FL ClearVoice Process

FL ClearVoice Process is the node that actually does the work in the filliptm/ComfyUI_FL-ClearVoice pack: you hand it a model handle from FL ClearVoice Model Loader, hand it audio, and it hands you back cleaner audio. No API calls, no keys, no cloud round-trip - everything runs locally through models downloaded to your machine. It's the "process" half of a deliberately minimal two-node pack, and it's the reason you installed the thing in the first place.

Reach for it whenever your voice audio sounds like it was recorded inside a jet engine. The obvious use case is fixing speech: denoise a voiceover, pull hiss off a TTS readout, or restore a scratchy old recording. It also does super-resolution, which is the trickier sell - feed it a 16 kHz voice track and it can upscale it to 48 kHz using MossFormer2_SR_48K or NovaSR. You don't get magic out of garbage, but for genuinely bandwidth-limited speech it's real. The pack comes from the same author as the big Fill-Nodes collection and the FL TTS packs, so if you already run ChatterBox or CosyVoice3 in your graph, this slots right into that audio pipeline.

How it works

The Process node doesn't know what model you picked - it just sees a CLEARVOICE_MODEL handle and reads the model's name off it, then routes to the matching backend: ClearVoice (Alibaba's open-source stack), Resemble-Enhance, VoiceFixer, or NovaSR. Each backend gets the same treatment: the audio is resampled to whatever sample rate the model expects, mixed to mono, processed, then stereo is duplicated back if your source had two channels. The output sample rate is set by the model, not your input, which is why the message output tells you exactly what happened.

That automatic resampling is the quietly useful part. ComfyUI's AUDIO format carries its own sample rate, so you can feed this a 48 kHz clip into a 16 kHz model and it converts silently rather than erroring. The flip side is worth knowing: these are mono-based speech models. Stereo in gets collapsed to mono, enhanced, then copied to both channels - you're not getting true stereo processing.

The inputs and outputs that matter

There are only two required inputs:

• model - the CLEARVOICE_MODEL handle from FL ClearVoice Model Loader. This is a type-specific wire, so ComfyUI won't let you plug a random checkpoint into it.
• audio - an AUDIO wire. Load it from any node that outputs AUDIO (a load-audio node, or the output of a TTS pack), and send the result downstream to whatever saves or plays audio.

Outputs: audio is your processed result, and message is a STRING that tells you what happened - something like Enhanced with MossFormer2_SE_48K | 16000Hz -> 48000Hz | Duration: 3.42s. For SR models it says "Super-resolved", for restoration models "Restored", for denoise-only "Denoised". Wire the message into a text display node to see it, or just read it in the console - the pack prints it there anyway.

Installation

Same story as the whole pack, so this is quick:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI_FL-ClearVoice.git
cd ComfyUI_FL-ClearVoice
pip install -r requirements.txt

Or just search "FL ClearVoice" in ComfyUI Manager and hit install. Restart ComfyUI, done. Note that requirements.txt pins transformers<4.46.0 - if your environment has a newer transformers pulled in by other nodes, installing this pack may downgrade it, which is the classic custom-node dependency dance. Also, the Resemble-Enhance and VoiceFixer models aren't installed by requirements.txt (they're commented out as optional) - pick those in the loader and you'll need pip install resemble-enhance or pip install voicefixer yourself.

Common issues

• Silence out, but no crash. The node catches its own errors and returns one second of empty audio plus a message starting with Error: .... If you get silence, check the message output - that's the pack telling you what broke.
• First run is slow. Models download on first use into ComfyUI/models/clear_voice/. The pack shows progress bars with speed and ETA, so it looks scary but it's working.
• Stereo surprise. Your two-channel source comes back as a duplicated mono signal. If you're processing music, this is the wrong tool; it's built for speech.
• CPU is usable but slow - 8 GB RAM is the floor and 16 GB is the sane minimum, per the README. On Mac, ClearVoice models use MPS but Resemble-Enhance falls back to CPU.

If your audio already sounds clean, this node won't make it meaningfully better - enhancement is about removing what's in the way, not adding fidelity that was never recorded. Start with the default MossFormer2_SE_48K, listen, and only branch out when you know what you're chasing.