signal energy of the subframe.

[0161]

[Equation 23]

Here, g _cn (j) is the calculated value of the random excitation signal gain of subframe j, r (l) is the l-th LP residual signal sample of subframe j, and l is the sample index (l = 0 ... 39). In equation (10), a scaling factor of 1.286 is used to match the comfort noise level with the background noise level encoded by the speech codec. However, the present invention is not limited to this particular value of the scaling factor.

Since the sub-frame excitation signal (pseudo noise) during the comfort noise generation period includes 10 samples having a non-zero value between -1 and +1, the calculated value of the energy of the LP residual signal is calculated as follows. By dividing by 10, the energy per random excitation signal pulse can be determined.
The calculated values of the random excitation signal gains were averaged and SP = “0” when updated CN parameters were needed.
Is updated in the first subframe of each frame n marked with.

[0164]

[Equation 24]

Here, g _cn (n) (1) is the calculated value of the random excitation signal gain in the first subframe of frame n, and g _cn (n−i) (j) is the past one frame (i =
The calculated value of the random excitation signal gain in subframe j of 1 ... 6). Since the random excitation signal gain of only the first subframe of the current frame is used in the averaging process, once the processing of the first subframe of the current frame is finished, an updated set of comfort noise parameters is set. It becomes possible to send.

[0166]

[Outside 9]

In calculating the RESC parameters, LP
Since the residual r (n) is slightly deviated from the flat spectrum characteristic, if the comfort noise is combined with the random excitation signal having the flat spectrum on the receiving side, the comfort noise quality is degraded (background noise). Difference between the noise and the comfort noise). To make the spectral characteristics better match, in the comfort noise averaging period, L
A second-order LP analysis is further performed on the P residual signal, and the obtained average LP coefficient is included in the comfort noise parameter message and transmitted to the receiving side. Let it happen. This method is called random excitation spectrum control (RESC). The LP coefficient obtained by this method is represented by the RESC parameter Λ.

The LP residual signal r (n) of each subframe of the frame is concatenated so that the autocorrelation r _res (k) (k = 0 ... 2) of the LP residual signal of the 20 ms frame is as follows. It is calculated by the formula.

[0169]

[Equation 25]

After calculating the autocorrelation using the above equation,
The obtained autocorrelation is normalized to obtain the normalized autocorrelation r ′.
_{Find res} (k).

Processing the first subframe of the current frame by averaging the newest frame of the comfort noise averaging period with the autocorrelation determined from only the first subframe. When is finished, it is possible to prepare updated comfort noise parameters to be sent next.

When it is necessary to update the comfort noise parameter, the calculated value of the normalized autocorrelation is calculated by using the following equation, and the first value of each frame n to which SP = "0" is added is calculated. Update in subset.

[0173]

[Equation 26]

Where r ′ _res (n) (1) is the normalized autocorrelation in the first subframe of frame n,
r ′ _res (n−i) is a past frame (i = 1 ...
6) Normalized autocorrelation of 1 frame, n is a frame index.

[0175]

[Outside 10]

[0176]

[Outside 11]

[0177]

[Outside 12]

The comfort noise coding algorithm produces 38-bit data as shown in Table 2 for each comfort noise parameter message. In addition, these 38
The bit data is represented by the vector cn [0 ... 37]. The comfort noise bit data cn [0 ... 37] is
It is supplied to the FACCH channel encoder in the order shown in Table 2 (no special bit alignment is done).

[0179]

[Table 2]

Regardless of its content (voice, comfort noise parameter message, other FACCH message, or no signal), the wireless receiver of the base station 30 has various processing functions depending on the three flags received. The designated traffic frames are sequentially supplied to the receiving side discontinuous transmission handler. The above three flags are a bad voice frame indication (BFI) flag, a bad comfort noise parameter indication (BFI_CN) flag,
Comfort Noise Update Flags (CNU), which are described below, such as in Table 3. These flags are used to distinguish traffic frames by their purpose. This partition allows the receiving discontinuous transmission handler to easily know how to process the received frame, as shown in Table 3.

[0181]

[Table 3]

The binary BFI and BFI_CN flags allow the traffic frame to contain meaningful information bits (BFI flag = "0" and B
FI_CN flag = "1") or does not contain a meaningful information bit (BFI flag = "1")
Also, it is indicated whether the BFI_CN flag = “1” or the BFI flag = “0” and the BFI_CN flag = “0”). In the present invention, the FACCH frame is not considered to contain meaningful bits unless it contains a comfort noise parameter message, so the BFI SP flag = "1" and B
The FI CN flag = “1”.

