arkennemasis fal · MiniMax Voice Clone · $1.50/voice

Hand it at least ten seconds of somebody talking, and it hands back a voice ID you can then use for any amount of speech. $1.50 per clone. No training run, no dataset, no GPU fan spin-up.

The KB's audio chapter makes the case that local cloning crossed the "good enough" line - Chatterbox came close enough that the community stopped treating the API as mandatory, and this pack even ships a local Qwen3-TTS clone node. So why route out? Two reasons: you want it done in one call with zero setup, or you want a specific high-quality voice and don't care that the sample went to a server.

What the node needs, and the 7-day rule

audio is the reference sample: any AUDIO value, at least 10 seconds of it. Clean speech, one speaker, no music. If your source is a ripped interview with a bed underneath, clean it up first - this pack also ships an ElevenLabs audio-isolation node for exactly that job, and the node's own noise_reduction toggle will help but won't perform surgery.

Then the part nobody reads until a voice disappears:

To retain the voice permanently, use it with a TTS (text-to-speech) endpoint at least once within 7 days. Otherwise it will be automatically deleted.

So the clone is generated, and it's provisional. Use it - wire custom_voice_id into a speech node and actually run something - inside a week, or it gets reaped and you're paying $1.50 again.

The inputs worth setting

• text - a preview line the clone speaks immediately. Defaulted to "Hello, this is a preview of your cloned voice! I hope you like it!" so you hear something the moment it finishes. Preview characters bill separately at $0.30 per 1,000 characters - three cents on a 100-character line, so use a real sentence from your script rather than a nonsense one; it tells you more.
• model - which MiniMax TTS generates the preview: speech-02-hd, speech-02-turbo, speech-01-hd, speech-01-turbo. It affects the preview only.
• noise_reduction / need_volume_normalization - off by default. Turn them on when the sample isn't studio-clean.
• accuracy - a text-validation threshold from 0 to 1, with -1 meaning "leave it to the model". If the clone keeps mis-reading your scripts, this is a dial rather than a mystery.

Outputs and how they wire

audio is the preview as a real ComfyUI AUDIO value with a player on the node. audio_path is that file on disk, under output/fal/fal-ai_minimax_voice-clone/. custom_voice_id is the point of the whole exercise - a STRING, which is exactly why it wires: voice_setting_voice_id on the pack's Speech 2.8 Turbo node is also a plain string, so you connect one to the other and your clone starts narrating. info is the JSON receipt (request id, estimate, files, fal's answer).

Install

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

Restart ComfyUI, or search arkennemasis in ComfyUI-Manager. requirements.txt is httpx alone - for the pack's Codex nodes - and the fal provider is standard-library only. Then your key:

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

in the .env next to run_nvidia_gpu.bat. No key widget exists on any node in this pack; that's a deliberate choice, because widget values are saved into the workflow JSON verbatim.

Where people get burned

• The clone sounds like a stranger. Sample quality beats sample length. Ten clean seconds of someone speaking directly at a mic outperforms two noisy minutes.
• It vanished a week later. The 7-day rule above. Generate speech with it once and it sticks.
• You fed it music, or two speakers. It clones the average of what you gave it, and the average of two people is nobody.
• You sent someone else's voice. Worth saying once, plainly: consent isn't a formality here, and the sample leaves your machine.
• You paid twice for the same clone. reuse_identical_run (default true) reuses the paid result for identical inputs, but a new reference sample is a new clone and a new $1.50 - that one's not a bug.