Voice AI glossary

AI judge

A language model that reads or listens to a finished call and scores it against a written rule, quoting the line that decided the score.

Example: "Fail. The agent booked the visit before reading the address back, at 00:27." See also rubric and judge agreement.

Backchannel

A short sound a listener makes to show they are following, such as "mm", "yeah" or "right". It is not a request to take the turn. An agent that stops every time it hears "mm" sounds jumpy.

See also barge-in. Read how voice agents handle interruptions.

Barge-in

The caller starting to speak while the agent is still talking. A good agent stops quickly and remembers only what the caller actually heard.

See also backchannel and turn detection.

Chained pipeline

A voice agent built from separate parts in a row: turn detection, speech-to-text, a language model and text-to-speech, with text passed between them. Read speech-to-speech or a chained pipeline.

Concurrent calls

How many calls run at the same moment. A run of 40 calls at 10 at a time takes about four call lengths to finish.

Diarisation

Working out who spoke when in a recording, so each stretch of speech is labelled caller or agent.

Endpointing

Deciding that the caller has finished speaking. Set it too eager and the agent cuts people off. Set it too patient and every reply feels slow. A caller reading a policy number in two halves is the classic victim of eager endpointing. On Twilio's <Gather> the default input timeout is 5 seconds, and speechTimeout="auto" stops recognising speech at the first pause, so test it with callers who pause mid-number (Twilio: Gather, read 28 October 2026).

See also turn detection.

Evidence report

The RingTru report for a run: a score for each behaviour, the quoted line behind every failed mark, and ranked findings with a suggested change.

G.711

The standard codec for telephone audio, set by the ITU. It carries 8,000 samples a second at 64 kbit/s. The UK and Europe use the A-law variant, North America and Japan use mu-law. See also narrowband audio.

Hallucination

The agent stating something as fact that is not in its instructions, tools or the caller's words. Example: telling a caller their policy covers a hotel stay when no such cover exists.

Human handover

Passing the call to a person. A warm handover also passes on what the caller has already said, so they do not have to repeat it.

Judge agreement

How often an AI judge gives the same mark as a trained person on the same calls. It is measured on a sample, and it should be published with the method and the sample size.

Large language model (LLM)

The model that decides what the agent says next. In a chained agent it reads and writes text. In a speech-to-speech agent it works on audio directly.

Narrowband audio

Audio sampled at 8 kHz, which keeps sound up to about 3.4 kHz. It is what most phone calls carry, and it loses detail that speech models rely on, such as the difference between "f" and "s". See also G.711.

p50 and p95

The median and the 95th percentile. A p95 latency of 1.8 seconds means one reply in twenty took longer than that.

Persona

The description of a simulated caller: who they are, how they speak, what line they are on and what they want from the call. Example: "Margaret, 71, Lancashire accent, radio on, reporting a burst pipe."

Regression set

Calls that once failed, saved so they can be rerun after a change to check the fault is fixed and stays fixed.

Rubric

The written rule a judge scores against, with what counts as a pass and a fail. Example: "Before booking anything, the agent reads the address and postcode back and waits for a yes."

Scenario

One test situation: the caller's goal, the persona and the rules the agent must keep. A run plays each scenario many times with different callers.

Signal-to-noise ratio (SNR)

How much louder the voice is than the background, in decibels. At 20 dB speech is clear. At 5 dB the noise is nearly as loud as the caller.

SIP

Session Initiation Protocol, the common way to set up voice calls over the internet. A SIP trunk connects a voice platform to the ordinary phone network.

Speech-to-speech model

A model that takes the caller's audio in and produces the agent's audio out, with no separate transcription or voice step in between. Read speech-to-speech or a chained pipeline.

Speech-to-text (STT)

Turning speech into text. Also called automatic speech recognition, or ASR. Streaming STT returns words while the caller is still talking. See also word error rate.

Text-to-speech (TTS)

Turning the agent's reply into a spoken voice. Streaming TTS starts speaking before the whole reply is written.

Time to first audio

How long a voice service takes to return the first sound after it receives text or audio. It is one part of the delay a caller hears, not the whole of it. See also voice-to-voice latency.

Tool call

The agent asking another system to do something, such as look up a policy or book an assessor. Tool calls are where a wrong word turns into a wrong action.

Turn detection

Deciding whose turn it is to speak. Simple versions wait for silence. Newer ones also listen to the words and the tone to judge whether the caller sounds finished. See also endpointing and voice activity detection.

Voice activity detection (VAD)

A small model that tells speech apart from silence and noise, many times a second. Turn detection is built on top of it.

Voice-to-voice latency

The gap the caller hears, from the end of their speech to the agent's first sound. It is the latency figure that matters most on a call.

Word error rate (WER)

The share of words a transcript gets wrong: words swapped, missed or added, divided by the number of words actually said. Results depend on how the text is normalised first, so compare WER only under the same rules. Example: 3 errors in a 50-word answer is a WER of 6 percent.