`
`LONDON
`
`November 18 - 22, 1996
`
`
`
`GLOBECOM 96 Lomlon
`
`Communications: The Key to Global Prosperity
`
`GLOBAL |NTERNET'96
`
`Conference Record
`
`Sponsored by the IEEE Communications Society, the IEE and
`
`the UKRI Communications Chapter
`
`ICC
`
`GLOBECOM !!flIITTff!| E
`IEE O
`@@
`
`RPXExhibit 1133
`
`RPX Exhibit 1133
`RPX v. DAE
`
`
`
`IEEE 1996 Global Telecommunications Conference
`
`IEEE catalog number:
`
`96CH35942
`
`ISBN numbers:
`
`softbound edition:
`
`0-7803-3336-5
`
`casebound edition:
`
`0-7803-3337-3
`
`microfiche edition:
`
`0-7803-3338-1
`
`Library of Congress number: 87-640337
`
`Copyrights and Reprint Permissions:
`
`Abstracting is permitted with credit to the source. Libraries are permitted to photocopy beyond the limits
`of U.S. copyright laws for private use of patrons those articles in this volume that carry a code at the
`bottom of the first page, provided the per-copy fee indicated in the code is paid through the
`
`Copyright Clearance Center
`20 Congress Street
`Salem, Massachusetts 01970, USA
`
`Instructors are permitted to photocopy isolated articles for non-commercial classroom use without fee.
`For other copying, reprint, or republications permission, write to
`
`Director, Publishing Services
`IEEE, 445 Hoes Lane, P.O. Box 1331
`
`Piscataway, NJ 08855-1331, USA
`
`All rights reserved. Copyright © 1996 by The Institute of Electrical and Electronics Engineers, Inc.
`
`Additional Proceedings may be ordered from:
`IEEE Service Center
`
`Publications Sales Department
`445 Hoes Lane, P.O. Box 1331
`
`Piscataway, NJ 08855-1331, USA
`
`
`
`This material may be protected by Copyright law (Title 17 U.S. Code)
`
`
`
`49
`
`of the A,, in the output signal, the candidate for A,,
`is chosen from the tolerance interval, whose leftmost
`L’ samples resemble as close as possible AL, where L’
`denotes the length of AL. A computationally efficient
`similarity measure, the cross-correlation coeflicient
`
`L’—1
`
`cc(1/, 6) = Z a:(tx,,_1 + T3, + k + 6) -a:(t$,,-1 + Ty + k)
`k=0
`
`is used
`
`Then the desired time instance is
`
`t3“, = tw,,_1 + T3 + 6*,
`
`6* = argmax(;{cc(z/,
`
`The output step size is fixed to T = %L (L being the
`length of the A,,) and a fixed length hanning window
`is employed for overlap.
`
`input signal x
`search region for ix,
`Txi‘5max'/ Ty
`A1'
`’x1*Txl‘5max Ty*’x1 A2‘
`
`
`
`‘J for comparision with A1
`/ and A2, used to find the
`best positions ix, and txg
`
`pitch period of
`audio signal
`
`
`
`2 Audio Packet Loss Recovery
`through Waveform Substitution
`
`Previous proposals for receiver-only concealment of lost
`
`audio packets substitute the missing signal segment by
`repeating a prior segment. The “Pattern Matching”
`([6],
`technique repeats a correctly received signal
`segment, of which maximum similarity with the lost
`
`segment is assumed. This is accomplished by match-
`ing a sample pattern immediately preceding the gap to
`series of samples received earlier. As entire signal seg-
`ments of at least one packet duration are completely
`repeated, this may cause echoing sounds.
`
`Echoes can be avoided by “Pitch Waveform Replica-
`tion” ([6], [7], [10]), where only one pitch period found in
`the most recently received packet is repeated through—
`out the missing packet. An extension of this technique,
`called “Phase Matching”, provides for synchronization
`
`on both edges of the substitute, reducing a clicking
`distortion caused by the other two methods
`In
`the latter two cases the multiple repetition of the same
`small signal segment can lead to tinny sounds.
