JP2845960B2 - Voice packet interpolation method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a packet switch, and more particularly, to a voice packet interpolation method for a lost voice packet in a voice packet switch.

(Prior Art) In voice packet communication, real-time communication is an important factor of communication quality. For this reason, if a voice packet does not arrive even after a predetermined time limit, the voice packet is treated as being lost, and real-time communication is guaranteed. Loss of voice packets also occurs, for example, when a load that exceeds the processing capacity of the exchange occurs.
When such a voice packet is lost, the voice of that part is not reproduced, and the communication quality is degraded. Therefore, when a voice packet is lost, it is desirable to interpolate the voice packet in the missing part, assuming the lost voice packet, to prevent deterioration of communication quality.

For example, DIGoodman, et al, “Waveform Substitution Tech”
niques for Recovering Missing Speech Segments in P
acket Voice Communications ", Proceedings ICASSP 86,
Pages 105 to 108 (1986). In this method, input voice packets are sequentially stored in a fixed amount in a buffer, and the stored voice packets are used as a search window to find a suitable voice packet to interpolate a lost voice packet. Is what you do. That is, if a voice packet is lost,
The waveform of the immediately preceding audio packet is used as a template, and the audio waveform that is most similar to this template is found from the search window of the buffer. When a voice waveform closest to the template is detected, the voice waveform following the voice waveform is interpolated as a missing voice packet.

In this prior art, a standardized difference equation shown in Expression (10) is used to detect a speech waveform that is closest to the template, and the waveform that minimizes this solution is the waveform that is closest to the template. That is, when the lack of the audio waveform is detected, NK + 1 D (n) of D (1) to D (NK + 1) are obtained from the equation (10), and the minimum D is obtained from these.
Select (n). Then, the voice waveform following the waveform having the smallest calculated value is interpolated as a missing voice packet.

(N = 1, 2,..., N−K + 1) x _t (k): each sample value in the template K: the number of samples in the template y (n + k−1): each sample value in the search window N: sample in the search window However, in such a conventional technique, since the waveform of one packet immediately after the waveform closest to the template in the search window is directly interpolated into the missing part, the leading sample of the missing packet is used. The interval between the search window and the sample on the side of the most missing portion of the search window must be separated by at least one packet. For this reason, when the audio waveform changes in the area of one packet between the search window and the missing packet, no waveform similar to the template exists in the search window, and the waveform detected by the matching is inappropriate. As a result, the waveform to be interpolated into the missing part is likely to be inappropriate. The probability that this interpolated waveform is inappropriate increases as the packet length increases,
This causes quality degradation of the reproduced sound subjected to the interpolation processing.

The present invention solves such disadvantages of the prior art, and in the area of one packet between the search window and the missing packet,
When a change in the audio waveform occurs, there is no waveform approximating the template in the search window, and the processing to reduce the probability of audio quality deterioration due to the inappropriate waveform to be interpolated in the missing part is performed in real time. It is an object of the present invention to provide a voice packet interpolation processing method which is executed while maintaining the above.

(Means for Solving the Problems) In order to solve the above-mentioned problems, the present invention provides a voice packet interpolation method for interpolating a missing voice packet into a missing portion by a signal of a voice packet input before the loss. Each time a corresponding sample value is input, the calculation of the optimal matching waveform with the signal of the voice packet input before the loss is performed by the non-standardized difference calculation process. The process of setting the time interval of the optimum matching waveform to the pitch period is executed regardless of whether or not the input packet is a missing packet. For this reason, at the time of input of the first sample of an arbitrary packet, the pitch period is always detected. If the input packet is a missing packet, a process of sequentially substituting the value of one sample before the pitch period of each sample of the missing packet is performed. repeat. If the detected pitch period is smaller than the number of samples of the packet, the sample value used for the interpolation once is repeatedly used.

