Simple RVC

A simple, high-quality voice conversion tool focused on simplicity and ease of use.

Key Features

• 20+ pitch extraction methods: pm, dio, crepe, fcpe, rmvpe, harvest, yin, pyin, swipe, djcm and more
• Hybrid f0 methods: combine multiple pitch extractors (e.g. hybrid[rmvpe+fcpe]) for robust results
• Powerful CLI for single-file and batch processing
• Multiple embedder models: contentvec, hubert (multilingual), spin
• Advanced features: formant shifting, noise reduction, autotune with adjustable strength, proposal pitch
• REST API: FastAPI-based HTTP server for integration into any application
• Eager model loading: Hubert and RMVPE models preload at startup for faster first inference
• ONNX export for optimized inference
• Multi-backend support: NVIDIA CUDA, AMD OpenCL, Apple MPS, CPU fallback

Quick Start

Try it in Google Colab

Click the badge below to open a ready-to-run demo notebook — no local installation required:

Install

pip install git+https://github.com/uziproj/rvc.git

CLI

rvc -i input.wav -o output.wav -m model.pth -p 12

Python API

from rvc.infer.infer import run_inference_script
from rvc.lib.config import Config

# Hubert and RMVPE models load eagerly at Config initialization
config = Config(embedder_model="contentvec_base", f0_method="rmvpe")

run_inference_script(
    config=config,
    input_path="input.wav",
    output_path="output.wav",
    pth_path="model.pth",
    pitch=12,
    f0_method="rmvpe",
)

Installation

Requires Python 3.10-3.12 and FFmpeg installed

Option 1: Install from Git (Recommended)

pip install git+https://github.com/uziproj/rvc.git

Option 2: Install from Source (Development)

git clone https://github.com/uziproj/rvc.git
cd rvc
pip install -e .

Python Version Compatibility

This project supports Python 3.10, 3.11, and 3.12. The code has been updated to be compatible with newer numpy versions (2.0+) which resolves compatibility issues that previously existed with Python 3.11+ installations.

Platform-Specific Notes

NVIDIA GPU:

• PyTorch with CUDA support is automatically installed
• Ensure you have CUDA drivers installed

AMD GPU:

• May require additional setup for OpenCL support
• Consider using ROCm PyTorch if available for your hardware

CPU Only:

• The package will automatically use CPU inference
• Note: CPU inference will be slower than GPU

Command Line Interface

After installation, you can use the RVC CLI tool for voice conversion:

rvc -i input.wav -o output.wav -m model.pth

Basic Options:

Flag	Description	Default
-i, --input	Path to input audio file or directory	(required)
-o, --output	Path to output audio file	./output.wav
-m, --model	Path to .pth model file	(required)

Common Options:

Flag	Description	Default
-p, --pitch	Pitch shift in semitones	0
-f0, --f0_method	F0 prediction method	rmvpe
-idx, --index	Path to .index file	None
-ir, --index_rate	Index rate for feature retrieval	0.5
-em, --embedder	Embedder model name	contentvec_base

Advanced Options:

Flag	Description	Default
-split, --split_audio	Split audio into chunks for processing	False
-clean, --clean_audio	Apply noise reduction to output	False
-cs, --clean_strength	Noise reduction strength (0.0-1.0)	0.7
-fa, --f0_autotune	Enable F0 autotune	False
-fas, --f0_autotune_strength	Autotune strength (0.0-1.0)	1.0
-fs, --formant_shifting	Enable formant shifting	False
-fq, --formant_qfrency	Formant quefrency	0.8
-ft, --formant_timbre	Formant timbre	0.8
-pp, --proposal_pitch	Enable proposal pitch	False
-ppt, --proposal_pitch_threshold	Proposal pitch threshold	255.0
-fr, --filter_radius	Filter radius for pitch extraction	3
-hl, --hop_length	Hop length for pitch extraction	64
-rs, --resample_sr	Resample output sample rate (0=disabled)	0
-fmt, --format	Output format (wav, flac, mp3, ogg)	wav

Available F0 Methods

Method	Quality	Speed	Description
pm	Low	Fastest	Parselmouth (Praat) autocorrelation — Boersma (1993)
dio	Low	Fast	PyWorld DIO — Morise & Kawahara (2010)
harvest	Medium	Medium	PyWorld Harvest — Morise (2017)
yin	Medium	Medium	YIN — de Cheveigné & Kawahara (2002)
pyin	Medium	Medium	pYIN — Mauch & Dixon (2014)
swipe	Medium	Medium	SWIPE — Camacho & Harris (2008)
rmvpe	High	Slow	RMVPE (U-Net) — Wei et al. (2023, Inter Speech)
rmvpe-legacy	High	Slow	RMVPE with StoneMask pitch refinement
fcpe	High	Slow	FCPE (Lynx-Net) — CNChTu (2025)
fcpe-legacy	High	Slow	FCPE with higher voicing threshold
crepe-*	Varies	Varies	CREPE CNN — Kim et al. (2018) — 5 sizes: tiny/small/medium/large/full
mangio-crepe-*	Varies	Varies	CREPE (Mangio fork) — no periodicity filter, raw Viterbi output
djcm	High	Slow	DJCM — Wei et al. (2024, ICASSP)
hybrid[m1+m2]	High	Slow	Weighted geometric mean of multiple methods

Example Usage

# Simple conversion
rvc -i input.wav -o output.wav -m model.pth -p 12

# Batch conversion from directory
rvc -i ./audio_folder -m model.pth -p 12 -f0 rmvpe

# With index file and autotune
rvc -i input.wav -m model.pth -idx model.index -ir 0.75 -fa

# With noise reduction and formant shifting
rvc -i input.wav -o output.wav -m model.pth -clean -fs -fq 0.9 -ft 0.7

# Using CREPE large for highest quality
rvc -i input.wav -o output.wav -m model.pth -f0 crepe-large

# Hybrid method combining RMVPE and FCPE
rvc -i input.wav -o output.wav -m model.pth -f0 "hybrid[rmvpe+fcpe]"

For more options, run:

rvc --help

REST API

RVC includes a built-in REST API server powered by FastAPI, allowing you to integrate voice conversion into any application over HTTP.

