A circuit-switched call had a relatively stable delay and no packet loss in the voice path (faults were analog noise or clipping). A packet voice path shares queues with other traffic. Three impairments dominate: packets that arrive late (latency), packets that arrive with uneven spacing (jitter), and packets that never arrive (loss). Echo, clock drift, and acoustic problems add a second layer that packet counters will not show.
HD Voice and a better codec cannot rescue a path that already violates delay or loss budgets. Measure the transport first, then the codec.
The three packet impairments
- Latency (delay)
- One-way mouth-to-ear time. Includes capture, encode, packetization, network, jitter buffer, decode, and playout.
- Jitter
- Variation in packet inter-arrival time. A jitter buffer turns jitter into extra delay, up to a limit, then packets are discarded as late.
- Packet loss
- RTP sequence gaps. Random loss and burst loss affect codecs differently. 1% is already noticeable on many encodings.
These are not independent. Congestion raises queueing delay and jitter, then tail-drop loss. An oversized jitter buffer hides jitter by adding latency, which in turn disrupts conversational turn-taking.
Latency budgets (ITU-T G.114)
ITU-T G.114 recommends keeping one-way mouth-to-ear delay below 150 ms for most applications. 150-400 ms is increasingly harmful to interactivity (talkers collide, or they insert awkward pauses). Above about 400 ms, delay is generally unacceptable for a conversation, though it may still work for one-way paging. Satellite hops and hairpin recording architectures consume this budget quickly.
| One-way delay | Typical effect on a conversation |
|---|---|
| Under 150 ms | Feels immediate for most talkers (G.114 preferred region) |
| 150-400 ms | Talk-over and delayed responses increase. Still used on long international paths. |
| Over ~400 ms | Half-duplex behavior. Users start saying 'hello?' and colliding. |
A 20 ms packetization plus 20-40 ms jitter buffer plus 10-30 ms encode/decode already spends a large fraction of 150 ms before the WAN. Keep local capture latency low (avoid large OS audio buffers) and do not hairpin media through a distant data center unnecessarily.
MOS and how quality is scored
Mean Opinion Score methods in ITU-T P.800 ask listeners to rate speech on a 1-5 scale (Bad to Excellent). A long-run average around 4.0 is 'good' telephone quality. Objective models estimate MOS without a panel: PESQ (P.862) for narrowband, POLQA (P.863) for narrowband through super-wideband. RTCP XR and vendor MOS estimators on SBCs are approximations. Treat them as trends, not as lab P.800 scores.
R-factor (E-model, ITU-T G.107) is another planning tool that maps delay, loss, and equipment impairment into a transmission rating. Some monitoring tools still display R-factor beside MOS.
Jitter buffers and loss concealment
The receiver cannot play a packet the instant it arrives if the next one might be 30 ms late. A jitter buffer holds packets and emits a steady sample stream. Adaptive buffers grow when jitter is high and shrink when the path is stable, trading delay for smoothness. If a packet is still missing at playout time, the codec's packet-loss concealment (PLC) invents a short substitute. PLC works for isolated losses. It cannot reconstruct 200 ms of silence after a Wi-Fi roam.
Clock skew between sender and receiver is handled by the same buffer (or by asynchronous resampling). Unchecked skew produces gradual buffer underflow or overflow: a slow fade into distortion or a sudden skip.
QoS markings and network treatment
On managed networks, voice RTP is marked Differentiated Services Code Point (DSCP) Expedited Forwarding, decimal 46 (RFC 3246). SIP signaling is often AF31 or CS3. Markings only work if every hop honors them: Wi-Fi WMM, switch trust boundaries, WAN policers. The public internet does not honor your DSCP. There, the available controls are reducing congestion (bitrate, call admission) and choosing better paths (regional media servers, TURN relays near the user for WebRTC).
- Do not let bulk backups share an unmarked uplink with voice during business hours.
- Cap Wi-Fi retries and band-steer voice clients off crowded 2.4 GHz where possible.
- Apply call admission control: a tenth concurrent HD call on a small DSL link will degrade the other nine.
- Measure what the listener hears: RTP loss/jitter from RTCP, plus a recording or MOS sample, not only ICMP ping.
What to inspect when a call sounds bad
| Symptom | Likely layer | Where to look |
|---|---|---|
| Talkers collide, 'you first' | Delay | One-way latency, jitter buffer size, geographic trombone, satellite |
| Robot / warble / slow fade | Jitter or clock | Inter-arrival jitter, buffer underflow, Wi-Fi roam |
| Gaps, syllables missing | Loss | RTP sequence gaps, Wi-Fi, oversubscription, bad PLC on the codec |
| Hollow, tinny, no consonants | Bandwidth / codec | Unexpected G.711 or G.729 instead of HD; transcoding chain |
| Echo | Acoustic or hybrid | Speakerphone, delayed tail, disabled echo canceller, analog FXO |
| Menus ignore keypresses | DTMF transport | In-band through G.729 vs RFC 4733; see DTMF |