Speech Property- -Based Based FEC FEC Speech Property for Internet Internet Telephony Telephony for Applications Applications Nguyen Tuong Long Le and Henning Sanneck GMD FOKUS, Berlin www.fokus.gmd.de/glone {le,sanneck}@fokus.gmd.de MMCN 2000, San Jose, CA January 24, 2000
Overview Overview • Voice over IP (VoIP) – Improved quality for VoIP • Approach – Performance of the G.729 loss concealment – Speech Property-Based FEC (SPB-FEC) • Evaluation – Reference FEC schemes – Network model – Objective speech quality measurement • Conclusions
Voice over over IP IP Voice • Main drivers: – current economical incentives (Internet flat rate ➾ pricing) ➾ Internet Telephony – service integration, unified packet-switching infrastructure • One of the main problems: – satisfaction of real-time QoS demands in a packet- switched network (fundamental tradeoff: statistical ➾ multiplexing vs. reliability ➾ packet loss )
QoS for Voice over IP IP flows flows QoS for Voice over – high compression (backward adaptive coding: ITU-T G.729, G.723.1) • no further sender adaptation / network adaptation (transcoding) possible • amplifies high perceptual impact of burst losses (error propagation) + tolerance to isolated losses (speech stationarity ➾ ➾ ➾ loss concealment) extrapolation of coder state ➾ ➾ enhance the loss resiliency of high-compressing codecs with open-loop error control (FEC)
Structure of an Internet an Internet Audio Audio Tool Tool Structure of
Additional components: Sender : Sender Additional components • G.729 coder used both for the payload and the redundancy • Side information available at the encoder is used • Decoder concealment process is taken into account
Additional components: Receiver : Receiver Additional components • No generic (PCM-level) concealment
Performance of of G.729 G.729 loss concealment loss concealment Performance Unvoiced Voiced 40 35 30 Signal to noise ratio 1-frame loss 25 2-frame loss 3-frame loss 20 4-frame loss 15 10 5 0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 frame position • Decoder fails to conceal losses at unvoiced/ voiced transition due to lack of state (synthesis filter parameters, excitation)
Speech Property- -Based Based FEC FEC Speech Property • Adjust amount of redundancy adaptively to loss concealment performance: overhead � 40-50% • Comparison with two reference FEC schemes overhead 100% FEC scheme 1 Packet (n) Packet (n+1) Packet (n+2) Packet (n+3) Packet (n+4) 50% FEC scheme 2 Packet (n) Packet (n+2) Packet (n+3) Packet (n+4) Packet (n+1) Audio data Redundant data
Network Model Network Model p 1-p q 0 1 1-q Network loss Network loss Network loss Network loss Network loss condition 1 condition 2 condition 3 condition 4 condition 5 p=0.05, p=0.1, p=0.15, p=0.2, p=0.25, q=0.2 q=0.3 q=0.4 q=0.5 q=0.6
Results: : Application Application- -level loss level loss rate rate Results 0,5 0,4 FEC scheme1 Application loss rate FEC scheme 2 0,3 SPB-FEC 0,2 No FEC 0,1 0 1 2 3 4 5 Network loss condition
Results: : Auditory Distance Auditory Distance Results • Application of recent advances in objective speech quality measurement: ITU P.861A 3 2,5 A uditory distance 2 FEC s c heme 1 FEC s c heme 2 1,5 SPB-FEC No FEC 1 0,5 0 1 2 3 4 5 Network loss condition
Results: : Perceptual Distortion Perceptual Distortion Results • Enhanced Modified Bark Spectral Distortion (EMBSD; Temple University) 8 6 Perceptual distortion FEC scheme 1 4 FEC scheme 2 SFB-FEC No FEC 2 0 1 2 3 4 5 Netw ork loss condition
Conclusions Conclusions • SPB-FEC exploits differences in „concealability“ to adjust the amount of added redundancy • simple network model & objective speech quality measures showed the reduction of necessary redundancy to maintain a good output quality • Speech samples: – www.fokus.gmd.de/glone/products/voice/spb-fec • end-to-end operation: add network adaptivity • mapping to network prioritization (DiffServ)
Recommend
More recommend
