arkennemasis fal · Qwen Audio 3 TTS · $0.05/1k chars

You need a voice-over. Locally you have real options - the KB's audio overview rates Chatterbox as the open model that made open TTS competitive with paid services, F5-TTS as the faster one, Kokoro and Piper as the tiny ones that run on CPU. What none of them give you for free is a wall of named voices across eleven languages in one combo box.

ArkFal_alibaba_qwen_audio_3_tts is Qwen Audio 3 TTS (the Flash endpoint) running on fal, wrapped as a node. Text in, an AUDIO value and an MP3 path out, billed at $0.05 per 1,000 characters. No weights, no VRAM, nothing to download - which is the actual reason to reach for it: you can narrate a whole video while your GPU is busy sampling something else.

Why an API node for this

Same trade as every other wrapper in external-api-nodes.md: it's the right call for a model you can't run and the wrong default for one you can. TTS is the cheap end of that trade - narration is a few thousand characters, so a five-minute voice-over lands in the cents, not dollars. The KB's own line is that ElevenLabs remains the reliability and language bar for the API path; this node exists for when you want that class of voice quality with a much wider language spread and a per-character bill you can predict.

And per-character is exactly how it bills: the character count of your text is the quantity the cost is computed from. A 3,000-character script is $0.15. You can do that math before you press Run.

What you set

• text - the script. Multiline, and this is the only thing that drives the price.
• voice - defaults to Cherry; the list includes Serena, Ethan, Chelsie, Momo, Vivian, Moon, Maia, Kai, and a long tail of others. They are not interchangeable across languages - the node's tooltip points at Alibaba's voice list, and you should actually look at it, because a voice that's excellent in English may not cover Japanese at all.
• language - Auto by default. Setting it explicitly is the single biggest quality win available here: the tooltip says it improves pronunciation and intonation, which in practice means your German narration stops being pronounced with English phonemes. Options are Chinese, English, Spanish, Russian, Italian, French, Korean, Japanese, German, Portuguese.

The optional four are the pack's shared fal controls: max_cost_usd (default $20, stops the run before anything is sent if the estimate is over it, 0 disables), reuse_identical_run (default on - same text, voice and language returns the already-paid-for MP3 instead of billing you twice), extra_json (a JSON object merged into the request for any fal field the node has no box for) and max_concurrent (how many paid fal calls may run at once in the whole run; 1 by default).

Outputs: audio, a real waveform you can wire into a lip-sync node or the pack's own narration-length utilities; audio_path, the saved MP3 on disk; and info, a JSON blob with the request id, the cost estimate, the saved files and fal's full answer.

Install

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

Restart, or install through ComfyUI Manager by searching arkennemasis. Then put your key in a .env file in the folder holding run_nvidia_gpu.bat (the ComfyUI folder on a normal install):

FAL_KEY=your-key-from-fal.ai/dashboard/keys

There is deliberately no key box on the node. The argument for that is good: a key typed into a widget gets saved verbatim into every workflow JSON you share. Nothing here ships a model file, so install is genuinely just "clone and restart".

Where people get burned

The number one support question for any arkennemasis fal · node is the .env file being in the wrong folder, or saved as .env.txt by Notepad. The nodes read that file and nowhere else.

Then reuse_identical_run. It defaults to true, so if you tweak a comma and run again you'll get the cached line back and wonder why the delivery didn't change. Turn it off for a fresh take.

And the boring one: text goes to fal, so it leaves your machine - no different from any other API node, but worth knowing before you paste client material into a text box.

Results land in output/fal/alibaba_qwen-audio-3-tts/, and every request is logged to output/fal/_requests.jsonl.