ComfyUI-Audio-ToolBox

ComfyUI Audio ToolBox

A compact audio toolkit for ComfyUI. Inspect, resample, cut, merge, stretch, denoise, separate and voice-detect audio with regular V3 nodes under the audio/toolbox category.

Nodes

| Group | Nodes | | --- | --- | | Basics | ATB Resample, ATB Audio Info, ATB List Get, ATB Batch Make, ATB Batch Merge | | Editing | ATB Split On Silence, ATB Split Milli, ATB Trim Audio, ATB Pad To Length, ATB Speed, ATB Stretch To Duration, ATB Pitch Shift | | Models | ATB Denoise, ATB Separate, ATB VAD |

Highlights:

• Millisecond-accurate cutting with MM:SS.milli (01:23.456) or plain seconds (83.456).
• List outputs for splits, VAD segments and batches. Connect them straight to Preview Audio to audition every item, pick one with ATB List Get, or join them back with ATB Batch Merge.
• DeepFilterNet3 ONNX noise suppression, Demucs stem separation (including the 6-source guitar/piano model) and Silero VAD, all running locally.

Installation

1. Clone or copy this folder into ComfyUI/custom_nodes/ComfyUI-Audio-ToolBox.

2. Install the dependencies:

   pip install -r requirements.txt
   

3. Restart ComfyUI. The nodes appear under audio/toolbox, and the model folder ComfyUI/models/audio_toolbox is registered automatically.

torch, torchaudio, numpy and comfy_api come with ComfyUI and are deliberately not pinned here.

Dependencies

| Package | Needed by | Notes | | --- | --- | --- | | onnxruntime | ATB VAD, ATB Denoise (DeepFilterNet) | CPU provider is enough. | | demucs | ATB Separate | Loads .th Demucs checkpoints. | | torch, torchaudio, numpy | Everything | Already shipped with ComfyUI. |

Models

Put model files under ComfyUI/models/audio_toolbox/:

ComfyUI/models/audio_toolbox/
├─ htdemucs/
│  ├─ htdemucs.th          # 4 stems: vocals, drums, bass, other
│  └─ htdemucs_6s.th       # 6 stems: adds guitar and piano
├─ vad/
│  └─ silero_vad.onnx
└─ DeepFilterNet3/
   └─ onnx/
      ├─ enc.onnx
      ├─ erb_dec.onnx
      ├─ df_dec.onnx
      └─ config.ini        # optional overrides

| Model | Download | Used by | | --- | --- | --- | | htdemucs | https://dl.fbaipublicfiles.com/demucs/hybrid_transformer/955717e8-8726e21a.th | ATB Separate | | htdemucs_6s | https://dl.fbaipublicfiles.com/demucs/hybrid_transformer/5c90dfd2-34c22ccb.th | ATB Separate | | Silero VAD | https://raw.githubusercontent.com/snakers4/silero-vad/master/src/silero_vad/data/silero_vad.onnx | ATB VAD | | DeepFilterNet3 | Any DeepFilterNet3 ONNX export (enc.onnx, erb_dec.onnx, df_dec.onnx) | ATB Denoise |

Models all in one: https://pan.quark.cn/s/ee526e881983

Notes:

• All editing nodes and the spectral_gate backend of ATB Denoise work without any model files.
• The ATB Separate model list contains every .th file in the folder.
• The ATB VAD model list contains every .onnx file, including the DeepFilterNet parts. Pick vad/silero_vad.onnx.
• Model lists are built when ComfyUI starts, so restart after adding or removing files.
• Demucs models that only produce 4 stems still return all 6 outputs; the missing stems are silent.

Quick start

1. Split a long clip: Load Audio → ATB Split Milli (split_times = 0.5, 1.25, 2.750) → Preview Audio plays every segment; ATB List Get picks one.
2. Rebuild a batch: several clips → ATB Batch Make → ATB Batch Merge → one continuous audio.
3. Cut a line: ATB Trim Audio with start_time = 00:12.480, end_time = 00:18.000.
4. Voice only: Load Audio → ATB VAD → Preview Audio or ATB Batch Merge.
5. Karaoke: Load Audio → ATB Separate (htdemucs_6s) → vocals output.

Documentation

• English: docs/document.md
• 中文: docs/document_zh.md

License

Licensed under either of Apache-2.0 (LICENSE-APACHE) or MIT (LICENSE-MIT), at your option.

Bundled model files are not included. Demucs is MIT licensed, Silero VAD is MIT licensed and DeepFilterNet is MIT / Apache-2.0 licensed.