📦 RVC Dataset Prep

Training a voice model is 90% dataset prep and 10% pressing go. This node is the 90%. If you want to train an RVC (Real-time Voice Conversion) model - a .pth that turns any input voice into a target voice - you can't just point a trainer at a folder of clips. The audio has to be sliced into training-sized chunks, resampled to a consistent rate, run through a feature extractor (HuBERT), and pitch-analyzed (F0). This node does all of that and hands you a clean training_dataset for the 🎓 Model Training node.

It's the first node in the suite's training chain: RVC Dataset Prep → Model Training → Load RVC Character Model. Get the dataset right and the rest is mostly waiting.

How it works

You give it your source audio and a target sample rate, and it does the mechanical work: slicing long recordings into short chunks (with a little overlap so nothing gets cut mid-word), normalizing volume, extracting HuBERT features and F0 pitch curves, and caching the result so you don't reprocess the same data every time you tweak a training setting. That caching is the quiet win - dataset prep is slow, and reusing it across training runs saves real time.

The inputs and outputs that matter

• TTS_engine - required; wire in an RVC engine node. This tells the training chain which pipeline you're on.
• model_name - a name for what you're building (default "MyVoice"). It's how your dataset and eventual model are labeled.
• dataset_source - a path, a zip, or a folder of audio. Alternatively, feed audio directly into opt_audio1 (an AUDIO input) if it's already in your graph.
• training_sample_rate - 32k, 40k, or 48k. 40k is the standard default and a safe pick; 48k is higher fidelity but heavier, 32k is lighter. Match whatever you plan to train at.

The optional set is your quality/speed tuning: chunk_seconds and overlap_seconds control the slicing, max_volume and mute_ratio handle normalization and silence, cpu_workers parallelizes the (CPU-bound) processing, and reuse_existing (on by default) reuses that prep cache.

Two outputs: training_dataset (TRAINING_DATASET) → straight into the Model Training node, and dataset_info (text) telling you how many chunks it produced and whether it reused the cache.

Installing it

Part of the pack. 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

The HuBERT and RMVPE feature/pitch models this node needs are auto-managed - they download on first prep. On Linux you'll also want the system audio libs the pack asks for (portaudio19-dev, libsamplerate0-dev) so resampling works cleanly.

Common issues & troubleshooting

Garbage dataset from messy audio. RVC training is only as good as its input. Use clean, single-speaker audio with no background music or noise - the slicing and feature extraction can't fix a bad source, they just chop it up. A few minutes of clean speech beats an hour of noisy speech.

It reprocessed everything again. If prep is re-running when you only changed a training parameter, check reuse_existing is on - that's the flag that lets you reuse the cache across runs instead of paying the slicing/feature cost every time.

Wrong sample rate mismatch. Pick training_sample_rate to match your intended training config; a dataset prepped at 40k and a trainer expecting 48k is asking for trouble. Decide the rate once, up front.

Prep is slow. It's CPU-bound feature extraction. Raise cpu_workers to use more cores - that's the dial that actually moves prep time.