DeepFuze Save Audio

The DeepFuze TTS node hands you an AUDIO object, but a live object isn't a file. This node is the "make it permanent" step: it takes any incoming AUDIO and writes it out as a timestamped wav into ComfyUI/input/audio, so you can actually find and reuse what the voice cloning produced. It's a small node, and it's honest about being one - there's no magic here, just persistence. You'll usually chain it right after TTS generation, and often into Play Audio afterwards to hear the saved result.

How it works

Mechanically it's simple: the node takes the incoming audio (a waveform + sample rate pair), saves it via torchaudio.save into the input folder with a time-based filename, and reports the filename in the UI. Decoding uses ffmpeg under the hood, which the pack resolves from imageio-ffmpeg or a system install. It's marked as an output node, so it also passes the AUDIO back out - you can chain save → play without losing the data.

Inputs and output

• audio (required) - the clip to save, typically the output of DeepFuze TTS or the audio output of a Lipsync/FaceSwap run.
• start_time and end_time - dropdowns 0–9999 that the README bills as trimming the clip. Here's the honest bit: in the shipped source these two fields aren't actually applied - the node saves the whole clip regardless. If you need a trim, cut it externally or do it in the audio editor; don't fight this widget.
• AUDIO output - a pass-through of what you saved, so it stays usable downstream.

Install

It's part of the DeepFuze pack, so one install covers it:

cd ComfyUI/custom_nodes
git clone https://github.com/SamKhoze/ComfyUI-DeepFuze

then pip install -r requirements.txt inside the pack folder and restart ComfyUI. Windows users can skip the terminal and install via ComfyUI Manager (search "DeepFuze"). Two heavy prerequisites lurk in the README: Visual Studio C++ Build Tools on Windows, and optional CUDA 11.8 + cuDNN if you want GPU speed - CPU works, just slower. Model files for the pack's face/video nodes go in ComfyUI/models/deepfuze/, though this node itself needs none of them.

Troubleshooting

The classic failure here isn't the save - it's the audio not arriving in a sane format. The README's standing advice is to load MP3s and route audio through the AudioScheduler node to avoid sample-rate errors. And since the pack is unmaintained (the README itself admits it isn't compatible with the latest ComfyUI), if you see odd audio errors, that's usually the underlying cause. If you only ever wanted the file and not the node, you can skip this entirely - but it's the difference between a voice clone you can reuse and one that evaporates when the queue clears.