README.md

---
pretty_name: "Rogue RVC Voice Dataset"
task_categories:
  - audio-to-audio
size_categories:
  - n<1K
tags:
  - audio
  - speech
  - voice-conversion
  - rvc
  - rvc-v2
  - applio
  - contentvec
  - rmvpe
  - 48khz
  - mono
  - cyberpunk-2077
  - rogue
---

# Rogue RVC Voice Dataset

Audio dataset prepared for training a **Retrieval-Based Voice Conversion (RVC)** model of **Rogue** with **Applio**.

Repository:

**https://huggingface.co/datasets/0xra/Rogue-Datasets**

This is a **voice-conversion training dataset**, not a text-to-speech corpus. The supplied archive contains segmented WAV clips and training metadata; it does **not** contain transcriptions.

## Dataset summary

| Property | Value |
|---|---:|
| Audio clips | **626** |
| Total duration | **35:00.422** |
| Sample rate | **48,000 Hz** |
| Channels | **1 / mono** |
| PCM sample width | **16-bit** |
| Mean clip length | **3.355 s** |
| Median clip length | **2.900 s** |
| Shortest clip | **1.210 s** |
| Longest clip | **11.732 s** |
| Archive | `rogue_datasets.zip` |
| Archive size | **135,605,530 bytes / 129.32 MiB** |
| Uncompressed archive contents | **201,715,880 bytes / 192.37 MiB** |
| Internal dataset folder | `mara/` |
| Internal model name | `mara` |
| Target / donor voice | `rogue` |
| Dataset build timestamp | `2026-09-04 02:19` |
| Build seed | **27** |

## Archive contents

The supplied archive uses this structure:

```text
rogue_datasets.zip
└── mara/
    ├── SETTINGS.txt
    ├── manifest.json
    ├── rogue_q005_f_*.wav
    ├── rogue_q101_f_*.wav
    ├── rogue_q103_f_*.wav
    ├── rogue_q108_f_*.wav
    ├── rogue_q115_f_*.wav
    ├── rogue_q116_f_*.wav
    ├── rogue_q203_f_*.wav
    ├── rogue_q307_f_*.wav
    └── rogue_sq031_f_*.wav
```

`SETTINGS.txt` records the Applio/RVC training configuration used with the dataset.

`manifest.json` records the dataset build settings and every retained clip with its duration.

## Dataset construction

The source audio was extracted from game audio rather than recorded from a microphone specifically for this training run.

The dataset builder applied a duration filter:

| Filter | Value |
|---|---:|
| Minimum retained duration | **1.2 s** |
| Maximum retained duration | **12.0 s** |
| Clips dropped for being too short | **35** |
| Clips dropped for being too long | **2** |
| Final retained clips | **626** |

The shortest retained clip is approximately 1.210 seconds and the longest is approximately 11.732 seconds, which is consistent with the configured filter boundaries.

No additional transcript or text annotation files are present in the archive.

## Filename source groups

The filenames contain source-group identifiers. The dataset contains the following groups:

| Filename group | Clips |
|---|---:|
| `q115` | 249 |
| `sq031` | 178 |
| `q103` | 72 |
| `q108` | 60 |
| `q101` | 29 |
| `q203` | 23 |
| `q307` | 7 |
| `q005` | 5 |
| `q116` | 3 |
| **Total** | **626** |

These names are retained from the source-extraction workflow. This dataset card does not assign additional semantic labels to those identifiers.

## Audio format

All 626 retained WAV files were verified to use the same basic format:

```text
Container:     WAV
Sample rate:   48000 Hz
Channels:      1 (mono)
Sample width:  16-bit PCM
```

This uniform format makes the archive directly suitable as an Applio/RVC source dataset without per-file sample-rate or channel conversion.

## Intended use

The dataset was prepared primarily for:

- training an **RVC v2** voice-conversion model;
- training through **Applio**;
- reproducing or continuing the corresponding Rogue voice-model experiment;
- voice-conversion research and testing where use of the source material is permitted.

It was **not** prepared as:

- an automatic speech-recognition dataset;
- a transcribed speech corpus;
- a general-purpose text-to-speech dataset;
- a speaker-identification benchmark;
- an objective voice-quality benchmark.

## Original Applio / RVC configuration

The included `SETTINGS.txt` records the following training setup.

### Preprocess

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

### Why effects and noise reduction were disabled

The source was already produced game audio rather than untreated microphone recordings. The original training notes intentionally disable another effects/denoise pass to avoid removing breath, sibilance, and other voice characteristics that can be useful to a voice-conversion model.

### Feature extraction

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

### Training

| 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` |

The settings file records the original headless training command as:

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

## Using the dataset with Applio

### 1. Download the repository

With the current Hugging Face CLI:

```bash
hf download 0xra/Rogue-Datasets \
  --type dataset \
  --local-dir ./Rogue-Datasets
```

You can also download the files directly from the repository page.

### 2. Extract the dataset archive

```bash
unzip Rogue-Datasets/rogue_datasets.zip -d ./rogue_rvc_dataset
```

After extraction:

```text
./rogue_rvc_dataset/mara/
```

is the dataset directory containing the WAV files.

### 3. Select the folder in Applio

Use:

```text
./rogue_rvc_dataset/mara
```

as the dataset path in Applio's training interface.

The archive also contains `SETTINGS.txt`, which can be used to reproduce the original preprocessing, extraction, and training configuration.

