starrydark/BanG_Dream_Qwen3_TTS

BanG Dream! character voices: Qwen3-TTS text-to-speech in Japanese, Mandarin and English

An unofficial, fan-made text-to-speech model with 49 BanG Dream! character voices across eight bands: Poppin'Party, Afterglow, Pastel*Palettes, Roselia, Hello, Happy World!, Morfonica, RAISE A SUILEN, MyGO!!!!! and Ave Mujica (persona variants such as 大魔姫あこ, 月ノ森 Soyo / Sakiko and Mortis have their own rows).

• Each voice is called by its speaker id. No reference audio is needed.
• Three languages: Japanese (the characters' own language) plus Mandarin and English. Language ids, native replay and in-voice pseudo-Mandarin keep the Japanese accent out of Mandarin.
• Built on Qwen3-TTS-12Hz-1.7B-Base, full fine-tune.

Code: github.com/starryark/bangdream-tts: data pipeline, fixed trainer, per-voice RVC route for native Mandarin, and every pre-registered plan and result.

Also known as: BanG Dream! Girls Band Party · バンドリ！ ガールズバンドパーティ！ · ガルパ · 邦邦 · 少女乐团派对.

日本語：『BanG Dream!』49 キャラクターボイスの非公式ファンメイド TTS（Qwen3-TTS 微調整）。話者 ID で呼び出し、参照音声は不要。日本語・中国語・英語に対応。公式・キャスト・運営とは無関係です。

中文：《BanG Dream!》49 位角色的非官方粉丝向语音合成（TTS）模型，基于 Qwen3-TTS 微调。按说话人 ID 直接合成，无需参考音频；支持日语、普通话和英语。与官方、运营及声优无关。

Voices	49 speaker rows (45 characters; 4 persona variants), table below
Languages	Japanese, Mandarin Chinese, English (language="Japanese" / "Chinese" / "English")
Type	Qwen3-TTS custom_voice checkpoint: text in, speech out, no reference audio
Base model	Qwen/Qwen3-TTS-12Hz-1.7B-Base, full fine-tune
Training data	9,268 Japanese voice lines of the 49 voices (from 22.1 h kept), plus 1,509 native Mandarin/English replay clips and 564 gated in-voice pseudo-Mandarin rows: 11,341 training rows, 23.4 h
Version	E1b epoch 2 (trained 2026-10-30, released 2026-11-06)
License	Unofficial fan model, no rights granted (details)
Code	starryark/bangdream-tts

Use (Python)

import torch, soundfile as sf
from qwen_tts import Qwen3TTSModel   # pip install qwen-tts

tts = Qwen3TTSModel.from_pretrained("starrydark/BanG_Dream_Qwen3_TTS",
                                    device_map="cuda", dtype=torch.bfloat16)
wavs, sr = tts.generate_custom_voice(
    text="今日の練習、思ったより早く終わったね。",
    speaker="ppp_kasumi",            # see the table below
    language="Japanese",             # or "Chinese" / "English"
)
sf.write("kasumi.wav", wavs[0], sr)

Usage notes:

• Always pass language. Training used a language id on every row; it is what keeps Mandarin and English free of the Japanese accent.
• Keep the default sampling (temperature 0.9, top-k 50). Greedy decoding makes codec language models loop.
• Try a few seeds for Mandarin and English on the voices marked "generic" below.
• Generated audio is synthetic: label it as AI-generated wherever you share it.

Voices

speaker ids are <band>_<name>. "Mandarin identity" is the share of 32 Mandarin test renders that an independent speaker probe attributes to the right voice among all 49 ("Japanese identity" is the same over the Japanese test renders. generic means the voice speaks Mandarin with a native accent but in a timbre closer to the base model's default voice than to the character; partly generic means that happens on some seeds).

