RU2402879C2 - Method for adaptation of codec parametres and transfer of voice signal in network with switching of packets (versions) - Google Patents

Abstract

FIELD: information technologies.

SUBSTANCE: tolerance is established for change of voice signal quality, network is tested by higher and lower load, following the network testing results, optimal parametres of codec and voice signal transfer are found for current connection and established as current, change of voice signal quality is calculated with a certain periodicity on the basis of E-model rating factor (R-factor) calculation. In case calculated change of voice signal quality exceeds established tolerance for its for change, network is again tested by higher and lower load, following the network testing results, optimal parametres of codec and voice signal transfer are found for current connection, found parametres are established as current, and again change of voice signal quality is calculated with a certain periodicity.

EFFECT: improved quality of speech transfer in IP-telephony under conditions of unknown and unstable throughput capacity of communication channel and availability of accidental losses of transmitted packets in communication channel.

2 tbl

Description

The invention relates to IP-telephony and can be used to optimize the quality of voice transmission.

In known methods for transmitting voice information, adaptive quality control in IP networks is realized by dynamically changing the encoding and transmission of the speech signal by terminal devices depending on the current characteristics of the communication channel. For this, it is necessary to choose a method for collecting statistical information about a communication channel, a method for adapting encoding parameters and transmitting a speech signal.

The collection of statistical information on a communication channel is usually carried out using RFC 3550 recommendation, which defines the RTCP protocol (Real Time Control Protocol, a protocol that controls real-time transport protocol), which allows terminal devices to obtain statistical information on the characteristics of the communication channel.

The main difficulty is the choice of how to adapt the encoding and transmission of the speech signal based on the received information about the communication channel. There are two problems here.

The first problem is the need for an objective measurement of the quality of transmitted speech, since it is possible to talk about the success of the adaptive adjustment method only when new coding parameters under these conditions provide better speech quality than previously used. The classical definition of the quality of a speech signal on the MOS scale requires specially equipped studios with the involvement of a large number of experts. Obviously, this method of assessing quality is unacceptable in most applications, and therefore other methods of assessing speech quality have been developed that have eliminated the "human factor". The most interesting for practical application was the E-model (G.107. ITU-T Recommendation G.107 E-model - a computational model used for transmission planning [Text]. - Approved - Geneva: ITU-T, 2005. - 28 p .) and its variants for IP-telephony (Lustosa, L.C. G. E-Model Utilization For Speech Quality Evaluation Over VoIP-Based Communication Systems / LCG Lustosa, LSG Carvalho, PHARodrigues, ESMota. - 22nd SBRC, 2004 . - P. 933-938).

The E-model proceeds from the assumption of the additive effect of distorting factors on the quality of communication and does not require an initial signal to evaluate it. Instead, the E-model derives a decision based on the characteristics of the communication channel, terminals, background noise level, etc. Thus, it is assumed that by measuring the characteristics of the terminal devices and the communication channel, it is possible to obtain an estimate of the quality of the output speech signal.

The second problem is to determine which parameters and how you can manage to achieve better quality. Common speech codecs (for example, G.729 or G.723.1) do not provide the possibility of any manipulation of the encoding parameters and are not suitable for these purposes. However, codecs are known in which there is control of the encoding mode (for example, Speex and GSM 06.90 AMR (Adaptive Multi-Rate)).

Voice quality in IP-telephony is also influenced by the speech signal transmission modes: blocking coefficient (packetization, otherwise, the number of frames in a packet) and the number of transmissions of the same packet per channel. An increase in the number of frames in a packet can reduce the overhead of their transmission by saving on packet headers, but it increases the overall signal delay. An increase in the number of parcels of the same packet per channel with a margin on the bandwidth of the communication channel allows you to effectively deal with random packet loss. These parameters can and should also be controlled.

