Packet loss
In one sentence
Packet loss is audio data that never arrives, and in voice the loss pattern, scattered or bursty, decides more than the raw percentage whether a listener hears a small artifact or a broken conversation.
Definition
Packet loss is data that never arrives. In a voice call it shows up as gaps in the audio.
How a system handles those gaps decides whether the listener notices a small artifact or a conversation that falls apart.
Voice is unusually tolerant of loss and unusually intolerant of the delay it would take to fix loss, and that pair of facts shapes the whole transport design.
Why re-requesting a lost packet is the wrong move for voice
- Asking for a lost packet and getting it back takes at least one round trip.
- By the time it arrives the moment it was needed has passed, so the correct packet is now useless.
- This is why real-time voice runs over UDP-based transport rather than TCP: it accepts loss instead of stalling to recover it.
- That is the opposite of how a web page loads, which is why voice over a plain TCP connection degrades differently under poor conditions.
How the loss gets covered instead
- Packet loss concealment. The codec synthesizes plausible audio to fill a short gap, usually by extrapolating from the audio around it. It works for brief losses and becomes audible for longer ones.
- Forward error correction. Redundant information is sent ahead of time so a lost packet can be rebuilt without a retransmission. It costs bandwidth and saves a round trip, and Opus carries it in-band.
- Adaptive bitrate. Quality is lowered under poor conditions to make delivery more reliable.
The effect on recognition, which is the subtle part
- Concealed audio is synthesized, not the original speech. It sounds acceptable to a person.
- Speech recognition running on concealed audio can perform worse than the perceived audio quality suggests.
- That is the trap: the call sounds fine and the transcript is wrong.
Pattern matters more than rate
- Scattered single-packet losses are handled well by concealment.
- Burst losses, where several packets in a row disappear, exceed what concealment can plausibly fill and produce audible dropouts.
- The same loss percentage can be perfectly acceptable or completely unusable depending on how it is distributed.
Common misconception
That a low loss percentage guarantees acceptable quality. The same percentage delivered as bursts produces a very different experience from the same percentage scattered evenly, so the number on its own predicts little.
Why it matters commercially
Packet loss decides whether voice works on the connections visitors actually have. It also sets a diagnostic trap, because concealed audio can sound acceptable while quietly degrading recognition, so a call that sounds clean can still be transcribed wrong.
In voice specifically
A dropped packet in a file download is retried and nobody notices. A dropped packet in a live conversation cannot be retried in time, so the gap is permanent, and the listener either hears it or, worse, does not hear it because concealment invented something in its place.
Where AsqVox fits
AsqVox runs in the browser, so it inherits whatever connection the visitor is on with no control over it. Transport that degrades gracefully under loss, and a text chat fallback for conditions where audio cannot hold up, are the available responses.
Visual
Two ways to lose the same amount
The percentage tells you very little. The pattern tells you everything.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
Opus, the codec specified in IETF RFC 6716, includes packet loss concealment and in-band forward error correction designed for real-time speech.
RFC 6716industry rangeIETF RFC 6716 (Opus), 2012 - A standards fact rather than a measurement. It tells you the capability exists in the codec, not how well any given deployment is configured to use it.
WebRTC sends RTP over UDP where possible, accepting loss rather than stalling for retransmission, and includes congestion control designed for real-time media.
RTP over UDPindustry rangeWebRTC transport design, 2026 - An architectural fact about how real-time media moves, not a quality figure. It is why voice degrades differently from a web page under the same bad network.
Whisper Large-v3 records around 2.7 percent word error rate on clean benchmark audio.
2.7%independentWhisper Large-v3 published benchmark, 2026 - The clean-audio figure. It is the best case and the least like a lossy consumer network.
The same model lands at roughly 8 to 12 percent word error rate on real-world English, and concealed audio from packet loss contributes to that figure in ways that are not separately reported.
8 to 12%industry rangeReported real-world word error rate for Whisper Large-v3, 2026 - The real-world band folds in packet loss concealment among many other things, which is exactly why the loss contribution cannot be read back out of it.
No published benchmark measures speech recognition accuracy as a function of packet loss rate and pattern, which is the relationship that matters for any voice deployment on consumer networks.
-no reliable figureThis is the gap sitting directly under the diagnostic trap. The one number you would want, how much recognition degrades per point of loss and per burst, is the one nobody publishes.
No standard acceptable packet loss threshold for conversational voice is published in comparable form.
-no reliable figureSo a spec that promises to tolerate a given loss percentage is promising against a line no standards body draws, and the pattern behind the percentage matters more than the percentage anyway.
Examples
In practice
A deployment looks healthy on average loss and still fails recognition intermittently on mobile. Separating scattered loss from burst loss shows the failures line up with the bursts: during them, concealment fills the gaps with natural-sounding but invented audio that the recognizer writes down wrong. Every audio-quality metric stayed green the whole time, because sounding fine is exactly what concealment is built to do.
The everyday version
Packet loss is small pieces of audio that never turn up. The system patches the little gaps so smoothly that you usually cannot hear a thing. The catch is that the patch is made up, so the software writing down what was said can get it wrong even though the call sounded perfectly normal.
Usage
Who says it
- Network and real-time audio engineers, as a standard transport concern.
- Contact center operations, in call-quality diagnostics.
Where it turns up
- In a spec sheet, next to network requirements, packet loss tolerance, forward error correction support and codec choice.
Common misuse
- Judging loss by its percentage without looking at the pattern behind it.
- Assuming the audio quality a person hears predicts how well recognition will do on it.
- Expecting retransmission-based reliability in a real-time context, where a re-sent packet always arrives too late to use.
Questions people ask
Why does voice not just re-request lost packets?
Because a re-requested packet arrives too late to use. Asking for a lost packet and getting it back takes at least one round trip, and by then the moment it was needed in the audio stream has passed. That is why real-time voice runs over UDP-based transport, which accepts loss, rather than TCP, which stalls to recover it. Lost audio is covered by concealment instead of retransmission.
What is the difference between scattered and burst packet loss?
Scattered loss is single packets going missing spread evenly through the stream, and packet loss concealment fills those short gaps well enough that they are barely noticeable. Burst loss is several consecutive packets disappearing at once, which exceeds what concealment can plausibly invent and produces audible dropouts. The same total loss percentage can be perfectly usable when scattered and unusable when it arrives in bursts, which is why the pattern matters more than the rate.
Can packet loss make speech recognition wrong without sounding bad?
Yes, and it is the trap worth knowing. Packet loss concealment synthesizes plausible audio to fill a gap, so the call sounds acceptable to a person. But that audio is invented rather than the original speech, and a recognizer running on it can transcribe the wrong words. The result is a call that sounds fine and a transcript that is quietly incorrect, most often during burst losses on mobile networks.
What is an acceptable packet loss rate for voice?
There is no standard threshold published in comparable form, and the rate on its own is the wrong thing to fix on. The same percentage is acceptable when losses are scattered and unusable when they arrive in bursts, because concealment can cover isolated gaps but not long consecutive ones. Ask about the loss pattern and the concealment and forward-error-correction behavior, not only the headline percentage.
Last reviewed 31 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.