`
`All of these techniques have in common that, with
`
`an increased length of the lost segment, the perceived
`quality deteriorates severely. We believe that this de-
`crease in quality is to a large extent due to the spe-
`cific distortions introduced by the different concealment
`techniques. Therefore, in the following sections a new
`method, based on time-scale modification, is proposed.
`
`3 Description of the new method
`
`3.1 Time-scale Modification with
`
`the
`
`Figure 2: Enhanced speed version of the WSOLA algo-
`rithm.
`
`WSOLA Algorithm
`
`An appropriate algorithm must perform time-scale
`modification in real-time, and it may not change the
`pitch frequency to preserve natural sound. The “Wave-
`form Similarity Overlap-Add”
`(WSOLA) algorithm
`presented in [9] meets these requirements.
`
`As shown in Fig. 2, time—scale expansion is achieved by
`extracting segments A,, from the input signal at time
`instances tan} and superimposing them in the output
`
`signal at larger spaced time instances ty,,. The time in-
`stances tm, are selected from a tolerance interval around
`
`t$,,_1 + Tw, where the input step size Tm controls the
`
`time—scaling factor. To achieve a synchronized overlap
`
`3.2 Selection of Parameters
`
`1 shows the principle of error concealment using
`Fig.
`time-scale modification of correctly received packets.
`Distortions due to discontinuities at the boundary of
`the substituted and the next received packet are re-
`duced by overlapping them in the range of 2 ms using
`hanning windows.
`
`As the WSOLA algorithm was originally developed for
`time—scale modification of long speech sequences,
`the
`parameters Tx, 6, Ty and the lengths of A,, and AZ,
`
`
`
`50
`
`must be chosen carefully to preserve the good sound
`
`quality when WSOLA is applied to short signal seg-
`
`certain number of packets, but to withhold these pack-
`ets from the playout buffer.
`
`ments as shown in Fig. 1.
`
`least one pitch period, so that
`include at
`AL must
`the correct synchronization can be found (L’ 2 T,,,m,m,
`where Tp7ma$ denotes the maximum pitch period of a
`speech signal).
`In contrast to [9], we chose the maxi-
`mum possible interval [—T$, Ty — T3,] for the parameter
`6. The length of this interval must as well include at
`least one pitch period (Ty = L/ 2 2 Tpymax). We obtain
`L’ 2 L/ 2 for the relation of the lengths of A,, and AL.
`In our experiments we set L’ = L / 2 ([9]: L’ = L), which
`results in an enhanced speed version (Fig. 2) with no
`
`pereeivable deterioration of the output signal.
`
`In addition, there is the following relation between the
`length of the input series of samples lm and the estimate
`1;” of the number of samples used by the algorithm,
`
`!
`lg” z (N—— 1)Tx +L 3 lm
`
`where N is the number of extracted segments (0 _<_ 1/ <
`N). Another constraint exists between the number of
`samples lam; needed to replace the missing speech seg-
`ment and the length of the output series of samples lfmt
`received from the algorithm:
`
`llout:(N—1)Ty+L:(N+1)
`
`For long speech sequences we have L << lm, so the time-
`scale ratio B = Ty/Tm is approximately equal to the
`expansion ratio 04 = lout/l,-,1. However, for very small
`input lengths the above condition doesn’t hold, which
`results in ,6‘ >> a. This implies extreme time expansion
`
`and poor output quality. Thus it is necessary to chose oz
`as a compromise of output quality and additional delay
`
`introduced. Additionally, an optimization algorithm for
`out _
`the parameter 5
`'
`lam might be employed to
`minimize loss of information. A further suggestion for
`the choice of these parameters can be found in [11].
`
`3.3 Discussion of Delay and Computational
`Complexity
`
`While other waveform substitution techniques repeat in-
`
`formation from correct packets, our new scheme also
`
`modifies these packets themselves. Thus the computa-
`tional effort to execute the algorithm is higher. Addi-
`tionally,
`it. is necessary not only to keep a copy of a
`
`Considering the parameter relations described in the
`
`previous paragraph, we chose to conceal one lost packet
`(160 samples PCM 8 kHz (G.711) ——> 20ms audio) with
`three following packets. Four packets are needed to
`recover from the loss of two consecutive packets.