An example of the development of a method for adapting the parameters of a speech codec can be a known method, namely, that they provide an adaptive codec capable of transmitting a continuous voice stream and having information about the source data rate and channel bandwidth; one quality parameter, determine at least one restriction associated with the transmission of the voice stream, adapt the data rate source and the channel bandwidth as a function of the quality parameter and a limiting factor for the maximum value of received signal quality while transmission of the continuous stream of voice. The parameter for evaluating the quality of speech transmission in this method uses the maximum number of media characters in the codeword, the maximum length of the codeword, network delay, delay degradation factor, packet loss factor, measured signal distortion factor, R-factor for calculating the MOS parameter, which is used to match the speed transmit data source and bandwidth of the communication channel (Application No. US2004160979 H04L 1/00; H04L 29/06; H04M 7/00; H04L 12/56, publ. 08/19/2004).

To apply the known method, it is necessary to be able to control the bandwidth of the communication channel, which is impossible in most IP-telephony tasks. Also, to assess the quality of speech transmission, you need to know such parameters as the maximum number of media characters in the code word and the maximum length of the code word, which requires additional costs for their assessment and is, in turn, a difficult technical task. In addition, communication by IP-phone does not always correspond to the task of transmitting a continuous voice stream, since a conversation between two subscribers is accompanied by natural pauses. The known method does not have means to compensate for random packet loss in the communication channel and reduce overhead during packet transmission. All this complicates the practical implementation of the method for use in IP-telephony.

The technical result of the proposed solution is to improve the quality of voice transmission in IP-telephony in the conditions of unknown and unstable bandwidth of the communication channel and the presence of random loss of transmitted packets in the communication channel.

To achieve this result, two variants of a method for adapting codec parameters and transmitting a speech signal in a packet-oriented network are proposed.

Option 1. A method of adapting the codec parameters and transmitting a speech signal in a packet-switched network, in which the initial parameters of the codec and transmitting the speech signal, as well as the tolerance for changing the quality of the speech signal are set, search for the optimal codec and speech signal transmission parameters for the current connection, the found codec and speech signal transmission parameters are set as current and, with a certain frequency, a change in the quality of the speech signal is calculated. If the calculated change in the quality of the speech signal exceeds the established tolerance for changing the quality of the speech signal, they again search for the optimal parameters of the codec and the transmission of the speech signal, set them as current and again change the quality of the speech signal with a certain frequency.

The difference of the proposed method is that the calculation of the change in the quality of the speech signal is carried out on the basis of the calculation of the rating factor (R-factor) of the E-model, and the parameters of the codec and speech signal transmission, optimal for the current connection, are searched by the coordinate descent method, using as the objective function R factor.

Option 2. A method of adapting codec parameters and transmitting a speech signal in a packet-switched network, in which a tolerance for changing the quality of a speech signal is established, testing the network with increased and reduced load, and using network testing results, the codec and speech signal transmission parameters that are optimal for the current connection are found the found codec and speech signal transmission parameters are set as current and, with a certain periodicity, the change in the quality of the speech signal is calculated. If the calculated change in the quality of the speech signal exceeds the established tolerance for changing the quality of the speech signal, the network is again tested with increased and reduced load, according to the results of network testing, the codec and speech transmission parameters that are optimal for the current connection are found, the found codec and speech transmission parameters are established signal as current and again with a certain frequency calculate the change in the quality of the speech signal.

The difference of the proposed method is that when testing a network with an increased load, the parameters of the codec and voice signal transmission are selected so as to exceed the actual throughput. In the established connection, when testing the network with a reduced load, the current parameters of the codec and voice signal transmission are selected so as not to exceed the actual bandwidth ability B of the established connection, and the calculation of the change in the quality of the speech signal is performed based on the calculation of the rating factor (R-factor) E-model . When calculating the codec and speech signal parameters optimal for the current connection, the real throughput of the established connection is determined according to the statistics of the RTCP protocol according to the following formula:

B = (B1-L1) / (1-dL),

where B1 is the throughput for which testing is carried out with increased load, and B1> B;

L1 - losses corresponding to increased load obtained according to the statistics of the RTCP protocol;

dL = L2 / B2;

B2 - throughput, for which testing is carried out with a reduced load, with B2 <V;

L2 - losses corresponding to a reduced load, obtained according to the statistics of the RTCP protocol.

