# RVC Command-Line Training and Offline Inference

[中文文档](../cn/cli.md)

This document covers RVC training and offline voice conversion. PyMSS/UVR5 source separation uses a separate CLI.

## Working directory

Run every command from the project's actual installation root. Replace the placeholder with your own path:

```powershell
Set-Location "<RVC_PROJECT_ROOT>"
```

Select either the bundled Python runtime or a system Python that has the project dependencies installed. Keep one assignment:

```powershell
$PYTHON = "runtime\python.exe"  # Bundled Python
# $PYTHON = "python"            # System Python
& $PYTHON --version
```

## Single-speaker and multi-speaker relationship

The training architecture always includes a speaker-ID embedding. At the model level, multi-speaker training therefore includes the single-speaker case, and a multi-speaker manifest containing only one ID is valid.

The project still maintains two explicit data conventions:

| Area | Single-speaker | Multi-speaker |
| --- | --- | --- |
| Dataset input | Audio files directly inside one directory | `Name_ID_Repeat` subdirectories below a root directory |
| Speaker ID | One ID assigned to every training record | One ID per manifest entry |
| Small-model metadata | No `speaker_info` key | Named IDs stored in `speaker_info` |
| Inference selection | ID 0 by default; another valid ID may be explicit | Names and IDs are exposed; the smallest ID is the default |
| Total feature file | `total_fea.npy` | One `total_fea_spkid<ID>.npy` per speaker |
| Added index | Standard index filename | One `_spkid<ID>.index` per speaker |

Use single-speaker mode for one voice. Use multi-speaker mode when several named voices must be selectable from one model.

## Training workflow

The stages are dataset preprocessing, F0 extraction, HuBERT feature extraction, model training, and index training.

### Single-speaker preprocessing

```powershell
& $PYTHON train\preprocess.py "DATASET_DIRECTORY" 40000 8 "logs\EXPERIMENT" False 3.7
```

The arguments are input directory, sample rate, worker count, experiment directory, disable-multiprocessing flag, and slicing parameter.

### Multi-speaker manifest and preprocessing

Subdirectory names use `Name_ID_Repeat`, and IDs range from `0` through `109`.

```powershell
& $PYTHON -c "from tools.multispeaker import build_manifest_from_root,write_manifest; write_manifest(r'logs\EXPERIMENT',build_manifest_from_root(r'DATASET_ROOT'))"

& $PYTHON train\preprocess.py "" 40000 8 "logs\EXPERIMENT" False 3.7 "logs\EXPERIMENT\multispeaker_manifest.json"
```

The repeat count duplicates final `filelist.txt` records only. Preprocessing, F0 extraction, and HuBERT extraction are not repeated.

### F0 extraction

CPU with PM:

```powershell
& $PYTHON train\dataset\extract_f0.py cpu "logs\EXPERIMENT" 8 pm
```

CPU with RMVPE:

```powershell
& $PYTHON train\dataset\extract_f0.py cpu "logs\EXPERIMENT" 8 rmvpe
```

CUDA with RMVPE:

```powershell
& $PYTHON train\dataset\extract_f0.py cuda 1 0 0 "logs\EXPERIMENT" true
```

The CUDA arguments are total process count, current process index, GPU ID, experiment directory, and half-precision flag. Start one command for each process index when using several processes.

DirectML with RMVPE:

```powershell
& $PYTHON train\dataset\extract_f0.py dml "logs\EXPERIMENT"
```

### HuBERT feature extraction

CUDA:

```powershell
& $PYTHON train\dataset\extract_hubert_feature.py cuda:0 1 0 0 "logs\EXPERIMENT" v2 true
```

CPU:

```powershell
& $PYTHON train\dataset\extract_hubert_feature.py cpu 1 0 "logs\EXPERIMENT" v2 false
```

DirectML:

```powershell
& $PYTHON train\dataset\extract_hubert_feature.py privateuseone:0 1 0 "logs\EXPERIMENT" v2 false
```

### Prepare `config.json` and `filelist.txt`

Before starting `train.py`, these files must exist:

```text
logs\EXPERIMENT\config.json
logs\EXPERIMENT\filelist.txt
```

