A voice agent is three systems in a chain. Understanding the chain explains most of
what feels wrong on a bad call.
The chain
Speech to text turns what the caller said into words.
A language model decides what to say back, following your prompt.
Text to speech turns that reply into audio.
Each link takes time, and the caller experiences the total as the pause before the
agent answers. Every provider choice you make is a trade between how good a link is
and how fast.
Latency is the thing callers notice
People forgive an agent that says something slightly wrong. They do not forgive an
agent that leaves a two second silence, because silence on a phone call reads as a
dropped connection.
Fluents records timings per link on every call, so when an agent feels sluggish you
can see which part of the chain is responsible rather than guessing. See
Calls.
If an agent has to do something slow, such as calling your own systems mid
conversation, let it say something first. A filled pause is fine. An unexplained one
is not.
Knowing when the caller has finished
The agent has to decide when a caller has stopped speaking rather than merely paused.
Too eager and it interrupts. Too patient and it feels unresponsive.
There is a related question of what to do when nobody says anything at all. The agent
waits, checks in, and eventually ends the call rather than holding an open line
forever.
Both are tunable. See Agents.
Calls have a ceiling
A call has a maximum length, so a conversation that goes nowhere cannot run
indefinitely and bill you for it.
Separately, each environment has a concurrency limit, which is how many calls can
run at once. This is the constraint people hit and misdiagnose most often. A campaign
that appears to be dialing slowly is usually queued against this number rather than
malfunctioning. See Settings.
Models are a choice, not a given
You can change which language model runs the conversation. Larger models reason better
about complicated instructions. Smaller ones reply faster.
For most phone conversations, faster wins. The reasoning demands of a well written
prompt are usually modest, and the latency cost of a large model is paid on every
single turn.
Recording and privacy
Recording is a setting, not a default, and it interacts with your obligations rather
than replacing them. Compliance modes restrict what is retained.
Compliance settings restrict what Fluents stores. They do not tell you what you are
legally required to do, and they are not a substitute for consent where consent is
required.
Where to go next
