README.md

---
pipeline_tag: audio-to-audio
tags:
  - rvc
  - rvc-v2
  - applio
  - voice-conversion
  - retrieval-based-voice-conversion
  - contentvec
  - rmvpe
  - hifigan
  - 48khz
  - cyberpunk-2077
  - rogue
---

# Rogue — RVC v2 Voice Conversion Model for Applio

RVC v2 voice-conversion model for **Rogue**, trained with **Applio** from extracted in-game voice audio.

> **Important:** this is a **voice conversion** model, not a standalone text-to-speech model. It converts an existing speech or singing performance into the target timbre while retaining much of the source timing, phrasing, and pitch contour.

## Model files

The supplied Applio/RVC package contains:

| File | Purpose | Size |
|---|---|---:|
| `rogue.pth` | RVC model weights | 57,532,660 bytes / 54.87 MiB |
| `rogue.index` | Retrieval index used to reinforce the target timbre | 333,505,619 bytes / 318.06 MiB |

The uploaded package `rogue.zip` contains both files.

### SHA-256

```text
rogue.pth
b13ed63734886b2894cd667eafa26e4e966c6aea51e4422b6ee7457a71cdc210

rogue.index
dbe27e2d354f4e690b1459a2872a7f65a4d9c9f9efadc822d361abd06264bb9e

rogue.zip
d556db0d73b0e56ee48012b8e180d2e4c19a5b3b841ae7561c409b6a26435a21
```

## Model details

| Property | Value |
|---|---|
| Model type | Retrieval-Based Voice Conversion (RVC) |
| RVC version | **v2** |
| Training / inference app | **Applio** |
| Packaged model filename | `rogue.pth` |
| Packaged index filename | `rogue.index` |
| Internal training model name | `mara` |
| Target / donor voice | `rogue` |
| Sample rate | **48,000 Hz** |
| F0 / pitch guidance | **Enabled** |
| Pitch extractor used for feature extraction | **RMVPE** |
| Embedder | **ContentVec** |
| Vocoder | **HiFi-GAN** |
| Number of speakers | **1** |
| Published checkpoint epoch | **200** |
| Published checkpoint step | **26,200** |
| Embedded dataset length | `00:35:00` |
| Model creation timestamp | `2026-09-09T10:45:38.675607` |
| Embedded model hash | `802d9890b58d28f51309e9db116ff921280d948615abfcbb0e9069c2fb530720` |

### Naming note

The original Applio training sheet uses **`mara`** as the training model name and **`rogue`** as the donor/target voice. The distributed files were packaged as **`rogue.pth`** and **`rogue.index`**. The internal metadata inside `rogue.pth` still reports `model_name: mara`.

## Training dataset

The supplied dataset archive was verified as:

| Property | Value |
|---|---:|
| WAV files | **626** |
| Total duration | **35:00.422** |
| Sample rate | **48,000 Hz** for all 626 files |
| Channels | **Mono** for all 626 files |
| PCM sample width | **16-bit** for all 626 files |
| Shortest clip | **1.210 s** |
| Median clip | **2.900 s** |
| Mean clip | **3.355 s** |
| Longest clip | **11.732 s** |

The archive stores the clips under:

```text
mara/
  rogue_*.wav
```

The provided dataset contains audio files rather than a transcription dataset.

### Dataset preprocessing rationale

The source audio was extracted from a game rather than recorded from a microphone for this training run. The training sheet therefore keeps both **Process Effects** and **Noise Reduction** disabled. The source was already produced/mastered audio, and an additional cleanup pass could remove breath, sibilance, and other speaker-specific details useful to voice conversion.

## Applio training configuration

### 1. Preprocess

| Setting | Value |
|---|---|
| Model Name | `mara` |
| Sample Rate | `48000` |
| Cut Preprocess | `Automatic` |
| Chunk Length | `3.0` s |
| Overlap Length | `0.3` s |
| Process Effects | **OFF** |
| Noise Reduction | **OFF** |
| CPU Cores | Default |

### 2. Extract

| Setting | Value |
|---|---|
| Model Name | `mara` |
| Sample Rate | `48000` |
| Pitch Extractor / F0 | **RMVPE** |
| Embedder Model | **ContentVec** |
| GPU | `0` |

### 3. Train

| Setting | Value |
|---|---|
| Model Name | `mara` |
| Sample Rate | `48000` |
| Vocoder | **HiFi-GAN** |
| Batch Size | **8** |
| Configured Total Epochs | **300** |
| Save Every Epoch | **10** |
| Save Only Latest | **OFF** |
| Save Every Weights | **ON** |
| Pretrained | **ON** |
| Cache Dataset in GPU | **OFF** |
| GPU | `0` |

### Published checkpoint vs. configured training length

The training configuration was set to **300 total epochs** and to save weights every 10 epochs. However, the `rogue.pth` file published in this package identifies itself as **epoch 200, step 26,200**.

For that reason, this model card describes the distributed model as the **epoch-200 checkpoint** rather than calling it a 300-epoch model.

## Using the model in Applio

Applio expects an RVC model to use the `.pth` weights file and, when available, the matching `.index` retrieval file.