## Recommended Hugging Face repository layout

For an **Applio-focused download repository**, keeping the prepared archive is simple:

```text
README.md
rogue_datasets.zip
```

The archive already contains:

```text
mara/SETTINGS.txt
mara/manifest.json
mara/*.wav
```

### Optional: enable a more native Hugging Face audio dataset layout

Hugging Face can directly recognize audio files stored at the repository root or inside a subdirectory.

If you want the repository to behave more like a conventional Hugging Face audio dataset and make Dataset Viewer integration easier, an alternative layout is:

```text
README.md
audio/
  rogue_q005_f_....wav
  rogue_q101_f_....wav
  ...
manifest.json
SETTINGS.txt
```

or:

```text
README.md
train/
  rogue_q005_f_....wav
  rogue_q101_f_....wav
  ...
manifest.json
SETTINGS.txt
```

For the current RVC use case, however, `rogue_datasets.zip` is convenient because it preserves the exact dataset folder expected by the original training workflow.

## Python inspection example

The archive can be inspected without extracting every file first:

```python
from pathlib import Path
import io
import wave
import zipfile

archive = Path("rogue_datasets.zip")

with zipfile.ZipFile(archive) as zf:
    wav_files = [
        info for info in zf.infolist()
        if info.filename.lower().endswith(".wav")
    ]

    total_seconds = 0.0

    for info in wav_files:
        with zf.open(info) as src:
            data = src.read()

        with wave.open(io.BytesIO(data), "rb") as wav:
            total_seconds += wav.getnframes() / wav.getframerate()

print("clips:", len(wav_files))
print("minutes:", total_seconds / 60)
```

Expected values for this archive:

```text
clips:   626
minutes: 35.007038...
```

## Manifest

`mara/manifest.json` contains machine-readable build metadata and a list of every retained audio clip.

The top-level metadata includes:

```json
{
  "model": "mara",
  "speaker": "rogue",
  "built": "2026-09-04 02:19",
  "seed": 27,
  "sample_rate": 48000,
  "channels": 1,
  "files": 626,
  "minutes": 35.01,
  "filter": {
    "min_sec": 1.2,
    "max_sec": 12.0,
    "dropped_short": 35,
    "dropped_long": 2
  }
}
```

The complete file also contains the filename and duration of each of the 626 retained clips.

## Integrity

SHA-256 of the supplied archive:

```text
rogue_datasets.zip
e5531e041a4d3e33baca6879509c58e03d6b7524239c918d800126eeba321fff
```

Archive size:

```text
135605530 bytes
```

This checksum can be used to verify that a local copy is identical to the dataset used to produce these statistics.

On Linux:

```bash
sha256sum rogue_datasets.zip
```

Expected result:

```text
e5531e041a4d3e33baca6879509c58e03d6b7524239c918d800126eeba321fff  rogue_datasets.zip
```

## Dataset limitations

- The dataset contains approximately **35 minutes** from a single target voice.
- It is derived from produced game dialogue, so its acoustic characteristics reflect the original recording, performance, mastering, and extraction pipeline.
- The dataset does not provide text transcriptions.
- The source material may not cover every phoneme, speaking style, emotion, intensity, or pitch range equally.
- The filename groups should not be treated as validated semantic labels.
- No formal train/validation/test split is provided.
- No objective dataset-quality or speech-recognition benchmark is provided.
- The dataset is intended for voice-conversion training rather than general speech-model benchmarking.

## Rights and responsible use

The dataset consists of audio derived from game content. This dataset card does **not** grant rights to the underlying game recordings, character, performance, trademarks, or other source material.

No license is declared in the Hugging Face YAML metadata because the supplied dataset files do not include a license establishing redistribution or downstream-use rights.

Users are responsible for determining whether their intended download, redistribution, training, generation, or publication use is permitted by the applicable rights, platform rules, and local law.

Voice-converted outputs should not be represented as authentic recordings of the original performer or as official game content when doing so could mislead people.

## Dataset card metadata notes

This card intentionally uses:

```yaml
task_categories:
  - audio-to-audio
size_categories:
  - n<1K
tags:
  - audio
  - rvc
  - applio
```

The `audio` tag makes the dataset modality explicit, while `audio-to-audio` reflects its intended voice-conversion use.

The card intentionally does **not** declare a `language` or `license` field because neither was established by the supplied dataset metadata.

## Hugging Face documentation

Useful references:

- Dataset cards: https://huggingface.co/docs/hub/datasets-cards
- Audio datasets: https://huggingface.co/docs/hub/datasets-audio
- Dataset repository structure: https://huggingface.co/docs/datasets/repository_structure
- Adding datasets to the Hub: https://huggingface.co/docs/hub/datasets-adding

---

## Quick technical summary

```text
Repository:          0xra/Rogue-Datasets
Purpose:             RVC / Applio voice-conversion training
Target voice:        rogue
Internal model name: mara
Clips:               626
Duration:            35:00.422
Audio:               WAV, mono, 16-bit PCM
Sample rate:         48000 Hz
Clip filter:         1.2 s - 12.0 s
Dropped short:       35
Dropped long:        2
Build seed:          27
Pitch extractor:     RMVPE
Embedder:            ContentVec
Vocoder:             HiFi-GAN
Configured epochs:   300
Archive:             rogue_datasets.zip
SHA-256:             e5531e041a4d3e33baca6879509c58e03d6b7524239c918d800126eeba321fff
```