> For the complete documentation index, see [llms.txt](https://docs.concurrence.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.concurrence.com/design/voice-families.md).

# Voice Families as Swappable Runtimes

Why a Concurrence service selects its voice family as a setting, what stays the same when you switch, what changes, and how to compare families yourself.

## The Problem

Speech models differ in latency, voice quality, languages, regions, and the call controls they support, and they change quickly. A workflow written for one speech provider is expensive to move: its prompts, tools, transfers, and test evidence are tied to that provider's session format.

## Our Design

A Concurrence service separates the workflow from the voice runtime. The agent, Context Graph, tools, forwarding block, and caller-verification setting live on the service. The voice family is one setting, `voice_config.session_provider`, with a settings block for that family where it has one. The service's value takes precedence over the agent's, and with neither set the call runs on `concurrence`. `eleven_realtime`, `gpt_live`, and `grok_voice` are selected on the service only.

* `concurrence` - the cascade: separate speech recognition, Concurrence reasoning, and speech generation.
* `gpt_realtime` - a speech-to-speech model on one action state or a graph of 2 to 12 action states.
* `eleven_realtime` - a speech-to-speech session in an ElevenLabs voice on exactly one action state.
* `gpt_live` - realtime speech that delegates tool requests to Concurrence reasoning on multi-state graphs.
* `grok_voice` - a speech-to-speech session in a built-in Grok voice on exactly one action state.

Three parts make the swap safe to attempt:

* **Discovery.** `GET /v1/{workspace_id}/voice-families` lists every family with its availability for the workspace, capabilities, models, and the JSON Schema for its settings block.
* **Save-time refusals.** A save that a family can't run is refused with `422` and a typed reason, for example a multi-state graph on `eleven_realtime`, warm transfer on a family without it, or another family's settings block left in place.
* **Run attribution.** Each run reports `session_provider` and a `manifest` of the versions and models its first session started with, so evidence stays attached to the runtime that produced it.

## What Stays the Same

* The service, its agent and Context Graph versions, and its version sets.
* The tools the graph binds, and the integration endpoints behind them.
* The forwarding block, its approved destinations, and the line's forwarding allowlist.
* The AI and recording notice at the start of every call.
* The run record, including `session_provider` and `manifest`.

## What Changes

| Capability                 | `concurrence` cascade     | `gpt_realtime`            | `eleven_realtime`                | `gpt_live`                                | `grok_voice`                     |
| -------------------------- | ------------------------- | ------------------------- | -------------------------------- | ----------------------------------------- | -------------------------------- |
| Multi-state Context Graphs | Yes                       | 2 to 12 action states     | One action state                 | Yes                                       | One action state                 |
| Dynamic behaviors          | Yes                       | Yes                       | No                               | Yes                                       | No                               |
| Phone transfers            | Cold and warm             | Cold and warm             | Cold                             | Cold                                      | Cold                             |
| Outbound calls             | Yes                       | Single-state graphs       | No                               | No                                        | No                               |
| Operator takeover          | Supported phone calls     | No                        | No                               | No                                        | No                               |
| Regions                    | Subject to data residency | Subject to data residency | Requires `data_residency: "any"` | US only, requires `data_residency: "any"` | Requires `data_residency: "any"` |

The model behind a call's answers also differs. On `gpt_live`, GPT-Live is the voice front end, and Concurrence reasoning answers every delegated request, including tools, policy, and the answer text, with the workspace's `voice.text_model`, or the platform default model when it is unset. On `gpt_realtime`, the realtime model itself answers.

Safety coverage also differs by runtime. The emergency script that cuts off a reply is documented for `gpt_realtime`, and the safe exit and the voice reply checks for platform-managed voice and `gpt_realtime`. GPT-Live and Grok Voice play the opening notice and emergency scripts from prerecorded audio. `gpt_realtime` doesn't read a state's `turn_policy`, `preload`, `wait_for`, `surface_spec_template`, or `channel_overrides`, even when a save accepts them.

## What It Guarantees

* A save whose configuration doesn't fit the service's family is refused with `422` and a reason.
* A run's `session_provider` and `manifest` are read from what its first session started with, not from the service's current configuration.

Language handling follows one rule on every family: a service serves the languages it declares, and a call starts in the agent's default language. A service that serves English and Spanish from an English default runs a bilingual session on each realtime family: it starts in English and follows the caller's clear choice of English or Spanish. The session does not cover a Spanish default, another language pair, or a third served language. See [Languages a Service Serves](https://docs.concurrence.com/developer-guide/platform-api/conversations/voice-agent#languages-a-service-serves).

A successful save doesn't establish that the new family behaves the same as the old one. Model behavior, timing, and voice differ between families, so evidence from one family doesn't carry over to another.

## Verify It Yourself

1. Call `GET /v1/{workspace_id}/voice-families` and read `availability` and `capabilities` for each family in your workspace.
2. On a test service, set `voice_config.session_provider` to another family and save. Read any `422` reason and the fields it names.
3. Place a test call on each family you are considering, with the same synthetic caller and the same request.
4. Read each run's `session_provider` and `manifest`, and compare the transcripts, tool calls, and timing.
5. Run your simulation suite and your safety cases again before you move a production service.

See [Voice Configuration](https://docs.concurrence.com/developer-guide/platform-api/workspaces/voice-configuration#choosing-a-realtime-voice-family) for the full comparison and settings, and [Phone](https://docs.concurrence.com/channels/voice.md) for the call lifecycle.


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation by asking a question.

Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters:

```
GET https://docs.concurrence.com/design/voice-families.md?ask=<question>&goal=<user_goal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.