The codec and speech signal parameters that are optimal for the current connection are selected for the found parameters B, L1 and L2 using the following condition:

min | B - [(P (m) × k + H) × F (k) / 8] × r |,

where m are the values of the controlled parameters of the codec;

k is the blocking coefficient (the number of frames in the packet);

P (m) - frame size for given m, bits;

H is the size of the packet headers, bits;

F (k) - packet sending frequency, 1 / s;

r is the number of parcels of the same packet in the communication channel;

min | f (m, k, r) | - the operation of finding the minimum of the absolute value of the function f (m, k, r) by controlled parameters.

The proposed method is implemented as follows. In the first variant of the method, the initial parameters of the codec and transmission of the speech signal are set, as well as the tolerance for changing the quality of the speech signal ΔQ _d .

The following controlled parameters are included in the speech transmission parameters:

- blocking coefficient k (another name - packetization coefficient or the number of frames in the packet, a positive integer);

- the number of parcels of the same packet in channel r (integer).

The specified parameters are used for any codecs used in IP-telephony.

Codec parameters depend on the codec used. Not all codecs can control encoding parameters. However, codecs are known in which there is a control of the encoding mode m. For example, Speex and GSM 06.90 AMR (Adaptive Multi-Rate, see http://en.wikipedia.org/wiki/Adaptive_Multi-Rate, as well as ETSI recommendation EN 301704 V7.2.1 at http://www.esti.org ) The encoding mode m of the Speex codec is called quality (encoding quality, integer from 1 to 8 inclusive). A similar parameter in the GSM codec 06.90 AMR is called mode, eight states and similar controls are also used. Therefore, in the future, everything is considered on the example of one codec - Speex codec.

Each of the controlled parameters in its own way affects the quality of the speech signal. Increasing the value of the parameter k increases the overall delay, which leads to a deterioration in the quality of the received speech signal, but reduces the load on the communication channel. The repetition of each packet in the presence of a reserve bandwidth of the communication channel allows you to compensate for random packet loss in the communication channel, but increases the load on the communication channel. Changing the value of the parameter m allows you to control the quality of speech encoding and the load on the communication channel: the higher the quality, the greater the load on the channel.

It is important to note that changing any of the above parameters changes the payload of the IP phone to the communication channel, which on narrow-band or congested channels leads to a change in the percentage of losses and delays and, as a consequence, to a change in the quality of the speech signal. Thus, a change in any of the parameters can lead to both an improvement in signal quality and its deterioration. It depends on the properties of the communication channel itself. Therefore, it is impossible to pre-select the encoding and transmission mode that is optimal for all cases.

The choice of the initial state of the parameters in this variant of the method is not fundamental. To speed up the search for optimal parameters, typical codec settings were chosen (for Speex it is m ₀ = 8), the blocking coefficient k ₀ = 1 and the number of transmissions of the same packet into the channel r ₀ = 1.

The tolerance for the change in the quality of the speech signal ΔQ _{d is} set in relative units (for example, 0.2, which is equivalent to a change in quality by 20%).

After setting the initial parameters of the codec and transmitting the speech signal, as well as the tolerance for changing the quality of the speech signal ΔQ _d , the optimal codec and speech signal parameters are searched for for the current connection. To do this, use the method of coordinate descent.

According to this method, from the initial state of the codec parameters and the transmission of a speech signal characterized by a triple of controlled parameters (k ₀ , r ₀ , m ₀ ), a search is made for the maximum quality of the speech signal Q along the direction of the k axis (blocking coefficient (packetization)) and determine the point ( k ₁ , r ₀ , m ₀ ), in which it is maximal. Then, a maximum is searched from this point in the direction of the r axis (the number of packets of the same packet per channel) and the following state is determined (k ₁ , r ₀ , m ₀ ). Similarly, the maximum is determined by the coordinate m (coding mode).

The quality of the speech signal Q is determined by calculating the R-factor of the E-model (G.107. ITU-T Recommendation G.107 E-model is a computational model used for transmission planning [Text]. - Approved - Geneva: ITU-T, 2005 . - 28 p.). The simplified formula for calculating the R-factor used in IP-telephony is presented as follows (Lustosa, L.C. G. E-Model Utilization For Speech Quality Evaluation Over VoIP-Based Communication Systems / LCG Lustosa, LSG Carvalho, PHARodrigues , ESMota. - 22nd SBRC, 2004. - P.933-938):

R = 93.4-Id (Ta) -Ie (codec, loss),

Where:

Ta is an absolute one-way delay;

Id - quality degradation caused by delay;

Ie - quality loss during encoding (codec), as well as due to packet loss in the communication channel (loss).

