<!-- source: https://docs.dreamlake.ai/agents/read-aloud @ 0.1.192 b2f973b1 -->
# Read agent replies aloud

Read aloud is available for built-in agent replies. The matching agent speech
endpoint and chat UI must both be deployed to enable playback.

Choose **Read aloud** beneath an assistant reply to listen. This opens inline
playback options for speaker voice, speech model, speed and volume. Options stay
open after stopping; closing them also stops playback. Voice/model choices share
Memo preferences and take effect on the next playback, stopping any current reply.
Speed (0.75×–2×) and volume apply immediately and to following chunks. Speed uses
Memo's pitch-preserving player and returns to 1× if unavailable. The browser-local
speed preference is shared with Memo playback. A voice catalog arriving after
playback starts does not change the active selection. If loading options fails,
playback remains available with the default selection and the panel offers a retry. Playback starts only
when you choose it. **Stop** cancels playback and the pending synthesis request.
Starting another reply stops the previous one, including in another agent panel.

You can start while a reply is being generated. Complete Markdown paragraphs
are read in order; the final incomplete paragraph is read when generation ends.
A long paragraph without a blank line waits until it finishes. Lists and blockquotes,
including the remainder of that reply, wait until generation ends so nested
paragraphs retain their Markdown context. This is streaming
playback by paragraph, not speech for each arriving token. After a paragraph's
audio is played, its buffer is released before the next chunk is synthesized.

The reader converts Markdown to spoken text, keeps link labels, and skips fenced
code and media. User messages, reasoning, tool receipts and system notices have
no read-aloud action. Switching conversations, closing the panel, opening the
session chooser or changing the signed-in identity stops playback. Returning to
the conversation does not restart it.

Preparing and waiting states appear beside the control. If synthesis fails,
the text remains available and the same action can retry from the beginning.
An AudioContext-capable browser and configured speech service are required.

## API boundary

`POST /agent/sessions/:id/speech` accepts `text` of 1–1500 characters.
`GET /agent/sessions/:id/speech/options` returns the shared Memo voice/model
catalog with the same private-session authorization. Speech POST responses return
streaming NDJSON containing 24 kHz signed 16-bit mono PCM, matching Memo speech.
The UI uses Memo's default or remembered voice/model selection. Choosing
**Server default** omits that override; without a configured server voice,
agent playback uses Memo's default Chris voice. Provider credentials remain
on the server. The existing optional `voiceId`, `modelId`, `documentText` and
`start` fields use the same validation as Memo speech.

Every request, including a cache hit, requires an active signed-in user with
built-in-agent team access, ownership of the private built-in agent session,
and its active personal namespace. Another user's conversation is unavailable.
The cache is scoped separately from Memo recordings. Both surfaces share the
same process-local per-user limits: two active requests and 60,000 requested
characters per hour; provider-side quotas remain the global ceiling. Responses
are `no-store`; closing the client connection aborts the provider stream.

Playback errors do not trigger another agent turn or repeat an agent action.
