arkennemasis fal · xAI Grok Voice · $0.00083/s

At $0.00083 per second this is the cheapest node in the pack by a mile - a ten-second exchange costs less than a cent - and it's also the one people install for the wrong job. It doesn't transcribe. You hand it a recording of someone speaking, and Grok listens and replies, out loud, with its own voice, and the transcript of what it said comes back on the text output.

So it's a conversational agent with a microphone socket: a recorded prompt in, a spoken answer out. In a ComfyUI graph that makes it weird and occasionally very useful - a node that gives your workflow an opinion it can say out loud.

Inputs

audio is required: the recording of the user's speech, up to 10 minutes and 50 MB, converted to 16-bit 24 kHz mono PCM before it's sent. Most common formats are fine.

prompt is optional and is not the thing you're saying - it's the system instruction. The tooltip describes it as the persona and conversation context. Leave it empty and you get Grok's default persona, which is chatty and a bit much. This is the field that makes it useful in a pipeline: "you are a continuity checker for a shot list; answer in one sentence" gets you a far more usable node than no prompt at all.

voice is a dropdown of xAI's built-in voices - carina, zagan, helix, orion, luna, iris, altair, zenith, perseus, helios, lux and more, plus (not set) which falls back to Grok's default.

tools is a JSON field that decides whether the agent can reach outside itself: x_search, web_search and mcp_servers, all disabled in the default blob. Switch a search on and the reply is no longer just a model's opinion, it's one grounded in something it looked up - and takes correspondingly longer and costs more seconds. Note what's deliberately absent: the tooltip says file_search and collections are unsupported here because of xAI's Zero Data Retention requirements, and caller-side function tools aren't in this configuration either.

Outputs: audio (the spoken reply, as an AUDIO you can save or mix), audio_path, text (the transcript of the reply), duration, and info with the request id, estimate and fal's full response.

Shared controls: max_cost_usd (default $20 - you will never hit this), reuse_identical_run, extra_json, max_concurrent.

Install

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

Manager → arkennemasis. Then in the .env next to run_nvidia_gpu.bat:

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

Restart ComfyUI. The node appears under arkennemasis/fal/Audio/xAI; results land in output/fal/xai_grok-voice/.

Where people go wrong

• Expecting a transcript of their own audio. text is what Grok said, not what you said. If you want a transcript, that's a speech-to-text endpoint, and there are several in this pack's audio menu.
• Leaving prompt blank and being surprised by the tone. The default persona is a general-purpose assistant; for anything structured, write the persona.
• Malformed tools JSON. It's a JSON object, edited by hand, with three keys. Break it and the node refuses the run before anything is uploaded.
• Enabling a search tool without thinking about latency. A web-search reply is several extra seconds of billed time, and the duration output shows you exactly how much.
• Re-running the same graph and getting the same reply. That's reuse_identical_run doing its job; turn it off for a second opinion on identical inputs.