arkennemasis fal · ElevenLabs TTS Turbo v2.5 · $0.05/1k chars

Five cents per thousand characters. In a pack where the music nodes charge sixty cents a minute and Seedance charges a dollar-plus a second of video, this is the node you point at volume work: thirty scenes of narration, a week of social clips, a batch of product videos where every listing gets a spoken description.

It's the same widget set as the Multilingual v2 node - all eleven of them, including the previous_text/next_text continuity pair - at half the price, on a model built for speed rather than for the last increment of naturalness. The tradeoff is exactly what the name says.

How it works

A generated fal node: FAL_KEY from the install's .env, input checks, a cost estimate from the character count, one billed call, polling with live status, and output into output/fal/fal-ai_elevenlabs_tts_turbo-v2.5/. Every request is appended to output/fal/_requests.jsonl, which is the honest way to find out what a hundred-scene batch cost - read the ledger rather than adding up node badges.

The inputs that matter

text is required; voice is a plain string defaulting to Rachel. Then the dials, in the order you'll actually touch them:

• stability - 0.5 by default. Higher for batch consistency, lower for performance.
• similarity_boost - 0.75 by default, pulling the output toward the reference voice at a small cost in naturalness.
• style and speed - style exaggeration (default 0 - leave it there unless you mean it) and 0.7–1.2 speed, with the tooltip's warning that extremes hurt quality. Speeding up a read to fit a runtime is the classic misuse; cut the script instead.
• previous_text / next_text - pass the neighbouring sentences in and the delivery stops resetting at every chunk boundary. The single biggest quality win available on this node, and it costs nothing.
• language_code (ISO 639-1, enforced - an unsupported code errors rather than falling back), apply_text_normalization (auto/on/off for numbers and dates), and timestamps for word-level timing you can drive captions with.

Outputs: audio, audio_path, timestamps, info. In a fan-out workflow - the pack's Shot Selector and Scene List nodes are built for exactly this - audio_path is what you pass along so a later stage can read the file without re-rendering.

Money

$0.05 per 1,000 characters. Ten thousand characters of narration is fifty cents. That's low enough that batching becomes practical: at the default max_concurrent of 1 the fal nodes run one paid call at a time across the whole run, which keeps you inside rate limits but serialises a big job. Raise it to 4 or 8 on the TTS nodes of a narration-heavy graph and a fifty-scene render stops being an afternoon.

Install

Manager → search comfyui-arkennemasis, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis

Restart, then FAL_KEY=your-key-from-fal.ai/dashboard/keys in the .env in the folder holding run_nvidia_gpu.bat (or the ComfyUI folder - those are the only two places it looks, and there's no key box on the node, deliberately).

Common issues & troubleshooting

It sounds flatter than the Eleven v4 nodes. It is. Turbo trades a little naturalness for speed and price; when the job is 200 short lines, that's the right trade, and when it's one hero narration, it isn't.

Every chunk sounds like a new speaker. previous_text and next_text are empty. Wire them and the seams disappear.

Numbers read as individual digits. apply_text_normalization - force it on.

Identical audio returned instantly with no charge. reuse_identical_run is on by default. For batch work that's a feature - a re-run of an unchanged branch costs nothing - and for auditioning it's the thing you have to remember to switch off.

A 429 or a stream that dropped mid-run. The pack retries rate limits, 5xx errors and interrupted streams with exponential backoff, so a blip usually resolves itself without a re-run. If it doesn't, the failure is logged with the request id, and fal's own answer is in info.

Cheaper still, locally. If the text is not sensitive and your voice isn't specific, Kokoro runs in real time on a CPU and Chatterbox runs well under 8 GB of VRAM - free after electricity. The honest case for this node is a pipeline that must not fail mid-batch, and a voice you don't have locally.