speaker	Character	Band	Japanese data	Japanese identity	Mandarin identity	Mandarin voice
ppp_kasumi	戸山香澄	Poppin'Party	35 min	0.84	0.25	generic
ppp_tae	花園たえ	Poppin'Party	31 min	1.00	0.94	own voice
ppp_rimi	牛込りみ	Poppin'Party	24 min	1.00	0.25	generic
ppp_saaya	山吹沙綾	Poppin'Party	47 min	1.00	1.00	own voice
ppp_arisa	市ヶ谷有咲	Poppin'Party	61 min	0.98	0.94	own voice
ag_ran	美竹蘭	Afterglow	26 min	1.00	0.91	own voice
ag_moca	青葉モカ	Afterglow	59 min	0.98	1.00	own voice
ag_himari	上原ひまり	Afterglow	45 min	1.00	0.88	own voice
ag_tomoe	宇田川巴	Afterglow	31 min	1.00	0.97	own voice
ag_tsugumi	羽沢つぐみ	Afterglow	30 min	1.00	0.88	own voice
pp_aya	丸山彩	Pastel*Palettes	52 min	0.98	0.84	partly generic
pp_hina	氷川日菜	Pastel*Palettes	34 min	1.00	0.97	own voice
pp_chisato	白鷺千聖	Pastel*Palettes	26 min	1.00	1.00	own voice
pp_maya	大和麻弥	Pastel*Palettes	45 min	1.00	0.94	own voice
pp_eve	若宮イヴ	Pastel*Palettes	46 min	1.00	0.62	partly generic
ro_yukina	湊友希那	Roselia	19 min	1.00	1.00	own voice
ro_sayo	氷川紗夜	Roselia	31 min	1.00	1.00	own voice
ro_lisa	今井リサ	Roselia	39 min	1.00	1.00	own voice
ro_ako	宇田川あこ	Roselia	8 min	1.00	0.88	own voice
ro_ako_maou	宇田川あこ	Roselia	5 min	0.88	0.31	generic
ro_rinko	白金燐子	Roselia	42 min	1.00	0.53	generic
hhw_kokoro	弦巻こころ	Hello, Happy World!	29 min	1.00	0.94	own voice
hhw_kaoru	瀬田薫	Hello, Happy World!	30 min	1.00	0.84	partly generic
hhw_hagumi	北沢はぐみ	Hello, Happy World!	23 min	1.00	0.81	partly generic
hhw_kanon	松原花音	Hello, Happy World!	36 min	1.00	0.66	partly generic
hhw_misaki	奥沢美咲	Hello, Happy World!	37 min	1.00	0.97	own voice
mo_mashiro	倉田ましろ	Morfonica	23 min	1.00	0.97	own voice
mo_touko	桐ヶ谷透子	Morfonica	22 min	1.00	0.91	own voice
mo_nanami	広町七深	Morfonica	34 min	1.00	0.94	own voice
mo_tsukushi	二葉つくし	Morfonica	25 min	1.00	1.00	own voice
mo_rui	八潮瑠唯	Morfonica	26 min	1.00	1.00	own voice
ras_layer	和奏レイ	RAISE A SUILEN	33 min	1.00	1.00	own voice
ras_lock	朝日六花	RAISE A SUILEN	34 min	1.00	0.50	generic
ras_masking	佐藤ますき	RAISE A SUILEN	23 min	1.00	0.81	partly generic
ras_pareo	鳰原令王那	RAISE A SUILEN	22 min	1.00	0.97	own voice
ras_chuchu	珠手ちゆ	RAISE A SUILEN	42 min	0.93	0.75	partly generic
mygo_tomori	高松燈	MyGO!!!!!	30 min	1.00	0.44	generic
mygo_anon	千早愛音	MyGO!!!!!	11 min	1.00	0.94	own voice
mygo_raana	要楽奈	MyGO!!!!!	13 min	1.00	0.91	own voice
mygo_soyo	長崎そよ	MyGO!!!!!	10 min	0.88	1.00	own voice
mygo_soyo_tsuki	長崎そよ	MyGO!!!!!	16 min	1.00	0.44	generic
mygo_taki	椎名立希	MyGO!!!!!	27 min	1.00	0.97	own voice
mujica_uika	三角初華	Ave Mujica	6 min	1.00	1.00	own voice
mujica_mutsumi	若葉睦	Ave Mujica	3 min	1.00	0.91	own voice
mujica_mortis	若葉睦	Ave Mujica	5 min	1.00	0.94	own voice
mujica_umiri	八幡海鈴	Ave Mujica	3 min	0.97	1.00	own voice
mujica_nyamu	祐天寺にゃむ	Ave Mujica	6 min	1.00	1.00	own voice
mujica_sakiko	豊川祥子	Ave Mujica	6 min	0.83	0.94	own voice
mujica_sakiko_tsuki	豊川祥子	Ave Mujica	13 min	1.00	0.78	partly generic

Summary: Japanese is in-voice for all 49 (pooled identity 0.986, real recordings 0.988). Mandarin is in-voice for 34 voices; the 15 marked generic or partly generic fall back toward a generic timbre in Mandarin and English.

