Technology Processing & SAR Report

INTRODUCTION

Voice over IP (VoIP) has been studied for number of years and recently has been deployed on a field trial basis in the Cable IP network. Important aspects of the service include capabilities to support end-to-end Quality of Service (QoS) and maintain optimum bandwidth utilization [4]. The term QoS applies to many policies and resource management activities in the VoIP service architecture. For example, a generic VoIP layered architecture would include: access node (e.g. HFC), regional area network with IP/MPLS connectivity, VoIP office with call control and signaling features, customer application domain, and resource space for creation and management of services. At each layer, there are number of parameters influencing QoS measure and assessment. This paper describes a subset of parameters effecting bandwidth utilization and QoS at access and network layers.

The selection of codecs depends on a number of factors, such as the protocol employed, compatible codecs between endpoints, network transcoding capability, and bandwidth availability and grant. If bandwidth were free, then a high-resolution sampled voice could be deployed in any given call. However, demand for bandwidth continues to drive optimization of applications and deployment of efficient speech codecs in the network. Voice codecs are typically deployed in terminal devices and possibly in edge devices. Deployment of voice codecs in edge devices is intended to optimize bandwidth usage in the managed IP network. Also, voice/speech based enhanced services are an integral part of a MS in the IP network. Media Server services include unified messaging, voicemail, conferencing, Interactive Voice Response (IVR), speech recognition and text-to-speech. For applications such as these, which require voice codecs in the network equipment, low bit rate codecs are recommended.

Voice codec characteristics have been quantified in terms of three parameters, bit rate, quality, and complexity. [5]Considering various packet overheads and codecs, a simple calculation is performed to evaluate the effective bit rate of IP packets in the cable network. Network performance has been quantified in term of packet loss, latency, jitter, echo, and transcoding. In this paper, we evaluate the effect of each parameter and outline objective models for voice QoS based on a given parameter’s measurement. The objective of this study is multifaceted. The information in the document will make recommendations to aid network performance modeling projects, serve as an engineering guideline for voice services planning/provisioning, and the measure of QoS would be used to tune network elements for optimum performance.

VOICE CODEC PERFORMANCE CHARACTERISTICS

During the last ten years, a number of voice codecs have been developed and deployed in Circuit Switched (CS), wireless networks and Voice Messaging (VM) services. Table 1 identifies a number of voice/speech codecs that we believe are candidates for voice services in the cable IP network. More advanced compression algorithms have produced low bit rate codecs requiring less transmission bandwidth, however, there are a number of performance factors to consider. These are:

· Large frame size (sampled voice) will be needed by the more advanced compression algorithms

· There is a look ahead time, which is needed for decoding or de-compression

· As a result of the above, additional codec delay (end-to-end delay) will be produced.

· Compression is not lossless, thus perceived voice quality will be compromised by more compression.

In conjunction with codecs, Voice Activity Detection (VAD) and Comfort Noise Generation (CNG) are features, which further reduce the average bit rate of transmission during silence periods. [5] Codec complexity refers to the amount of memory and processing time needed to decode each frame of compressed data. Processing time is measured in term of Millions of Instructions Per Second (MIPS), and memory is the amount of RAM needed to process a frame. For example, G.711 is estimated to require less than 1 MIPS and 1 byte of memory, in contrast to G.723.1, which requires almost 18 MIPS of processing power and 2 Kbytes of memory. Historically voice codecs are implemented in Multimedia Terminal Adapters (MTA), but their applications in IP network equipment has been realized, which will be covered later in this paper.

Voice code complexity directly translates to the cost of implementation. Another factor influencing the overall cost of codecs includes licensing charges (if any). Their inventors patent most codecs, therefore MTA manufacturers may need to pay royalties for the use of codecs in their equipment. The current G.711 however is royalty free, which is a major factor in the decision to launch initial deployments with this codec. Table 1 below summarizes commonly used codecs in wired and wireless applications. Dark blue entries in Table 2 indicate codecs that are recommended by the CableLabs for implementation in the devices in cable IP networks, which include PCM (Pulse Code Modulation), ADPCM (Adaptive Differential Pulse Code Modulation), LD-CELP (Low Delay Code Excited Linear Prediction), CS-ACELP (Conjugate Structure Algebraic CELP), CS-ACELP (Conjugate Structure Algebraic CELP), MP-MLQ (Multi-Pulse Maximum Likelihood Quantization) &ACELP (Algebraic Code Excited Linear Prediction)

