Every voice AI conversation is thick with acronyms, and most of them describe simple ideas. Here they are without the jargon, in the order they matter.
The pipeline
STT — speech to text. Turns what the person said into words. Also called transcription or ASR. This is where accents, background noise and code-mixed speech cause trouble.
LLM — large language model. Reads the conversation so far and decides what to say next. The "AI" people usually mean.
TTS — text to speech. Turns that reply into audio. Where the voice comes from.
Pipeline. The three above, chained, running live during a call. Each is a separate service, which is why a vendor can swap one without rebuilding the product.
Timing
Latency. The gap between the person finishing and the agent starting. Under about two seconds feels normal; over three feels broken. Ask what it is measured from — most quoted numbers exclude the largest part.
End-of-speech detection. Deciding the person has stopped talking. Harder than it sounds, because people pause mid-thought, and it is usually the biggest single chunk of latency.
Barge-in. Interrupting the agent while it is speaking, and it actually stopping. The single best indicator of whether a voice product is any good.
Telephony
DTMF. The tones a keypad makes. A "DTMF agent" plays a menu and reads which key was pressed — no speech recognition, so it works on any handset at any signal strength.
IVR. The traditional press-1-for-sales phone menu. Routes calls; does not hold conversations.
Caller ID. The number that shows on the recipient's phone. An asset with a reputation, not a formality.
Concurrency. How many calls can be live at once. Decides whether a batch takes two hours or two days.
Narrowband / 8 kHz. Phone audio quality. Roughly a sixth of what a laptop microphone captures, which is why laptop demos overstate transcription accuracy.
The workflow
Agent. The configured persona: prompt, voice, language, model, behaviour. What you build.
Prompt. The instructions the agent follows. The single largest determinant of call quality.
Knowledge base. Facts the agent may answer from, and a boundary on everything else.
Batch / campaign. A list of people to call. A batch is built by code; a campaign is built from a CSV upload.
Rating template. Your scoring rubric, applied to every call after it ends.
Extraction. Pulling defined fields — notice period, budget, city — out of free speech into columns you can filter.
Webhook. Your system being told the moment a call ends, instead of asking repeatedly.
The one to remember
If you take one term away, take barge-in. A voice agent that cannot be interrupted will fail on real calls no matter how good it sounds in a demo — and it is the easiest thing to test in thirty seconds on your own phone.