Speech-God Post Process

Raw TTS output is dry, uneven, and too quiet next to music. Speech-God Post Process is the pack's answer to that - a compact master bus in node form that sits between Generate and Export, and it's the difference between "voice acting demo" and "finished audio." Everything here runs on CPU, which is deliberate: it keeps the VRAM budget for the TTS model itself, and DSP like this is cheap enough not to need the GPU anyway.

It's a chain of toggleable modules, applied in a sensible order: trim silence → noise reduction → de-breath → EQ → compressor → loudness normalize → limiter. Because every module is a checkbox, the node doubles as your debugging tool - if something sounds wrong, you can isolate whether it's the compressor or the noise gate by flipping one switch.

The inputs that matter

• audio - in, then straight back out. It accepts the batched AUDIO from Generate, so all your takes get processed identically in one pass.
• loudness_normalize (default on) with target_db (default -18) - the one you'll actually reach for. It normalizes to an RMS target so dialogue sits at a consistent loudness instead of one line whispering and the next shouting. -18 dBFS is a reasonable podcast-ish default; -14 is louder, -23 is broadcast-style quiet.
• compressor (default on) and limiter (default on) - the dynamic-range pair. Compressor tames the peaks, limiter caps them so nothing clips. These two alone fix most "one take is way louder than another" complaints.
• noise_reduction (default off) - a noisereduce pass for hiss in the reference or the room. Off by default because it can smear the voice if the source is already clean; flip it on when the reference clip is buzzy.
• eq (default off) with eq_low_db / eq_mid_db / eq_high_db (-12 to +12) - a three-band shelf. This is the "brighten the voice" button; subtle values like +2 on high shelf do more than you'd think for intelligibility.
• de_breath (default off) - trims breath sounds between phrases. F5 and fish both produce audible inhales; this cleans them up when you're making something polished.
• trim_silence (default on) - strips leading and trailing silence so exported files start exactly on the voice.

Output is audio, ready for the Speech-God Export node.

Installing

Same pack, same path as every node here: ComfyUI Manager (search "Speech-God") or git clone https://github.com/realTNEU/ComfyUI-SpeechGod into custom_nodes, then install requirements.txt into the ComfyUI venv and restart. The DSP chain pulls in noisereduce and soundfile - both light, both in the pack's requirements, both install alongside f5-tts on the normal path.

Two small truths from the pack's own docs: if you hear robotic artifacts, that's almost always the generation profile (extreme age+tone pitch stacking), not this node - fix it upstream. And if your final export is a whisper, the limiter isn't the culprit; check target_db and the energy slider on the Character before you blame the master bus.