Table 1: Voice Codec Parameters

Coding Standard	Algorithm	Data Rate, Kbps	Frame Size, msec	Look Ahead, msec	Codec Delay, msec
G.711	PCM	64	0.125	0	0.25
G.726	ADPCM	16, 24, 32, 40	0.125	0	0.25
G.728	LD-CELP	16	0.625	0	1.25
G.729e	CS-ACELP	11.8	10	5	25
G.729	CS-ACELP	8	10	5	25
G.723.1	MP-MLQ	6.3, 5.3	30	7.5	67.5

NETWORK PERFORMANCE CHARACTERISTICS

In cable IP networks, analog voice is sampled, packetized, serialized, and transported over the Hybrid Fiber Coax (HFC), Regional Area Network (RAN), and gated to a party on the Public Switched Telephone Network (PSTN), or a party on the IP network. As a result, there are number of network parameters that affect voice quality, some of which are described here, including effective bit rate, delay, echo, and packet loss. The end-to-end service quality in cable IP networks include other parameters (e.g. piggybacking, concatenation, etc.) and features (e.g. Best Effort, UGS), which will be studied in the future releases of the document.

BANDWIDTH

As noted previously, voice samples are encapsulated into an IP packet having overheads from various layers of protocol as shown in Figure 1. The number of bits in the DOCSIS physical (PHY) layer’s overhead is not specified, because the actual number depends on number of parameters and conditions. These parameters and conditions also will be explained briefly in this section. Also in Figure 1, the RTP header does not include Contributing Source (CSRC) identifier.^[1]It is assumed only one stream (a voice call) is included in the RTP transport protocol.

DOCSIS

PHY

DOCSIS

MAC

(14)

ETH

(14)

(20)

UDP (8)

RTP (12)

Payload

(Variable size compressed samples)

FCS (4)

Figure 1 – A Pictorial representation of IP packet payload + headers

During a call, most of the overheads shown in Figure 1 are redundant from one packet to the next and may be suppressed to optimize bandwidth usage. This is known as Payload Header Suppression (PHS), which is listed as an “option” in the current DOCSIS 1.1 specification. [6] At the link layer, CM and CMTS negotiate through DOCSIS protocol to perform PHS both on the upstream and downstream during a call. Figure 2 is the pictorial representation of overheads with PHS enabled. Note that the PHS is transparent to the network, since CMTSs reconstruct the original header prior to forwarding to the RAN.

DOCSIS

PHY

DOCSIS

MAC

(14)

ETH

(0)

UDP (0)

RTP (12)

Payload

(Variable size compressed samples)

FCS (4)

Figure 2 – A Pictorial representation of IP packet payload with PHS

Tables 2 & 3 below provide the effective bit rates of various codecs without & with considering PHS, however, the DOCSIS Physical layer overheads are not considered in the calculation. The effective bit rate is calculated by taking into account headers (as shown in the Figure 1&2) in a frame plus the codec’s bit rate, and producing a number in terms of bits per second (bps).

Note that the 14-bytes DOCSIS headers shown in Figure 1 are for upstream, which include, 6 bytes of base header, 3 bytes of extended UGS header, 5 bytes of extended BPI+ header. Dark blue columns in Table 2 indicate codecs that are recommended by CableLabs.

Table 2: Effective bit rate of IP packets with different voice codecs without PHS

Packet size, ms	Effective bit rate (using G.711) bps	Effective bit rate (using G.728) bps	Effective bit rate (using G.729e) bps	Effective bit rate (using G.729) bps	Effective bit rate (using G.723.1) bps
10	121600	73600	69400	65600	62900
20	92800	44800	40600	36800	34100
30	83200	35200	31000	27200	24500

Table 3: Effective bit rate of IP packets with different voice codecs with PHS

Packet size, ms	Effective bit rate (using G.711) bps	Effective bit rate (using G.728) bps	Effective bit rate (using G.729e) bps	Effective bit rate (using G.729) bps	Effective bit rate (using G.723.1) bps
10	88000	40000	35800	32000	29300
20	76000	28000	23800	20000	17300
30	72000	24000	19800	16000	13300

PACKET LOSS

There are number of parameters contributing to packet loss, including network congestion or data corruption. Packet loss affects fidelity of voice quality and in most codecs more than 3% packet loss results in an unacceptable grade of service.

