IndexTTS 2.5 人声后处理 · T8star-Aix

IndexTTS 2.5's raw output is usually fine, but "usually" isn't "always" - and the fix for most voice issues is a filter, not a re-synthesis. This node takes any standard ComfyUI AUDIO (not just this pack's output) and applies one of five voice presets at a blendable strength, plus an optional peak normalization. No FFmpeg, no system audio tools, no extra model - it's pure PyTorch signal processing, which is a big deal on Windows where chasing FFmpeg builds is a hobby nobody wants.

The presets

• voice_clarity - the default and the one to try first. Cuts mud, makes the voice more distinct.
• clear_narration - same family, tuned toward spoken-word narration.
• deharsh - knocks back the harsh, sibilant edges (the "s" hiss) that TTS models love to add.
• warm - softens and adds body; the anti-deharsh, basically.
• normalize - loudness to a target peak.
• off - passes the waveform through untouched. The README's guarantee is the important one: off does not alter your audio at all, so you can wire this node in permanently and toggle presets without ever worrying about hidden processing.

strength (0–1) blends the processed result against the original - dial in 0.4 if a preset is almost right but too aggressive. target_peak_db (default −1, range −12 to −0.1) sets the peak normalization headroom when you want it.

What you get back

处理后音频 - a standard AUDIO you can chain onward (Save Audio, a mixer, a video). 后处理报告 JSON - what preset ran, at what strength, and the measured peak, so you can compare runs objectively instead of by feel.

The workflow trick

The Generate node has the same presets built in, so why use this standalone one? Two reasons. First, it processes any audio - another TTS model's output, an existing recording, a stem. Second, it lets you A/B in real time: keep the raw audio from Generate on one path and the post-processed version on the other, and preview both before committing. The README explicitly recommends exactly this - connect the raw and processed paths to compare, then pick. It's also the only way to try presets after a multi-role mix without re-running the whole generation, since the multi-role node bakes its postprocess in.

Gotchas

None of the exotic kind - no dependencies, no model downloads, no FFmpeg. The one thing to remember: the preset list here is the same as the Generate node's postprocess_preset list, so don't stack the same preset at both ends at full strength or you'll double-process (usually harmless, sometimes over-cooked). And target_peak_db is a peak, not a loudness spec - if you're chasing true loudness normalization you're in the wrong node, this one is about headroom and consistency.

Ships with the pack (ComfyUI Manager search IndexTTS 2.5 · T8star-Aix, or clone the repo). It's a pure filter node - no VRAM, no inference cost, runs in a fraction of a second. Install once and it's basically a permanent fixture in your audio chain.