🎞️ MOSS Clip Staging

This is step one of training your own MOSS-TTS voice adapter inside ComfyUI. Before you can teach a model a voice, you need a tidy pile of audio clips on disk in a known place. That's all this node does: it takes your audio and stages it - writes the clips into a training folder and packages them as a MOSS_CLIP_DATASET handle the rest of the training chain reads.

It's deliberately dull, and that's the point. TTS Audio Suite added MOSS LoRA training as a unified, node-based flow instead of a detached command-line script, and staging is the intake gate. Nothing downstream works until your clips are staged.

How it works

You feed it audio and a dataset name. It copies/writes the audio into a training subdirectory under ComfyUI/output/tts_audio_suite_training/moss_tts/ and emits a dataset object describing what it staged. The next node in the chain (Dataset Rows, then Dataset Prep) consumes that object rather than raw files, so the whole pipeline stays connected by wires instead of by fragile file paths you have to keep in sync by hand.

The training UX here is admittedly early - the author has said the dataset-building side still needs polish - but the end-to-end flow is functional.

The inputs and outputs that matter

• dataset_name (default MyMossDataset) - the label for this dataset. It becomes part of the folder path and how you refer to the set later. Give each voice project its own name.
• opt_audio1 (AUDIO, optional) - the audio to stage. Wire your source clip(s) in here.
• output_subdir (default tts_audio_suite_training/moss_tts/staged_audio) - where the staged files land. Leave it unless you have a reason to reorganize.
• overwrite (default on) - whether re-running replaces what's already staged under this name. On is convenient while iterating; turn it off if you're accumulating clips and don't want a re-run to wipe them.

Outputs: clip_dataset (MOSS_CLIP_DATASET) - the handle you pass to MOSS Dataset Rows next - and dataset_info (STRING), a summary of what got staged so you can sanity-check the count.

How to install it

Part of TTS Audio Suite. ComfyUI Manager: search TTS Audio Suite, install, restart. Manual:

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

then restart. On Linux, install portaudio19-dev and libsamplerate0-dev before running the installer. Staging itself downloads no model - but the training run it feeds into pulls the MOSS-TTS model (several GB) plus its tokenizer, so make sure you have the disk and VRAM headroom before you commit to the full pipeline.

Common issues & troubleshooting

Where's my staged data? Under ComfyUI/output/tts_audio_suite_training/moss_tts/ by default. Check dataset_info for the exact path and the clip count - if the count is zero, no audio actually reached opt_audio1.

A re-run wiped my clips. overwrite defaults to on. If you're building a dataset across several runs, turn it off so each run adds rather than replaces.

This is only step one. Staging alone doesn't train anything. The full MOSS training flow is: Clip Staging → Dataset Rows (pair clips with their transcripts) → Dataset Prep (encode features) → Training Config → the unified Model Training node. Skipping the middle steps is the usual reason "training won't start."

MOSS training scope is narrow right now. As shipped, this pipeline targets MOSS-TTS 8B (Delay) LoRA training specifically. If you're trying to train a different engine through it, that's not supported yet - RVC has its own separate training nodes.