Studies of voice grade of service in the context of PSTN (ITU-T Recommendation P.800) have produced a Mean Opinion Score (MOS) for toll (acceptable) quality at level 4 on a scale of 5. Thus QoS comparisons are made based on this measurement. A 3% packet loss rate results, on average, in a reduction in MOS scores of 0.5 point.

Interpolation is one approach to compensate for lost packets. The speech/voice decoder will predict what the missing packet (payload) should be based on the previous packet. This technique is known as Packet Loss Concealment (PLC). All codecs mentioned in this document have PLC algorithms built into their standards. The improvement trade-off is that latency resulting from PLC processing and the interpolation process may produce an audible artifact.

LATENCY

The end-to-end latency of a voice call over a cable IP network results from cumulative effect of algorithmic, packetization, serialization, propagation, and component delays. A simple problem arising from delay is called “talker overlap”. This happens when large gaps (delays) exist between received signals; thus encouraging the other party to speak (thinking the first talker has stopped). Given below, is a summary of delays in sequential order.

Codec processing (algorithmic)
Packetization of compressed data

Local network (DOCSIS) traversal
Routing to the backbone network
Backbone traversal

Far-end reception of packets and traversal of local access
Buffering of out-of-order and delayed packets
Decoding, decompression, and reconstruction of the audio stream

Control of overall latency requires a hand-in-hand effort by system resources and the VoIP application. ITU-T recommendation (G.114) defines a maximum end-to-end delay of 150 milliseconds. Typical end-to-end delays in IP networks range from 50 to 300 milliseconds.

DELAY VARIATION (JITTER)

IP networks have been optimized for reliable data transmission, thus allowing for a variation in delays across different packets in a transmission. Even though a source gateway generates voice packets at regular intervals (say, every 20 ms), a destination gateway will typically not receive these packets at regular intervals due to jitter. In order to correct for this, data needs to be stored in a buffer with a dynamic size, to allow for slowest packets to arrive and put into order. The ordering process contributes to end-to-end delay.

ECHO

When a transmitted signal is returned at some late time, usually at much lower power levels, echoes are produced in a transmitters’ ear. There are at least two common types of echoes in the wired communication. “Acoustic echo”, typically results from poor electro-acoustic coupling between earpiece and mouthpiece in the handset. “Hybrid echo”, typically results from the impedance mismatches in the 2-4 wire conversion processes along the PSTN network.

The degree to which echo is objectionable depends on echo loudness (level of returned signal power), measured in dB, and total delay (as explained above), measured in milliseconds. For example, if the returned signal power were very small (20 dB below the power of the transmitted signal), even with a large delay it would not be noticeable.

The same analogy holds true for short round trip delay, typically below 50 milliseconds. The ITU’s generalized recommendation is that connections with one-way delay greater than 25 msec should have echo control devices. Echo cancellers are the preferred and most commonly deployed mechanism, which is typically located as near to the source of the echo as possible.

TRANSCODING

Transcoding is a mechanism to convert between incompatible voice codecs, e.g. converting between G.711 and G.728 at an edge device. The transcoding process requires resources (MIPS & Memory) as well as adding additional latency into the connection. Codecs are not lossless, thus any transcoding results in a degradation of voice quality.

A combination of transcoding and packet loss could result in a signal quality that is below minimum acceptable grade of services. In that case, a higher bandwidth codec is typically employed, which will result in higher system bandwidth utilization. If a trade-off between bandwidth and quality is permissible, then Transcoding could be deployed at edge devices (CMTS) to reduce bandwidth utilization.

