Audio Post-Process (Resample/Fade/Norm)

This is the "mastering" node - the last thing in your audio chain before the file lands on disk, doing what a real mixing engineer does at the end of a session: resample, filter, fade, normalize. If you've been generating speech with CosyVoice or running voice conversion, you've probably noticed the audio doesn't sound bad, exactly, but it has a digital sheen and the occasional blip at cut points. That's precisely what this node exists to kill.

What it actually does

Four distinct jobs, all standard and all explainable:

• Resample - up-sample to 44.1kHz or 48kHz with a high-quality Sinc interpolation (sinc_interp_hann default; sinc_interp_kaiser is the alternative). Output at 22–24kHz just sounds cheap; this fixes that.
• LowPass filter - default 11000Hz, and it's the most interesting part. The README calls it a 6-stage cascade designed like a "brick wall" to smash the high-frequency aliasing artifacts that CosyVoice's resampling leaves behind. 0 disables it.
• HighPass filter - default 60Hz, a gentler single-stage roll-off that kills low rumble and DC offset without eating the bass warmth of a voice.
• Fade in/out + Normalize - small fades on the edges so joins and cuts don't pop, and a normalize-to-−1dB so the loudness is full but safe.

The inputs

• audio - whatever came out of your TTS/VC/enhance node.
• target_rate - 44100 (default), 48000, 24000, 22050, or Original. 44.1k is the standard call.
• lowpass_cutoff (11000 default, 0=off) and highpass_cutoff (60 default, 0=off) - the two filters above. Leave the defaults unless you hear something specific; they were tuned for this pack's own output.
• fade_length - seconds of fade on each end, 0.1 default.
• normalize - on by default.
• resampling_alg - the interpolation kernel; sinc_interp_hann is the default for good reason.
• splice_info (optional) - the passthrough again, so this node can sit inside a voice-conversion pipeline without dropping the seam metadata. It comes back out the other side.

Outputs

audio (the finished track) and splice_info (passthrough).

Where it goes in a graph

The README is specific: put it after Voice Conversion. The full "sounds recorded, not generated" chain is:

TTS → Audio Enhance (Resemble) → Audio Post-Process → save

or for the long-conversion path:

source → Smart Chunker → Voice Conversion → Audio Post-Process

Gotchas

• It can't fix what's not there. Post-processing is cosmetic; if the conversion itself landed a seam mid-word, this won't repair it - that's what the Silence Splitter/Splice Analyzer pair is for, upstream.
• Don't stack resamples. If you set target_rate but your save/preview node also resamples, you're doing two Sinc passes for nothing. Pick one place.
• The defaults are genuinely good for this pack's output. Tweak only on evidence - e.g. raise lowpass_cutoff if you're feeding it a real instrument recording and it's getting dull.