Method for mitigating the effects of speech segments losts in a packet based transmission

The invention relates to a method for concealing speech segment losses in packet-oriented transmission. In order to maintain the bit integrity of the data stream and the quality of speech transmission, measures must be taken to replace lost cells. A repeated request not or incorrectly received data, such. B. when file transfer, is in real-time services such as telephony is not possible because subsequently sent information can not be stored and output as long as desired.

Substitution methods for concealing lost, packet-oriented transmitted speech segments have long been known. In the following, known substitution methods will be described in more detail.

In the method of zero substitution, lost speech packets are substituted by simply replacing the samples of the missing packet with zero amplitudes or selecting the codewords so that the zero amplitudes result after decoding. See Jayant NS; Christensen SW<img file="EP0945853A2_D0001.tif" />Effects of Packet Losses in Waveform Coded Speech and Improvements Due to an ODD-Even Sample-Interpolation Procedure. IEEE Trans. On Communications, Vol. Com. 2 February 1981

Another known substitution method is the packet repetition method. In contrast to the zero substitution, the obfuscation of the lost speech segments is effected by a repeated output of the samples or code words received in the immediately preceding packet. The simple packet repetition, as with Wasem OJ; Goodman, DJ; Dvorak CA; Page HG in<img file="EP0945853A2_D0002.tif" />The Effect of Waveform Substitution on the Quality of PCM Packet Communications "IEEE Trans. on Acoustics, Speech and Signal Processing, Vol. 36, No. 3 March 1988 has the advantage that, with the exception of the storage of the last received packet and the detection of packet loss required in all substitution methods, no computational effort is necessary to replace the lost voice segment. A disadvantage of this method is that, however, usually arise discontinuities both at the boundaries between the preceding and the substituted speech signal as well as the transition to the subsequent segment. Relative to zero substitution, packet repetition achieves a maximum tolerable packet loss rate about 2.5 times greater.

Also known is a substitution method, which is referred to as fundamental frequency synchronous signal form replacement. In this method, it is assumed that the speech signal statistics change only slightly in time-successive packets and that the signal shape, especially with voiced sounds, remains largely unchanged with respect to the preceding and subsequent packets. As an indication, the average duration of a syllable (10 - 30 ms) can be used. For relatively short speech segments falling on the order of up to 10 ms, with a substitution of packet losses, spurious signal changes can be largely avoided if the waveform continues in sync with the speech fundamental frequency, see Goodman D. J .; Lockhart G. B .; Wasem O. J; Wong W.<img file="EP0945853A2_D0003.tif" />Waveform Substitution Techniques for Recovering Missing Speech Segments in Package Voice Communications. IEEE Trans. On Acoustics, Speech and Signal Processing, Vol. ASSP-34, no. 6, Dec. 1986. In contrast to the simple repetition of the entire preceding speech segment, the estimate of the waveform is made by the repeated output of the last T samples when T corresponds to the speech fundamental period of the previous signal.

Furthermore, DE 44 46 558 describes a digital transmission system with an improved decoder in the receiver. The invention relates to a digital transmission system, in particular a mobile radio system with at least one transmitter for transmitting codewords, which can be derived from a digital signal. If the disturbance of a codeword is detected, the most recent five samples of a prediction filter are replaced by very similar samples from previous signals, not by injecting the erroneous signal.

Furthermore, from DE 41 11 131 a method for transmitting digitized sound signals is known. For error concealment, the disturbed signal section is muted or replaced by a signal section preceding the disturbed signal section in the same channel or by a simultaneous undisturbed signal section of an adjacent channel.

The known method for error concealment of disturbed signals in packet-oriented transmitted voice data has in common that the substitution process according to the known substitution only begins when a packet loss is definitely suspected or festgetellt.

The object of the invention is to perform a part of the substitution process, but not the actual replacement, preventively before becoming aware of a loss and then gain time to reduce in particular the delay-induced packet losses and to increase the quality of the language.

The object is achieved by the characterizing features of the first claim. Advantageous embodiments and further developments emerge from the subclaims. Methods for obscuring speech segment losses in packet-oriented transmission are always based on the fact that lost, incorrectly or too late transmitted segments or Packets may not be replaced so that the perceptible error is minimal. The substitution is carried out either by known methods, such as zero substitution, block repetition, fundamental-frequency synchronous waveform continuation or a speech model-based signal form substitution described in more detail below. The possibility of recognizing a segment or Packet loss is always assumed. The inventive method is also based on the o. a. Premises, the fundamental difference from the known methods being based on the consideration of performing the estimation of the signal frame for each packet in advance, without there already being information at the time of estimation on whether the next packet is correctly received or not. The estimation of the speech frame of the subsequent not yet received packet takes place without access to side information exclusively on the basis of the previously received packets.