VOICE CODEC QOS MEASURES IN CABLE IP NETWORKS

There are two well-known and established measures of voice quality of service, namely subjective and objective models. Mean Opinion Score (MOS) ratings are a subjective measurement that has been developed by the ITU to characterize grade of service as well as aid in the development of objective models.

Figure 3: MOS for Voice Codecs

ITU-T P.800 MOS is a widely known standard for measure of voice quality. The model ranks voice quality based on 5-point scale. Using this scale, an average score of 4 and above is considered as toll-quality. Figure 3 depicts MOS for different voice coding algorithms. The dark blue color area means the rating for that codec is within that range. For example, G.711 is rated between 4.3 and 4.4. Under ideal network conditions, with the exception of G.723.1, other codecs are rated above 4, i.e. their quality meet or exceed toll grade requirements.

MOS subjective measurement is most salient since it is based on human opinion of grade of services. However, it may not be an economical and practical method of measurement as an ongoing process, since it involves human subject sampling on a large and geographically dispersed scale.

Often, limited MOS testing results are correlated with well-known mathematical objective models to assess quality. Most notable objective models include, Perceived Speech Quality Measure (PSQM, ITU-T P.861), Perceptual Evaluation of Speech Quality (PESQ, ITU P.862), and E-models (ITU-T G.107). [2-4] - The E-model, described in ITU-T G.107 is derived from an equivalent model that has been deployed in the assessment of voice quality in the PSTN. The E-model results in a transmission-rating factor, R, calculated from following equation:

R = Ro - Is - Id - Ie + A, where

R is the Transmission Rating Factor
Ro is the basic signal to noise ratio based on send, receive loudness, electrical and background noise
Is represents the sum of real-time voice transmission impairments, for example, loudness, sidetone, and PCM quantizing distortion
Id represents the sum of delayed impairments relative to the voice signal, for example, talker echo, listener echo, and absolute delay
Ie represents the Equipment Impairment factor for special equipment, for example, low bit-rate coding (determined subjectively for each CODEC, for each percentage of packet loss)
A is the Advantage factor (compensates for advantage of access, for example, satellite phone)

The transmission rating factor R takes on value from 0 to 100, with R=100 representing very high quality. The R-value is related to MOS value through following rule:

For R < 0 MOS = 1

For 0 < R < 100 MOS = 1+ 0.035R + R(R-60)(100-R)7x10^-6

For R > 100 MOS = 4.5

For example, R=80 is equivalent of MOS of 4.03. Provisional guide for relating R-value [3] to user satisfaction is shown in Table 4.

Table 4: Relation Between user satisfaction and R-value

R-Value	MOS calculated from R	User Satisfaction
90	4.34	Very Satisfied
80	4.03	Satisfied
70	3.6	Some users dissatisfied
60	3.1	Many users dissatisfied
50	2.58	Nearly all users dissatisfied

SUMMARY

This paper has reviewed three fundamental aspects of voice quality of service in the cable IP network. The first of which include voice codecs parameters, such as compression rate, processing delay, and look ahead period. The second aspect of voice quality of service is influenced by network parameters such as bandwidth, delay, echo, and transcoding. The third aspect includes incorporating these parameters into either a subjective and/or objective measure to assess performance and obtain an MOS model for voice quality of service in the cable IP network.

REFERENCES

[1] H. Schulzrinne, S. Casner, R. Frederick, and V. Jacobson, "RTP: A Transport Protocol for Real-Time Applications", IETF RFC 1889, January 1996.

[2] ITU-T Recommendation P.862 (2001), Perceptual Evaluation of Speech Quality (PESQ): An objective method for end-to-end speech quality assessment of narrowband telephone networks and speech codecs.

[3] ITU-T Recommendation G.107 (5/2000), The E-Model, a computational model for use in transmission planning.

[4] ITU-T Recommendation P.861 (2/1998), Objective quality measurement of telephone-band (300-3400 Hz) speech codecs.

[5] PacketCable™ Audio/Video Codecs Specification, PKT-SP-CODEC-I04-021018, October 18, 2002

[6] Data-Over-Cable Service Interface Specifications, DOCSIS 1.1, Radio Frequency Interface Specification, cableLabs, SP-RFIv1.1-I10-030730

Volume 5, Number 1, Fall 2004