The binary CNU flag marks CNU = "1" for traffic frames aligned according to the transmission channel quality information sent by the FACCH.

The receiving side discontinuous transmission handler is involved in the entire receiving side discontinuous transmission operation. The discontinuous transmission operation on the receiving side is performed as follows. If the speech frame is determined to be a good speech frame, the discontinuous transmission handler feeds the speech frame directly to the speech decoder.
If it is determined that the voice frame or the comfort noise parameter message is missing, the replacement and muting process is performed. When a valid comfort noise parameter message frame is detected, it is used to detect the occurrence of comfort noise by the next comfort noise parameter message (CNU = "1") or the next good speech frame. Continue until done. During the above processing, the receiving side discontinuous transmission handler ignores all the frames supplied from the wireless receiver if they are unusable. In addition, the following operations are possible as options. That is, if the first missing comfort noise parameter message is detected, that parameter is replaced with the parameter of the last detected valid comfort noise parameter message and the normal Perform processing. If the second missing comfort noise parameter message is detected, the muting process is performed.

In the LP parameter averaging process and the decoding process, when the decoder receives a voice frame, the LP parameters of the latest six voice frames are stored in the memory. The decoder counts the number of frames received after the encoder last updated the comfort noise parameter and provided it to the wireless transmitter. Based on this count, the decoder determines whether there is a hangover period at the end of the voice burst (when the first comfort noise parameter message after the voice burst is received, the comfort noise last is received). If at least 30 frames have elapsed since the parameters were updated, it may be considered that there was a hangover period at the end of the voice burst).

[0186]

[Outside 13]

The averaging process for obtaining the above-mentioned reference parameters is performed as follows.

When a voice frame is received, the LSF parameters are decoded and the result is stored in memory. If the comfort noise parameter message is received and it is determined that there is a hangover period at the end of the voice burst, the stored LSF parameters are as follows, as is done in the voice encoder. Average out.

[0189]

[Equation 27]

[0190]

[Outside 14]

[0191]

[Outside 15]

[0192]

[Equation 28]

[0193]

[Outside 16]

In each subframe, a fixed codebook excitation signal vector containing four non-zero pulses of a conventional speech decoder is a random excitation containing 10 non-zero pulses during the interruption of speech speech. It is replaced with the signal vector. The pulse position of the random excitation signal and its sign are locally generated using a uniform distribution pseudo-random number. The excitation signal pulse in the random excitation vector can take +1 and -1 as its value. The random excitation signal generation algorithm is executed according to the pseudo code shown below.

For (i = 0; i <40; i ++) code (i) = 0; for (i = 0; i <10; i ++) {j = random (4); idx = j * 10 + i; if ( random (2) == 1) code (idx) = 1; else code (idx) =-1;} where code [0 ... 39] is a fixed codebook excitation signal buffer and random ( k)
Is a pseudorandom integer uniformly distributed in the range of [0 ... k-1].

The received RESC parameter index is decoded and the received RESC parameter λ (i) (i =
1, 2) are obtained. The generated random excitation signal is filtered by a RESC synthesis filter having the characteristics given by:

[0197]

[Equation 29]

Preferably, the RESC synthesis filter is realized by using a lattice type filter. After passing through the RESC synthesis filter, the random excitation signal is subjected to scaling processing and then filtered by the LP synthesis filter.

The generation of comfort noise is performed using a method in which the speech decoding algorithm is modified as follows. That is, the fixed codebook gain value is replaced with the random excitation gain value obtained by the comfort noise parameter message, and the fixed codebook excitation signal is replaced with the locally generated random excitation signal as described above. . The random excitation signal is also filtered by the RESC synthesis filter, as described above. The value of the adaptive codebook gain of each subframe is set to 0. The pitch delay value of each subframe is set to 60, for example. The parameters received by the comfort noise parameter message are used as the LP filter parameters. The predictor memory used in conventional LP parameter quantization and fixed codebook gain quantization algorithms is reset when SP flag = "0" and when the speech is restarted again, the quantizer has its initial state. The operation can be started again from. The comfort noise is synthesized by the speech decoder performing the normal operation using the above parameters. Each time a valid comfort noise parameter message is received, the comfort noise parameters (random excitation signal gain, RESC parameters, LP filter parameters) are updated as described above. In order to make the change smoother in the comfort noise update, the above parameters are interpolated during the comfort noise update period. An unusable frame received when the receiving discontinuous transmission handler is experiencing comfort noise is called a missing comfort noise parameter message. If such a frame is received, a comfort noise message is inferred (comfort noise update flag CNU = "1").