According to the invention, immediately after receipt of a packet k, an estimate of the signal frame of the following packet (not yet received) k + 1 on the basis of the previously received packets, or based on the decoded PCM signal of the previously received packets. In this case, starting from the approach of LPC analysis and synthesis or from the model of speech generation, first an analysis of the previously received speech information regarding excitation and sound shaping takes place. The respective subsequently expected segment is then estimated by subsequent synthesis filtering of the fundamental-frequency extrapolated excitation signal. Furthermore, it is advantageous to perform an asymmetric weighting of the analysis segment, a linear extrapolation of the sound shaping coefficients with bandwidth expansion and an improved fundamental frequency analysis of the excitation signal. In this case, the estimation of the signal frame of the following not yet received packet k + 1 is carried out according to the following steps:<ul id="ul0001" list-style="dash" compact="compact"><li>Recording the last segment into a segment buffer for storing the last M speech frames,</li><li>Estimation of the L samples of the subsequently expected speech segment plus an overhang of the length HOV samples,</li><li>Separating and caching the overhang and</li><li>Save the estimated segment.</li></ul> If there is a suspicion or if the absence, defectiveness or delay of at least one segment or packet k + 1 is detected, an immediate replacement of the respective segments or of the entire packet k + 1 takes place according to the invention on the basis of the previously received speech signal estimated signal frame.

The inventive method is realized by means of a simulation of an ATM speech decoder. The above described displacement of the substitution process to the time between the cell arrival times and the resulting instantaneous output of the already pre-substituted speech segment saves time, which is preferably used to increase an input buffer to compensate for packet delay variations. This makes it possible to reduce the delays caused by packet loss and thus indirectly increase the quality of the language.

The inventive method will be explained in more detail with reference to an embodiment.<ul id="ul0002" list-style="none" compact="compact"><li>Fig. 1 shows the block diagram of an ATM decoder with embedded Fehlerverschleierungsverfahren.</li></ul>

The mode of operation of the ATM decoder is on the one hand from the information about the loss or the correct reception of the current voice packet and on the other hand from the time sequence successively received or lost language packs. The information necessary for controlling the program is provided to the ATM decoder in the form of a binary file and evaluated by the control unit. Due to the different switch settings, parts of the decoder are activated or disabled. Hereinafter, the inventive method or the algorithm of the method according to the invention is explained on the basis of the possible scenarios, whereby only the functional blocks which become effective in the corresponding situation are described.

The method according to the invention is described on the basis of the modes of packet reception and packet loss. Operating mode packet reception: In the packet reception mode, the starting point is the error-free reception of the ATM cells. The switching means SM1 is closed in this mode. The received packet is decoded in the standard decoder. The type of decoding is independent of the method according to the invention.

In the case of error-free reception, the switching means SM2 is at position 1 (top) = error-free reception. If the preceding packet before the receipt of the error-free packet has been substituted, ie lost, then the switching means SM3 and SM4 are switched to position 2 (below). The blending of the reinserting error-free signal with the overhang of the previously substituted packet is performed according to the equation<maths id="math0001" num=""><math display="block"><mrow><mtext>Obuf (i) = </mtext><apply>&lt;sum /&gt;<lowlimit><mtext mathvariant="italic">i</mtext><mtext>= 0</mtext></lowlimit><uplimit><mtext mathvariant="italic">HOV</mtext><mtext>-1</mtext></uplimit><mrow><mtext>(1-</mtext><mtext mathvariant="italic">w</mtext><mtext>)</mtext></mrow></apply><msub><mrow><mtext> Mbuf (i) + ws</mtext></mrow><mrow><mtext>k + 2</mtext></mrow></msub><mtext> (i) with w = 0.5-0.5 cos (</mtext><mfrac><mrow><mtext>π ·</mtext><mtext mathvariant="italic">i</mtext></mrow><mrow><mtext mathvariant="italic">HOV</mtext><mtext>-1</mtext></mrow></mfrac><mtext>)</mtext></mrow></math><img file="EP0945853A2_D0004.tif" /></maths><dl id="dl0001"><dt>S_{k + 2}</dt><dd>corresponds to the latch containing the PCM values ​​of the packet following the substituted block</dd><dt>mbuf</dt><dd>corresponds to the determined overhang of the estimated packet k + 1</dd><dt>OBUF</dt><dd>corresponds to the initial values ​​after the transition</dd><dt>HOV</dt><dd>corresponds to the length in samples used for crossfading</dd></dl>