In addition, the position of the sample farthest from the packet missing part in the search window area for searching for the optimal matching waveform with the template is determined from the maximum pitch cycle set by the average pitch cycle of the voice and the template length from the packet missing leading sample. The position of the sample closest to the packet missing part is set to the position separated by the minimum pitch period set by the average pitch period of the voice from the sample of the packet missing part, and the search window is set. Is set as close as possible to the missing portion so that the latest pitch cycle immediately before the missing portion can be detected.

(Operation) According to the present invention, every time a sample value corresponding to a template is input, detection of an optimal matching waveform portion with a signal of a voice packet input before the loss is performed by non-standardized difference calculation processing. Then, based on the calculation result of the non-standardized difference calculation processing, the processing of setting the time interval between the template and the optimal matching waveform to the pitch cycle is executed regardless of whether or not the input packet is a missing packet. For this reason, at the time of input of the first sample of an arbitrary packet, the pitch period is always detected. If the input packet is a missing packet, a process of sequentially substituting the value of one sample before the pitch period of each sample of the missing packet is performed. repeat. If the detected pitch period is smaller than the number of samples of the packet, the sample value used for the interpolation once is repeatedly used.

By repeatedly using the sample values, it is not necessary to separate the sample on the side of the most missing part of the search window from the first sample of the missing packet, and the latest pitch period immediately before the missing part can be detected.
Due to the change in the audio waveform, the probability that an appropriate interpolation waveform does not exist in the search window is reduced, and the audio quality after the interpolation processing is improved.

(Embodiment) Next, an embodiment of a voice packet interpolation system according to the present invention will be described in detail with reference to the accompanying drawings.

Referring to FIG. 1, there is shown an embodiment in which a voice packet interpolation system according to the present invention is applied to a PCM interpolation processing device provided on the packet receiving side of a voice packet switch. The interpolation processing device 1 is connected to an input terminal 101 for inputting a PCM signal of a voice packet and a control terminal 102 for notifying the loss of the voice packet. When the voice packet is lost, the portion is interpolated, and the interpolated PCM signal is converted. This is a device that outputs from the output terminal 103.

FIG. 2 shows various parameter values in this embodiment. As shown in the figure, in this embodiment, the voice packet is 64 bytes (8 ms), and the length of the template which is a basic signal for detecting a similar signal of a lost packet is 30 bytes (3.75 ms). The setting position of the search window, which is the search range for detecting the missing packet similar signal,
20 bytes (2.5 ms) on the packet drop side so that the interval from the first sample of the drop section is the minimum pitch cycle set by the average pitch cycle of the voice. On the side far from the packet drop section, the drop section is The interval from the beginning is the maximum pitch period set by the average period of the voice (128 bytes = 1
6 ms) and the interval (32 bytes = 4 ms) between the template top sample and the defective packet top sample, the interval is set to 160 bytes (20 ms).

With such a setting, the search window length becomes 140 bytes (17.5 ms). The search window length and its setting position are 20 to 128 bytes (2.5 to
16ms) and the amount of calculation are taken into account, and the template length is determined to be an appropriate value based on the prevention of erroneous detection of the optimal matching waveform and the amount of calculation. These numerical values show an example, and the present invention is not limited to these values.

Further, in this embodiment, the audio packet is 64 samples (1 byte = 1 sample), the m-th sample value of one packet is x (m), the k-th sample value of the template is x _t (k), Furthermore, (n + k) in the search window
The (-1) th sample value is indicated by y (n + k-1). In this embodiment, the latest input packet 33
The 30 samples from the sample to the 62nd sample were used as templates, and the 140 samples input from 160 samples to 21 samples before the first sample of the missing packet were used as the samples belonging to the search window.

