🤐 Voice Fixer

🤖 Voice Fixer is the cleanup node. Feed it degraded speech - a noisy phone recording, a clip with hiss and clipping, a rough reference you want to clone from - and it restores it toward clean, full-bandwidth voice. It's a wrapper around the well-known VoiceFixer restoration model, and it does one thing without a pile of settings, which is exactly what you want when you're just trying to salvage a clip. In this suite it earns its keep in two spots: cleaning a reference before you clone from it, and polishing TTS output that came out a little rough.

How it works

VoiceFixer is a trained restoration model that handles several kinds of degradation at once - noise, reverb, clipping, and low bandwidth - and resynthesizes clean speech rather than just filtering. That's the difference from a plain noise gate: it can actually rebuild missing high frequencies and smooth out damage instead of only subtracting hiss. You give it audio and a restoration mode, and it returns the cleaned version.

The inputs and outputs that matter

There are only three inputs, which is the whole appeal:

• audio - the clip to restore (AUDIO input).
• restoration_mode - the one real choice. Three options: 0 - Original (Default) is the standard restoration for typical degraded audio; 1 - With High-Freq Removal restores and then trims high frequencies, useful when the top end is harsh or noisy; 2 - Train Mode (Seriously Degraded) is the aggressive setting for badly damaged audio. Start with mode 0, escalate to 2 only if the clip is genuinely rough.
• use_cuda (default true) - run it on GPU. Leave it on unless you have a reason not to.

Two outputs: restored_audio (the cleaned clip - send it onward to cloning, conversion, or Save Audio) and info (a text summary).

Installing it

Comes with TTS Audio Suite. Easiest: ComfyUI Manager → search TTS Audio Suite → install → restart, which runs the pack's install.py and handles the dependency conflicts (NumPy, librosa, s3tokenizer) and Python 3.13. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/diodiogod/TTS-Audio-Suite.git, then python install.py from inside the folder with your ComfyUI venv active (run the script, not just a bare requirements install). Linux: portaudio19-dev libsamplerate0-dev first. The VoiceFixer model auto-downloads on first use. There's a "🤐 Voice Cleaning" example workflow that pairs Voice Fixer with the pack's noise/vocal removal for a full two-stage cleanup pipeline.

Common issues

• Over-smoothed, "underwater" result. VoiceFixer resynthesizes, so on already-clean audio it can soften detail and add a slightly processed quality. Only run it on audio that actually needs it - clean clips are usually better left alone.
• Mode 2 makes clean audio worse. Train Mode is aggressive and meant for seriously degraded input. On mildly noisy audio it over-processes. Match the mode to the damage; default 0 first.
• Still noisy after restoration. Voice Fixer restores speech but isn't a dedicated denoiser for heavy background noise or music bleed. Pair it with the pack's noise/vocal-removal node (that's what the Voice Cleaning workflow does) for the tough cases.
• Great as a pre-clone step. The best use is cleaning a reference clip before feeding it to 🎭 Character Voices - a clean reference clones far better than a noisy one, and a noisy reference poisons every generation.