StepAudioEditX - Edit ✏️

The Edit node in the Step Audio EditX pack is the half most people don't see coming. The Clone node gets the attention because "copy my voice" is a party trick, but Edit is the actually-useful one: you feed it an existing audio clip and a directive - angry, whisper, 1.5x faster, add laughter - and it hands back a new version of the line that still sounds like the same speaker. It's a second-pass neural re-read, not a filter.

Most TTS in ComfyUI works one way: type text, get audio, and if you want the line delivered differently you re-generate and hope. Step Audio EditX lets you take a take you already like and re-perform it. The killer workflow from the README is Clone → Edit: generate the line, then re-emote it. That's why the pack ships both.

How it actually works

Under the hood this is StepFun AI's Step-Audio-EditX, a 3B audio language model paired with a dual-codebook tokenizer and a CosyVoice vocoder outputting 24kHz. The crucial thing to understand: nothing here is audio post-processing. The model reads your clip, understands the words (which is why the transcript matters so much), then re-synthesizes the whole thing with the new emotion/style/speed baked in. The voice identity is preserved because it's conditioned on the source audio itself.

Two consequences follow. First, you must give it the exact transcript in audio_text - it uses the words to align content with the edit. Second, n_edit_iterations (1–5) is your strength dial: each iteration runs the clip back through the model to amplify the effect. 1 is subtle, 2–3 is the sweet spot, 4–5 can push into distortion - and community reports say heavy iterations also start to flatten accents into a generic American delivery. Worth knowing before you crank it.

The inputs that matter

• audio_text - exact transcript of the clip. The single biggest quality lever; get the wording and punctuation right.
• edit_type - the category: emotion, style, speed, paralinguistic, or denoising. Pick one.
• The value field matching your edit_type - emotion (happy, sad, angry, excited, …), style (whisper, gentle, serious, casual, formal, …), speed (faster/slower × 1.2x/1.5x), paralinguistic ([Laughter], [Breathing], [Sigh], …), or denoising (denoise / vad). There's a paralinguistic_text box if you want the effect inserted at a specific phrase instead of auto-appended to the end.
• n_edit_iterations - effect strength, as above.
• input_audio - the clip to edit. It's listed as optional, but the node errors if it's not connected, so treat it as required. Hard limit: 0.5–30 seconds. Longer audio won't work - split it first (the README pairs this with Whisper + a chunking step for a reason).

One honest quirk: temperature, do_sample, and max_new_tokens appear on the node, but the author's own tooltips say they're hardcoded (0.7 / true / 8192) and have no effect in Edit mode. Don't touch them; the interesting knobs are the ones above. Output is a single audio (AUDIO) - wire it into a save or preview node, or back into a second Edit pass.

Installing it

Same as the whole pack: ComfyUI Manager, search "Step Audio EditX TTS", install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Saganaki22/ComfyUI-Step_Audio_EditX_TTS
cd ComfyUI-Step_Audio_EditX_TTS
pip install -r requirements.txt

Two things trip people up. FFmpeg must be installed at the system level (not the Python package) - that's a hard requirement. And you need both the Step-Audio-EditX and Step-Audio-Tokenizer models, cloned from HuggingFace into ComfyUI/models/Step-Audio-EditX/. The README also pins transformers==4.53.3; newer versions have a known bug that produces silent or garbled audio. This pack is VRAM-hungry too: ~14–18GB for Edit in bfloat16, and quantization (int8/int4) is available but the author notes they're waiting on official quantized weights from StepFun.

Troubleshooting

• Weak edit - bump n_edit_iterations to 2–3, and double-check audio_text matches the clip exactly.
• Garbled/silent output - pip install transformers==4.53.3, and reinstall librosa + hyperpyyaml.
• "No models found" - model_path is a dropdown populated at startup, not a text box, so you can't type a path. Put the models in the right folder and fully restart ComfyUI.
• CUDA OOM - drop to int8, float16, use sage_attn, and turn off keep_model_in_vram.

One last honest note, from people who actually run this: style and emotion options vary wildly in how much they change the result. Some are transformative, some barely move the needle - that's the nature of the underlying model, not a broken install. Start with 2 iterations and a clear emotion, and you'll get a feel fast.