elevenlabs/text-to-speech-multilingual-v2

Local TTS got genuinely good - Chatterbox and F5-TTS run under 8GB and VibeVoice handles long multi-speaker reads - and for a lot of narration work that's now enough. What hasn't changed is the failure mode: a mispronounced product name in a client video, or a language your local model half-speaks. ElevenLabs Multilingual v2 is the expensive, boring, reliable answer, and this node is that model with all of its tuning knobs exposed.

It's the fullest node in the pack's Text to Speech folder. If you only want "type text, get voice," the Turbo 2.5 node next to it is simpler; this one is for when you're willing to trade a little latency for control over how the read lands.

How it works

You give it text and a voice; the node assembles a request for KIE's elevenlabs/text-to-speech-multilingual-v2 model, submits it as a hosted task, polls until it finishes, then downloads the audio and decodes it into ComfyUI's standard AUDIO object. That decode is why the output wires into Save Audio or any audio-driven video model without a converter in between. Nothing runs on your GPU here - you're paying per call, and the node reports what each one cost.

The inputs that actually matter

text is required, capped at 5000 characters. voice is required too, and it's a 68-entry dropdown of preset names and IDs - Rachel, Adam-style names plus raw IDs like N2lVS1w4EtoT3dr4eOWO (Callum, "Husky Trickster" in the tooltip's roster). The tooltip's roster is worth reading once; it's the only place the pack tells you what each voice sounds like. Same preview trick as the rest of the family: https://static.aiquickdraw.com/elevenlabs/voice/<voice_id>.mp3.

Then the four dials, all floats with sensible defaults:

• stability (0.5 default) - lower is more expressive and more variable, higher is more consistent and flatter. It's a 0–1 float here, unlike the three-position dropdown on the dialogue node.
• similarity_boost (0.75) - how tightly it clings to the reference voice. Push it up for a known voice, down if it starts sounding strained.
• style (0) - exaggeration. At 0 you get a clean read. Above that it starts performing, which is either exactly what you want for a character or exactly wrong for a corporate VO.
• speed (0.7–1.2) - keep it near 1.0. Extreme values "may affect quality," per the author's tooltip, and 1.2 sounds like a podcast ad read.

Two more worth knowing: previous_text and next_text are continuity hints. If you're splitting a long script into chunks, paste the neighbouring lines in and the model keeps the intonation from drifting between takes - that's the difference between a spliced narration and a seamless one. And timestamps (off by default) asks for word-level timing if you're syncing captions or lip movement.

The trap: language_code

It looks optional and harmless, and it is not on this model. The tooltip spells it out: language enforcement is supported by Turbo v2.5 and Flash v2.5, and for other models an error is returned if a language code is provided. Multilingual v2 handles languages natively and auto-detects; hand it de and you get a failed job, not German. Leave it empty here.

Outputs

audio is the one you wire - into Save Audio, into a lip-sync/video model, or into an audio-driven avatar graph. url gives you the hosted file if you want it elsewhere, all_urls_json carries every result URL, task_id and raw_json are for debugging a bad job, and credits_consumed / credits_left are floats so you can watch spend inside a long batch.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next

Or search "KIE.ai Nodes Next" in ComfyUI Manager. Restart, then set the key once under Settings → KIE.ai Nodes Next → Connection - it's validated before saving, kept out of your workflow JSON, and the frontend only ever sees a masked fingerprint. KIE_API_KEY still works for headless installs and takes priority.

Dependencies are requests, Pillow, numpy, PyYAML. No checkpoints to download; the voice lives on KIE's side.

Common issues

• "No KIE API key configured" - you skipped the Settings step. The error message tells you where to go; it's not a node bug.
• Jobs that hang - timeout_seconds defaults to 1200 and goes up to 7200. Long text on a busy provider can genuinely take minutes.
• Output sounds robotic halfway through - that's usually a missing previous_text/next_text pair rather than a bad voice.
• Anything about a language error on this node - see the language_code section. Empty is the correct value.