1. Field of the Invention
The present invention relates to a voice data processing method and device, and in particular to a voice data processing method and device for a VoIP communication system which mounts thereon the voice codec G.711 Appendix I with a packet loss compensating function and transmits voice data over an IP network.
2. Description of the Related Art
Also, the packet loss compensator 3 includes a pitch detector 30, which is composed of a coarse search processor 31 and a fine search processor 32. In this packet loss compensator 3, the pitch detector 30 sequentially executes coarse search (at step S100) and fine search (at step S200) as shown in
The generated compensating data C is weighted at the packet loss time to achieve smoothness. When packet losses sequentially occur, the compensating data is gradually attenuated.
Operations of
Firstly, by a packet loss flag G provided from an upper system, the packet loss compensator 3 recognizes a normal mode/packet loss mode (normal mode or packet loss mode). It is assumed in this description that “H” indicates the normal mode, while “L” indicates the packet loss mode.
The decoder 1 always performs decoding for every frame (10 ms), so that data decoded by the decoder 1 is stored in the history buffer 2 for every 80 samples (10 ms), as shown in
At the timing of a frame F6 where a packet loss has occurred, the packet loss compensator 3 executes packet loss compensation by using decoded data of the normal frames F1-F5 (for 390 samples) stored in the history buffer 2, and detects a pitch P to generate the compensating data C during the packet loss.
The hatched portions during the packet loss in
Namely, this pitch detection is performed, as shown in
An autocorrelation between a signal delayed by the maximum pitch (120 samples) from the reference signal L and a signal delayed by the minimum pitch (40 samples), and the cross-correlation between each of the delay signals R and the reference signal L are calculated, in which the calculation of the normalized cross-correlation is given by the following equation:
Normalized cross-correlation=cross-correlation/√{square root over (autocorrelation)} (1)
In order to reduce a pitch detection load in the pitch detector 30, the processing is separated into main two stages. Firstly, as shown in
Firstly, the reference signal L and the delay signal R are set (at step S1). An autocorrelation “energy” and a cross-correlation “corr” are calculated (at step S2_2) at the rate of once per two samplings (at step S2_3), and the product-sum calculation is respectively performed 80 times (for 160 samples) (at step S2_4) (at step S2: steps S2_1-S2_4).
From the calculated autocorrelation value “energy” and the cross-correlation value “corr”, based on the above-mentioned equation (1), a normalized cross-correlation value “corr” is obtained (at step S3). This value is set to a cross-correlation initial value “bestcorr” (at step S4). Also, the delay data value “bestmatch” is initialized to “0” (at step S4).
In the loop of the subsequent normalized cross-correlation calculation (j<PITCH_DIFF: at step S50), the reference signal L and the delay signal R are also used. While the delay signal R is shifted by every sample, the autocorrelation calculation (at step S6) and the cross-correlation calculation (at steps S7 and S8) are performed to obtain the normalized cross-correlation (at step S9). By 80 samples (at step S120), the peak value “bestcorr” of the normalized cross-correlation calculation value “corr” and the delay data value “bestmatch” at this point (j) are obtained (at steps S10 and S11).
In this case, the calculation is performed by the frequency of a difference PITCHDIFF between a Pmax (120) and a Pmin (40), that is the frequency (80 times) of loops required (at steps S14 and S120).
As another prior art technology, an error concealment apparatus and method are mentioned, by which a plurality of algorithms for concealing errors are prepared in order to enable various error concealment technologies to be dynamically selected and applied, the error concealment is performed by using any one of the algorithms, an algorithm to be selected is determined by a selection signal, and the selection signal is made based on various parameters indicating throughput of a computer and a characteristic of a voice signal (see e.g. patent document 1).
Also, as still another prior art technology, a pitch detection method and device in a packet loss compensation are mentioned, by which a correlation calculation is always performed by a pitch buffer, a correlation calculating portion, and a correlation buffer, a pitch is detected, and interpolating data is prepared for loss of a subsequent frame. When a frame loss occurs, lost voice data is immediately interpolated by interpolation processing for input data (see e.g. patent document 2).
[Non-patent document 1] ITU-T TELECOMMUNICATION STANDARDIZATION SECTOR OF ITU G.711
[Non-patent document 2] ITU-T TELECOMMUNICATION STANDARDIZATION SECTOR OF ITU G.711 Appendix I (09/99)
[Patent Document 1] Japanese Patent Application Laid-open No.2003-218932
[Patent Document 2] Japanese Patent Application Laid-open No.2004-239930
The whole processing amount in the above-mentioned packet loss compensator 3 is about 39 MHz. The pitch detection occupies 29 MHz, the 75% of the whole processing amount, in which especially only the coarse search processor occupies 23 MHz, a high rate of about 60% of the whole pitch detection amount.
This is affected by the fact that the product-sum calculation is performed 81 times, the product-difference calculation is performed once, and the division calculation is performed once in a single loop, as shown in
Since the processing amount is only about 1 MHz in the normal mode where no packet loss occurs, as for the throughput of G.711 Appendix I type decoder, there has been a possibility of affecting the operation during the packet loss depending on a system incorporated therein to cause a malfunction or an operation halt.