Persona rows. ro_ako_maou (Ako's 大魔姫 persona), mygo_soyo_tsuki and mujica_sakiko_tsuki (the 月ノ森 versions) and mujica_mortis (Mortis) are separate rows. Their real recordings are confused with the base character 89–97% of the time by the same probe, so expect them to be close to their base row.

Evaluation

Checkpoints were chosen by selection rules written down before each training run (eval/E1_PLAN.md, eval/E5_PLAN.md). Per checkpoint: 4,356 renders, i.e. 49 voices × (held-out game lines + 12 new Japanese sentences, 16 Mandarin, 8 English) × 2 seeds.

Measures:

• Identity (leak): the share of renders a codec-calibrated speaker probe (logistic regression on Qwen3-TTS speaker-encoder vectors, fitted on resynthesised real clips) assigns to a different voice. Lower is better.
• ja CER: Qwen3-ASR character error rate on held-out Japanese lines.
• Mandarin accent: SER (shared-character syllable error rate, human-validated on the sibling Kurisu project) and lid (a language-id log-ratio; more negative = more native).
• en WER / lid: the same for English.

Model	ja leak ↓	ja p_true ↑	ja CER ↓	zh SER ↓	zh lid ↓	zh leak ↓	Own-voice Mandarin (zh leak ≤ 0.15)	en lid ↓	en WER ↓
Real recordings, through the codec	0.012	0.777	0.023	—	—	—	—	—	—
Floor: untrained Base, same speaker vectors	0.100	0.567	0.032	0.077	−19.5	0.157	36/49	−12.2	0.001
E1a epoch 1: Japanese + language ids + replay, no pseudo-Mandarin (rejected: accented Mandarin)	0.009	0.810	0.027	0.177	−12.6	0.008	—	—	0.006
This model: E1b epoch 2	0.014	0.802	0.028	0.069	−18.6	0.155	34/49	−9.3	0.003
E2 epoch 2: + more self-generated rows (not adopted)	0.021	0.786	—	0.063	−19.7	0.186	34/49	−10.1	—
E5 epoch 2: + 218 RVC-converted Mandarin rows (not adopted)	0.014	0.815	0.025	0.078	−18.1	0.088	40/49	−9.5	0.003

How to read the table:

• Japanese identity is at the real-recording level (leak 0.014 against 0.012 for real clips through the same codec; the untrained floor is 0.100). The remaining errors are mostly inside persona pairs, plus ppp_kasumi (0.84).
• Mandarin accent stays at the floor (SER 0.069 vs 0.077). With language ids and 13% native replay but only Japanese in the characters' voices (E1a), Mandarin comes out Japanese-accented (SER 0.177); adding 5% gated in-voice pseudo-Mandarin (this model) removes it.
• Mandarin identity is the open problem. Voices that had no usable in-voice Mandarin rows fall back to the base model's timbre.
• Two follow-ups did not replace this model.
  • E2 (doubling the self-generated rows) diluted identity.
  • E5 trained per-voice RVC converters on each voice's Japanese clips, converted native Mandarin into 15 voices, and retrained. Mandarin identity improved (0.155 → 0.088, 34 → 40 voices; Eve, Kanon, Hagumi, Aya, Kaoru fixed), but it missed the adoption bar fixed before training (≤ 0.08, ≥ 43/49, targets ≤ 0.15). Following that rule, E5 is not released.
  • Details: eval/E5_RESULTS.md.

Training details

The code is in the GitHub repository. Training records (data counts, mix, reports) are in training_data/; per-clip transcripts are deliberately not published.

Data

• Japanese: 9,804 voice lines (22.1 h) of the 49 voices, re-cut and levelled per voice; transcripts checked with Qwen3-ASR; held-out test lines and validation are split off; 9,268 lines are in training. Poppin'Party to MyGO!!!!! come from the game Girls Band Party; Ave Mujica's (3–13 min per voice) mostly from the anime.
• Native replay: about 13% of training rows are native Mandarin (AISHELL-3) and English (LibriTTS-R) clips from 55 auxiliary speakers, so the model keeps native articulation in those languages.
• Pseudo-Mandarin (in-voice): gated Base-model renders in each voice's locked speaker vector (accent, text and identity gates), capped per voice.

Speaker vectors

• Each voice has a fixed vector: the mean Qwen3-TTS speaker-encoder embedding of up to 64 screened reference clips.
• It is written into codec-embedding rows 3000–3048 and never trained.

Optimisation

• Full fine-tune, fp32 master weights, bf16 autocast; AdamW, lr 5e-6, weight decay 0.01, cosine over a 3-epoch schedule with checkpoints at epochs 1, 2 and 3; this is the epoch-2 checkpoint (picked by the pre-registered rule); effective batch 32.
• Language id on every row, 25% of rows on the Auto prefix.
• Trainer: the fixed multi-speaker trainer from the Kurisu project (corrects the official sft_12hz.py's label shift, sub-talker frame and text_projection bugs).

Compute: one H100 80GB. The whole E1 study (smoke test, two arms × three epochs, floor and real-clip references, evaluation) took about 5 h, about $20; this model's own training took about 10 minutes. The follow-ups (E2, E5) are documented in the code repository.

Limitations

• Mandarin in a generic timbre for 15 voices: ppp_kasumi, ppp_rimi, ro_ako_maou, ro_rinko, ras_lock, mygo_tomori, mygo_soyo_tsuki (generic), and pp_eve, hhw_kanon, ras_chuchu, mujica_sakiko_tsuki, hhw_hagumi, ras_masking, pp_aya, hhw_kaoru (partly). Their Mandarin is native-accented but may not sound like the character.
• English is intelligible but mildly Japanese-accented for most voices (en lid −9.3; the untrained floor is −12.2), and clearly accented for the generic group. There were no English in-voice rows.
• Persona pairs are close. mujica_sakiko / mujica_sakiko_tsuki, mygo_soyo / mygo_soyo_tsuki and ro_ako / ro_ako_maou sometimes swap (Japanese identity 0.83–0.88 for the weaker side). ppp_kasumi has the lowest Japanese identity among the game voices (0.84).
• Ave Mujica voices have little data (3–13 min, mostly anime audio), so they are less stable than the game voices.
• Evaluation is automatic (speaker probe, ASR, accent metrics). Listening checks were used for intermediate data, not for a full blind test of this checkpoint.

Rights and responsible use

This is an unofficial fan project, with no affiliation to Bushiroad, Craft Egg, the BanG Dream! project, or the voice cast.

• The characters, the source audio and all related marks belong to their respective owners.
• No copyright, performer, publicity or voice rights are granted by this repository.
• No game audio, dataset, transcripts or audio samples are distributed here; only model weights and evaluation numbers.
• The official BanG Dream! fan-work guideline (https://bang-dream.com/bdp-guideline/) restricts derivative works that use the franchise's audio. Read it before using outputs anywhere public.

Do not use it to:

• impersonate the voice actors or anyone else;
• mislead people about authentic or official speech, or make anything that could be mistaken for official material;
• harm the image of the characters or the cast;
• suggest endorsement, harass, or make commercial use.

Disclose synthetic audio as synthetic. Users are responsible for lawful use in their jurisdiction. Rights holders and cast members may contact the repository owner (Hugging Face: starrydark; GitHub: starryark, issues) for removal, and the repository will be taken down on request.

Model weights derive from Qwen3-TTS (Apache-2.0).

Citation

@misc{bangdream_qwen3_tts_2026,
  title  = {BanG Dream! character voices: a cross-lingual Qwen3-TTS fine-tune},
  author = {starrydark},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/starrydark/BanG_Dream_Qwen3_TTS}},
  note   = {Code: \url{https://github.com/starryark/bangdream-tts}}
}
@article{qwen3tts2026,
  title   = {Qwen3-TTS Technical Report},
  journal = {arXiv preprint arXiv:2601.15621},
  year    = {2026}
}