Returning to FIG. 1, the waveform replacement switching unit 15 is connected to the input terminal 101 via the input 150 and to the control terminal 102 via the input 152.
Furthermore, they are connected to the amplitude correction unit 14 via the input 156, respectively. When a voice packet is lost, the switching unit 15 replaces the waveform with the interpolation signal input from the correction unit 14 to interpolate the missing portion of the voice packet. That is, the switching unit 15 determines whether the PCM signal input from the input terminal 101 is a missing signal in which a packet is missing, based on the LOSS signal indicating that the packet is missing from the control terminal 102.

In the case of a missing signal, the switching unit 15 performs a process of replacing the missing waveform with the signal input from the correction unit 14, and then outputs
To the output terminal 103 via When the PCM signal input from input terminal 101 is not a missing signal, switching section 15 outputs the input signal to output terminal 103 via output 154 as it is. The output 154 is also connected to the processing switching unit 11, and the signal output to the output terminal 103 is input to the switching unit 11.

The processing switching unit 11 calculates the sample number m of the input packet.
The non-normalized difference calculation processing unit 10, the cycle detection unit 17, and the amplitude correction coefficient derivation unit 16 are executed in accordance with. That is, the switching unit
Reference numeral 11 denotes a case where the sample of the input packet is sent to the buffer memory updating unit 12 and the sample number k of the template is k = 1 to 30 when the input signal belongs to the template.
In the case of (1), the difference calculation processing section 10 is executed, when k = 31, the cycle detection section 17 is executed, and when k = 32, the amplitude correction coefficient deriving section 16 is executed.

The difference calculation processing unit 10 is a processing unit that performs non-standardized difference calculation processing for detecting an optimal matching waveform of a template sample and a PCM signal of a packet stored in the memory update unit 12 described later. The arithmetic processing unit 10 includes a buffer memory pointer designating unit 18, a subtracting unit 19, and an absolute value arithmetic unit 2.
0, an adder 21 and a D (n) memory updater 22. (N = 1,..., N−K2 + 1) where x _t (k): each sample value in the template K: number of samples in the template K2: sample between the first sample of the template and the first sample of the missing part Number y (n + k-1): each sample value in the search window N: number of samples in the search window The arithmetic processing of equation (1) is executed.

The memory pointer designation unit 18 is a designation unit for designating the address (n + k-1) of the search window of the memory update unit 12. In this embodiment, n = 1 to 109, and the pointer designation unit 1
8 is 1 from 1 to 109 for each template sample
Change every step. The pointer designator 18 outputs the address (n + k-1) of the search window to the subtractor.

The subtractor 19 calculates the k-th sample value x _t (k) of the template via the input 160 and the (n + k-1) -th sample value y (n + k-1) of the search window via the input 172.
, Respectively, and calculates a difference between these sample values. Subtraction result _{x t (k) -y (n} + k-1) is
It is output to the absolute value calculation unit 20. The absolute value calculation unit 20 is a calculation unit that calculates the absolute value of the input numerical value.
_t (k) −y (n + k−1) |

The adding unit 21 and the D (n) memory updating unit 22 calculate
_{D k (n) = D k} -1 (n) + | x t (k) -y (n + k-1)
| That is, the adding unit 21 calculates | x _t (k) -y (n + k-) of the sample number k of the template.
1) When | is input from the absolute value calculation unit 20, the calculation result D _k-1 (n) of the sample number (k-1) is input from the memory update unit 22, and these are added to obtain D _k (n). Memory update unit
Output to 22. Adder 21 and D (n) memory updater
22 repeatedly performs this arithmetic processing from k = 1 to 30.

FIG. 3 shows an operation example of the calculation processing of the difference calculation processing unit 10. When difference operation unit 10 inputs the sample number k = 1 for the template, the processing unit 10, for example of up to 109-th from the left end of the sample search window of FIG calculate the equation (2), D ₁ (n ) = D ₀ (n) + | x _t (1) −y (n) | (n = 1 to 109) (2) D ₀ (n) = 0 These calculation results are output to the memory update unit 22.

