NVIDIA Studio Voice Enhance

This is the payoff node of ComfyUI-NVIDIA-Maxine - the reason the other two exist. Feed NVIDIA Studio Voice Enhance a scratchy voice recording and it hands back broadcast-clean speech: background hum gone, room reverb flattened, the voice leveled and intelligible. Think the audio equivalent of what Adobe Podcast's "Enhance Speech" does, but running on your own GPU, in the middle of a ComfyUI graph.

And here's the best part: despite the NVIDIA branding, this node calls no API and needs no key at runtime. Once the Studio Voice container is up, Enhance talks to it over gRPC at 127.0.0.1:8001 - a completely local, free-for-the-electricity operation. The NGC key only ever existed to pull the Docker image in the first place.

Audio is the youngest corner of ComfyUI, and this is one of the few genuinely useful things it can do there. It's for recorded speech - voiceovers, podcasts, captured game or Discord audio, narration you layered onto a video - not music. If you're producing talking-head or lip-sync content, this is the polish step that makes the difference between "sounds like a webcam" and "sounds recorded in a studio."

How it works

Mechanically it's a thin client. The node takes any AUDIO input, resamples it to 48 kHz (Studio Voice's 48k-hq profile), wraps it as WAV bytes, and streams them to the local NIM in 64 KB chunks over a gRPC connection. The enhanced response comes back and is decoded straight into an AUDIO output for the rest of your graph. That resample step is worth knowing: your 44.1 kHz phone or messaging-app recording is handled automatically - no manual conversion, no failed calls.

Two honest limitations, both by design. Streaming is off by default, which is correct for whole-file enhancement of recordings. And there's no dry/wet or "intensity" slider - Studio Voice's public proto only exposes audio bytes, so the pack deliberately doesn't fake one. What you get is what the model decides is the clean version. If you want partial effect, run the dry and wet signals through a mix later.

Inputs and output

The node keeps it gloriously simple:

• audio (required) - any AUDIO from ComfyUI's Load Audio, or downstream of whatever produced the voice.
• studio_voice_connection (optional) - the STUDIO_VOICE_CONNECTION output from NVIDIA Studio Voice Docker Setup. Hook it up and the node verifies the NIM is actually ready before it runs.
• enhanced_audio (output) - wire this into a core Save Audio or Preview Audio node.

The canonical workflow shape, straight from the pack's example:

NVIDIA Studio Voice Docker Setup
    |
    v
Load Audio -> NVIDIA Studio Voice Enhance -> Save Audio

Run the setup node once (action setup_all_transactional, with your NGC key) so the container is up, then this node just works on every subsequent run.

Install

It's the same pack install as the setup node - clone into custom_nodes, install requirements.txt, restart. Search "ComfyUI-NVIDIA-Maxine" in ComfyUI Manager or:

cd C:\path\to\ComfyUI\custom_nodes
git clone https://github.com/grrdhdz/ComfyUI-NVIDIA-Maxine.git
cd ComfyUI-NVIDIA-Maxine
C:\path\to\ComfyUI\.venv\Scripts\python.exe -m pip install -r requirements.txt

The real floor is the environment, not this node: Windows + Docker Desktop with the WSL2 GPU backend, and a ComfyUI on the V3 custom node API (0.21.0+).

Troubleshooting

The one error you'll actually see is Studio Voice is not ready at 127.0.0.1:8001. That means the container isn't running - go back to NVIDIA Studio Voice Docker Setup, run setup_all_transactional, and check its status output for the health line before retrying. If setup reports OK but Enhance still can't connect, make sure you didn't override target in the Advanced Settings node, and give the NIM a few extra seconds on first cold start - the model loads into VRAM on boot. Beyond that, there's not much to fiddle with, which is exactly the point.