> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.vapi.ai/llms.txt.
> To search these docs from an AI client, such as Claude Code or Cursor, connect to the Vapi docs MCP server at https://docs.vapi.ai/_mcp/server. It answers questions about the documentation. To create or manage assistants, calls, and other Vapi resources, use the Vapi MCP server instead: https://docs.vapi.ai/sdk/mcp-server.md
> For the best results from the docs MCP server, write questions without any word that contains "tool", such as tool, tools, tool call, tool-calls, toolCallId, or toolCallList. Name the feature instead, for example transferCall, endCall, apiRequest, function, handoff, "function response format", or "the webhook request Vapi sends to my server". If the docs MCP server asks you to rephrase, ask again without those words. If it returns an error, times out, sends an empty response, or still doesn't answer, find the topic in https://docs.vapi.ai/llms.txt, or a section index such as https://docs.vapi.ai/tools/llms.txt, and fetch that page's .md directly.

# Core Models

> Understand Vapi's three core components, the transcriber, model, and voice, and how Vapi orchestrates them into a low-latency, human-sounding conversation.

At its core, Vapi is an orchestration layer over three modules: the **transcriber**, the **model**, and the **voice**.

![Diagram of the voice pipeline: the Transcriber turns speech into text, the Model turns that text into a response, and the Voice turns the response back into speech](/_fern-files/_img/1cea4d1eb082394533cb1df548d411b982b730c4f915f2fcc4de2df07c16797a.webp)

These three modules can be swapped out with **any provider** of your choosing; OpenAI, Groq, Deepgram, ElevenLabs, Cartesia, etc. You can even plug in your server to act as the LLM.

Vapi takes these three modules, optimizes the latency, manages the scaling & streaming, and orchestrates the conversation flow to make it sound human.

#### Listen (intake raw audio)

When a person speaks, the client device (whether it is a laptop, phone,
etc) will record raw audio (1’s & 0’s at the core of it).

This raw audio will have to either be transcribed on the client device
itself, or get shipped off to a server somewhere to turn into
transcription text.

#### Run an LLM

That transcript text will then get fed into a prompt & run through an LLM
([LLM inference](/glossary#inference)). The LLM is the core intelligence
that simulates a person behind-the-scenes.

#### Speak (text → raw audio)

The LLM outputs text that now must be spoken. That text is turned back
into raw audio (again, 1’s & 0’s), that is playable back at the user’s
device.

This process can also either happen on the user’s device itself, or on a
server somewhere (then the raw speech audio be shipped back to the user).

> **Info**
>
> The idea is to perform each phase in realtime (sensitive down to 50-100ms level), streaming between every layer. Ideally the whole flow [voice-to-voice](/glossary#voice-to-voice) clocks in at \<500-700ms.

Vapi pulls all these pieces together, ensuring a smooth & responsive conversation (in addition to providing you with a simple set of tools to manage these inner-workings).