ElevenLabs Pro - Text to Speech

This is the node the whole pack is named for, and it's the one you'll reach for when you want ElevenLabs-quality narration inside a ComfyUI graph instead of bouncing between a browser tab and your editor. Text to Speech takes a string, a voice ID, and a model, and returns a decoded AUDIO tensor plus a JSON blob about the call. ElevenLabs' API has every parameter exposed here - which is genuinely rare for a ComfyUI wrapper, most integrations give you five knobs and hide the rest.

The mechanism is straightforward: it POSTs to /v1/text-to-speech/{voice_id} with your voice_settings bundled in, then decodes whatever bytes come back into ComfyUI's AUDIO format (PCM is decoded directly as int16 so it works even without an MP3 decoder; wav/mp3/opus go through soundfile first with a torchaudio fallback). What makes this node worth your time is what it does before the request: it validates text length against the model's limit, and it knows the API's quirks so you don't have to. Pick eleven_turbo_v2_5 or eleven_flash_v2_5 without an explicit language and it stops you before the inevitable 422; eleven_v3 silently forces text normalization to "off" because that's the only value the flagship accepts.

The three inputs you can't skip:

• text - supports v3 expression tags right in the prompt, so [whispers] or [laughs] actually do something.
• voice_id - connect the Voice Selector node's output here (or Fetch Voices / Get Voice By Name).
• model - default eleven_v3. Character limits differ per model: 5,000 for v3, 10,000 for multilingual v2, up to 40,000 for the flash/turbo lines.

After that, the tuning knobs most people actually set: stability (lower = more expressive, higher = more consistent monotone), similarity_boost (closer to the original voice, but too high introduces artifacts), speed (0.5–2.0), and output_format. The formats list has a gotcha baked into its tooltip: mp3_44100_192 and the opus options require Creator tier or better, so on a free plan they'll fail with a permission error that reads like a mystery until you remember tier limits.

There are also two context-only fields, previous_text and next_text, that are easy to misread. They don't generate audio - they give the model the surrounding text so you can chunk a long script without losing pacing. The tooltips are explicit about this, and it matters: people who skip them end up with per-paragraph reads that don't sound like one continuous narration.

Outputs are audio (wire it to Save Audio, a preview, Audio Concat, or any audio consumer) and raw_response_info, a JSON string with model, voice, format, byte count, and character count - handy for logging or a quick cost check.

Install is the usual clone-and-pip, and it's light:

cd ComfyUI/custom_nodes
git clone https://github.com/IxMxAMAR/ComfyUI-ElevenLabs-Pro.git
pip install -r ComfyUI-ElevenLabs-Pro/requirements.txt

Then restart ComfyUI, or use ComfyUI Manager to install "ComfyUI-ElevenLabs-Pro". Only requests and soundfile in requirements - no models to download, the synthesis happens on ElevenLabs' side. You'll want an API key in the node or the ELEVENLABS_API_KEY env var.

The big gotcha, same as every paid node in this pack: it's billed per character, and it's cached by input hash so identical inputs don't re-bill on re-queue. That cache was the whole point of the v2.1 rewrite - v2.0 re-billed you on every single queue. Change any input and you pay again, so make your text edits before you hit queue, not after.