In addition, when such a packet loss occurs immediately after signals decoded have continued at a silent level, the compensating data should be inevitably silent. However, in the prior art system, there has been a problem of unnecessary packet loss compensation being performed even when a signal decoded continues at a silent level.
It is accordingly an object of the present invention to provide a voice data processing method and device detecting a pitch based on history data during a packet loss and generating compensating data thereof, whereby a calculation amount in a packet loss mode is reduced and unnecessary packet loss compensation is avoided when a signal is a silence signal.
In order to achieve the above-mentioned object, a voice data processing method (device) according to the present invention comprises: a first step (means), in a normal mode, decoding input signal data, repeating a calculation in coarse search used for a pitch detection by a predetermined frequency of loops within a required frequency of loops, based on history decode data, and holding a peak value of a normalized cross-correlation obtained by the calculation and a delay data value corresponding thereto; and a second step (means), in a packet loss mode, executing the pitch detection by repeating a calculation of a normalized cross-correlation in the coarse search by a remaining required frequency of loops, by using the peak value of the normalized cross-correlation and the delay data value, thereby generating compensating data.
Namely, in a pitch detection during the packet loss, both of coarse search and fine search have been conventionally executed (at steps S100 and S200 of
This is schematically shown by a flowchart in
A peak value bestcorr_tmp of the normalized cross-correlation within the coarse search obtained by the calculation, and a delay data value bestmatch_tmp at this time are held in e.g. a buffer (not shown) as variables (at step S102). In the packet loss mode, with the variables (at step S103), the remaining coarse search is performed (at step S104), and then the processing is taken over to the fine search (at step S200).
As a result, by separating the processing into the normal mode, the processing amount in the packet loss mode can be reduced. Also, since the frequency of loops in the coarse search given in the normal mode can be variably set by a user or the like, the processing amount of the normal mode and the loss mode can be preliminarily adjusted to a request of the user.
Also, in the present invention, the first and the second step (means) respectively may include a third and a fourth step (means) determining whether or not the input signal data is silence signal data, and of invalidating the coarse search when the input signal data is determined to be the silence signal data.
Namely, since the processing amount in the pitch detection does not depend on a sound source inputted, packet loss compensation in the packet loss mode and a level determination of a signal inputted to the coarse search processor are added, thereby suppressing the processing amount in a case where a silent level continues in a signal to be decoded.
Furthermore, in the present invention, the first and the second step (means) respectively may include a fifth and a sixth step (means) invalidating and validating the third and the fourth step (means) respectively when the predetermined frequency of loops is a first value corresponding to a suppression request of a coarse search amount in the normal mode, and of contrarily validating and invalidating the third and the fourth step (means) when the predetermined frequency of loops is a second value corresponding to a suppression request of a coarse search amount in the packet loss mode.
Namely, when the suppression of the coarse search amount in the normal mode is desired by a user's request or the like, or when the same suppression in the packet loss mode is desired, a silence determination operation can be invalidated or disabled by using a first and a second predetermined frequencies of the loops, thereby enabling an unnecessary silence determination to be avoided.
As described above, the following effects can be obtained in the present invention:
The above and other objects and advantages of the invention will be apparent upon consideration of the following detailed description, taken in conjunction with the accompanying drawings, in which the reference numerals refer to like parts throughout and in which:
The flow of
In this embodiment, the frequency of loops at steps S5-S12 is changed by using a variable “x” newly shown in
Therefore, for the coarse search in the packet loss mode, as shown in
After the coarse search ends, the fine search is performed (at step S200), finishing the pitch detection.
It is supposed that there is a request from e.g. a system side for making a processing amount in the packet loss mode and a processing amount in the normal mode fixed. In this case, the predetermined frequency “x” shown in
α:PROCESSING AMOUND OF STEPS S1-S5 & S102 (ABOUT 1 MHz)
*DON'T CARE
In this case, the frequency of the normalized cross-correlation processing loops assumes PITCHDIFF−20=80−20=60 in the coarse search in the normal mode. Since the frequency of loops is added by 2 (at step S12), an actual frequency of loops of the normalized cross-correlation processing assumes 60/2=30 times. After the loop processing ends, the intermediate results of the normalized cross-correlation peak value “bestcorr” and the delay data value “bestmatch” are respectively held in the buffers bestcorr_tmp and bestmatch_tmp (at step S102).
Paying attention to the frequency of the normalized cross-correlation calculations in the coarse search of the normal mode, the frequency assumes 30×(product−sum of 81 times+product−difference of 1 time+division of 1 time)=product−sum of 2430 times+product-difference of 30 times+division of 30 times=2490 times. Since this processing is not performed in the normal mode of the prior art method, the frequency is increased by 2490×8 KHz (sampling frequency) cycle, that is 19.92 MHz.
