Music - Master Audio Enhancement

This is the flagship of the ComfyUI Music Tools pack, and the one node that could plausibly replace a whole chain of utility nodes by itself. Music - Master Audio Enhancement is a configurable mastering chain in a single box: denoise, three-band EQ, multiband dynamics, clarity, optional vocal processing, stereo shaping, loudness management and true-peak protection, all in series, all parameterized.

It's the node the pack's launch post led with, and the intended audience is obvious: people who generated audio (ACE-Step instrumentals, Suno tracks, TTS vocals) and want it to stop sounding like a rough render. Where Music Fix is a one-knob finish, this is the full dashboard.

The inputs that actually matter

There are fourteen inputs, which is a lot, but you can group them and you mostly touch three groups:

• Denoise first. denoise_mode (Hiss Only / Full Denoise / Off) and denoise_intensity (0–1). If your source is clean, set it to Off and skip the risk. If it's noisy, Hiss Only is the safer default.
• The vocal group. vocal_enhance, deesser_amount, breath_smooth, naturalize_vocal. These exist to humanize AI vocals - the de-esser cuts sibilance around 4–8 kHz, naturalize_vocal smooths the robotic pitch-quantization artifacts you get out of generated singing. This is genuinely the most interesting part of the pack: it's a deterministic artifact smoother, not pitch reconstruction. It won't undo a bad pitch, it just makes the "digital" edge less offensive.
• The loudness/peak group. target_loudness (default −14 LUFS) plus the built-in true-peak protection. Set the streaming target, let the limiter worry about overs.

The EQ trio (eq_low_gain, eq_mid_gain, eq_high_gain) and clarity_amount are the tone-shaping layer; defaults are a gentle high-shelf presence boost. reverb_amount adds a little multi-tap ambience if the track is too dry.

The ai_enhance trap

ai_enhance (default false) routes the audio through a SpeechBrain MetricGAN+ neural enhancer, with ai_mix (default 0.6) blending it back with the DSP result. Here's the thing the README states plainly: this is for speech and vocal material, not transparent full-mix mastering. Enable it on a full song and it'll do weird things to the instrumental bed. Enable it on a podcast or a voiceover and it's genuinely good.

It's also the only input with a real install cost. MetricGAN+ comes from requirements-ai.txt (speechbrain + huggingface-hub) and pulls a model on first use. The pack is careful here: heavy AI imports are lazy-loaded, so if the optional stack is missing or a torch/torchaudio version is mismatched, the node just runs its DSP chain and never crashes your ComfyUI session. Use the torchaudio build that already matches ComfyUI's torch - don't independently upgrade either.

How it works, briefly

Under the hood it's a serial DSP pipeline: spectral subtraction for noise, peaking EQs, compression, stereo shaping, LUFS normalization via pyloudnorm, and an oversampled true-peak limiter at the end (4x oversampling, the same measurement the whole pack leans on). All NumPy/SciPy, CPU-friendly, near real-time. The AI stage only joins if you flip the toggle and have the deps.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/jeankassio/ComfyUI_MusicTools.git
cd ComfyUI_MusicTools
python -m pip install -r requirements.txt   # core, always
python -m pip install -r requirements-ai.txt  # only if you want MetricGAN+

Restart ComfyUI and the node appears under music. Defaults are sane enough to hit run and be done - which is exactly how the good mastering tools should behave.