vocal

Rust implementation scaffold of RVC (Retrieval-based Voice Conversion) in ~/repos/maolan/vocal.

Status

This project implements a functional RVC-style pipeline in pure Rust:

1. WAV I/O
2. Content feature extraction (HuBERT-like proxy features)
3. F0 extraction (frame-wise proxy)
4. Retrieval blending (nearest-neighbor feature mixing)
5. Voice conversion synthesis (baseline resynthesis/pitch-shift proxy)
6. Output post-processing (RMS mix)
7. Burnpack model artifact loading (.bpk) in the same style used by maolan/generate

It is designed to mirror RVC's architecture and CLI flow while remaining dependency-light and easy to extend.

CLI

Create config

cargo run --release -- new-config --output config.json

Inference

cargo run --release -- infer \
  --input-wav input.wav \
  --output-wav output.wav \
  --burnpack rvc_model.bpk \
  --config config.json \
  --pitch-shift-semitones 3 \
  --index-mix 0.2 \
  --rms-mix-rate 0.3 \
  --target-sr 48000

Inspect a burnpack

cargo run --release -- inspect-burnpack --model rvc_model.bpk

Convert PyTorch checkpoint to burnpack

cargo run --release -- convert-pth \
  --input model.pth \
  --output rvc_model.bpk \
  --top-level-key state_dict

Project layout

• src/main.rs: CLI (infer, new-config)
• src/config.rs: RVC config schema
• src/audio.rs: WAV I/O + linear resampler + RMS
• src/rvc.rs: model components (HubertEncoder, F0Extractor, VoiceConverter)
• src/retrieval.rs: retrieval index + nearest-neighbor blending
• src/pipeline.rs: end-to-end inference pipeline

Notes on parity with upstream RVC

This is a Rust-native baseline and not yet bit-exact to upstream Python/PyTorch RVC. To reach full parity, next steps are:

1. Full RVC synth graph in Burn modules loaded from burnpack tensors.
2. RMVPE/Harvest/DIO-grade F0 extraction.
3. FAISS-compatible retrieval index loading and querying.
4. Exact frame slicing, protection, and infer-time blending behavior.

The current code is a clean foundation to implement those pieces incrementally in Rust.