README.md

---
pretty_name: Makise Kurisu Voice Dataset
language:
- ja
license: other
task_categories:
- text-to-speech
- automatic-speech-recognition
tags:
- audio
- speech
- japanese
- steins-gate
- anime
- games
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_dir: data
  drop_labels: true
---

# Makise Kurisu Voice Dataset

A single-character Japanese speech dataset containing 10,469 labeled clips
(11.055 hours) associated with Makise Kurisu, including the Amadeus Kurisu
variant. The audio samples were extracted and curated from games and anime in
the **STEINS;GATE series**.

This is an unofficial, fan-created dataset. It is not affiliated with or
endorsed by the STEINS;GATE creators, publishers, animation studios, cast, or
other rights holders.

## Dataset summary

| Property | Value |
|---|---:|
| Clips | 10,469 |
| Duration | 11.055 hours |
| Language | Japanese |
| Speaker label | Makise Kurisu |
| Character variants | Kurisu, Amadeus Kurisu |
| Audio | WAV, PCM 16-bit, mono, 24 kHz |
| Train split | 9,299 clips |
| Validation split | 652 clips |
| Test split | 518 clips |

### Source media

The samples come from the following STEINS;GATE series source groups:

| Source group | Clips |
|---|---:|
| Game — STEINS;GATE (`game:SG`) | 2,831 |
| Game — STEINS;GATE 0 (`game:SG0`) | 903 |
| Game — Linear Bounded Phenogram (`game:PHENO`) | 1,737 |
| Game — STEINS;GATE RE:BOOT (`game:SGRE`) | 2,860 |
| Anime — STEINS;GATE (`anime:SG`) | 1,511 |
| Anime — STEINS;GATE 0 (`anime:SG0`) | 597 |
| Anime — Divide By Zero (`anime:SG_DivideByZero`) | 30 |

## Dataset structure

The repository follows the Hugging Face `AudioFolder` layout. Each split keeps
its audio and `metadata.jsonl` together:

```text
data/
├── train/
│   ├── metadata.jsonl
│   └── *.wav
├── validation/
│   ├── metadata.jsonl
│   └── *.wav
└── test/
    ├── metadata.jsonl
    └── *.wav
```

Each metadata row contains:

| Field | Description |
|---|---|
| `file_name` | Audio filename, relative to the split directory |
| `transcription` | Training label; Japanese text with any retained `[gasp]` or `[sigh]` tag |
| `text` | Japanese transcript without the added expressive tag |
| `speaker_id` | Normalized speaker ID (`kurisu`) |
| `speaker_name` | Display label (`Makise Kurisu`) |
| `character_variant` | `kurisu` or `amadeus` |
| `source_medium` | `game` or `anime` |
| `source_series` | Short source identifier |
| `source_group` | Combined provenance category |
| `source_quality_tier` | Upstream curation tier (`A`, `B`, or `C`) |
| `duration_seconds` | Clip duration in seconds |
| `expressive_tags` | Zero or more retained expressive labels |
| `from_quarantine` | Whether the source clip came from the upstream short/non-lexical review pool |
| `sha256` | SHA-256 digest of the published WAV file |

There are 38 `[gasp]` labels and 157 `[sigh]` labels. These bracketed strings
are dataset conventions, not standardized speech tokens.

## Loading

```python
from datasets import load_dataset

dataset = load_dataset("starrydark/Kurisu_Voice")
print(dataset["train"][0])
```

For text-to-speech training, use the decoded `audio` feature with
`transcription`. Use `text` instead if expressive tags are not desired.

## Curation and quality

- Game samples were aligned to their scenario text; anime samples were aligned
  to Japanese subtitle text and speaker-attributed during upstream curation.
- Audio was standardized to 24 kHz mono PCM-16 WAV.
- Source groups were loudness-aligned using one gain value per source group;
  no per-clip loudness normalization was used.
- The released split contains only accepted records. A separate 137-item
  uncertain review pool, build scripts, model outputs, checkpoints, caches,
  credentials, and local paths are not included.
- The local pre-publication verifier reported zero failures and zero warnings.

## Intended uses

This dataset may be useful for speech research, Japanese ASR/TTS experiments,
alignment research, and non-commercial study, subject to applicable law and
permission from the relevant rights holders.

Do not use it to impersonate the performer, mislead people about authentic or
official speech, suggest endorsement, harass anyone, or evade copyright,
publicity, personality, or voice-protection rights. Clearly disclose synthetic
audio produced from models trained on this dataset.

## Limitations

- The speech is acted character dialogue and is not representative of general
  conversational Japanese.
- Anime clips can retain residual music, effects, ambience, or separation
  artifacts; game clips are generally cleaner.
- Some speaker attribution and expressive labels were inferred
  algorithmically and can be incorrect.
- The source distribution is uneven, and train/validation/test splits are not
  guaranteed to prevent all scene-level or near-duplicate leakage.
- The dataset has not received exhaustive human listening review.

## Rights and licensing

The underlying audio, dialogue, characters, names, and STEINS;GATE-related
marks remain the property of their respective rights holders. No copyright,
trademark, performer, publicity, personality, or voice rights are transferred
or licensed by this repository. The repository therefore uses Hugging Face's
`other` license designation rather than claiming an open-data license.

Users and redistributors are responsible for determining whether their access
and intended use are lawful and authorized in their jurisdiction. Rights
holders may contact the repository owner through the Hugging Face repository
for correction or removal requests.

## Attribution

If you use the dataset in research, cite this dataset page and identify the
source material as audio from STEINS;GATE series games and anime. This request
does not replace any attribution or permission required by the rights holders.