🔊 Demo 5 — Text → speech snippet

Goal: in one chat turn, ask the agent to read a paragraph aloud and save the result as a WAV file you can play.

Time to complete: ~5 minutes once the model is downloaded (the first run downloads ~5.5 GB into <Storage:RootPath>\models\qwen-tts\). Tools used: text_to_speech. Surface: Chat (or Tools page → text_to_speech → Direct Invoke).

🎯 Skip the typing — use the built-in template. Open Jobs → Job Templates and pick Text-to-speech snippet (id text-to-speech-snippet). Note the prerequisites — first run downloads ~5.5 GB and the tool requires approval by default.

💡 Why this scenario? It's the simplest possible "approve and run" demo. text_to_speech is gated behind tool approval (RequiresApproval = true), so this is also a quick way to see the approval flow.

Step 1 — Send the prompt#

Web UI → Chat:

Please read the following paragraph aloud using the text_to_speech tool with
speaker="serena" and language="english". After the WAV is generated, give me
the absolute path you wrote to.
 
Text:
"OpenClaw .NET is an agent framework that runs entirely on your machine.
Today we'll show how a single chat turn can produce real audio."

Step 2 — Approve the tool call#

Because text_to_speech requires approval, the agent will pause and show an approval card with the exact arguments it's about to send. Click Approve.

You'll see the synthesis run for a few seconds (CPU-bound), then a result line like:

Saved to: c:\openclawnet\storage\binary\text-to-speech\20260421-160212-345-serena.wav
Speaker: serena
Language: english

Step 3 — Play it#

# On Windows, default audio app:
Invoke-Item 'c:\openclawnet\storage\binary\text-to-speech\20260421-160212-345-serena.wav'

Or open the file in Explorer and double-click it.

Step 4 — (Alternative) Direct Invoke from the Tools page#

Open the Tools page → click the text_to_speech card.
Switch to the Direct Invoke tab.
Click Defaults to pre-fill text, speaker=ryan, language=english.
Edit the text if you want, then click Run.

The output viewer shows the same Saved to: path. Direct Invoke skips the agent and skips the approval card — useful when iterating on the audio itself.

Variations#

Try other voices: vivian, aiden, dylan. The full list is in 20-tools.md#text_to_speech.
Pipe a summary out of Demo 1 — extend that job's prompt with: "Then call text_to_speech with the bullet summary as text so the user can listen to it." Make sure text_to_speech is in the job's Tools allowed list and that auto-approval is enabled for the job (jobs run unattended, so they need approvals turned off for these tools — see 30-jobs.md for the policy).
Combine with Demo 3 — read the winning snippet of a search out loud.

Where to go next#