Next, when the processing unit 10 inputs the template sample number k = 2, the equation (3) is calculated for the second to 110th samples from the left end of the search window, and D ₂ (n) = D ₁ ( n) + | x _t (2) −y (n + 1) |
(3) (n = 1 to 109) Output to the memory update unit 22 as in the case of the sum number k = 1. Hereinafter, similarly, the above-mentioned arithmetic processing is executed from k = 3 to 30. In this way, the memory update unit 22 finally determines D ₃₀
(1) to D ₃₀ (109) are input, and these calculation results D _k (n)
Is output to the cycle detection unit 17 via the output 180.

Returning to FIG. 1, when the cycle detection unit 17 is activated by the process switching unit 11, the cycle detection unit 17 receives D _k (n),
That is, the minimum value of D ₃₀ (1) to D ₃₀ (109) is selected.
Then, when the selected value of n is subtracted from 129, the value is the time interval between the template and the optimal matching waveform, that is, the pitch period PITCH which is the basic period of the audio waveform.
In this example, a human average pitch period of 20 to 128 bytes (2.5 to 16 ms) can be detected. The detecting unit 17 outputs the basic period PITCH to the buffer memory pointer designating unit 13 via the output 182 and outputs to the amplitude correction coefficient deriving unit 16 via the output 186.

Normally, the amplitude of the envelope of a voice waveform changes, so that by performing waveform replacement after correcting the amplitude at the time of the waveform replacement process, an effect that the voice does not feel unnatural can be expected.

The amplitude correction coefficient KLVL used for the amplitude correction may be, for example, a ratio of the sum of the absolute values of the amplitudes of the template and the optimum matching waveform as shown in Expression (4).
However, when the denominator of equation (4) is small, a maximum value is set for KLVL so that KLVL does not become an abnormally large value.

Returning to FIG. 1, in this case, when the coefficient deriving unit 16 is activated by the process switching unit 11, the basic period input from the detecting unit 17
Sample value y (α) of the optimal matching waveform in the search window corresponding to PITCH and sample value of the template
Enter x _t (β). The deriving unit 16 calculates the amplitude correction number KLVL so that the sample value y (α) of the optimal matching waveform in the search window is suitable for the envelope of the speech waveform with respect to the sample value x _t (β) of the template.

The derivation unit 16 outputs the calculated correction coefficient KLVL to the amplitude correction unit 14 via the output 184.

Note that the switching unit 11 does not perform the above processing for the first two packets. That is, the switching unit 11 is connected to the control terminal 102 via the input 152, and receives the PSYNC signal of the packet therefrom. The PSYNC signal is included in the control signal corresponding to the input sample value together with the LOSS signal. When the PSYNC signal is 1, it indicates that the input sample is the first sample of the packet. By counting the PSYNC signal, the switching unit 11 monitors whether three or more consecutive packets have been input. If any one of the first three packets is lost, interpolation is performed using the initial value of the basic cycle PITCH. The processing switching unit 11 is connected to the buffer memory updating unit 12 via the output 160, and outputs a signal x (m) to the memory updating unit 12.

The buffer memory update unit 12 is a storage unit that can store a search window for 128 samples and the latest input sample, for example, 128 samples and a voice packet PCM signal for one sample. The updating unit 12 updates the sample value every one sample input, and discards the stored PCM signals in the order of input when the sample value exceeds the storage volume.

The buffer memory pointer designating unit 13 searches for a PCM signal for interpolating the missing voice packet from the memory updating unit 12 based on the basic cycle PITCH input from the cycle detecting unit 17, and uses this as a signal x (mopt). Output to the amplitude correction unit 14.

