ChupaVoice: Japanese Chupa-sound Dataset (Voice-Acted Anime Oral-Slurp SFX)

A focused subset of the OOPPEENN/Galgame_Dataset containing チュパ音 (chupa-sounds): the exaggerated oral-slurp SFX commonly heard in Japanese R-18 visual novels (eroge).
Each clip comes with its original transcription and game and speaker identifier.

Ethical Considerations & Age Warning

This dataset contains explicit erotic content (oral sex SFX). Do NOT use if you are a minor or if local regulations prohibit adult material.

Dataset Description

Dataset Summary

Domain: voice-acted Japanese chupa-sounds (oral blowjob SFX)
Source: snapshot of OOPPEENN/Galgame_Dataset taken on 2025-04-29
Size:
- 133,003 audio files
- 389 hours total duration
Language: Japanese
File Format: Not changed from OOPPEENN/Galgame_Dataset, mainly .ogg with some .opus and .wav
Audio Quality: Not changed from OOPPEENN/Galgame_Dataset (good enough since they're the same audio quality as the original Visual Novel)
Metadata: Game identifier, speaker identifier, and transcription for every clip
License: GPL-3.0

Why Chupa?

Researching chupa-sounds poses unique challenges:

These are extreme non-verbal vocalisations, hard to model compared with ordinary speech.
Transcription style varies widely:
- A clip written as ch u p a ch u p a can sound nothing like the usual pronunciation of those phonemes.
- Scenario writers adopt different onomatopoeic conventions.
- Voice actresses interpret the same transcription very differently.

As a result, tasks such as TTS or ASR on chupa-sounds remain under-explored.
We hope this dataset accelerates work on:

TTS: Synthesize Voices containing chupa sounds from text (or something else)
ASR: Chupa-to-Text(?) recognition (or robust onomatopoeia handling)
Speaker Recognition: I don't know if this is possible (but some erogamers can!)
Analysis: phonetic / stylistic study of chupa-sounds and their transcriptions

Dataset Structure

The archive chupa-voice.zip contains a single metadata.jsonl.
Example entries:

{"file_name":"0204f6ddf8/27cc689d64/07737edfadd86e7e.ogg","text":"ふぶうぅ、ぶっ、ぷ、むうううぅぅ～～……っ！","duration":7.450896}
{"file_name":"0204f6ddf8/27cc689d64/0cb658482af1bb1a.ogg","text":"鈴々で、ひゃんろっ、出ぅ？せーえき……出ひてぇっ、んふうぅ、ぶ、ふっ、む、ン","duration":13.060958}
{"file_name":"0204f6ddf8/27cc689d64/129ba948b95b30f9.ogg","text":"ん～～～嫌じゃないのだぁ、んぷぅ、ぺろっ、ぺろ……見て、いっぱい舐めひゃう","duration":11.441729}

Directory Layout

If the original path in Galgame_Dataset is {game_name}/{speaker}/{file_name}, it is mapped here to {sha1(game_name)[:10]}/{sha1(speaker)[:10]}/{sha1(file_name)[:16]}.{ext}

The SHA-1 helper used:

import hashlib

def sha1_hex(text: str) -> str:
    return hashlib.sha1(text.encode("utf-8")).hexdigest()

Limitations & Biases

Some non-Blowjob sounds (kissing, licking) are included.
A small number of clips feature male performers.
Certain files include ordinary speech with only short chupa segments
- e.g. いつまでだんまりを決め込んでるつもりなの？むちゅ……っ
Transcriptions are copied verbatim from the source dataset and may contain errors such as んっ、むぅ……▼ん、ぅ……ん～……▼.
- So it's recommended to normalize or filter transcriptions before using transcriptions