If such a drop occurs in a single comfort noise parameter message, then that parameter is replaced with the parameter of the previous valid comfort noise parameter message, and for this parameter the above Processing is performed. Furthermore, when the comfort noise parameter message is lost, the comfort noise is muted, the output level is gradually reduced (-3 dB / frame), and finally the output of the decoder is zero. The muting is performed by reducing the random excitation signal gain to 0 at a constant rate of -3 dB per frame. This value is maintained as it is when a further dropout occurs.

As described above, specific values such as the frame length and the number of frames are assumed, and the specific message type (F
The present invention has been described by taking the ACCH) as an example. And various changes can be made according to requirements. In addition, in the above, FIG.
Although the present invention has been described with reference to circuit block diagrams such as (b), 3 (a), 3 (b), 4, 5, 10 and the like, some of these circuit blocks may be digital cellular telephones 1.
It can also be implemented using a suitably programmed digital data processor (eg, controller 18 of FIG. 12 is an example) included as part of a zero. As one of the examples, for example, the selector 307,
Although 319 and 410 are shown as switches in FIGS. 4 and 5, these selectors can also be completely implemented in software.

Further, depending on the system, when comfort noise is generated, the comfort noise parameter message (SID frame) used for transmitting the RESC parameter from the transmission side to the reception side may include a spare bit. is there. In such cases, the RESC filter of the present invention can be replaced by a synthesis filter with fixed coefficients. The fixed filter coefficients are optimized such that the frequency response characteristic of the synthesis filter is the average frequency response of a conventional RESC filter whose coefficients are determined by the value transmitted. Alternatively, the filter coefficients may be selected so that the filter has a frequency response characteristic suitable for obtaining a perceptually preferable comfort noise.

Although the present invention has been described above with reference to specific embodiments, the present invention is not limited to these, and various modifications can be made without departing from the scope and spirit of the present invention. It will be obvious.

[Brief description of drawings]

FIG. 1a is a block diagram illustrating a conventional circuit for generating a comfort noise parameter at a transmitting side.

FIG. 1b is a block diagram showing a decoder conventionally used to generate comfort noise at the receiving side.

1c is a diagram representing the spectrum of a signal in some parts of the conventional decoder of FIG. 1 (b).

FIG. 1d shows the averaging block of FIG. 1 (a) in more detail.

FIG. 2a is a block diagram showing a circuit for generating comfort noise parameters at the transmitter side according to the present invention.

FIG. 2b is a block diagram showing a circuit for generating comfort noise at the receiving side according to the present invention.

FIG. 2c is a diagram showing a spectrum in the decoder of FIG. 2 (b).

FIG. 3a is a block diagram showing a circuit for generating comfort noise parameters at a transmitter side according to a second embodiment of the invention.

FIG. 3b is a block diagram showing a decoder at a receiving side according to a second embodiment of the present invention.

FIG. 4 is a block diagram showing a circuit for calculating comfort noise parameters at a transmitter side of a discontinuous transmission digital communication system according to an embodiment of the present invention.

FIG. 5 is a block diagram showing a circuit for calculating comfort noise parameters at a transmitter side of a discontinuous transmission digital communication system according to another embodiment of the present invention.

FIG. 6 is a block diagram of a speech encoder according to the prior art.

7 is a timing diagram showing an example of the output of the speech encoder according to the prior art of FIG.

8 is a timing diagram showing another example of the output of the speech coder according to the related art of FIG.

FIG. 9 is a block diagram showing a speech decoder according to the related art.

FIG. 10 is a block diagram showing a speech decoder according to an embodiment of the present invention.

FIG. 11a shows a typical frequency response of a RESC filter.

FIG. 11b is a diagram showing a typical frequency response of a RESC filter.

FIG. 11c shows a typical frequency response of a RESC filter.

FIG. 11d shows a typical frequency response of a RESC filter.

FIG. 11e shows a typical frequency response of a RESC filter.