The amplitude correction unit 14 calculates the signal x (mopt) and the correction coefficient KLVL
KLVL × x (mop), which is multiplied by the input signal and subjected to amplitude correction according to the audio waveform envelope
Calculate t). The correction unit 14 outputs the interpolation signal KLVL × x (mop
t) is output to the waveform replacement switching unit 15 via the output 156. When the voice packet is a missing signal, the switching unit 15
KLVL × x (mopt) is interpolated and output terminal 103 via output 154
Output to

FIG. 4 shows an operation flow of the PCM interpolation processing device 1. The operation will be described with reference to FIG. 1 and FIG. The processing switching unit 11 determines whether or not the input sample is the head of the packet, that is, whether or not PSYNC = 1 (4
00). When determining that the input sample is not the head of the packet, the switching unit 11 checks whether the input packet is a missing packet (406). If the input sample is the head of the packet, the switching unit 11 determines whether three or more packets have been input continuously (402), and if three or more packets have not been input continuously. In this case, the counter PCKT whose initial value indicating the number of consecutively input packets is 3 is decremented by 1 (40
Four). Then, similarly to the above, it is confirmed whether or not the input packet is a missing packet (406).

When a packet is missing, the waveform replacement switching unit 15
The interpolation signal KLVL × x (mopt) input from the correction unit 14 is interpolated into a missing signal and output to the output terminal 103, and at the same time, the process switching unit 11 inputs this sample value to the memory (40
8). If no packet is missing, the waveform replacement switching unit 15 outputs the sample value x (m) of the packet to an output terminal.
At the same time, the process switching unit 11 inputs the sample value to the memory (410). At this time, the counter PCKT
Is larger than 0, the sample value in the buffer memory updating unit 12 is updated one sample at a time. Otherwise, m>
32, that is, if the input sample belongs to the template, steps 414 to 422 are executed (412, 424), and then the sample value in the buffer memory updating unit 12 is updated one by one. That is, the sample number k of the template of the third packet counted from the first input packet is “1 to 3”.
0 ”,“ 31 ”, or“ 32 ”(41
4,418), when the sample number k is smaller than 31, the non-standardized difference calculation processing is executed by the difference calculation processing unit 10 (416).

FIG. 5 shows an arithmetic processing flow in the calculation of the optimum matching waveform detection in the arithmetic processing shown in step 416. That is, by executing the calculation shown in the figure from D _k (1) to D _k (109) in the case of the present embodiment, the time interval between the template and the audio waveform closest to the template in the search window is calculated. The data is used to determine the basic cycle PITCH.

Returning to FIG. 4, when k = 31, the process of the cycle detection unit 17 is executed. The detection unit 17 calculates the minimum value from the calculation results D _k (1) to D _k (109) obtained in the flow 416, and obtains the basic period PITCH (420). When the basic period PITCH is calculated, it is output from the detection unit 17 to the buffer memory pointer specification unit 13.

When k = 32, the process switching unit 11 executes the process of the amplitude correction coefficient deriving unit 16. When the deriving unit 16 is executed, the basic period PI
The correction coefficient KLVL is obtained from the TCH, and is output to the amplitude correction unit 14 (422).

As described above, according to the present invention, every time a sample value corresponding to a template is input, the detection of the optimum matching waveform portion with the audio packet signal input before the loss is performed by the non-standardized difference calculation processing. Then, based on the calculation result of the non-standardized difference calculation processing, the processing of setting the time interval between the template and the optimal matching waveform to the pitch cycle is executed regardless of whether or not the input packet is a missing packet. For this reason, at the time of input of the first sample of an arbitrary packet, the pitch period is always detected. If the input packet is a missing packet, a process of sequentially substituting the value of one sample before the pitch period of each sample of the missing packet is performed. repeat. If the detected pitch period is smaller than the number of samples of the packet, the sample value used for the interpolation once is repeatedly used.

By repeatedly using the sample values, it is not necessary to separate the sample on the side of the most missing part of the search window from the first sample of the missing packet, and the latest pitch period immediately before the missing part can be detected.
Due to the change in the voice waveform, the probability that an appropriate interpolation waveform does not exist in the search window is reduced, and the voice quality after the interpolation processing is improved.

