refine-vocal-with-rvc

name	refine-vocal-with-rvc
description	Use when a rough DiffSinger vocal is ready and you want to optionally improve its timbre and naturalness using an RVC voice conversion model before mixing. This is a refinement step, not the core singing step.
allowed-tools	Read Bash Grep Glob Write
argument-hint	["rough-vocal-wav"]
effort	medium

Skill 4: refine_vocal_with_rvc

Purpose

Optionally improve the DiffSinger vocal by running it through an RVC/SVC voice conversion model.

This is a refinement step, not the core singing step.

Inputs

rough_vocal.wav

Config:

{
  "rvc_model": "models/rvc/singer_b.pth",
  "rvc_index": "models/rvc/singer_b.index",
  "transpose": 0,
  "output": "refined_vocal.wav"
}

Tools

Use an RVC CLI or Python wrapper.

RVC converts existing vocal audio into a target voice/timbre. It does not create singing from symbolic notes. That is why this skill comes after DiffSinger.

Responsibilities

1. Validate input WAV.
2. Run RVC/SVC inference.
3. Preserve timing and length as much as possible.
4. Detect clipping/silence/artifacts.
5. Write refined_vocal.wav and rvc_log.json.

Outputs

refined_vocal.wav
rvc_log.json

Example log:

{
  "input": "rough_vocal.wav",
  "output": "refined_vocal.wav",
  "model": "models/rvc/singer_b.pth",
  "transpose": 0,
  "duration_seconds": 23.42,
  "warnings": []
}

Why this output is useful

DiffSinger solves structure:

notes + lyrics → sung vocal

RVC solves timbre/texture:

rough synthetic vocal → more natural/pleasant vocal

Keep this step optional because it can sometimes:

• blur consonants
• add artifacts
• reduce lyric intelligibility
• over-color the singer identity

The pipeline should always preserve both:

rough_vocal.wav
refined_vocal.wav

for A/B comparison.

Suggested CLI

refine_vocal_with_rvc rough_vocal.wav \
  --model models/rvc/singer_b.pth \
  --index models/rvc/singer_b.index \
  --transpose 0 \
  --out refined_vocal.wav \
  --log rvc_log.json

Failure modes

Fail if:

• input WAV missing or silent
• RVC model missing
• inference exits nonzero
• output WAV missing

Warn if:

• output duration differs significantly from input
• output clips above -1 dBFS
• spectral/noise checks suggest severe artifacts