If the previous packet has been received without errors then the switches SM3 and SM4 are switched to position 1 (above). The switching means SM5 and SM6 are switched depending on the number of lost packets E.<ul id="ul0003" list-style="dash" compact="compact"><li>The lost counter E for previously lost packets is decremented when the number of lost packets is E&gt; 0.</li><li>If the lost consecutive packet error counter indicates that the number of lost packets is E&gt; E1, that is, the threshold for stepping down the output signal has previously been exceeded, then the switching means SM5 and SM6 are switched to position 2. The level adjustment of the reinserting transmission takes place in dependence on the number of lost packets E by stepwise insertion of the signal.</li><li>If the lost packet error counter indicates that the number of lost packets is E &lt;E1, then the switching means SM5 and SM6 are switched to position 1 (above). About the switching means SM6, the speech segment is written to the output memory. During further processing, the output memory is read out at the sampling frequency and the memory contents are output. The waveform of the subsequently expected speech segment is estimated until the arrival of the next packet. The estimation of the speech segment is subject to the following steps:</li><li>Recording the last segment in the segment buffer to store the last M speech frames,</li><li>Estimation of the L samples of the subsequently expected speech segment plus an overhang of the length HOV samples. This overhang is required for transitions. The type of estimation of the L samples of the subsequently expected speech segment can be selected as desired, provided that it fulfills the above-mentioned features.</li><li>Separating and caching the overhang and</li><li>Save the estimated segment.</li></ul>

If a backward-controlled coding method, for example according to ITU-T G.726, is used for the coding, a receiver-side coding of the replacement segment is advantageous, which proceeds according to the following steps:<ul id="ul0004" list-style="dash" compact="compact"><li>Initializing the receiver-side encoder status variables with the decoder status variables after decoding the last received packet. For this purpose, the switching means SM8 for the parameter transfer from the decoder to the encoder is closed for a short time.</li><li>Coding of the replacement segment to estimate the status variables of the decoder or Encoders after a possible packet loss and subsequent storage of the estimated status variables. For example, using a backward coded method according to ITU-T G.726, it has been found that the consequent errors in the reconstructed speech signal that occur after a packet loss due to the divergent status variables between the standard encoder and decoder could be reduced if the estimated Segment is coded again. The goal of the receiver-side encoding is to provide the encoder state variables after the substituted segment's encoding as starting values ​​for the standard decoder after a packet loss, i. H. to use the reinserting correct reception.</li></ul>

Package loss mode: The packet loss mode assumes that a packet loss (invalid, lost or not yet arrived packets) has already been detected. The switching means SM1 is open. There is no decoding. The switching means SM2 is switched to position 2 (below) so that the error concealment can be done by the previously estimated (substituting) segment. The switching means SM3 and SM4 are switched to position 1 (above).<ul id="ul0005" list-style="dash" compact="compact"><li>If the lost packet error counter indicates that the number of lost packets is E &lt;E2, the error counter is incremented. E2 represents the upper error threshold (maximum threshold).</li><li>If the number of consecutively lost packets is E &lt;E1, then the switching means SM5 and SM6 are switched to position 1 (above).</li><li>If the number of consecutively lost past packets is E1≤E≤E2 then the switching means SM5 and SM6 are switched to position 2 (below). This results in a gradual suppression (scaling) of the substituted signal.</li><li>If the number of packets lost consecutively applies: <maths id="math0002" num=""><math display="inline"><mrow><mtext>E = E2</mtext></mrow></math><img file="EP0945853A2_D0005.tif" /></maths>, that is, the output signal after the skip processing is already scaled to zero, the switching means SM6 is switched to position 3. A noise signal is displayed via the noise module. The speech segments are written to the output memory via the switching means SM6.</li></ul>

During further processing, the output memory is read out at the sampling frequency and the data is output. To take over the immediately previously estimated status variables, the switching device SM8 is closed for a short time. The transfer of the parameters takes place from the encoder to the decoder. This process takes place only in the event that a reverse-controlled coding method is used. The saved overhang from the estimated package is transferred to a second memory. For this purpose, the SM7 switching device is closed for a short time. The signal curve of the subsequently expected speech segment is estimated in the period until the arrival of the subsequent packet, whereby, as already in the mode packet reception<ul id="ul0006" list-style="dash" compact="compact"><li>the last output (estimated) segment is included in the segment buffer for storing the last M speech frames,</li><li>an estimation is made of the L samples of the subsequently expected speech segment plus the overhang of the length of the HOV samples, the type of estimation being arbitrary, provided that it fulfills the above-mentioned features and</li><li>separation and buffering of the overhang from the estimated packet and storage of the estimated segment. When using a backward-gated encoding method, as already described, a receiver-side encoding of the replacement segment for estimating the status variable of the decoder or encoder is performed after a further possible packet loss and the storage of the estimated status variables.</li></ul>

List of reference signs

<dl id="dl0002" compact="compact"><dt>k</dt><dd>Package / Pack</dd><dt>M</dt><dd>Number of language frames</dd><dt>L</dt><dd>Segment length in samples</dd><dt>HOV</dt><dd>Length of overhang (merge hangover) in samples</dd><dt>e</dt><dd>Counter reading indicating the number of packets lost consecutively</dd><dt>E1</dt><dd>Counter reading at which the gradual blanking of the estimated segment starts, or the insertion / scaling of the reimplementing signal ends.</dd><dt>E2</dt><dd>Counter reading at which the gradual blanking of the estimated segment ends and the replacement signal noise is displayed.</dd><dt>ATM</dt><dd>Asynchronous transfer mode</dd><dt>SM1 - SM8</dt><dd>Switching means of the ATM decoder</dd></dl>