A jitter buffer schedules against a media clock while absorbing bounded arrival variation

RTP sequence numbers help identify packet ordering and losses, while RTP timestamps represent media sampling time in the payload's declared clock. A packet can contain several codec frames, and gaps or reordered arrivals require explicit handling. RFC 7587 specifies a forty-eight-kilohertz RTP clock for Opus even when the audio representation or decoder output uses another supported rate. Thus a twenty-millisecond media interval advances the timestamp by 960 ticks, not necessarily by the count of samples returned by a sixteen-kilohertz decoder. Handle timestamp wrap and stream restarts under a defined unwrapping policy.

Map the source media timeline into the receiver's playout clock using an origin and rate estimate. A fixed schedule needs a starting cushion, while an adaptive schedule can adjust timing through an explicit policy. Source and sink clock mismatch gradually changes occupancy even on a network with constant delay. Reordering can fix packet order but cannot itself correct that rate mismatch. Long sessions need clock-aware adaptation, such as a suitable time-scale or resampling strategy, with quality and latency measurements rather than arbitrary repeated sample drops.

RFC 3550 defines an interarrival-jitter estimator based on changes in relative transit time, with arrival and media differences expressed in compatible units. It smooths variation; it does not measure an absolute one-way delay or directly prescribe a playout target. A burst, a heavy delay tail, and a gradual clock offset can need different responses even when a recent average statistic looks similar. Evaluate deadline misses and occupancy distributions under realistic traces rather than selecting a buffer solely as a fixed multiple of one smoothed number.

Packets arriving after their useful playout deadline may be discarded, concealed, or handled by another codec-specific policy; waiting for every missing packet can make delay grow without bound. Loss concealment can maintain continuity without reconstructing the exact missing speech. A larger target cushion often reduces late loss while increasing conversational latency, and an overflow policy can create another loss mode during bursts. Measure startup delay, steady-state delay, late discard, concealment, and underflow separately so an average packet-loss statistic cannot hide avoidable interaction stalls.

Replay controlled traces with reordered packets, isolated losses, long bursts, timestamp wrap, clock drift, and a route change. Preserve the original source timestamps when measuring outcomes. Compare actual audible delay and deadline loss, including startup and adaptation periods, and test the consequences for endpoint detection and interruption. Buffer targets that work for prerecorded playback can feel poor in conversation. A correct packet counter alone cannot reveal duplicated decoded frames, discontinuities after concealment, or a buffer that slowly accumulates latency over a long call.

The queue lab isolates a constant-rate mismatch rather than network jitter. Reset, explicitly set BOTH λ=3 and μ=3, predict q(2)=4 from the initial backlog four, then Play or Step four times. Reset and re-set BOTH λ=3.5 and μ=3; predict q(2)=5. Edits retain t and recompute the original fluid trajectory. This model has no packet timestamps, bursts, finite capacity, deadlines, or concealment. A constant mismatch can grow occupancy even with smooth arrivals; actual jitter-buffer behavior needs the richer trace tests just described.