In the absence of an echo effect, the parameter Id is completely determined by the one-way propagation delay of the signal Ta and its calculation does not represent a fundamental difficulty. Calculation of the delay is usually performed using special control signals that are built into the Speex codec packets (analogue ping).

The parameter Ie requires experimental determination. Such measurements were carried out for a number of the most popular codecs, while for Speex, this work has not been previously performed. It should be noted that the Ie parameter is affected by both the quality degradation associated with Speex coding and the quality degradation due to losses in the communication channel. These parameters are not independent, because Speex decoder uses packet loss concealment algorithms, which are very effective for single losses, but their efficiency decreases with increasing percentage of losses.

The experimental design for estimating the parameter Ie consists in the following sequence of actions:

- for a given set of Speex encoding modes and a given set of loss levels, encoding and then decoding of a set of speech files are performed; sound fragments resulting from decoding are compared with the original ones according to the PESQ algorithm (P.862. ITU-T Recommendation P.862 Perceptual evaluation of speech quality (PESQ): An objective method for end-to-end speech quality assessment of narrow-band telephone networks and speech codecs [Text]. - Approved 02/01/2001. - Geneva: ITU-T, 2001. - 21 p.), which evaluates the quality of MOS in the range from 0 to 5;

- the MOS score is converted to the R-factor using the formula given in ITU-T Recommendation G.107;

- taking the Id value equal to zero, Ie is calculated for each set of codec and channel parameters.

The results of the experimental determination of the parameter Ie are shown in table 1.

In total, about 20 sound files with a length of 15 to 120 seconds were tested. When experimenting with other sound files, the results may vary. This, in particular, is due to the features of the Speex codec, which is optimized for male voices. Therefore, there will be differences when dividing speech into the voices of adults / children, male / female.

Table 1 Ie value for various Speex codec modes and loss levels Packet Loss Rate (%) Codec mode 0 one 2 3 four 8 16 one 53.49 53.85 54.17 54.00 54.15 55.01 56.48 8 43.98 44.41 44.73 45.07 45.54 47.07 50.32 2 33.04 34.53 35.63 36.77 38.07 41.54 46.62 3 27.86 30.30 31.86 33.59 35.38 39.72 45.58 four 23.14 26.44 28.52 30.92 33.03 37.67 43.93 5 21.22 24.82 27.17 29.78 31,99 36.79 42.96 6 17.85 22.65 25.28 28.02 30.58 35.93 42.87 7 15.41 20.94 24.22 27.32 29.90 35.49 42.41

The period between the change of parameters and quality checks is determined by the duration of transients in IP-telephony. Because of this, the period must be at least 1 s. Empirically, you can choose the optimal value of this parameter for a specific implementation of the IP-phone. Typically, the period is 2-4 s. Large values of the period lead to an increase in the inertia of the control and untimely reaction of the adaptation system to changes in the parameters of the communication channel. This also applies to the second variant of the method described below.

The set of parameters found in this way is considered optimal for this channel, it is set as the current one, the corresponding maximum value of the R-factor Q _{opt is} used in the future to calculate the quality change.

The search for optimal parameters is carried out on the side of the receiver of the speech signal, however, the parameters of the codec must be adjusted on the side of the transmitter. Transfer of control signals is carried out by means of the built-in signaling (in-band signalling) implemented in the Speex codec.

The main idea of the built-in signaling is that the receiver has the ability to send specially shaped "pseudo-frames" to the transmitter - Speex packets, whose speex mode field is set to 14, and the encoded control is located in the data field. The party that received the packet with a special mode number determines the control action code from the first four bytes, and interprets all other data in the packet as its parameters. The Speex codec programming interface provides the ability to set handlers for events of the form “received control action with code X”.