A single-speaker F0 record has this format:

```text
wav_path|HuBERT_feature_path|coarse_F0_path|continuous_F0_path|speaker_ID
```

A multi-speaker F0 record adds the speaker name:

```text
wav_path|HuBERT_feature_path|coarse_F0_path|continuous_F0_path|speaker_ID|speaker_name
```

Remove both F0 path fields for a non-F0 model. A multi-speaker `config.json` also needs:

```json
{
  "model": {
    "spk_embed_dim": 110
  },
  "speaker_info": [
    {"id": 0, "name": "Speaker A"},
    {"id": 1, "name": "Speaker B"}
  ]
}
```

The WebUI Model Training and One-click Training actions generate both files automatically. A pure CLI workflow prepares them before invoking `train.py`.

### Model training

```powershell
& $PYTHON train\train.py -e "EXPERIMENT" -sr 40k -f0 1 -bs 8 -g 0 -te 200 -se 5 -pg assets\pretrained_v2\f0G40k.pth -pd assets\pretrained_v2\f0D40k.pth -l 0 -c 0 -sw 1 -v v2
```

Common options:

| Option | Meaning |
| --- | --- |
| `-e` | Experiment name under `logs` |
| `-sr` | `32k`, `40k`, or `48k` |
| `-f0` | `1` enables F0; `0` disables it |
| `-bs` | Batch size per GPU |
| `-g` | GPU IDs such as `0` or `0-1`; omit for CPU training |
| `-te` | Total epochs, up to 1200 |
| `-se` | Checkpoint interval in epochs |
| `-l` | `1` keeps only the latest large checkpoint |
| `-c` | `1` caches the dataset in GPU memory |
| `-sw` | `1` also saves inference-ready small models |
| `-v` | `v1` or `v2` |

### Index training

Single-speaker:

```powershell
& $PYTHON train\train_index.py "EXPERIMENT" v2 "assets\indices" 8 single
```

Multi-speaker:

```powershell
& $PYTHON train\train_index.py "EXPERIMENT" v2 "assets\indices" 8 multi
```

Multi-speaker mode does not merge every speaker's features into one index. The script groups features by speaker ID and separately creates `total_fea_spkid<ID>.npy`, `trained_..._spkid<ID>.index`, and `added_..._spkid<ID>.index` for each speaker. Every speaker in the model therefore has its own index; after selecting a speaker for inference, use the added index for that ID.

Automatic manifest detection:

```powershell
& $PYTHON train\train_index.py "EXPERIMENT" v2 "assets\indices" 8 auto
```

## Offline inference CLI

The entry point is `infer/cli.py`.

Show help:

```powershell
& $PYTHON infer\cli.py --help
```

### List model speakers

```powershell
& $PYTHON infer\cli.py --model "assets\weights\MODEL.pth" --list-speakers
```

Multi-speaker models with `speaker_info` print IDs and names. Older or single-speaker models print the valid ID range.

### Single-speaker inference

```powershell
& $PYTHON infer\cli.py `
  --model "assets\weights\SINGLE.pth" `
  --input "D:\input.wav" `
  --output "D:\output.wav" `
  --pitch 0 `
  --f0-method rmvpe `
  --index-rate 0.75
```

Single-speaker models default to ID 0. Use `--speaker-id` for an older model that contains several unnamed IDs.

### Multi-speaker inference

When `--speaker-id` is omitted, the CLI selects the smallest declared ID and automatically matches its `_spkid<ID>.index`:

```powershell
& $PYTHON infer\cli.py `
  --model "assets\weights\MULTI.pth" `
  --input "D:\input.wav" `
  --output "D:\output.wav" `
  --index-rate 1
```

Select a specific speaker:

```powershell
& $PYTHON infer\cli.py `
  --model "assets\weights\MULTI.pth" `
  --speaker-id 1 `
  --input "D:\input.wav" `
  --output "D:\output.wav" `
  --index-rate 1
```

### Batch inference

Directory input scans direct child files. Add `--recursive` to scan subdirectories and preserve their relative layout:

```powershell
& $PYTHON infer\cli.py `
  --model "assets\weights\MODEL.pth" `
  --speaker-id 0 `
  --input "D:\input_audio" `
  --output "D:\output_audio" `
  --format flac `
  --recursive `
  --overwrite
```

Output formats are `wav`, `flac`, `mp3`, and `m4a`. Input extension matching is case-insensitive.