FIG. 11f shows a typical frequency response of a RESC filter.

FIG. 11g shows a typical frequency response of a RESC filter.

FIG. 12 shows a mobile station suitable for implementing the present invention.

FIG. 13 is a diagram of a mobile terminal for wirelessly communicating with a ground station, suitable for implementing the present invention.

FIG. 14 is a diagram showing a normal hangover procedure, in which N
elapsed represents the number of frames after the comfort noise (CN) parameter was last updated, and Nelapsed causes hangover only when 24 or more.

FIG. 15 is a timing diagram showing processing of a short voice burst when Nelapsed is smaller than 24.

[Explanation of symbols]

100 audio signals 101 Linear Predictive Coding (LPC) Analysis Block 103 Inverse filter 104 residual signal 105 pump gain calculation block 107, 108 averaging block 110 Random Excitation Signal (RE) Generator 112 Speech synthesis filter 200 Random excitation linear predictive coding analysis block 202 Random excitation control inverse filter

─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Besar Oppira Finland, 33720 Tampere, Biruppuruppoluk 20 Be (72) Inventor Yani Rotora Pukkyra Finland, Tampere, Fin-33820, Mannicon Kats 3 Ar 22 ( 56) References JP-A-8-195700 (JP, A) International Publication 96/32817 (WO, A1) International Publication 96/32823 (WO, A1) (58) Fields investigated (Int.Cl. ⁷ , DB name) ) H04B 1/10 G10L 19/08

Claims (25)

