Other media-independent FEC mechanisms proposed involve exclusive-OR operations, whereby a redundant parity packet is send every n^th data packets, by exclusive ORing the other n packets.¹² This method allows the recovery from a single loss in an n-packet message. FEC mechanisms in general are desirable when lost packets are dispersed throughout the stream of packets. Their advantage is that they are independent of the underlying media and they yield an exact replacement of the lost packet. Their computational requirements are relatively small and generally simple to implement. On the other hand, they lead to an increase in bandwidth as well as an added delay at the decoder side.

C. Content-dependent FEC In content-dependent FEC approaches, each frame of speech is transmitted in more than one packet with each copy represented in different compressed format. The first copy is referred to as the primary encoding and the subsequent copies as the secondary encoding.

For instance, in the method proposed by Bolot,² the first frame is PCM-encoded and sent in packet n and secondary encoding of the same frame is done with a low bit rate coding such as the LPC (2.4 to 5.6 kbit/s) or GSM coding (13.2 kbit/s) and sent in packet n+1 (Figure 5).

In Bolot's architecture, the choice of the primary and secondary encoding is a function of the computational cost, the available bandwidth and the degree of error robustness. Using GSM encoding for example is computationally demanding but is more robust to the type of errors experienced over the Internet. In addition, the amount of redundancy can be adjusted dynamically as the characteristics of the IP network changes. Thus during high loss periods, the secondary GSM encoding for packet n may be sent in packets n+1 and n+2 or in packets n+1, n+2, and n+3. The tradeoffs between the rewards of better information recovery and the added bandwidth and complexity are illustrated in reference 2 below for a number of combinations.

2. Adaptive Wideband Coding
IP networks provide potentially wide bandwidth and this in turn offer the possibility of sending high-quality speech through wideband coding. However, network congestions or other impairments also result in a high variance in the available bandwidth. As a result, a speech coder must ideally be able to exploit the high bandwidth to transmit higher fidelity speech (7 kHz instead of 4 kHz) yet at the same time be able to drop the bit rate (and gradually compromise on the fidelity) during congestion or whenever the available bandwidth on the IP network or the access network is no longer guaranteed.

Adaptive multirate wideband coders have been proposed in the context of wideband coding. The coder in highlighted in reference 14 below operates in five different modes—and bit rates—ranging from 24 to 9.1 kbit/s.

The goal is to provide at the higher rate a speech quality that equals or exceeds the quality of G.722 wideband coder (48 kbits). Thus, the coder scheme described in reference 14 exploits human auditory perception in that the lower band (0 to 6 kHz) is coded with a variable rate ACELP and the higher 1 kHz (representing 1 critical band in the auditory log scale) uses either a bandwidth expansion scheme or ADPCM coding, depending on the availability of BW and application. Most of the bits are reserved for the lower band with the upper band using as low as 6 bits per 20-ms frame (160 samples) or as much as 2 bits per sample when the overall bit budget is sufficiently big (24 kbit/s mode).

3. Transcoding Across Networks
Smart transcoding refers to the ability of providing a transparent and quality-wise effective way to map the various coefficients between two speech coders at the boundaries of a VoIP network.¹⁷. For instance, the scheme proposed in reference 15 below, maps a G.723 to an EVRC coder. The 2 coders have inherently different bit rates: 5.3 or 6.3 kbit/s for G.723 and 8 kbit/s for EVRC as well as frame size (30 ms with 7.5 lookahead delay vs. 20 ms with 10 ms lookahead for EVRC).

The line spectral pairs (LSP) are converted by translating 2 sets of G.723 information into 3 sets of LSP parameters for EVRC using an interpolation scheme over 3 frames. After the LSP conversion, the open-loop pitch of EVRC is computed using the closed-loop pitch of G.723 using the perceptually weighted speech.

The closed-loop pitch of G.723 is compared with the one from the previous EVRC subframe. If the distance of the 2 values is less than 10 samples, the closed-loop pitch of G.723 is determined as the open-loop pitch of EVRC. Otherwise, a pitch smoothing method is applied whereby a pitch value is searched in a range of +/-3 samples around the closed loop pitch of G.723 and EVRC. The 2 maxima are compared and a decision is made based on pitch value and the pitch gain in the previous subframe.

4. Background Noise Reduction
The aim of noise reduction is to minimize the effect of noise on the performance of voice communications systems. This means improving the perceived quality to the human listener as well as providing a more appropriate signal for estimating crucial signal parameters such as spectral content, pitch and voicing. There are a variety of methods for achievement speech enhancement. A detailed survey on the subject is found in reference 20 below:

• Wiener filtering: enhance speech by spectral subtraction and optimal linear filters. These filters are derived by minimizing the MSE or other criteria.
• Comb filtering: reinforcing the harmonic structure of the speech by combing through the spectrum and enhancing the periodic peaks.
• Maximum likelihood estimation: involved an estimation of the speech envelope or the magnitude spectrum based on a statistical model of the speech and noise.
• Psychoacoustics methods: which consist of special filtering that takes into account the peculiarities of perceptually important speech parameters or acoustic criteria of human hearing.

5. Low Delay Modems,br> An analysis by B. Goodman¹⁰ about the total delay in a typical VoIP call using dialup modems concluded that the component added by an analog modem significantly exceeds the theoretical lower limit. This limit is determined analytically, given the data rate of the modem, the number of speech frames per packet and the bit rate of the speech coder. In modems such as V.34, the actual measured delay can be up to 3 times that lower limit.

Further analysis showed that data compression, though resulting in higher effective rates, adds delay due to the buffering process, whose size depends on the compression ratio. The error correction and detection adds significant delay due to added framing required and the retransmission of errored blocks. This retransmission is effectively useless for VoIP applications that cannot tolerate additional waiting for a retransmitted packet.

Furthermore, the block size used in error correction is not optimized for speech frames. Since procedures vary across modems with respect to when a partial buffer is transmitted, the delay impact is unpredictable. Other features in typical modems, such as the equalizer filters, the interleaving of data as well as the trellis modulation adds more delay to speech frames.

While the problem is somewhat alleviated with high-speed modems, such as DSL or cable, there is still room to optimize the operations of the lower layers in order to keep the overall delay as small as possible in a VoIP call. This is particularly important if higher data rate, such as wideband coders are used, thus necessitating larger speech frames. Particular functions such as the bit error correction and the channel equalization need to be revised and adapted to the delay and error tolerance of speech transmission.

Final Thoughts
The merging of telecom carriers with other service providers, such as cable and Internet is becoming the norm, as the business arguments for bundling services and reducing operational costs become more and more compelling. Some of the challenges remaining in offering a competitive VoIP-based telephony are the service quality as well as the speech quality that consumers naturally expect to be equal, or even exceed that of the PSTN system. While most of the hurdles that are inherent to the VoIP context have been tackled to some degree, more robust and optimized solutions remain to be developed.

The speech coders currently used were not originally developed for today's IP telephony applications or for a high available bandwidth. As such, they do not fully exploit the available features and do not optimally address the problems of this new IP context. It is clear however that as these problems are properly addressed, IP-based telephony will be a natural progression to its current PSTN counterpart and will eventually provide a higher voice quality and service quality, at a competitive cost to all parties involved.

Editor's Notes: