starrydark/Kurisu_Voice

Makise Kurisu Multilingual Voice Dataset

13,999 labelled clips (17.1423 hours) of Makise Kurisu, including the Amadeus Kurisu variant, in six languages, cut from the STEINS;GATE games, the anime, and three character songs.

Every clip carries the transcript, a measured acoustic profile, an independent speaker-identity check, and — where it could be earned rather than guessed — an expressive tag. This is an unofficial, fan-made dataset with no affiliation to the STEINS;GATE rights holders.

Read this first: six languages means six performers

speaker_id is kurisu for every row because the character is the same. The voice is not. Each dub was recorded by a different actor, so a model trained on all six rows at once will learn a blend of six people.

Language	Performer	Clips	Hours
Japanese (ja)	Asami Imai	10,039	11.914
English (en)	Trina Nishimura	1,485	1.95
Spanish (es)	Spanish dub cast	775	0.971
German (de)	German dub cast	735	0.968
French (fr)	French dub cast	591	0.862
Chinese (zh)	Chinese dub cast	374	0.479

Filter on voice_id (or equivalently language) for single-speaker training. Japanese covers games, anime and songs and is by far the largest single voice.

Summary

Property	Value
Clips	13,999
Duration	17.1423 h
Languages	6
Audio	WAV PCM 16-bit, mono, 24 kHz
Character variants	kurisu 12,521 · amadeus 1,478
Clips carrying an expressive tag	1,309 (9.4%)
Cross-lingual parallel lines	5,124 clips in 1,497 groups
Human-reviewed clips	1,475
Median / p05 / p95 duration	3.552 s / 0.81 s / 11.001 s

Splits

Split	Clips	Hours
train	12,832	15.8493
validation	607	0.6627
test	560	0.6303

Splits are cut between performances, not between files. Whole episodes and whole game scenario files move together across all six languages, every cross-lingual parallel line stays on one side, and any line whose lexical content repeats within a language is bound to a single split — which matters here because RE:BOOT re-records thousands of STEINS;GATE script lines.

Verified on the published files: 0 substantive lines shared between train and validation, 0 between train and test, 0 parallel groups straddling a boundary, across 344 leakage components.

Sources

Source	Clips
SGRE	2,868
SG	2,837
PHENO	1,746
Steins;Gate	1,661
SG0	908
Steins;Gate 0	765
Steins;Gate German Dub	735
Steins;Gate Spanish Dub	727
Steins;Gate French Dub	555
Load Region of Deja Vu	531
Steins;Gate 0 Chinese Dub	374
Steins;Gate OVA	86
Steins;Gate OVA Spanish Dub	48
Divide By Zero	46
Steins;Gate OVA French Dub	36
Cognitive Computing	24
Labyrinth of Idea	18
A Promising Paradigm	17
World-Line	17

Expressive tags

Two kinds of tag reach transcription. Event tags mark a non-verbal vocalisation and sit where it happens. Condition tags describe the whole utterance and sit at the front, in a fixed order.

Tag	Kind	Clips	From human review	Automatic
[pause]	structural	814	0	814
[digital]	condition	188	132	56
[sigh]	event	170	22	149
[gasp]	event	75	41	34
[singing]	condition	53	53	0
[breath]	event	31	31	0

A tag needs at least 30 examples to ship at all; anything rarer was demoted to metadata rather than published as a control nobody can steer.

How a tag is earned

The manual review of the anime corpus had already written bracket annotations into 1,478 transcripts. That is a character-specific gold set, so this build spends it on calibration instead of treating those annotations as the only labelled data.

Non-verbal vocalisations are named by laion/vocalburst-classifier-single with fine-class probabilities aggregated by family before ranking, checked by mispeech/ced-base, and — for anything inside a spoken line — required to fall where laion/vocalburst-locator sees a burst and Qwen3-ForcedAligner-0.6B says no transcript unit reaches. Thresholds come from a sweep against the gold positives and 2275 ordinary speech clips used as a negative control:

Family	Scope	Mode	score ≥	margin ≥	2nd model must agree	precision	recall	false positives
gasp	whole_clip	specificity	0.2	0.0	yes	—	—	0/2275
sigh	whole_clip	specificity	0.25	0.0	yes	—	—	11/2275
sob	whole_clip	specificity	0.2	0.0	yes	—	—	0/2275
gasp	in_speech	specificity	0.6	0.0	yes	—	—	1/180
giggle	in_speech	specificity	0.55	0.0	yes	—	—	1/180
scream	in_speech	specificity	0.2	0.0	yes	—	—	0/180
sigh	in_speech	specificity	0.75	0.0	yes	—	—	1/180

Families with gold examples that did not reach the floor and are therefore never auto-assigned: breath, moan.

Delivery and playback conditions have no off-the-shelf model, so a small logistic regression per language is fitted on the deterministic acoustic features and scored leave-one-out, on the reviewed subset only — the one place where the absence of a tag means a human looked and decided no.

Tag	Fitted on	Gold positives	Reviewed clips	LOO AUC	precision	recall	Outcome
[digital]	German	22	162	0.952	0.875	0.955	usable
[digital]	English	54	391	0.819	0.88	0.407	usable
[digital]	Spanish	17	220	0.994	0.895	1.0	usable
[digital]	French	10	103	0.763	—	—	floor not reached
[digital]	Japanese	29	428	0.886	—	—	floor not reached
[digital]	pooled: dubs	103	876	0.865	0.852	0.505	usable
[whisper]	Japanese	17	428	0.753	—	—	floor not reached
[whisper]	pooled: all	25	1426	0.779	—	—	floor not reached
[yell]	Japanese	13	428	0.872	—	—	floor not reached
[yell]	pooled: all	18	1426	0.836	—	—	floor not reached
[echo]	pooled: all	29	1426	0.793	—	—	floor not reached
noise	German	19	162	0.552	—	—	floor not reached
noise	English	29	391	0.607	—	—	floor not reached
noise	Spanish	8	220	0.233	—	—	floor not reached
noise	Japanese	44	428	0.632	—	—	floor not reached
noise	Chinese	20	122	0.559	—	—	floor not reached
noise	pooled: all	123	1426	0.556	—	—	floor not reached

Each language then uses its own fit when that clears the floor, and the pooled fit it belongs to otherwise:

Tag	Language	Model used	precision	recall	In transcript
[digital]	German	language:de	0.875	0.955	yes
[digital]	English	pooled:dubs	0.852	0.505	yes
[digital]	Spanish	language:es	0.895	1.0	yes
[digital]	French	pooled:dubs	0.852	0.505	yes
[digital]	Chinese	pooled:dubs	0.852	0.505	yes

Where a fit could not clear the precision floor, the human labels still ship and nothing is added. [digital] is the clearest case: the European dubs render Amadeus through an audible low-pass that a detector finds easily, while the Japanese original uses a thin, bass-cut EQ that the reviewers heard as narrative context — so Japanese keeps its 29 human labels and receives no automatic ones.

Whispering, shouting and room reverberation are physiological or physical rather than linguistic, so those detectors pool their gold clips across all six languages — the only way [yell] (18 gold clips) and [echo] (29) have enough positives to be fitted at all. [digital] pools only across the four dubs whose signature points the same way.

[pause] is purely measured: an internal silence the aligner places between two words, long enough to be in the upper decile for its own language (Japanese dialogue breathes differently from Spanish, and a fixed threshold would tag every sentence boundary in one and nothing in the other), at most two per clip, and only where the transcript does not already write an ellipsis — …… says it perfectly well on its own. Thresholds used: German 1.24s, English 1.2s, Spanish 1.2s, French 1.44s, Japanese 1.28s, Chinese 1.2s.

[nonverbal] is the honest coarse label. Naming a sound is harder than noticing there is one, so a separate gate is calibrated for the weaker claim "this is a non-lexical vocalisation, family unknown" — the classifier must be confident a burst is present and CED must not think it is ordinary speech. It is calibrated for specificity: 0 of 2275 ordinary-speech clips would pass it. It exists so that a take whose script line is only …… can enter the corpus with a label that says exactly what is known, instead of being thrown away or shipped with a guessed family.

Labels that deliberately do not reach the transcript

The reviewers also marked non-character sound and performance notes. Those are useful, but a TTS model should not learn to produce a doorbell, so they were moved out of transcription and into columns:

• contamination — noise, footstep, knock, ringtone: audible non-character sound in the clip. 13,999 clips carry a contamination_score from the same calibrated framework, so the whole corpus can be filtered, not just the reviewed part.
• emotion_notes — serious, playful, angry, embarrassed, scared, mumbled, foreign_accent: too rare to be a control, kept for curation.

Dataset structure

Hugging Face AudioFolder layout:

data/
├── train/       metadata.jsonl + *.wav
├── validation/  metadata.jsonl + *.wav
└── test/        metadata.jsonl + *.wav

Columns

Field	Meaning
file_name	audio file, relative to the split directory
transcription	the training label — text plus any earned tag
text	the same line with every tag removed
language / voice_id	ja en zh de fr es; also the performer identity
voice_actor	who is speaking
speaker_id / speaker_name	kurisu / Makise Kurisu
character_variant	kurisu or amadeus
medium / series / scene	game · anime · song; title; scenario file or episode
style	speech or singing
tags / event_tags / condition_tags	tags present in transcription
human_tags / auto_tags	which of them a person wrote and which a detector earned
human_reviewed	a person adjudicated this clip upstream
from_quarantine	recovered from the game's non-lexical quarantine pool
parallel_group	id shared by the same line across dubs, or null
contamination / contamination_score	non-character sound, human and automatic
emotion_notes / measured_pauses	performance notes; measured internal pauses
quality_tier / quality_flags	A/B/C and why
duration_seconds sample_rate active_rms_dbfs peak_dbfs lufs	level and length
speech_ratio lead_silence_s tail_silence_s internal_pause_max_s	activity
snr_est_db noise_floor_dbfs	what is left between the words
f0_median_hz f0_range_st voiced_ratio hnr_db	pitch register, movement, periodicity
bandwidth_roll995_hz spectral_tilt_db reverb_decay_db_per_s	playback path and room
speaking_rate_units_per_s lexical_units	rate over speaking time; chars for ja/zh, words otherwise
speaker_similarity speaker_rival_similarity speaker_margin speaker_rival	independent identity check
alignment_coverage	share of the clip the forced aligner attributes to the transcript
transcript_source	where the words came from
sha256	digest of the published WAV

Quality tiers

Tier	Clips	Meaning
A	8,197	no quality flag, and a trusted transcript source
B	5,363	usable, with at least one soft flag
C	439	contamination, low SNR, clipping, or a contradicted speaker check

Flags observed: speech_at_edge (2,969), low_alignment_coverage (836), speaker_outlier (283), nonspeech_contamination (131), very_short (111), very_long (55), low_snr (37).

Reference clips

reference/<voice_id>.wav is one clean, untagged, 4–10 s take per voice, chosen by measurement rather than by ear: the highest margin over the closest other STEINS;GATE character among tier-A clips of that voice. reference/reference.json records which clip each one is and why.

voice_id	Language	File	Duration (s)	Margin over rival	Closest rival
de	German	reference/de.wav	4.4	0.4434	urushibara_ruka
en	English	reference/en.wav	7.33	0.6505	shiina_kagari
es	Spanish	reference/es.wav	5.33	0.426	amane_suzuha
fr	French	reference/fr.wav	9.26	0.4532	amane_suzuha
ja	Japanese	reference/ja.wav	6.53	0.8132	reyes
zh	Chinese	reference/zh.wav	5.659	0.2586	urushibara_ruka

This matters for voice-cloning fine-tunes that read a reference on every training row and write its embedding into a speaker table — a reference that is not actually the target voice poisons the run silently.

Loading

from datasets import load_dataset

ds = load_dataset("starrydark/Kurisu_Voice")

# single-speaker Japanese, clean tier, with the expressive tags
ja = ds["train"].filter(lambda r: r["language"] == "ja" and r["quality_tier"] != "C")

# the same line in every dub
groups = ds["train"].filter(lambda r: r["parallel_group"] is not None)

# no tags at all
plain = ds["train"].map(lambda r: {"label": r["text"]})

Use transcription to train tag control, text to ignore it.

Curation

Audio. One standardised 24 kHz mono PCM-16 WAV per clip. Game audio is resampled from 48 kHz with soxr VHQ. Nothing is denoised, gated or trimmed.

Dynamics were put back. The upstream anime and song builds normalised every clip independently to a fixed active-speech RMS, which flattens the difference between a whisper and a shout. Each clip recorded the gain it was given, so this build inverts it exactly and then applies one gain per source group to a common -26.0 LUFS target, measured to ITU-R BS.1770-4 by pooling the K-weighted blocks of the whole group. Loudness relationships inside a group survive untouched. The result is an active-level spread of 10.41 dB p05–p95 across the corpus (7.98 dB within the anime portion, which was flat before).

Source group	Measured LUFS	Gain applied
game:PHENO	-18.931	-7.069
game:SG	-14.142	-11.858
game:SG0	-19.765	-6.235
game:SGRE	-20.085	-5.915
anime:ja	-26.965	0.965
anime:en	-26.875	0.875
anime:zh	-27.013	1.013
anime:de	-27.942	1.942
anime:fr	-27.365	1.142
anime:es	-27.15	1.15
song:ja:labyrinth_of_idea	-19.104	-6.896
song:ja:promising_paradigm	-16.119	-9.881
song:ja:world_line	-15.167	-10.833

Speaker identity was re-checked on the published audio with speechbrain/spkrec-ecapa-voxceleb, which took no part in either upstream decision. Embeddings are centred before scoring — raw cosines between any two young female anime voices sit above 0.9 and separate nothing — and every clip is scored against the closest other STEINS;GATE character recorded in the same language and medium, drawn from 7,882 negative-control clips. speaker_margin is that difference and it is published per clip.

The game quarantine pool was relabelled, not discarded. 272 takes whose scenario text is only …… were set aside upstream for having no usable transcript. The line is the sound. Where the calibrated gate could name it, the transcript became that tag: 162 of them are now in the corpus — 130 with a named family and 16 as [nonverbal]. The rest are held back rather than shipped with a guessed control token.

Held back: 113 clips with no nameable content are not published.

Limitations

• Six languages are six performers. Train on one voice_id at a time unless a blended voice is what you want.
• The speech is acted character dialogue and does not represent ordinary conversation in any of these languages.
• Anime clips are source-separated and can retain music, effects, or separation artefacts. snr_est_db, contamination and quality_tier are there to filter on; game clips are studio-clean by construction.
• The source mix is very uneven — Japanese is 11.914 h of 17.1423 h, and Chinese covers only the STEINS;GATE 0 dub.
• Automatic tags are precision-optimised, so recall is low by design: an untagged clip is not evidence that nothing happened in it.
• Tag strings are a convention of this dataset, not an official Qwen3-TTS, XTTS or VITS feature. Evaluate controllability separately from audio quality.
• No exhaustive human listening pass has been done on the automatic labels.

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, which is why it uses Hugging Face's other designation rather than claiming an open-data licence.

Do not use this dataset to impersonate any of the performers, to mislead anyone about authentic or official speech, to suggest endorsement, to harass, or to evade copyright, publicity, personality, or voice-protection rights. Disclose synthetic audio produced from models trained on it. Users and redistributors are responsible for determining whether their access and intended use are lawful in their jurisdiction. Rights holders may contact the repository owner here for correction or removal.

Attribution

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