Skip to content

Text to Speech

Use Text to Speech when the agent should turn a written answer into spoken audio during the conversation.

This tool is useful when the user asks to listen instead of read, when the answer will be easier to review by ear, or when your channel experience is designed around audio follow-up.

When Text to Speech Is the Right Tool

Use it when:

  • The user explicitly asks for an audio version of an answer
  • The answer is stable enough to read aloud
  • Hearing the response is more helpful than reading it

Do not use it when:

  • The agent is still gathering facts or deciding what to say
  • The user only needs a normal text answer
  • The spoken content would include secrets, payment data, or other sensitive details that should not be turned into audio

Step 1: Add Text to Speech

In Editor, open Tools, click Add Tool, and choose Text to Speech.

Step 2: Choose a Speech Model

Select the ElevenLabs speech model the agent should use.

If the model list is empty, the workspace or environment may not have ElevenLabs configured yet. Ask an admin to confirm the provider setup before testing the tool.

Step 3: Choose Language Behavior

Leave Language on auto-detect unless you want to pin the spoken language.

Pinning a language is useful when:

  • The agent should always speak in one target language
  • The user may type in multiple languages, but the audio should stay consistent
  • You want to reduce surprises in a teaching, coaching, or support flow

Step 4: Choose a Voice

Use Voice Gender only to narrow the list. The final behavior depends on the specific Voice you choose.

After choosing a voice, play the preview if one is available. A preview sample may be in another language even when the multilingual model can speak your selected language.

Step 5: Write a Narrow When to Use

Treat speech generation as a deliberate output mode, not a default response style.

Example:

Use this tool when the user asks to hear the answer as audio, or when the agent has finished a stable explanation that would be easier to review by listening. Do not use it while collecting sensitive information or when a normal text reply is enough.

That works because it defines:

  • The user requests that should trigger audio
  • The point in the conversation when audio is safe
  • The situations where the tool should stay off

Step 6: Test with Audio-Specific Prompts

Try prompts such as:

  • Can you read that answer out loud for me?
  • Please give me an audio version of the next-step summary.

Then confirm:

  • The tool is called only when audio is clearly useful
  • The selected voice and language match the expected experience
  • The agent still provides the right text answer before or alongside the audio behavior

Operator Tips

  • Keep the trigger narrow so the agent does not generate audio for every answer.
  • Pick a voice that matches the service tone before sharing the agent with real users.
  • Do not use the tool for one-time secrets, verification codes, payment card details, or other content that should stay text-only.
  • If the user sends an audio attachment, the agent's LLM model must still support audio input. This tool controls generated speech output; it does not make every model able to understand audio input.

Common Mistakes

Using speech as a default reply style

If every answer becomes audio, the conversation can feel slower and harder to scan. Use audio only when the scenario benefits from it.

Picking a voice without testing it

Always preview the voice when a sample is available, then test a real prompt before publishing.

Confusing audio output with audio input

Text to Speech creates spoken audio from text. Audio uploads and audio attachments still depend on the model and channel support described in LLM Model Selection.

Next Steps