### Index behavior

- Omit `--index` to match an added index by model name and speaker ID.
- Supply `--index` to select a path explicitly. A trained filename is converted to its added counterpart.
- If `--index-rate` is greater than zero and the index is missing, the command exits with an error instead of silently disabling retrieval.
- Use `--index-rate 0` to disable retrieval.

Common inference options:

| Option | Default | Meaning |
| --- | --- | --- |
| `--speaker-id` | Single: 0; multi: smallest declared ID | Inference speaker |
| `--pitch` | `0` | Pitch shift in semitones |
| `--f0-method` | `rmvpe` | `pm` or `rmvpe` |
| `--index-rate` | `0.75` | Retrieval blend ratio |
| `--resample-sr` | `0` | Output sample rate; 0 keeps the model rate |
| `--rms-mix-rate` | `1.0` | Output RMS-envelope blend ratio |
| `--protect` | `0.33` | Consonant and breath protection, from 0 to 0.5 |
| `--overwrite` | Off | Replace existing output files |

Successful completion returns exit code 0, an inference failure returns 1, and user interruption returns 130.

## CLI Inference For The Current WebUI PyMSS Models

This section only covers the five models exposed by the WebUI's Vocal/Instrumental Separation and Dereverb page. Invoke PyMSS as a module:

```powershell
& $PYTHON -m tools.pymss.cli infer --help
```

The WebUI labels map to CLI catalog model names as follows:

| WebUI action | CLI model name |
| --- | --- |
| Dereverb | `dereverb_mel_band_roformer_less_aggressive_anvuew_sdr_18.8050.ckpt` |
| Dereverb (aggressive) | `dereverb_mel_band_roformer_anvuew_sdr_19.1729.ckpt` |
| Remove instrumental | `model_bs_roformer_ep_368_sdr_12.9628.ckpt` |
| Remove instrumental (aggressive) | `model_bs_roformer_ep_317_sdr_12.9755.ckpt` |
| Extract lead vocal | `model_mel_band_roformer_karaoke_aufr33_viperx_sdr_10.1956.ckpt` |

`--input` accepts one audio file or a directory, while `--output` is an output directory. Missing files in the CLI model cache are downloaded automatically.

### Dereverb

```powershell
& $PYTHON -m tools.pymss.cli infer `
  "dereverb_mel_band_roformer_less_aggressive_anvuew_sdr_18.8050.ckpt" `
  --input "<input audio or directory>" `
  --output "<output directory>" `
  --device auto `
  --format flac
```

### Dereverb (Aggressive)

```powershell
& $PYTHON -m tools.pymss.cli infer `
  "dereverb_mel_band_roformer_anvuew_sdr_19.1729.ckpt" `
  --input "<input audio or directory>" `
  --output "<output directory>" `
  --device auto `
  --format flac
```

### Remove Instrumental

```powershell
& $PYTHON -m tools.pymss.cli infer `
  "model_bs_roformer_ep_368_sdr_12.9628.ckpt" `
  --input "<input audio or directory>" `
  --output "<output directory>" `
  --device auto `
  --format flac
```

### Remove Instrumental (Aggressive)

```powershell
& $PYTHON -m tools.pymss.cli infer `
  "model_bs_roformer_ep_317_sdr_12.9755.ckpt" `
  --input "<input audio or directory>" `
  --output "<output directory>" `
  --device auto `
  --format flac
```

### Extract Lead Vocal

```powershell
& $PYTHON -m tools.pymss.cli infer `
  "model_mel_band_roformer_karaoke_aufr33_viperx_sdr_10.1956.ckpt" `
  --input "<input audio or directory>" `
  --output "<output directory>" `
  --device auto `
  --format flac
```

For directory batches, add `--save-as-folder` to create a separate result folder for each input. On CUDA, replace `--device auto` with `--device cuda --device-id 0`; for CPU, use `--device cpu`. Output formats are `wav`, `flac`, `mp3`, and `m4a`.