🎨 Step Audio EditX - Audio Editor

Most TTS comes out flat. You clone a voice, the words are right, the delivery is a robot reading a grocery list. This node is the fix: it's a second pass that takes finished speech audio and re-shapes it - pushes an emotion into it, changes the speaking style, inserts a laugh or a sigh, speeds it up. It doesn't generate new speech from scratch; it edits the audio you already have. That's the whole idea, and it works on output from any engine in the suite, not just Step's own.

The author's own framing when he shipped it: the interesting part isn't Step Audio EditX as a cloner, it's that you can "add paralinguistic noises like laughter, breathing, emotion and style to any other TTS you generated that you think it's lacking in those areas." Generate with ChatterBox, then add a natural laugh here and a bit of warmth there.

How it works

Under the hood it's an LLM-based audio editor (a 3B Step model) that takes your audio plus a transcript of what's said, and regenerates it conditioned on the edit you asked for. That's why it needs the text: it re-synthesizes against the words. You pick an edit_type and the matching control, and it does one editing pass.

Be honest with your expectations here. In the author's own testing a lot of the emotion and style options "don't feel like they change the audio all that much," while some land really well. So treat it as a knob you audition, not a guarantee.

The inputs and outputs that matter

• input_audio (AUDIO) and audio_text (STRING) - the clip you're editing and its transcription. The text has to match the speech; that's how the model knows what it's re-voicing.
• edit_type - the mode: emotion, style, speed, paralinguistic, denoise, or vad. This decides which of the option dropdowns actually matters.
• emotion (16 choices like happy, sad, angry, excited, calm), style (35 choices - whisper, serious, child, sweet, radio, and more), speed (faster / slower / more faster / more slower). You set the one that goes with your edit_type.
• n_edit_iterations (1–5) - how hard to push. More passes = stronger effect, but this is a trap past 2–3 (see below).
• tts_engine (TTS_ENGINE, optional) - wire in the Step Audio EditX Engine node here. If you leave it off, the editor will need the engine loaded some other way.

Out come edited_audio (AUDIO) - send it to your Save Audio / preview node - and edit_info (STRING), a text log of what it did.

How to install it

Easiest path is ComfyUI Manager: search TTS Audio Suite, install, restart. It runs the pack's own install.py, which sorts out the messy dependency conflicts for you. Manual works too:

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

then restart ComfyUI. On Linux install portaudio19-dev and libsamplerate0-dev first. The Step Audio EditX model (~7GB) auto-downloads into ComfyUI/models/TTS/step_audio_editx/ the first time you run it - so the first execution is slow, later ones aren't.

Common issues & troubleshooting

Nothing seems to change. That's often the model, not you - many emotion/style presets are subtle by design. Bump n_edit_iterations a notch, or try a style that's more obviously distinct (whisper, shout) to confirm it's wired up.

The voice stops sounding like the person. More than about 3 iterations degrades the voice's resemblance - that's a documented limit. If you overcook it, back the iterations off, or (in the inline-tag workflow) use the suite's <restore> tag to pull the original voice character back.

Weird artifacts or an error on long clips. Step Audio EditX expects segments roughly 0.5s to 30s. Feed it a 2-minute file and it won't be happy - split first.

Wrong language mangles it. The model only really handles Mandarin, English, Sichuanese, Cantonese, Japanese, and Korean. Other languages lose their accent and distort. Stick to those.