Voice AI glossary

AI receptionist: A voice agent set up to answer a business's phone line. It handles greetings, FAQs, message taking, appointment booking and call transfers. It's usually sold as a monthly plan with a set number of minutes or calls included.
Voice agent platform: Software for building and running AI phone or web voice agents. Developer platforms expose APIs. No-code platforms offer visual builders.
STT (speech-to-text): Turns the caller's speech into text for the language model. It's often billed per audio minute; Deepgram's streaming models, for example, cost under a cent per minute.
TTS (text-to-speech): Turns the agent's reply into audio. It's often the largest model cost per minute, and premium or cloned voices cost more.
LLM: The large language model that decides what the agent says and which tools it calls. It's billed per token, so cost per minute depends on prompt length and how much the agent talks.
Speech-to-speech model: A model that takes audio in and produces audio out directly, without separate STT and TTS steps. Examples include OpenAI's realtime models and Ultravox.
Platform fee / hosting fee: What the voice platform itself charges per minute, separate from model and phone costs. Vapi lists $0.05/min, for example.
Pass-through pricing: Model or telephony costs the platform bills at the provider's own rate, or that you pay to the provider directly with your own API keys.
All-in per-minute rate: A single rate per minute that bundles platform, models and sometimes telephony. Check exactly what's included.
Telephony: The phone network side: phone numbers plus per-minute carrier charges for inbound and outbound calls. Twilio lists US local inbound at $0.0085/min, for example.
SIP trunking: A standard way to connect a voice agent to an existing phone system or carrier, instead of buying numbers from the platform.
WebRTC: Browser and app real-time audio. It lets users talk to an agent without a phone call, and is often cheaper than phone minutes.
Concurrency: How many calls can run at the same time. Plans often cap it, and extra concurrent lines can cost more.
Latency: The delay between the caller finishing speaking and the agent starting to reply. Lower feels more natural; under about a second is a common target.
Barge-in / interruption handling: Lets the caller interrupt the agent mid-sentence and have it stop and listen.
Turn detection: Working out when the caller has finished speaking. Good turn detection avoids talking over people and awkward pauses.
Warm transfer: Handing a live call to a human, usually with context passed along. Some vendors bill transfer time separately; Bland lists $0.04–0.05/min, for example.
BAA (Business Associate Agreement): A contract required under HIPAA when a vendor handles protected health information for a US healthcare business. Get one signed before handling patient calls.
Overage: The per-minute (or per-call) rate charged once you go past the minutes or calls included in a monthly plan.
Knowledge base / RAG: Documents the agent can look up during a call, such as FAQs, policies and price lists. RAG stands for retrieval-augmented generation.