⚙️ RVC Engine

First, the thing that trips everyone up: RVC Engine does not make speech from text. It's a voice conversion engine - it takes audio and re-voices it. Plug it into 🔄 Voice Changer, never into TTS Text or TTS SRT (it'll just error there). RVC - Retrieval-based Voice Conversion - is the same tech behind most of the "AI covers" you've heard, and it's excellent: give it a trained character model and a source clip, and it rebuilds that clip in the target voice while preserving the original delivery. This node bundles the model loading and all the conversion quality controls into one config that outputs a TTS_ENGINE for the Voice Changer.

How it works

RVC works by extracting the content of the source audio (the phonemes and pitch, via a HuBERT content encoder) and re-synthesizing it with the timbre of a trained target model. A FAISS index (loaded alongside the .pth) does the "retrieval" part - matching the source's content features to the target speaker's, which is what makes conversions sound like the target rather than a smeared average. Pitch extraction runs separately and is where a lot of the quality comes from; that's configurable via the optional 🔧 RVC Pitch Extraction Options node.

The inputs and outputs that matter

The four required knobs are your main quality dials:

• pitch (−14 to +14 semitones, default 0) - pitch shift. The critical one for cross-gender conversion: male-to-female or female-to-male target usually needs about ±12. Leave at 0 for same-range voices.
• index_ratio (0–1, default 0.75) - how much the FAISS index (target-voice retrieval) influences the result. Higher tracks the target accent/timbre harder; too high can introduce artifacts. 0.75 is a sane start.
• consonant_protection (0–0.5, default 0.25) - protects consonants and breath sounds from being over-smoothed, which keeps speech intelligible. Raise it if consonants turn mushy.
• volume_envelope (0–1, default 0.25) - how much the source's loudness dynamics are mimicked vs. the target's. Low keeps it steady.

Optional: hubert_model picks the content encoder - auto is fine; the recommended explicit choice is content-vec-best, with Japanese/Korean/Chinese HuBERT variants for those languages. rvc_pitch_options takes a bundle from the pitch-options node for finer pitch control. output_sample_rate (0 = keep native, or force 40k/48k etc.) and device round it out.

Output: a single TTS_engine - wire it into 🔄 Voice Changer's TTS_engine input.

Installing it

Ships with TTS Audio Suite. Easiest is ComfyUI Manager: search TTS Audio Suite, install, restart - Manager runs the pack's install.py, which handles the RVC dependencies along with the usual conflicts (NumPy, librosa, s3tokenizer) and Python 3.13. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/diodiogod/TTS-Audio-Suite.git, then python install.py from inside the folder with your ComfyUI venv active - run the script, or the RVC stack won't install. Linux: sudo apt-get install portaudio19-dev libsamplerate0-dev first. RVC base models (HuBERT, RMVPE) auto-download; character .pth models you load separately with 🎭 Load RVC Character Model, from ComfyUI/models/TTS/RVC/.

Common issues

• "It won't connect to TTS Text." By design - RVC converts, it doesn't synthesize. Use 🔄 Voice Changer. The node's own description says as much.
• Converted voice sounds off-pitch or chipmunky. Wrong pitch. Cross-gender needs roughly ±12 semitones; get this right before touching anything else.
• Robotic / artifact-heavy output. index_ratio too high, or a low-quality character model. Pull index_ratio toward 0.5, and remember RVC quality is capped by how well the .pth was trained.
• Muddy consonants. Nudge consonant_protection up toward 0.5.
• No target voice. The .pth model isn't loaded on this node - it comes from 🎭 Load RVC Character Model into the Voice Changer's narrator_target. If conversions sound generic, that link is probably missing.