> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cobalt.peoplereign.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Voice

> Dictation, spoken conversations, and read-aloud — where voice works and how to control it.

# Voice

Voice in Cobalt is three separate things, each independently available:

| Capability         | What it is                                                                                              | Where it works                              |
| ------------------ | ------------------------------------------------------------------------------------------------------- | ------------------------------------------- |
| **Dictation**      | Speak instead of type — the recording is transcribed into the composer for the user to review and send. | Web widget (and full-screen), admin Preview |
| **Voice mode**     | A real-time spoken conversation — the user talks, the agent talks back, live.                           | Web widget (and full-screen), admin Preview |
| **Read-aloud**     | The agent's written answer, spoken on demand.                                                           | Web widget, admin Preview                   |
| **Voice messages** | Slack voice clips, transcribed and answered as text.                                                    | Slack channel                               |

## The widget experience

In the [web widget](/channels/web-widget), dictation and voice mode sit in the
composer, and each answer offers read-aloud. Read-aloud uses the channel's
**preferred voice**, which you pick in the channel's branding settings. Voice
mode needs microphone permission from the browser — on embedded intranets
(SharePoint), the page must allow microphone access; see the
[SharePoint guide's troubleshooting](/channels/sharepoint) if the mic prompt
never appears.

## Slack voice messages

Slack's own voice clips are supported on the [Slack channel](/channels/slack):
when transcription is enabled in the channel settings, an employee can send a
voice message and the agent answers it as a normal text turn. Long clips over
the size cap are declined with a short explanation rather than silently
dropped.

## Anonymous visitors and abuse control

On anonymous widget channels (no sign-in), voice is rate-limited so a public
embed can't be abused:

* **Per-visitor limits** on dictation, uploads, and voice-mode sessions.
* **A per-tenant daily cap** on anonymous voice-mode usage — when the day's cap
  is reached, voice mode pauses until the next day while text chat continues
  normally.

Signed-in channels (embedded token, federated identity) aren't subject to the
anonymous caps.

## Usage

Voice turns are metered like any other conversation turn — one credit per
answered request, regardless of whether it was typed or spoken. See
[Billing & usage](/operate/billing-and-usage).