Starting the Server

# Using the CLI entry point
rvc-api --host 0.0.0.0 --port 8000

# Or with uvicorn directly
uvicorn rvc.api.app:app --host 0.0.0.0 --port 8000

The interactive API docs are available at http://localhost:8000/docs (Swagger UI) and http://localhost:8000/redoc (ReDoc).

Quick Example

# 1. Load a voice model
curl -X POST http://localhost:8000/api/v1/models/load \
  -H "Content-Type: application/json" \
  -d '{"model_path": "/path/to/model.pth"}'
# Response: {"model_id": "abc123def456", "model_path": "...", "version": "v2", ...}

# 2. Convert audio (upload file)
curl -X POST http://localhost:8000/api/v1/convert \
  -F "audio=@input.wav" \
  -F "model_id=abc123def456" \
  -F "pitch=12" \
  -F "f0_method=rmvpe" \
  -o output.wav

# 3. Convert audio (server file path)
curl -X POST http://localhost:8000/api/v1/convert/file \
  -H "Content-Type: application/json" \
  -d '{"model_id": "abc123def456", "input_path": "/path/to/input.wav", "pitch": 12}'

# 4. Unload model
curl -X DELETE http://localhost:8000/api/v1/models/abc123def456

API Endpoints

Method	Endpoint	Description
GET	/api/v1/health	Health check & system status
GET	/api/v1/config	Current server configuration
GET	/api/v1/methods	List available F0 methods
GET	/api/v1/embedders	List available embedder models
POST	/api/v1/models/load	Load a voice model into memory
GET	/api/v1/models	List all loaded models
GET	/api/v1/models/{model_id}	Get info about a loaded model
DELETE	/api/v1/models/{model_id}	Unload a model from memory
POST	/api/v1/convert	Convert audio (file upload, returns audio stream)
POST	/api/v1/convert/file	Convert audio (server paths, returns output path)

For full request/response schemas, see the interactive docs at /docs.

Models

Using CLI:

• Specify the model path with the -m option
• Specify the index path with the -idx option (recommended for better voice quality)

Model Download: Predictor and embedder models are automatically downloaded from HuggingFace. When using the Python API, models are preloaded at Config() initialization for faster first inference. The CLI and API server also preload models at startup.

Note: Pre-trained RVC models (.pth files) can be downloaded from various sources. Ensure you have the right to use any model before converting audio with it.

Recent Bug Fixes

The following bugs have been identified and fixed in this repository:

Critical Fixes

• HiFi-GAN Generator: forward() method was incorrectly nested inside __init__(), making the vocoder completely non-functional
• ONNX Export: SynthesizerONNX class didn't exist, causing ImportError on import
• FCPE Model: SelfAttention was missing use_norm parameter, causing TypeError during model creation
• F0 Method Lookup: compute_f0() was splitting method names on -, causing all CREPE and legacy methods to silently fall back to PM
• Noisereduce Import: Wrong import path rvc.lib.tools.noisereduce (should be rvc.tools.noisereduce) — noise reduction was broken
• DJCM OpenCL: Wrong import from main.library.backends.utils import STFT — DJCM on OpenCL devices was broken
• PyWorld DLL: Was attempting to open the assets directory as a file instead of assets/models/world.bin
• Memory Cleanup: VoiceConverter.cleanup() double-deleted net_g causing AttributeError
• DJCM Window Length: WINDOW_LENGTH reassignment inside if svs: block caused UnboundLocalError when svs=False

Significant Fixes

• Autotune Strength: f0_autotune_strength defaulted to False instead of 1.0, making autotune ineffective by default
• FCPE Gaussian Blur: Operator precedence bug caused * to bind before &, producing boolean tensors instead of float probabilities
• FCPE Decoder: Both "argmax" and "local_argmax" decoder options mapped to the same decoder function
• GELU Activation: gelu_accurate() returned None on all calls after the first
• Fairseq Assertion: assert src_len, key_bsz == value.shape[:2] was parsed as a tuple assertion (always True)
• Cross-Attention: SelfAttention.forward() left out undefined when cross_attend=True

For the complete list, see DOCUMENTATION.md.

Disclaimer

• The RVC project is developed for research, educational, and personal entertainment purposes. I do not encourage, nor do I take any responsibility for, any misuse of voice conversion technology for fraudulent purposes, identity impersonation, or violations of privacy or copyright belonging to any individual or organization.

• Users are solely responsible for how they use this software and must comply with the laws and regulations of the country in which they reside or operate.

• The use of voices of celebrities, real people, or public figures must be authorized or ensured not to violate any applicable laws, ethical standards, or the rights of the individuals involved.

• The author of this project holds no legal liability for any consequences arising from the use of this software.

Terms of Use

• You must ensure that any audio content you upload and convert through this project does not infringe upon the intellectual property rights of any third party.

• This project must not be used for any illegal activity, including but not limited to fraud, harassment, or causing harm to others.

• You are fully responsible for any damages resulting from improper use of the product.

• I am not liable for any direct or indirect damages arising from the use of this project.

Documentation

For detailed documentation, including API references, troubleshooting guides, and advanced usage, see DOCUMENTATION.md.

Project-based construction

• Algorithm: Vietnamese RVC