`In
`
`Fig. 1, packets 1 — 5 have to be queued for WSOLA
`and packets 1 — 3 for the other techniques, if overlap-
`add with correct neighbor packets is performed (for the
`other algorithms, copies of packets preceding packet 1
`
`have to be stored dependent on the length of the search
`area). In communication systems that artificially intro-
`duce audio delay, such as video conferencing systems
`([12]), the extra delay is acceptable. Furthermore, au-
`dio tools like vat and Ne VoT include delay adaptation
`mechanisms ([13]), which add some delay to cope with
`packet delay jitter.
`
`twice the number of additions
`In the above example,
`and four times the number of multiplications have to be
`performed, compared to the other algorithms (see [14]
`for a detailed analysis). Taking into account the abso-
`lute number of operations (m 8- 104 multiplications and
`z 4- 104 additions for the example) and the execution
`time usually available (because of delay adaptation) as
`well as the computing power of today’s average hard-
`ware, the computational effort can still be considered
`as low.
`
`4 Experimental Results
`
`4.1 Test Environment
`
`Four speech signals with different pitch frequencies (two
`male and two female speakers), sampled at 8 kHz,
`were used. The new Time—scale Modification technique
`(TM) was compared to Silence Substitution (S), Pat-
`tern Matching (PM) and Pitch Waveform Replication
`(PWR) using 40 test conditions of 10 seconds each.
`Thirteen non-expert listeners were asked to judge over-
`all quality on a five—category (Mean Opinion Score)
`scale, comparable to test schemes used in [1],
`[2],
`[7]
`and
`Additionally the presence of the disturbance
`components “tinny, metal”, “interrupted, clicking” and
`“echoing, reverberating” for each condition was judged.
`
`loss by suppressing one
`We modelled single packet
`packet within five (“1 of 5”) and double packet loss by
`
`
`
`suppressing two packets within six (“2 of 6”). This de-
`terministic order of dropping packets provides for equal
`comparison conditions. As the sound quality after error
`concealment heavily depends on which information has
`been lost, it is important to suppress the same audio
`
`segments over all comparisons.
`
`“Anchoring” (introducing the quality scale) was done
`by presenting the original signal and a “worst case”
`signal (which contained all disturbance components) to
`the listenener.
`Immediately after that,
`the test sig-
`nals for one speaker were presented to the listeners in
`a random sequence, with pauses of only a few seconds
`between two signals.
`
`4.2 Results
`
`of Subjective
`
`Performance
`
`Tests
`
`Mean Opinion Scores (MOS) were calculated for each
`of the four test signals. Fig. 3 shows the average of
`these values for the different algorithms. A quality en-
`hancement of TM compared to all other techniques can
`be observed (especially for “2 of 6”).
`
`-1015
`
`Elzote
`
`51
`
`percentage of listeners
`100
`
`tinny/metal
`echoing/reverberating
`interrupted/clicking
`
`B0
`
`60
`
`40
`
`Figure 4: Disturbance components
`
`is achieved.
`
`In addition to the subjective performance tests, mea-
`surements in the Internet ([14]), comparable to those
`in [15], were carried out. These measurements showed
`a loss distribution, which allows the application of the
`TM error concealment in the range of 55% to 75% of
`all “loss events” (2 1 consecutive packets are lost).
`
`5 Conclusions
`
`A new error concealment technique for lost audio pack-
`ets based on time-scale modification has been proposed,
`
`and the parameters of the WSOLA time—scaling algo-
`rithm have been adapted to the context of short signal
`segments. Experiments show that typical disturbance
`components of other techniques are reduced and overall
`quality is improved.
`
`worst
`case
`
`PM
`
`PWR
`
`TM
`
`original
`
`References
`
`Figure 3: Mean Opinion Score of overall quality.
`
`Fig. 4 summarizes, how often a specific disturbance
`component was noticed.
