elevenlabs/text-to-speech-turbo-2-5

If you're going to keep a paid text-to-speech node wired into a workflow and hit it ten times while you iterate on a line of dialogue, it should be the cheap fast one. That's Turbo 2.5 - same family as the Multilingual v2 node in this pack, same tuning knobs, but built for latency rather than the last few percent of delivery quality. For placeholder VO, temp tracks, and the "does this line even work" pass, it's the one I'd reach for first.

The honest framing, since this pack sits in the middle of an argument the local TTS scene is still having: Chatterbox and F5-TTS are close enough that plenty of people stopped paying for voice work entirely, and the KB's audio doc says exactly that - what changed is cloning locally, not the reliability bar. ElevenLabs is still the reliable one. Turbo is the affordable way to use it.

How it works

The node takes your text, asks KIE to run elevenlabs/text-to-speech-turbo-2-5, polls the job until it's done (the timeout_seconds budget, 1200 by default), then downloads the returned audio and decodes it into ComfyUI's standard AUDIO dict. Your GPU does nothing; the audio is decoded with torchaudio and handed back as a normal Comfy audio object, so it wires into Save Audio, a preview node, or straight into an audio-driven video model.

Because it's a hosted call, the two things that matter are the text you send and what you're paying per attempt. The node shows credits spent and remaining after each run, and reports both as float outputs.

Inputs worth setting

text is the only required input, capped at 5000 characters.

voice is optional here, which is the notable difference from the Multilingual v2 node next door - leave it alone and you get Rachel, a preset name rather than an ID. The dropdown mounts 68 entries, a mix of friendly names (Rachel, Adam) and raw voice IDs with descriptions in the tooltip, and you can audition any of them in a browser with the pattern from the author's tooltip:

https://static.aiquickdraw.com/elevenlabs/voice/<voice_id>.mp3

Then the shared dials: stability (0.5), similarity_boost (0.75), style (0, i.e. no exaggeration), speed (1.0, keep it inside the 0.7–1.2 range - the tooltip warns that extremes cost you quality). timestamps is off by default; flip it on if you need word-level timing for captions. previous_text and next_text are continuity hints for when you're stitching a long script out of several runs - paste the adjacent lines and the intonation stops jumping between chunks.

language_code is the one place this node is more capable than its bigger sibling, not less. Multilingual v2 errors out if you send a language code; Turbo v2.5 is one of the two models that supports language enforcement, per the tooltip. So if you need a German or Japanese read pinned to an exact language rather than auto-detected, this is the node in the family that does it - leave it empty otherwise.

Outputs

audio goes downstream. url is the hosted file, all_urls_json is every result URL the job returned, task_id and raw_json are for debugging, and credits_consumed / credits_left let a long batch stay honest about cost.

Install

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

ComfyUI Manager can do it too - search "KIE.ai Nodes Next". It's a light install: requests, Pillow, numpy, PyYAML in requirements.txt, no model files. Restart, then paste your KIE API key once under Settings → KIE.ai Nodes Next → Connection. The key is validated against KIE before it's saved, lives in Comfy's user config rather than your workflow, and KIE_API_KEY in the environment overrides it for server installs.

Common issues

• It fails before doing anything with a message about no API key configured - that's the Settings step, not the node.
• The read is too flat or too theatrical - stability and style, in that order. Lower stability for performance, higher for consistency; style above 0 starts acting.
• You want your own cloned voice - not available here. voice is a fixed 68-entry list; there's no free-text field for a voice ID from your own ElevenLabs account.
• Node missing after an update - the pack builds its node list from KIE's live documentation catalog, so a model KIE retires can vanish from the menu until they publish it again.