### Option 1 — Applio model downloader

If `rogue.zip` is uploaded to this Hugging Face repository, its direct file URL can be used in Applio's **Download Model** panel.

The URL format is:

```text
https://huggingface.co/<USERNAME>/<REPOSITORY>/resolve/main/rogue.zip
```

Replace `<USERNAME>` and `<REPOSITORY>` with the actual Hugging Face repository path.

### Option 2 — Manual installation

Extract `rogue.zip` and place both files in one model folder, for example:

```text
Applio/
└── logs/
    └── Rogue/
        ├── rogue.pth
        └── rogue.index
```

Then refresh the model list in Applio and select the matching `.pth` and `.index`.

Official Applio model-installation documentation:

https://docs.applio.org/getting-started/installing-inference-models/

## Example Applio CLI inference

Current Applio versions expose inference through `core.py`.

Example starting point:

```bash
python core.py infer \
  --input_path input.wav \
  --output_path rogue_output.wav \
  --pth_path logs/Rogue/rogue.pth \
  --index_path logs/Rogue/rogue.index \
  --pitch 0 \
  --index_rate 0.75 \
  --volume_envelope 1 \
  --protect 0.5 \
  --f0_method rmvpe \
  --embedder_model contentvec
```

The inference values above are **starting values, not training parameters**. Adjust them for the source recording.

Useful notes:

- Keep the embedder as **ContentVec**, matching training.
- **RMVPE** is a sensible default for this model because it was also used during feature extraction.
- `pitch 0` preserves the source pitch by default.
- The retrieval index can improve target-timbre similarity, but an excessively high index influence may also introduce artifacts.
- Source audio quality, pitch range, delivery, noise, and pronunciation can materially affect the result.

Official Applio CLI documentation:

https://docs.applio.org/reference/cli/

## Reproducing the training setup

The supplied training sheet also records a headless helper command:

```bash
tools/rvc_train.sh mara
```

The recorded expected Applio outputs were:

```text
/mnt/sata/Applio/logs/mara/mara.pth
/mnt/sata/Applio/logs/mara/*.index
```

These are local paths from the original training environment and are not required for inference after the files have been packaged.

## Evaluation

No formal objective evaluation results, AB listening test, benchmark scores, or reference inference samples were supplied with the uploaded files.

The package therefore does **not** claim that epoch 200 is objectively superior to another saved checkpoint. Users should evaluate the model by listening on representative speech/singing inputs and tuning inference parameters for their use case.

## Limitations

- Trained from approximately **35 minutes** of a single target voice.
- Training material is game dialogue, so the model may inherit characteristics of that recording, performance style, mastering, or available pitch range.
- Voice conversion quality depends strongly on the input speaker and recording.
- Pronunciation, emotion, singing range, very high/low pitch, background noise, and aggressive processing can reduce similarity or create artifacts.
- No multilingual or cross-language evaluation was provided.
- No formal quality benchmark was provided.
- This model performs voice conversion; it does not generate speech from text on its own.

## Responsible use

Use the model transparently and responsibly.

- Do not present generated audio as an authentic recording of the original performer or as an official game asset.
- Do not use generated audio to mislead, defraud, harass, or impersonate someone deceptively.
- Clearly label synthetic or voice-converted audio when context could otherwise make its origin unclear.
- Respect applicable rights, platform rules, and local law.

## Rights and redistribution

The training material was extracted from game audio. This model card does **not** grant rights to the original game recordings, characters, performances, trademarks, or other underlying assets.

No repository license is declared in the YAML metadata above because the supplied files did not include a license specifying what redistribution or reuse rights can be granted.

If the raw dataset is going to be made public, verify that you have the right to redistribute the source audio. If you do not, a safer public repository layout is to publish the trained model package and documentation without publishing the extracted source WAV files.

This is an unofficial community model and is not presented as affiliated with or endorsed by the game's developers, publishers, performers, or other rights holders.

## Integrity information

Additional uploaded-source checksums:

```text
rogue_datasets.zip
e5531e041a4d3e33baca6879509c58e03d6b7524239c918d800126eeba321fff

SETTINGS.txt
5e6ba830f607c276c3cd1e1acbf0184140e0916c6f8492959df26c282db60fae
```

## Recommended Hugging Face repository layout

For a compact repository that supports Applio's ZIP downloader:

```text
README.md
rogue.zip
SETTINGS.txt
```

If you also want people to access the two RVC files directly, you can additionally upload:

```text
rogue.pth
rogue.index
```

Be aware that keeping both the ZIP and the extracted model files duplicates the model storage.

## Frameworks and references

- Applio: https://github.com/IAHispano/Applio
- Applio documentation: https://docs.applio.org/
- Hugging Face model cards: https://huggingface.co/docs/hub/model-cards

---

### Quick technical summary

```text
RVC version:       v2
Sample rate:       48000 Hz
Pitch guidance:    yes
Pitch extractor:   RMVPE
Embedder:          ContentVec
Vocoder:           HiFi-GAN
Dataset:           626 mono 16-bit WAV files
Dataset duration:  35:00.422
Batch size:        8
Configured epochs: 300
Published epoch:   200
Published step:    26200
Weights:           rogue.pth
Index:             rogue.index
```