VoIP jitter

VoIP jitter is the variation in arrival times of voice packets over an IP network. In a healthy voice call, packets are sent at fixed intervals — typically one every 20 milliseconds — and should arrive at the same cadence. When packets arrive at inconsistent intervals, the receiver experiences jitter. Enough jitter and the audio becomes choppy, delayed, or drops out entirely.

Jitter matters for AI voice systems specifically because automatic speech recognition and downstream voice agent logic depend on clean, continuous audio. A jittery stream forces the ASR model to work harder to reconstruct phonemes, increases word error rate, and can cause endpointing logic to fire early or late.

How jitter is measured

Jitter is not just variance in absolute latency — it is the difference in latency between consecutive packets. The standard formula (RFC 3550) computes an exponentially weighted running average of the difference in transit delays between successive packets. In practice, this produces a millisecond value that most VoIP monitoring tools report as "current jitter."

A healthy VoIP call has average jitter under 30 milliseconds. Anything below 20 ms is inaudible to callers. Between 30 and 50 ms, sensitive listeners start to notice occasional glitches. Above 50 ms, the call sounds broken to most people, and ASR accuracy degrades noticeably.

What causes jitter

Jitter has three main sources. Network congestion is the most common — when a router's queue fills, some packets are held longer than others before forwarding, so their delay varies. Wi-Fi and cellular last-mile links are second, because wireless spectrum is inherently variable and retransmissions from lost frames introduce microsecond-to-millisecond gaps. Path changes are the third: when a route flips because of a link failure or BGP update, packets in flight can take different routes with different delays.

Codec choice can amplify or mask jitter. A codec with a longer frame size like G.711 at 30-millisecond frames experiences fewer but larger jitter events, while a codec with shorter frames like Opus at 5 or 10 ms experiences more frequent but smaller events. Choice of audio codec and network conditions interact strongly.

Jitter buffers

The primary defense against jitter is the jitter buffer — a small block of memory on the receiving side that holds arriving packets for a configurable amount of time before releasing them to the decoder in order. A larger buffer smooths out more jitter but adds latency; a smaller buffer preserves latency but drops or plays out-of-order packets that arrive late.

Modern VoIP endpoints use adaptive jitter buffers that expand under adverse network conditions and contract when the network stabilizes. The buffer size is typically expressed in milliseconds and ranges from 20 to 200. For an AI voice agent optimized for low latency, an initial buffer of 40 to 60 ms strikes a workable balance between call quality and responsiveness.

Why AI voice agents are especially sensitive

Human listeners tolerate more jitter than machines. A brief gap or micro-glitch is auditorily annoying but the listener's brain reconstructs the missing sound from context. An ASR model has no such benefit — a corrupted phoneme is a corrupted phoneme, and the transcription downstream will reflect that. Studies of production voice AI systems show WER degradation of 15 to 40% when jitter exceeds 50 ms compared to jitter under 20 ms on the same audio content.

The other AI-specific concern is endpointing. If a jittery stream causes a packet to arrive 100 ms late, the endpointing model may see it as a natural pause in speech and cut off the caller mid-sentence. This produces one of the most frustrating voice AI failure modes: the agent starts talking over the caller who was in the middle of an important word.

Mitigation in production

The most durable jitter mitigations are network-level rather than application-level. Prioritizing voice traffic with QoS markings, minimizing hops in the media path, and routing calls to media servers geographically close to callers all reduce jitter far more than any tuning of the jitter buffer or codec parameters. For AI voice teams, the most cost-effective jitter reduction is often selecting a telephony provider with a modern IP backbone and edge presence in the caller's region.

Application-level tuning still matters at the margins. Voice AI teams should measure jitter continuously per call, alert on sustained values above 30 ms, and correlate jitter metrics with WER and endpointing metrics to catch quality regressions early. A jitter spike that lasts more than a few seconds is nearly always a network-side problem that no amount of buffering will fully solve.