OpenClaw TTS: The Voice Automation Nobody's Talking About

Voice automation sits mostly unused in AI agent setups — not because it's hard, but because nobody explains what it's actually good for. The TTS skill in OpenClaw isn't about making your chatbot sound human. It's about creating audio output channels that don't require screen time: a spoken briefing in your earbuds during a commute, a voice alert when a trading threshold triggers, a podcast episode narrated from a draft without any manual recording.

Why TTS Transforms Agent Output

Text output has a ceiling. It requires the recipient to be looking at a screen, actively reading. Voice output removes that requirement entirely. It reaches people when their eyes are busy, when they're driving, exercising, cooking, or in a meeting where pulling up a screen isn't appropriate.

The teams getting the most out of OpenClaw TTS aren't building fancy voice interfaces. They're doing three specific things: morning briefings delivered as audio to Telegram while getting ready for work, alert notifications that speak the critical metric aloud instead of sending a silent push notification, and content repurposing that converts written articles into listenable audio without recording a single word manually.

Here's what we've seen consistently across production deployments: the agents that include voice output channels have dramatically higher engagement than those that don't. A spoken alert at 9 AM gets responded to within minutes. A Telegram text notification from the same agent gets read three hours later if you're lucky.

We'll get to the exact configuration in a moment — but first you need to understand the provider tradeoffs, because they dictate which use cases are cost-viable at your volume.

Supported TTS Providers and Tradeoffs

Three providers cover the full quality-cost spectrum for OpenClaw TTS integrations.

ElevenLabs produces voice quality that's indistinguishable from a human narrator in many cases — ideal for customer-facing audio, podcast narration, or any use case where the listener's experience matters. The cost is 20x higher than OpenAI TTS, so reserve it for high-value output.

OpenAI TTS is the practical default for most internal automation. The tts-1-hd model produces excellent quality at a fraction of ElevenLabs' price. The standard tts-1 model is faster and cheaper — use it for notifications and briefings where a natural-but-not-perfect voice is fine.

Google Cloud TTS has the deepest language support — over 220 voices across 40+ languages. If your use case involves multilingual voice output, Google is the practical choice. WaveNet voices are high quality; the standard voices are noticeably robotic at longer text lengths.

Skill Configuration

Add the TTS skill to your agent's agent.yaml under the skills block. Here's a complete configuration with multiple provider options:

# agent.yaml
name: voice-agent
model: gpt-4o
system_prompt: |
  You are a voice-enabled assistant. When asked to narrate, announce,
  or send audio, use the TTS skill to generate audio output.
  Always confirm audio generation before sending.

skills:
  tts:
    enabled: true
    providers:
      openai:
        api_key: ${OPENAI_API_KEY}
        model: tts-1-hd
        default_voice: alloy
        speed: 1.0
      elevenlabs:
        api_key: ${ELEVENLABS_API_KEY}
        voice_id: "EXAVITQu4vr4xnSDxMaL"
        model: eleven_multilingual_v2
        stability: 0.5
        similarity_boost: 0.75
      google:
        credentials_file: ${GOOGLE_APPLICATION_CREDENTIALS}
        language_code: en-US
        voice_name: en-US-Neural2-D
    default_provider: openai
    output_dir: ./outputs/audio/
    output_format: mp3
    cache_audio: true
    cache_ttl_hours: 24
    max_chunk_chars: 4096

The cache_audio setting is critical for cost control. With caching enabled, the agent checks whether it has already generated audio for the exact same text within the TTL window before making a new API call. For daily briefings and standard notification templates, this cuts your TTS API spend by 60–80%.

The max_chunk_chars setting controls how the agent handles text longer than the provider's per-request character limit. Set it to 4,096 for OpenAI TTS. The agent automatically splits longer text at sentence boundaries, generates audio for each chunk, and concatenates the results into a single output file.

Real Use Cases for OpenClaw TTS

Four use cases consistently deliver concrete value in production setups as of early 2025.

Morning briefings via Telegram. The agent runs at 7 AM, pulls your calendar events, top emails, key metrics, and any overnight alerts, compiles them into a two-minute briefing script, generates the audio with OpenAI TTS, and sends it to your Telegram as a voice message. You listen while getting ready. No screen required.

Alert narration for trading and monitoring. When a price threshold triggers, a server goes down, or a key metric crosses a boundary, the agent generates a spoken alert. "BTC crossed 98,000. Portfolio value up 12%. Three positions near stop-loss." That spoken alert gets your attention faster than a silent push notification.

