The major challenges of real-time voice over IP are the strict QoS requirements, such as low latency and loss rate, versus the best-effort services provided by current network infrastructure. Despite of the vulnerability and stringent QoS requirements of voice streams, the data rate of voice traffic is very low compared other types of data and streaming multimedia traffic. For this reason, there have been techniques designed to transmit voice stream with added redundancy as protection against packet loss in hostile channels.
A common redundancy-adding method is forward error correction (FEC), which transmits redundant copies of each packet across over multiple packets [5][4][12]. In this sender-based scheme, a lost packet can be recovered from the copies piggybacked in subsequent packets should they be received successfully. Obviously, in this scheme, loss recovery is performed at the cost of higher delay [4]. In many cases, however, the losses of successive packets are correlated, due to the way packets are dropped as networks get congested and router buffers are becoming full. A packet loss may usually be followed by bursts of losses, which decreases the efficiency of FEC schemes[3]. In order to combat burst loss, the FEC scheme has to increase the span of packing redundant copies across over subsequent packets following the original one, which introduces even higher delay. Hence, the repair capability of FEC is limited by the acceptable delay.
Another sender-based loss recovery technique, interleaving, which does not increase the date rate of transmission, also faces the same dilemma. The efficiency of loss recovery depends on over how many packets the source packet is interleaved and spread, and greater span of spreading increases latency[15].
One important QoS related issue which should deserve our attention is that, in delay-sensitive applications such as interactive voice communications, packet loss is not only a result of channel erasure, but also a result of the delay variation (also known as jitter) of the network. At the receiving end, minimal buffering delay and playout deadline are usually desired in order to reduce the total end-to-end latency. With traditional playout scheduling mechanisms, packet loss occurs when the network delay increases quickly while the playout schedule does not respond accordingly in a timely manner. In this situation, packets are discarded and lost since they arrive later than the playout deadline [10]. Therefore, delay jitter greatly impairs the QoS of real-time applications.
In this work we address all these problems and QoS concerns from a different approach. We send multiple copies of the voice stream over different independent paths. Since the channels are independent, or nearly independent, the statistical behavior of the two paths are largely uncorrelated. The probably of a negative disturbance, such as erasure or high jitter, affecting both channels at the same time will be small.
The multiple streams to be delivered to different paths are formed by multiple description coding (MDC), which generates multiple descriptions of the source signal in equal importance, which can be decoded independently at the receiver. If all descriptions are received, the source signal can be reconstructed in full quality. If only one copy is received, the quality of the reconstructed is degraded, but still better than that resulting from losing all the descriptions. In this work we study the scheme of transmitting two MDC streams.
In previous literature, multi-path has been used for reliable video communication over lossy networks using multiple state encoding [1]. With the averaged path behavior, or path diversity, burst losses are isolated and outage probability decreased. Multi-path transmission also alleviates the problem when the default path chosen by the routing algorithm is not optimum, which might be often according to [13].
In the context of delay-sensitive applications, the novelty and the key point of using path diversity lies in that the behavior of delay jitter is also uncorrelated over different paths. The latency can be reduced by always playing out the description copy with lower delay if full audio quality is not a priority. In case of the stream from one path experiences sudden high delay, packets from the other stream can be used in substitute in time, to avoid any loss.
Multi-path transmission over the Internet can be realized by future peer-to-peer architectures, in which each peer receiving service also serves as a relay. In this way voice streams can be lead to follow different routes.
This work is organized as follows. We first introduce our receiver playout scheduling algorithm for multiple streams in Section 2. Section 3 presents multi-path measurements performed in the Internet. Using the measured traces we compare single-path and multi-path transmissions and show that considerable improvements can be obtained for multi-path voice transmission. With a more detailed simulation in Section 4 we analyze and elaborate the benefits that multi-stream multi-path transmission provides in more general settings.