`It is shown that the echoing
`sound produced by PM is eliminated completely and
`
`the tinny sound of PWR is reduced by the new TM
`technique. The component “interrupted/clicking” was
`estimated about the same for all techniques and might
`further be improved for TM,
`if phase matching ([8])
`would be incorporated into the TM technique. This
`
`in
`can be done by controlling the time instances tan,
`the WSOLA algorithm such that phase matching be-
`tween the right boundary of the stretched signal and
`the left boundary of the next correctly received packet
`
`[1]
`
`J. Suzuki and M. Taka. Missing packet recovery
`techniques for low-bit—rate coded speech.
`IEEE
`Journal on Selected Areas in Communications,
`SAC—7(5):707—717, June 1989.
`
`[2] M. Yong. Study of voice packet reconstruction
`methods applied to CELP speech coding. In Pro-
`ceedings ICASSP-.92, pages II/ 125-128, March
`1992.
`
`[3]
`
`L.A. DaSilva, D.W. Petr, and V.S. Frost. A class-
`oriented replacement
`technique for lost speech
`packets. IEEE Transactions on Acoustics, Speech
`
`and Signal Processing, ASSP-37(10):1597-1600,
`October 1989.
`
`
`
`52
`
`[14]
`
`H. Sanneck. Fehlerverschleierungsverfahren fur
`
`Sprachiibertragung mit Paketverlust. Master’s
`thesis, Telecommunications Department, Uni-
`versity of Erlangen—Nuremberg, Erlangen, June
`1995.
`
`[15]
`
`J.—C. Bolot, H. Crépin, and A.V. Garcia. Analysis
`of audio packet loss in the Internet.
`In Proceed-
`
`ings of the 5th International Workshop on Net—
`work and Operating System Support for Digital
`Audio and Video, pages 163-174, Durham, NH,
`April 1995.
`
`Effects
`Jayant and S.W. Christensen.
`N.S.
`of packet
`losses
`in waveform coded speech
`and improvements due to an odd—even sample-
`
`IEEE Transactions on
`interpolation procedure.
`Communications, COM-29(2):101—109, February
`1981.
`
`V. Hardman, M. Sasse, M. Handley and A. Wat-
`son. Reliable Audio for Use over the Internet In
`
`Proceedings Inet 95, http://info.isoc.org/HMP/
`PAPER/070/abst.html
`
`D.J. Goodman, G.B. Lockhart, O.J. Wasem, and
`
`W. Wong. Waveform substitution techniques
`for recovering missing speech segments in packet
`voice communications.
`IEEE Transactions on
`
`Acoustics, Speech and Signal Processing, ASSP—
`
`34(6):1449—1464, December 1986.
`
`D.J. Goodman, O.J. Wasem, C.A. Dvorak, and
`
`H.G. Page. The effect of waveform substitution
`on the quality of PCM packet communications.
`IEEE Transactions on Acoustics, Speech and Sig~
`nal Processing, ASSP—36(3):342—348, March 1988.
`
`R.A. Valenzuela and C.N. Animalu. A new voice
`
`In Proceedings
`packet reconstruction technique.
`ICASSP-89, pages 1334-1336, May 1989.
`
`W. Verhelst and M. Roelands.
`
`An overlap-
`
`add technique based on waveform similarity
`(WSOLA) for high quality time—scale modifica-
`tion of speech. In Proceedings ICASSP—93, pages
`554-557, April 1993.
`
`[101
`
`L.R. Rabiner and R.W. Schafer. Digital Process-
`
`ing Of Speech Signals. Prentice Hall, Englewood
`Cliffs 1978.
`
`[11]
`
`A. Stenger, K. Ben Younes, R. Reng and B.
`Girod. A new Error Concealment Technique for
`
`Audio Transmission with Packet Loss to be pub-
`lished in EUSIPCO 96.
`
`[12]
`
`[13]
`
`International Telecommunication Union. Narrow-
`
`Band Visual Telephone Systems and Terminal
`
`Equipment.
`
`ITU-T Recommendation H.320,
`
`Helsinki, March 1993.
`
`H. Schulzrinne. Voice communication across the
`
`internet: A network voice terminal. Technical
`
`Report TR 92-50, Dept. of Computer Science,
`University of Massachusetts, Amherst, MA, July
`1992.