Hereinafter, the processing in the packet loss mode will be described. In the above-mentioned normal mode, values held in the buffers bestcorr_tmp and bestmatch_tmp are respectively initialized to the bestcorr and the bestmatch (at step S103). Since the frequency of loops in the normalized cross-correlation is the remaining frequency “x”, “20” is set. Since the frequency of loops is added by 2, similar to the frequency of loops in the above-mentioned normal mode (at step S120), the frequency of loops assumes 10 times.
Paying attention to the frequency of the calculations in the normalized cross-correlation in the coarse search of the packet loss mode, the frequencies of the calculations according to the present invention and the prior art example are as follows:
Thus, the present invention can achieve the effect of 75% of cycle reduction (−19.92 MHz) compared with the prior art example, so that the processing amount in the packet loss mode assumes 39 MHz−19.92 MHz=19.08 MHz.
As a result, as shown in Table 1,
It is now supposed that the present invention is mounted on a system where numerous calls are one-way calls such as voice guidance. In such a case, a silence part of data largely occupies input data, so that the processing is also performed to the silence data. In order to prevent this, a mechanism of performing a silence determination for the silence data and bypassing the coarse search and the packet loss compensation is provided, thereby enabling the processing to be efficiently performed.
In the history buffer 2, a signal decoded by the decoder 1 is stored, regardless of presence/absence of the packet loss. The packet loss compensator 3 performs the pitch detection and the generation of the packet loss compensating data C or the like from the decode data stored in the history buffer 2. However, when a signal level for 390 samples (390×125 μs) of the size of the history buffer 2 is at a silence by adding the silence determining portion 8 of the signal level in front of the packet loss compensator 3, the packet loss compensation is not performed.
Also, in the coarse search in the normal mode, the pitch detection is performed from the signal stored in the history buffer 2. When the signal level for the 390 samples (390×125 μs) of the size of the history buffer is at a silence by adding the silence determining portion 7 of the signal level in front of the coarse search in the normal mode, the coarse search is not performed.
As mentioned above, in the presence of a request of suppressing a processing load as much as possible in the normal mode and in the system where numerous calls are one-way calls from the system side, “x”=“80” is rendered, as shown in Table 1, in order to suppress the processing amount of the normal mode as much as possible. Also, the processings of only steps S1-S5 and S102 in
However, by adding only the silence processors 7 and 8 as shown in
In the embodiment [3] of the present invention, silence determination executing portions 9 and 10 are respectively connected to the silence determining portions 7 and 8 added in the embodiment [2], and a predetermined frequency “x” of loops is provided to the silence determination executing portions 9 and 10, thereby further determining whether or not the silence determination should be performed. Therefore, the predetermined frequency “x” of loops includes the first value x1 and the second value x2.
In operation, when a packet loss flag G designates the normal mode, the data decoded by the decoder 1 is stored in the history buffer 2. Based on the data stored in the history buffer 2, the silence determining portion 7 performs a silence determination (detection), and validates or invalidates the coarse search processor 6. However, before the validation or invalidation, whether or not the silence determination itself should be performed is determined by the silence determination executing portion 9.
In the silence determination executing portion 9, e.g. the frequency “x” of loops during the pitch detection provided from a user is inputted as a parameter. In the presence of the request of suppressing the processing amount in the normal mode, as shown in Table 1, the frequency “x” of loops is set with “80” as the first value x1. In case of x1=80, the silence determination executing portion 9 makes the silence determination portion 7 do a through-operation, so that the decode data of the history buffer 2 is switched over so as to be provided as it is to the coarse search processor 6. Thus, the operation of the silence determining portion 6 is not executed, thereby enabling the processing amount to be suppressed to α.
Contrarily, in the presence of a request of suppressing the processing amount (pitch detection amount including fine search amount in this case) in the packet loss mode, from the value shown in Table 1 in the same way as the above, the frequency “x” of loops is set with “0” as the second value x2. In case of x2=0 in the silence execution determining portion 10, it is devised that the silence execution determining portion 10 makes the silence determining portion 8 do a through operation, and that the decode data of the history buffer 2 is switched over so as to be transmitted as it is to the packet loss compensator 3. Thus, steps S6-S11 and S120 of
Namely, when the processing amount of the silence determining portion 7 is larger than that of the packet loss compensator 3 in the packet loss mode, and also when data is voiced data, the data is validated or enabled. When the data is silence data for example, the data is passed through the silence determining portion 8 as it is, so that the packet loss compensation is performed without fail. In such a case, the processing amount assumes 13.4 MHz also from Table 1. However, when the data is passed through the silence determining portion 8 (x2=0) as it is, the packet loss compensation is bypassed with the determination result (silence). Therefore, the processing amount assumes only the processing amount of the silence determining portion 8.
It is to be noted that the present invention is not limited by the above-mentioned embodiments, and it is obvious that various modifications may be made by one skilled in the art based on the recitation of the claims.
Number | Date | Country | Kind |
---|---|---|---|
2005-304871 | Oct 2005 | JP | national |