Content narration for blogs and newsletters. Feed a draft article to the agent and it produces a narrated MP3 suitable for embedding as a podcast episode or audio version of the post. At ElevenLabs quality, the output is genuinely listenable. At OpenAI TTS quality, it's acceptable for audiences who prioritize access over audio fidelity.

Podcast automation for AI-generated content. Pair the TTS skill with a content generation agent that drafts scripts based on RSS feeds, trending topics, or your note files. The pipeline runs end-to-end: topic research → script drafting → TTS generation → file delivery. One trigger, one podcast episode, zero manual work.

Production Setup and Delivery

Generating audio files is only half the job. Getting those files to the right person at the right time through the right channel is where production setups diverge from toy demos.

The most practical delivery chain for personal use: TTS skill generates the MP3, the telegram skill sends it as a voice message. The Telegram API accepts voice messages up to 50 MB — enough for 30–40 minutes of speech. For longer content, split into parts or upload as a document instead.

For team use, an S3 delivery chain works better. The agent generates audio, uploads to S3 with a public or presigned URL, and sends the URL to a Slack channel or email list. Recipients click through to listen in their browser without needing any special app.

For podcast distribution, the pipeline needs an RSS feed update after each new audio file lands in your media host. OpenClaw can trigger this automatically — generate the audio, upload to your media host via the file skill, then call your podcast platform's API to publish the new episode. From trigger to published episode: under 10 minutes, zero manual steps.

Common Mistakes in OpenClaw TTS

• Not chunking long text — OpenAI TTS has a 4,096-character per-request limit. Without chunking, requests for long articles fail silently. Set max_chunk_chars and let the agent handle splitting automatically.
• Using ElevenLabs for every request — ElevenLabs costs 20x more than OpenAI TTS per character. Use it only where quality justifies the cost. Route internal notifications through OpenAI TTS.
• Skipping audio caching — if your agent sends the same briefing template daily, you're paying for the same TTS generation every single day. Enable cache_audio and set a 24-hour TTL for templates that don't change frequently.
• Not handling TTS failures gracefully — when the TTS API fails, the agent should fall back to sending the text version of the message rather than failing silently. Build explicit fallback logic in your agent's system prompt.
• Ignoring speaking rate for long content — the default speed of 1.0 can feel slow for informational content. Many users prefer 1.15–1.25x for briefings. Set speed in your provider config to match your listeners' preferences.

Frequently Asked Questions

Which TTS providers does OpenClaw support?

OpenClaw supports ElevenLabs, OpenAI TTS (tts-1 and tts-1-hd), and Google Cloud Text-to-Speech as of early 2025. Each is configured as a provider under the tts skill block in agent.yaml. You can set a default provider and override it per request when needed.

How do I configure the TTS skill in OpenClaw?

Add a tts block to your agent's skills section in agent.yaml. Specify provider, API key via environment variable, default voice, output format, and output directory. The agent then generates audio from any text it produces or receives on request.

What is the best TTS provider for voice notifications?

OpenAI TTS is the best starting point — fast, cheap at $0.015 per 1,000 characters, and requires no additional credentials beyond your existing OpenAI key. Use ElevenLabs when voice quality is critical for customer-facing audio or professional podcast narration.

Can OpenClaw narrate blog posts or documents automatically?

Yes. Feed text content to the TTS skill via a pipeline trigger. The agent splits long documents into chunks under 4,096 characters, generates audio per chunk, and concatenates results into a single MP3. The entire process runs without human involvement once configured.

How do I send TTS audio files through Telegram?

Enable both the tts and telegram skills in your agent config. When the agent generates a TTS file, chain the telegram skill to send it as a voice message using telegram.send_voice with the generated file path. Recipients hear a playable audio message directly in Telegram.

What audio formats does the TTS skill output?

MP3 by default, compatible with all major platforms. WAV is available for workflows requiring uncompressed audio. Some providers also support OPUS for lower file sizes. Set the output_format key in your skill config to specify your preferred format.

How do I keep TTS costs low in a high-volume setup?

Cache generated audio for frequently repeated text. Use OpenAI TTS tts-1 (not tts-1-hd) for internal notifications where quality is secondary. Reserve ElevenLabs for customer-facing audio where realism matters. Caching alone cuts costs by 60–80% for briefing-style workflows.