⚙️ Step Audio EditX Engine

This is the loader node. On its own it doesn't make a sound - it loads the Step Audio EditX model into memory, sets how it runs, and hands out a TTS_ENGINE that the actual working nodes plug into. In TTS Audio Suite the pattern is always the same: an Engine node configures a model, and a downstream node (TTS Text, TTS SRT, or the Audio Editor) does the work. This is the Engine half for Step.

Step Audio EditX is a two-in-one deal. Feed this engine into a TTS node and it does zero-shot voice cloning. Feed it into the Step Audio EditX - Audio Editor node and it does the emotion/style/paralinguistic second-pass editing that's the model's real party trick. Same engine, two jobs - you pick by what you wire it into.

How it works

You set the model, the device, and the precision; the node loads a ~7GB, 3B-parameter Step model and keeps it resident. The TTS_engine output is just a handle to that loaded model. Downstream nodes read it and run inference against it, so you only load once even if several nodes use it.

The inputs and outputs that matter

Honestly, most beginners can leave this node on defaults. The few worth knowing:

• runtime_mode - Shared Runtime, Dedicated Runtime, or Main Environment. This is the suite's isolation system for engines with fussy dependencies. Step is a modern engine that's happy in Main Environment, but if you hit dependency conflicts, the shared/dedicated modes give it its own sandbox instead of forcing your whole ComfyUI backward.
• quantization - none, int4, or int8. If 7GB of model is straining your VRAM, int8 (and especially int4) shrink the footprint. Expect a small quality cost; often worth it to make it fit.
• torch_dtype - leave it on bfloat16 unless your GPU is old enough to prefer float16.
• temperature (default 0.7), do_sample, max_new_tokens (1024) - generation randomness and length. Higher temperature is more varied, lower is more deterministic. Fine at defaults to start.
• device - auto picks your GPU. Leave it.

The single output is TTS_engine (TTS_ENGINE). Run it into a TTS Text/SRT node for cloning, or into the Audio Editor for editing.

How to install it

It ships in TTS Audio Suite, so install the pack. Via ComfyUI Manager, search TTS Audio Suite, install, restart - the Manager runs the pack's install.py and handles the dependency mess. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/diodiogod/TTS-Audio-Suite.git
cd TTS-Audio-Suite
python install.py

then restart. On Linux install portaudio19-dev and libsamplerate0-dev first. The Step model auto-downloads to ComfyUI/models/TTS/step_audio_editx/ on first run, so budget time (and disk) for that ~7GB pull the first time.

Common issues & troubleshooting

Out-of-memory on load. 7GB plus overhead is a lot on an 8–12GB card, especially with other models loaded. Drop quantization to int8 or int4, and consider running Step on its own rather than alongside a big image model in the same graph.

Dependency conflicts on install. Step is meant to be Main-Environment friendly, but if the pack's install trips over something, that's what the runtime_mode shared/dedicated options exist for - set isolation on the engine instead of downgrading your base ComfyUI.

First run hangs for ages. It's downloading the model, not frozen. Watch the console for the download progress; subsequent runs load from cache and are quick.

It "works" but the voice is bland. That's the model's character, not a config bug - Step's cloning is decent but a bit neutral. Its strength is the editing pass; lean on the Audio Editor node for expressiveness rather than expecting drama from the raw clone.