(57) [Claims] 1. A method for generating comfort noise (CN) in a digital mobile terminal using a discontinuous transmission method, which comprises buffering a set of speech coding parameters when speech is interrupted, In the averaging period, the step of replacing the speech coding parameters that do not well represent the background noise in the set with speech coding parameters that well represent the background noise, and the average of the set of speech coding parameters And a step of seeking. 2. The step of performing the substitution included in the step includes the step of measuring the distances of the speech coding parameters from each other in the individual frames in the averaging period, and the other step in the averaging period. Identifying the speech coding parameter with the greatest distance from the parameter, and if the distance is greater than a predetermined threshold, the identified speech coding parameter is averaged with the averaged value. the method of claim 1 further comprising the step of Toto replacing speech coding parameters measured distance to the other speech coding parameters is the minimum in the period. 3. The replacing step included in the above step includes the steps of measuring the distances of speech coding parameters from each other in the individual frames within the averaging period, and other steps within the averaging period. Identifying the speech coding parameter with the greatest distance from the parameter, and if the distance is greater than a predetermined threshold, the identified speech coding parameter with the median value the method of claim 1 further comprising the step of Toto replacing the speech coding parameter having. 4. The step of obtaining the average includes the average excitation signal gain gmean and the average short-term spectrum coefficient fmean.
(I) The method of claim 1 further comprising calculating a. 5. The step of permuting included in the step of forming a set of buffered excitation signal gain values during an averaging period; and a step of forming the set of buffered excitation signal gain values. Among the excitation signal gain values in which the difference between the aligning step and the median value of the set of excitation signal gain values exceeds a predetermined threshold value, the L excitation signal gain values having the largest difference from the median value. Replacing the median with the median.
The method described in 1 . 6. The length N of the averaging period is an odd number,
The method of claim 5 , wherein the median of the aligned set of excitation signal gain values is the value of the ((N + 1) / 2) th element of the set. 7. A buffered line spectrum pair (LSP) coefficient f (k) (k =
1, ..., M), and the LSP coefficient fi of the i-th frame in the averaging period
The LS of the j-th frame in the same averaging period of (k)
The method of claim 1, further comprising the step of determining the spectral distance to P coefficient fj (k). 8. The step of obtaining the spectral distance is performed by the following equation: (Where M is the order of the LPC model, fi (k) is the k-th LSP of the i-th frame in the averaging period.
Method according to claim 7, carried out according to (parameters). 9. In the averaging period of length N, all other frames (j = 1, ..., N, j) of the LSP coefficient fi (k) of the i-th frame in the same averaging period.
8. The method according to claim 7 , further comprising the step of determining the spectral distance .DELTA.Si for the LSP coefficient of .noteq.i). 10. The step of determining the spectral distance is performed by the following equation: According to the sum of the spectral distances ΔRij for all i = 1,
..., The method of claim 9 performed by determining for N. 11. A step of, after obtaining a spectral distance ΔSi for each LSP vector fi in the averaging period, aligning the spectral distances according to the value, the averaging period (i = 1, 2, ..., N). The vector fi having the smallest distance ΔSi in is defined as the median vector fmed of the averaging period, and the distance is represented by ΔSmed, and the median value of P (0 ≦ P ≦ N−1) LSP vectors fi is The method of claim 9 , further comprising the step of replacing with the median vector fmed. 12. The method of claim 2 , wherein the identifying and replacing steps are performed independently for each of the excitation signal gain value g and the line spectrum pair (LSP) vector fi. 13. The identifying step and the step of replacing the excitation signal gain value g and the line spectral pair (LSP) vectors simultaneously The method of claim 2, wherein the performed on both fi. 14. When it is determined that the speech coding parameter of a frame should be replaced by the median value of the parameters, both the excitation signal gain value g and the LSP vector fi of the frame are set to the median parameter. 14. The method of claim 13 including the step of substituting with parameters corresponding to each of the frames containing. 15. The distance ΔTij between the i-th frame and the j-th frame in the averaging period is calculated by the following formula: (Where M is the order of the LPC model, fi (k) is the kth LSP parameter of the ith frame in the averaging period, and gi is the excitation signal gain parameter of the ith frame). 14. The method according to 14 . 16. Distance ΔSi for all other frames j = 1, ..., N (j ≠ i) of the speech coding parameters of frame i within the averaging period of length N.
, For all i = 1, ..., N, the following equation: 16. The method of claim 15 , further comprising the step of determining according to. 17. A step of arranging the distances in order of the values after obtaining the distance ΔSi for each frame within the averaging period, and the averaging period (i = 1, 2, ..., N). 17. The method of claim 16 , wherein the frame with the smallest distance ΔSi in) is the central frame of the averaging period, the distance is ΔSmed, and its speech coding parameters are denoted by gmed and fmed. 18. An averaging period (i = 1, 2, ...,
The median value substitution process for the speech coding parameters of the frames in N) is performed by converting the parameters gi and fi of L (0≤L≤N-1) frames into the parameter gme of the central frame
18. The method according to claim 17, which is carried out by substituting d and fmed. 19. The difference between the individual distance and the median distance is calculated by:
18. The method of claim 17 , wherein the individual distances are represented by the quotient ΔSi / ΔSmed divided by the median distance. 20. The difference between the individual distances and the median distances,
A method according to claim 11 , characterized by the quotient ΔSi / ΔSmed divided by the median distance of the individual distances. 21. An apparatus for generating comfort noise in a system including a digital mobile terminal which performs discontinuous communication with a communication network, wherein a set of speech coding parameters is used for averaging period when speech is interrupted. Buffering and replacing the speech coding parameters of the set that do not represent well the background noise with speech coding parameters that represent well the background noise, the data processing means comprising: An apparatus for obtaining an average value of a set of coding parameters and transmitting the obtained set of average values of the voice coding parameters to the communication network. 22. The data processing means replaces the speech coding parameters of the set by aligning the set of speech coding parameters with each other of the speech coding parameters between individual frames within an averaging period. Measure the distance and identify the speech coding parameter that has the greatest distance from other parameters within the averaging period, and if the maximum distance exceeds a predetermined threshold, then 22. The apparatus of claim 21 , wherein speech coding parameters identified as having the greatest distance are replaced with speech coding parameters measured to have a minimum distance to other speech coding parameters within the averaged distance. . 23. The data processing means replaces the speech coding parameters of the set by arranging the set of speech coding parameters such that the speech coding parameters between individual frames within an averaging period are mutually equal. Measure the distance and identify the speech coding parameter that has the greatest distance from other parameters within the averaging period, and if the maximum distance exceeds a predetermined threshold, then 22. The apparatus of claim 21 , wherein speech coding parameters identified as having a maximum distance are replaced with speech coding parameters having a median value. 24. The data processing means comprises an excitation signal gain value g of a speech coding parameter and a line spectrum pair (L).
The apparatus according to claim 21, wherein the SP) vector fi is identified and replaced independently. 25. The data processing means comprises an excitation signal gain value g of a voice coding parameter and a line spectrum pair (L
22. The apparatus of claim 21 , wherein both SP) vectors fi are identified and permuted simultaneously.