Then, with a certain periodicity, the change in the quality of the speech signal ΔQ = | (QQ _opt ) / Q _opt |, where | x | is the absolute value of the number x, Q is the current quality calculated using the R-factor. The period is chosen for the same reasons as when searching for the optimal parameters of the codec and transmitting the speech signal. If the ΔQ value exceeds the tolerance ΔQ _{d, they} again search for the codec and the speech signal optimal for the current connection (this time from the previously found optimum point). Set them as current and again with a certain frequency calculate the change in the quality of the speech signal.

In fact, all the time of a two-way connection, they work in a cycle of the search mode for the optimal parameters of the codec and the transmission of the speech signal, as well as the monitoring mode for changing the quality of the speech signal.

In the second variant of the method, the tolerance for changing the quality of the speech signal ΔQ _{d is set} , and, as in the first variant of the method, ΔQ _{d is} set in relative units (for example, 0.2, which is equivalent to a quality change of 20%).

Next, the network is tested with increased and reduced load. Testing starts at the start of two-way connection of subscribers. At the same time, VAD (Voice Activity Detector, speech activity detector) should not be turned on, otherwise you can pause and test with increased load.

Testing allows you to determine the causes of packet loss during transmission over the network and the actual network bandwidth. This is important because packet loss is a major source of degradation in voice restoration quality.

There are two main reasons for packet loss in a channel:

- limited bandwidth of the communication channel B, tied to the entrance to the channel (maximum bandwidth of the communication channel); if the packet stream exceeds the bandwidth of the communication channel, then some of the packets are lost;

- the proportion of independent random losses dL in the communication channel observed for any band of the real packet stream Bi≤В transmitted to the channel (i = 1 when testing the network with increased load, i = 2 when testing the network with reduced load).

Based on this simplified model, to search for the real bandwidth of communication channel B and the fraction of independent random losses dL, they make up a system of two equations.

The first equation is constructed when testing a channel with a known overload B1> B, for example, by increasing the number of sendings of the same packet to the channel:

L1 = B1- (B × (1-dL)),

where L1 is the total flow loss in the channel when testing the network with increased load.

The second equation is constructed when testing the network with a load of B2 <B, not exceeding the channel capacity. B2 is actually the accepted payload in the first enhanced test. Then

L2 = B2- (B2 × (1-dL)) = B2 × dL,

where L2 is the total flow loss in the channel when testing the network with a reduced load.

Where from

dL = L2 / B2,

in its turn

B = (B1-L1) / (1-dL).

All values necessary for the calculation are found directly from the statistics of the RTCP protocol.

Based on the network testing results, the codec and speech signal transmission parameters that are optimal for the current connection are found (the task is to fit into the measured B band) as follows:

- in a cycle from a larger to a smaller frame size P (m), the corresponding encoding mode m of the Speex codec, the blocking (packetization) coefficient k and the number of packets of the same packet in channel r are selected so as to fit in the measured band B; moreover, with packet losses greater than hdL (hdL value can be set, for example, equal to 1%) and a band more than hBand (hBand value can be set, for example, equal to 2500 bytes / s), packet retries are used;

- if it was possible to fit into the measured band B, then the selection of parameters ends, otherwise a transition to a smaller frame size P (m) is performed;

- if for all possible frame sizes P (m) it was not possible to fit in the strip, then at the upper level of the IP-telephony system, a decision is made to break the two-way connection.

The calculation of the required band for a given k, r and m is performed according to the formula:

B (m, k) = [(P (m) × k + H) × F (k) / 8] × r,

where m is the encoding mode of the Speex codec;

k is the blocking coefficient (packetization, otherwise, the number of frames in the packet);

In (m, k) - the required band, Kb / s;

P (m) - frame size for a given m, bits;

H is the size of the headers (usually 224), bits;

F (k) - packet sending frequency, 1 / s;

r is the number of parcels of the same packet in the communication channel.

