Slide 1Speech Adaptation Over Multi-Hop Wireless NetworksMotivationSlide 4Caption Embedding SchemeSlide 6Slide 7Slide 8Adapting To Very High Loss using SR/TTSConclusions1UCLA CS Wireless Adaptive Mobility Lab Speech Adaptation Over Multi-Hop Speech Adaptation Over Multi-Hop Wireless NetworksWireless NetworksTsuwei Chen, Mario Gerla,Tsuwei Chen, Mario Gerla, Manthos Kazantzidis,Manthos Kazantzidis,Yuri Romanenko, Ilya SlainYuri Romanenko, Ilya Slainhttp://www.cs.ucla.edu/NRL/wirelesshttp://www.cs.ucla.edu/NRL/wirelessJune 1999June 1999{tsuwei,gerla,kazantz,yuri,ilyas}@cs.ucla.edu{tsuwei,gerla,kazantz,yuri,ilyas}@cs.ucla.eduMini Quality Of Service Conference during ICC 992UCLA CS Wireless Adaptive Mobility Lab Speech Adaptation Over Multi-Hop Speech Adaptation Over Multi-Hop Wireless NetworksWireless NetworksMario Gerla, Manthos Kazantzidis et alMario Gerla, Manthos Kazantzidis et alhttp://www.cs.ucla.edu/NRL/wirelesshttp://www.cs.ucla.edu/NRL/wireless3UCLA CS Wireless Adaptive Mobility Lab MotivationMotivation•Wireless networks characteristics:Wireless networks characteristics:–high packet losshigh packet loss–highly variable packet delivery timehighly variable packet delivery time–constantly changing bandwidthconstantly changing bandwidth•Audio delivery applications characteristics:Audio delivery applications characteristics:–Sensitive to packet lossSensitive to packet loss•~ 20% pkt loss ~ 20% pkt loss  sharp perceptual quality degradation sharp perceptual quality degradation–Sensitive to delay jitterSensitive to delay jitter•Existing audio/video delivery Internet Existing audio/video delivery Internet applications (e.g. VIC, VAT) & IP telephony:applications (e.g. VIC, VAT) & IP telephony:–Few are adaptive to changing network conditionsFew are adaptive to changing network conditions–Not suitable for wireless environmentNot suitable for wireless environment4UCLA CS Wireless Adaptive Mobility Lab Wireless NetworkUCLA Adaptive Speech ExperimentUCLA Adaptive Speech ExperimentMultihop Testbedclient• Adjustable Parameters - sampling rate - packet size• QoS Monitoring: - packet loss - jitter Audio(UDP)Control(TCP)A d a p t a t I o n S t r a t e g y :Audio source adapts to QoS feedbackIncrease in Packet loss packet size is reducedsampling rate is reducedIncrease in jitternetwork congestedchannel noise/interferencePiggybacked Text Stream(UDP)serverTTSSyncSpeechRecognition5UCLA CS Wireless Adaptive Mobility Lab Caption Embedding SchemeCaption Embedding Scheme0 sec 2 secVoice file (.wav)Question: When you are driving @ at fifty-five miles per hour …0 sec2 sec……..Caption file (.cpt)For reliable caption delivery, each 2-second segment of text is replicated in several packets.#0 #1 #2 #3 #4Wireless packets6UCLA CS Wireless Adaptive Mobility Lab • Packets are forwarded using routing tables• In order to force multihop in confined environment, nodes are partitioned into subnets• WaveLan I, CSMA/CA, 915Mhz, 2Mbps, 700ft (200m) open, 100ft (30m) closedMultihop RoutingMultihop RoutingVirtual Interfaces131.179.25.26131.179.26.26131.179.27.26131.179.27.174131.179.25.175131.179.26.2526 subnet25 subnet27 subnetTestBedServer NT/95Intermediate nodeLinuxClient NT/95Additional Traffic Linux7UCLA CS Wireless Adaptive Mobility Lab No AdaptationNo Adaptation0204060801000 20 40 60 80 100 120 140 160 180 200tim e (sec)LossJitterSampling Rate / 250• The audio stream cannot adaprt to the loss rates observed.• Significant quality degradation perceived with increased loss rate8UCLA CS Wireless Adaptive Mobility Lab Adaptation on Both Packet Size and Adaptation on Both Packet Size and Sampling RateSampling Rate0204060801000 20 40 60 80 100 120 140 160 180 200tim e (sec)Loss Jitter Sampling Rate / 250•Adaptation on both packet size and sampling rate results in the lowest loss rates (best perception quality)9UCLA CS Wireless Adaptive Mobility Lab Adapting To Very High Loss Adapting To Very High Loss using SR/TTSusing SR/TTS02 04 06 08 01 0 01 2 011 12 13 14 15 16 17 18 19 11 0 11 1 11 2 11 3 11 4 11 5 11 6 11 7 11 8 11 9 12 0 1T i m e•At very high loss rates, beyond a threshold, Speech Recognition is applied at the server and Text-To-Speech synthesis at the client side.•The speech is recognized at packet loss rates up to 99%10UCLA CS Wireless Adaptive Mobility Lab ConclusionsConclusions•speech quality is substantially enhanced by payload size speech quality is substantially enhanced by payload size and rate adaptationand rate adaptation•Recognizable speech is delivered in extreme conditions Recognizable speech is delivered in extreme conditions by Text-To-Speech translationby Text-To-Speech translationWork in progressWork in progress•Real/Synthetic Speech synchronizationReal/Synthetic Speech synchronization–explicit relative byte-to-word mappingexplicit relative byte-to-word mapping•Production of Speech CODEC tracesProduction of Speech CODEC traces–CELP, SBC, A/u-Law, DSP TSpeech, GSM 6.10, G723, MP3 etc.CELP, SBC, A/u-Law, DSP TSpeech, GSM 6.10, G723, MP3 etc.•Large scale network hybrid simulationLarge scale network hybrid simulation–using GlomoSim platformusing GlomoSim platform•Adaptive Video-On-DemandAdaptive Video-On-Demand•Network Feedback (QoS routing)Network Feedback (QoS routing)•Multicast Stream AdaptationMulticast Stream