Speech Property-Based FEC (SPB-FEC)


Introduction

The main idea of SPB-FEC is to adjust the amount of added redundant data adaptively to the properties of speech signal. In contrast to other FEC schemes, SPB-FEC only uses FEC mechanisms to protect speech frames that are essential to the speech quality and cannot be concealed well by the frame-based coders's concealment algorithm while relying on this algorithm when other speech frames are lost.

Evaluation

The SPB-FEC scheme is compared against other FEC schemes that equally distribute the amount of redundant data under five network loss conditions which have been measured in the Internet. The following schemes are employed to send voice packets through a simulated network:
  • No FEC: Voice packets without any redundant data are sent (0% redundancy overhead).
  • FEC scheme 1: The two frames of packet (n) are piggy-packed on the packet (n+2). This scheme has a redundancy overhead of 100%.
  • FEC scheme 2: The four frames of the packet (n) and (n+1) are XOR'ed and the result is piggy-backed on the packet (n+2). This scheme has a redundancy overhead of 50%.
  • SPB-FEC: This scheme is similar to the FEC scheme 1 but the FEC mechanisms are only turned on to protect speech frames essential to the speech quality, resulting in a redundancy overhead of approximately 40% (for the speech sample used).
A simple but well known Gilbert model is used to drop voice packets. Each voice packet contains two G.729 speech frames (coresponding to 80 speech samples or 20 ms). The Gilbert has two states reflecting whether the previous packet is received (state 0) or lost (state 1). In the Gilbert model, p is the probability for the network model to go from state 0 to state 1 (i.e., the probability for a packet to be dropped given that its previous packet has arrived) and q is the probability for the network model to stay in state 1 (i.e., the probability for a packet to be dropped given that its previous packet was also dropped).

The speech material is taken from a speech sample of Digital Voice Systems, Inc.

Original signal
Decoded signal without frame loss


SPB-FEC FEC 1 FEC 2 No FEC
Network loss condition 1:
p=0.05, q=0.20		 decoded signal decoded signal decoded signal decoded signal
Network loss condition 2:
p=0.10, q=0.30		 decoded signal decoded signal decoded signal decoded signal
Network loss condition 3:
p=0.15, q=0.40		 decoded signal decoded signal decoded signal decoded signal
Network loss condition 2:
p=0.20, q=0.50		 decoded signal decoded signal decoded signal decoded signal
Network loss condition 2:
p=0.25, q=0.60		 decoded signal decoded signal decoded signal decoded signal



Download

Software for the simulation of SPB protection for VoIP streams is available in this package. Please refer to the README file.