Table 2 shows the results of calculating the required band for various values of m and k at r = 1.

table 2 The required band for different values of m and k at r = 1 F (k) fifty 25 16.66 12.5 10 8.33 7.14 k one 2 3 four
In (m, k) 5 6 7 m P (m) one 48 1700 1000 766 650 580 533 500 8 80 1900 1200 966 850 780 733 700 2 120 2150 1450 1216 1100 1030 982 950 3 160 2400 1700 1466 1350 1280 11233 1200 four 224 2800 2100 1866 1750 1680 1633 1599 5 304 3300 2600 2366 2250 2180 2132 2099 6 368 3700 3000 2766 2650 2580 2532 2499 7 496 4500 3800 3565 3450 3380 3332 3299

The found codec and speech signal transmission parameters are set as current. After that, the quality of the transmitted speech signal Q _{opt is} calculated using the R factor and stored. The period between the change of parameters and quality checks, as well as the R-factor is determined similarly to the first variant of the method.

Then, with a certain frequency, the change in the quality of the speech signal ΔQ = | (QQ _opt ) / Q _opt |, where | x | is the absolute value of the number x, Q is the current quality calculated using the R-factor. The period is chosen for the same reasons as in testing. If the ΔQ value exceeds the tolerance ΔQ _d , the network is again tested with increased and reduced load, according to the network test results, the codec and speech signal parameters that are optimal for the current connection are found, the found codec and speech signal transmission parameters are set as current and the quality change is calculated again with a certain frequency speech signal.

Repeated tests with increased load are carried out at a slightly increased channel load in Ktest (for example, 1.25 times), which makes it possible in principle to carefully probe the channel capacity.

In fact, all the time of a two-way connection, they work in a test mode cycle with determination of the optimal parameters of the codec and voice signal transmission, as well as a mode for monitoring the change in the quality of the speech signal.

The proposed technical solution allows to improve the quality of speech transmission in the conditions of unknown and unstable bandwidth of the communication channel by selecting the optimal parameters of the codec and transmitting the speech signal. In the presence of random packet losses in the communication channel, packet retransmission (speech transmission parameter) is used to compensate for them.

Claims (1)

A method of adapting codec parameters and transmitting a speech signal in a packet-switched network, in which a tolerance for changing the quality of a speech signal is established, testing the network with increased and reduced load, using the network testing results, find the codec and speech transmission parameters that are optimal for the current connection, establish the found parameters of the codec and speech signal transmission as current, with a certain periodicity calculate the change in the quality of the speech signal, if This change in the quality of the speech signal exceeds the established tolerance for changing the quality of the speech signal, the network is again tested with increased and reduced load, according to the results of network testing, the codec and speech signal transmission parameters that are optimal for the current connection are found, the found codec and speech signal transmission parameters are set as current and again, with a certain periodicity, the change in the quality of the speech signal is calculated, characterized in that when testing the network the increased load the parameters of the codec and voice signal transmission are selected so as to exceed the actual bandwidth B of the established connection, and when testing the network with a reduced load, the current parameters of the codec and voice signal transmission are selected so as not to exceed the actual bandwidth B of the established connection, while calculating the quality change the speech signal is produced based on the calculation of the rating factor (R-factor) of the E-model, when calculating the codec and transmission parameters optimal for the current connection the speech signal is determined according to the statistics of the protocol RTCP real throughput of the established connection according to the following formula:
B = (B1-L1) / (1-dL),
where B1 is the throughput for which testing is carried out with increased load, and B1>B;
L1 - losses corresponding to increased load obtained according to the statistics of the RTCP protocol;
dL = L2 / B2;
B2 - throughput, for which testing is carried out with a reduced load, with B2 <V;
L2 - losses corresponding to reduced load, obtained according to the statistics of the RTCP protocol;
despite the fact that the parameters of the codec and voice signal transmission that are optimal for the current connection are selected for the found parameters B, L1 and L2 using the following condition:
min | B - [(P (m) · k + H) · F (k) / 8] · r |,
where m are the values of the controlled parameters of the codec;
k is the blocking coefficient (the number of frames in the packet);
P (m) - frame size for given m, bits;
H is the size of the headers, bits;
F (k) - packet sending frequency, 1 / s;
r is the number of parcels of the same packet per channel;
min | f (m, k, r) | - the operation of finding the minimum of the absolute value of the function f (m, k, r) by controlled parameters.