In the present embodiment, the embodiment in which the present invention is applied to the PCM interpolation processing apparatus 1 for interpolating a lost voice packet has been described. However, the present invention is not limited to this. The present invention can also be applied to a device in which the pitch period of the audio waveform is always determined at the timing of the execution.

(Effects of the Invention) As described above, according to the present invention, every time a sample value corresponding to a template is input, the detection of the optimum matching waveform portion with the signal of the voice packet input before the loss is performed by the non-standardized difference calculation. Performed by processing. Then, based on the calculation result of the non-standardized difference calculation processing, the processing of setting the time interval between the template and the optimal matching waveform to the pitch cycle is executed regardless of whether or not the input packet is a missing packet. For this reason, at the time of input of the first sample of an arbitrary packet, the pitch period is always detected. If the input packet is a missing packet, a process of sequentially substituting the value of one sample before the pitch period of each sample of the missing packet is performed. repeat. If the detected pitch period is smaller than the number of samples of the packet, the sample value used for the interpolation once is repeatedly used.

By repeatedly using the sample values, it is not necessary to separate the sample on the side of the most missing part of the search window from the first sample of the missing packet, and the latest pitch period immediately before the missing part can be detected.
Due to the change in the audio waveform, the probability that an appropriate interpolation waveform does not exist in the search window is reduced, and the audio quality after the interpolation processing is improved.

Further, the present invention is not limited to the interpolation of the audio waveform, and can be applied to a device in which the pitch period of the audio waveform is always determined at the timing when the first sample of the packet is input in the audio packet communication.

[Brief description of the drawings]

FIG. 1 shows an embodiment of a voice packet interpolation system according to the present invention.
FIG. 2 is a functional block diagram applied to the PCM interpolation processing device, FIG. 2 is a parameter configuration diagram showing the configuration of various parameters used in FIG. 1, and FIG. 3 is a calculation process of the non-standardized difference equation in FIG. Conceptual diagram of calculation processing showing an operation example, FIG. 4 is an operation flow chart showing the operation of the PCM interpolation processing device in FIG. 1, and FIG. 5 is a diagram of the optimum matching waveform detection calculation in the calculation processing shown in FIG. It is a flowchart. Description of Signs of Main Parts 1 PCM interpolation processing device 10 Non-standardized difference calculation processing unit 11 Processing switching unit 12 Buffer memory update unit
13: Buffer memory pointer designating unit, 14: Amplitude correction unit, 15: Waveform replacement switching unit, 16: Amplitude correction coefficient deriving unit, 17: Period detection unit, 18: Buffer memory pointer designating unit, 19 ... Subtractor, 20... Absolute value calculator, 21... Adder, 22... D (n) memory update unit.

Claims (2)

(57) [Claims] In a voice packet interpolation method for interpolating a missing voice packet into a missing portion using a signal of a voice packet input before the loss, each time a sample value corresponding to a template is input, Processing for detecting the optimal matching waveform portion with the input voice packet signal by non-standardized difference calculation processing, and using the calculation result of the non-standardized difference calculation processing to set the time interval between the template and the optimum matching waveform to the pitch period Is executed regardless of whether or not the input packet is a missing packet. At the time of inputting the first sample of an arbitrary packet, the pitch period is always detected, and if the input packet is a missing packet, each of the missing packets is This is a method of sequentially substituting the value of one pitch period before the sample, and the detected pitch period is a packet. If from the number of samples smaller voice packet interpolation method, characterized by repeatedly using the sample value used at a time interpolation. 2. The method according to claim 1, wherein a position of a sample farthest from a packet missing portion in a search window area for searching for an optimal matching waveform with a template is determined from an average of voices from a packet missing portion leading sample. The maximum pitch period set by the pitch period and the template length are set apart from each other, and the position of the sample closest to the packet missing part is set by the average pitch period of the sound from the packet missing part top sample. A voice packet interpolation method characterized by being located at positions separated by a minimum pitch period length.