Latency and performance

Character error rate

Character error rate (CER) is transcription accuracy measured over individual characters rather than whole words, using the same substitution, deletion and insertion formula as word error rate, and it is the right metric where word boundaries are unclear or where exact spelling matters.

End-to-end latency

End-to-end latency is the total elapsed time across a system from one named point to another, and in voice it usually means from the moment a visitor stops speaking to the moment they hear a reply, a number that means nothing until its start and stop points are stated.

Interruption handling latency

Interruption handling latency is how quickly a voice agent stops its own speech once a person starts talking over it, targeted at roughly 60 milliseconds and judged about an order of magnitude more strictly than the delay before it answers.

Jitter

Jitter is variation in the timing of arriving packets, and because audio playback needs a steady stream, a packet that arrives late is as damaging as one that never arrives at all, since the moment it was needed has already passed.

Mean opinion score

Mean opinion score (MOS) is a 1 to 5 quality rating for synthesized or transmitted audio, produced by asking human listeners to score samples and averaging their answers, where 5 is excellent.

Naturalness score

A naturalness score rates how human synthesized speech sounds on a roughly 1 to 5 scale averaged from listener ratings, and it has become a weaker way to choose a voice as leading systems cluster near the human ceiling of 4.5 to 4.7.

Packet loss

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.

Percentile latency

Percentile latency describes how latency is distributed rather than averaged, so a p95 figure means 95 percent of requests were faster than it and the slowest 5 percent, the turns a visitor actually remembers, were slower.

Real-time factor

Real-time factor (RTF) is how long a system takes to process audio divided by the duration of that audio, so a value below 1 is faster than real time and a value above 1 is falling behind, and it governs cost and concurrency rather than conversational responsiveness.

Round trip time

Round trip time (RTT) is how long a packet takes to reach a server and come back, the network floor beneath every other latency number, and the one component no amount of software optimization can push below.

Silence threshold

The silence threshold is the amount of trailing quiet a voice agent waits for before deciding the speaker has finished, usually set between 150 and 300 milliseconds, and it is the single most-felt setting in the pipeline because it is paid on every turn.

Time to first audio

Time to first audio (TTFA) is the end-to-end wall-clock delay from the moment a person stops speaking to the moment the reply becomes audible, and it decides more than any other number whether a voice agent feels alive.

Time to first byte

Time to first byte (TTFB) in voice AI is how quickly a speech synthesizer starts emitting audio after it receives text, which is the moment the listener's wait ends, and it shares its name with an unrelated web-performance metric that has to be qualified apart from it.

Time to first token

Time to first token (TTFT) is how long a language model takes to produce its first piece of output after a request, measuring the model alone rather than the full pipeline, which makes it one component of time to first audio and the most frequently misquoted latency number in voice AI.

Word error rate

Word error rate (WER) is the standard measure of transcription accuracy: the words a speech recognition system got wrong, counted as a percentage of the words that were actually said, where lower is better.