Nodes/Step Audio EditX TTS/StepAudioEditX - Edit ✏️
ComfyUI Node

StepAudioEditX - Edit ✏️

A ComfyUI node in audio/step_audio with 20 inputs and 1 output.

By Saganaki22·Created 9 months ago·Updated 8 months ago· 62
StepAudioEditX - Edit ✏️
  • input_audio
  • audio
audio_text
model_path(No models found - download Step-Audio-EditX)
devicecuda
torch_dtypebfloat16
quantizationnone
attention_mechanismsdpa
edit_typeemotion
emotionnone
stylenone
speednone
paralinguisticnone
denoisingnone
paralinguistic_text
n_edit_iterations1
temperature0.7
do_sampletrue
max_new_tokens8192
seed0
keep_model_in_vramtrue
Categoryaudio/step_audio

Inputs (20)

NameTypeDefaultDescription
audio_textSTRINGExact transcript of what the input audio says. The model needs this to understand the content before editing the audio characteristics (emotion, style, speed, etc).
model_pathCOMBO(No models found - download Step-Audio-EditX)Step-Audio-EditX model location. Place models in ComfyUI/models/step_audio/ or select from discovered models.
deviceCOMBOcudaHardware device: 'cuda' for GPU (10-100x faster, ~8GB VRAM) or 'cpu' (very slow, no VRAM needed). Always use CUDA if available.
torch_dtypeCOMBObfloat16Model precision: bfloat16 (best quality, stable, 8GB VRAM), float16 (good quality, 6GB VRAM), float32 (max quality, 16GB VRAM), auto (selects best for your GPU).
quantizationCOMBOnoneVRAM reduction: 'none' (best quality, 8GB VRAM), int8 (good quality, 4GB VRAM), int4 (acceptable quality, 3GB VRAM), int4_awq (optimized int4, requires pre-quantized model). Use if low on VRAM.
attention_mechanismCOMBOsdpaAttention layer: sdpa (fastest, good VRAM, default), eager (slowest, most stable), flash_attn (fastest, needs RTX 30xx+), sage_attn (best VRAM efficiency).
edit_typeCOMBOemotionEdit category: emotion (happy, sad, angry), style (whisper, formal), speed (faster/slower), paralinguistic (laughter, breathing), denoising (clean audio). Only one type per edit.
emotionCOMBOnoneTarget emotion (only if edit_type=emotion): happy, sad, angry, excited, calm, fearful, surprised, disgusted. Changes the emotional tone while keeping voice identity.
styleCOMBOnoneSpeaking style (only if edit_type=style): whisper, gentle, serious, casual, formal, friendly. Changes delivery style while keeping voice and emotion.
speedCOMBOnoneSpeed adjustment (only if edit_type=speed): faster (1.2x), slower (0.8x), more faster (1.5x), more slower (0.6x). Changes tempo without pitch shift.
paralinguisticCOMBOnoneSound effect (only if edit_type=paralinguistic): [Laughter], [Breathing], [Sigh], [Gasp], [Cough]. Inserts natural non-speech sounds. If paralinguistic_text is empty, auto-appends effect to end of audio.
denoisingCOMBOnoneNoise removal (only if edit_type=denoising): 'denoise' (remove background noise), 'vad' (voice activity detection, remove silence). Cleans up audio quality.
paralinguistic_textSTRINGText location for effect insertion (only if edit_type=paralinguistic). Example: 'I love this' - effect inserts before/after this phrase. Leave empty to auto-append effect to end of audio_text.
n_edit_iterationsINT11–5Edit strength through iteration: 1 (subtle change), 2-3 (moderate, recommended), 4-5 (strong, may degrade quality). Each iteration re-applies the edit to amplify the effect.
temperatureFLOAT0.70.1–2Voice variation: 0.1-0.5 (consistent), 0.6-0.8 (natural, recommended), 0.9-1.5 (expressive), 1.6-2.0 (creative). Note: Edit mode hardcoded to 0.7, this parameter has no effect.
do_sampleBOOLEANtrueSampling mode: True (natural speech) or False (deterministic, greedy decoding). Keep True. Note: Edit mode hardcoded to True, this parameter has no effect.
max_new_tokensINT8192256–16384Maximum audio tokens to generate: 2048 (~10s), 4096 (~20s), 8192 (~40s). Higher = more VRAM + time. Note: Edit mode hardcoded to 8192, this parameter has no effect.
seedINT00–18446744073709550000Reproducibility seed: 0 (random output each time) or any number (same input = same output). Use fixed seed to reproduce exact voice outputs.
keep_model_in_vramBOOLEANtrueModel caching: True (keep loaded, fast repeated use) or False (unload after use, frees ~8GB VRAM). Disable if switching between multiple models frequently.
input_audiooptAUDIOSource audio to modify (0.5-30 seconds). The AI will apply the selected edit (emotion/style/speed/etc) while preserving voice identity and content.

Outputs (1)

NameTypeDescription
audioAUDIO