starrydark/BanG_Dream_Qwen3_TTS

BanG Dream! (邦邦 / バンドリ) character voices: Qwen3-TTS text-to-speech in Japanese, Mandarin and English

BanG Dream! (邦邦) character-voice TTS: 49 voices + 25 opt-in Our Notes voices (MyGO!!!!!, Ave Mujica, Girls Band Party, Our Notes). They speak Japanese, Mandarin and English with no reference audio. A Qwen3-TTS fine-tune. Unofficial fan project, not affiliated with Bushiroad, Craft Egg, the BanG Dream! project, the official 夢ノ結唱 voicebanks or the voice cast.

This revision (main) holds the 49 Girls Band Party, MyGO!!!!! and Ave Mujica voices. 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.
• Try it online: demo Space (ZeroGPU; pick a voice, type text). Its output is AI-generated, under the same terms as this model.
• 25 more voices, opt-in: the ournotes branch is a separate checkpoint (E9) with voices trained on BanG Dream! Our Notes story audio: 16 new characters (一家Dumb Rock!, 夢限大みゅーたいぷ, millsage, 沢海奏多) and the Our Notes era of nine MyGO!!!!! / Ave Mujica characters (*_ournotes). Load it with revision="ournotes" for those 25 voices only, and keep using this revision for the 49 here. The demo Space serves both.

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: 邦邦 · 邦多利 · BanG Dream! 少女乐团派对! (Girls Band Party) · 《迷途之子!!!!!》 (MyGO!!!!!) · 《颂乐人偶》 (Ave Mujica) · BanG Dream! 交织的乐章 (Our Notes) · バンドリ！ ガールズバンドパーティ！ · ガルパ.

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.

中文说明

非官方 BanG Dream!（邦邦）角色语音合成模型，基于 Qwen3-TTS 微调。按说话人 ID 直接合成，无需参考音频；会说日语、普通话和英语。粉丝自制，与 Bushiroad、Craft Egg、BanG Dream! 官方、官方 AI 声库“梦的结唱”及声优均无关。

本模型（main）包含少女乐团派对 7 个乐队、MyGO!!!!!（《迷途之子!!!!!》）和 Ave Mujica（《颂乐人偶》）共 49 个声线；《BanG Dream! Our Notes》的 25 个声线在可选分支 ournotes。在线体验：演示 Space（74 个声线，可直接输入中文名搜索）。

import os; os.environ.setdefault("HF_ENDPOINT", "https://hf-mirror.com")   # 中国大陆网络
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="今天的排练真开心，明天也一起加油吧！", speaker="mygo_anon", language="Chinese")
sf.write("out.wav", wavs[0], sr)

普通话音色（说中文时像不像本人）：

程度	数量	声线
保持角色音色	34	下表中标为 own voice 的声线
部分偏通用	8	丸山彩（pp_aya）、若宫伊芙（pp_eve）、濑田薰（hhw_kaoru）、北泽育美（hhw_hagumi）、松原花音（hhw_kanon）、佐藤益木（ras_masking）、珠手知由（ras_chuchu）、丰川祥子（月之森）（mujica_sakiko_tsuki）
偏通用音色	7	户山香澄（ppp_kasumi）、牛込里美（ppp_rimi）、宇田川亚子（大魔姬）（ro_ako_maou）、白金燐子（ro_rinko）、朝日六花（ras_lock）、高松灯（mygo_tomori）、长崎爽世（月之森）（mygo_soyo_tsuki）

“偏通用”的意思是：口音仍是标准普通话，但音色更接近基础模型的默认声音，而不是角色本人。

使用规则：

• 生成的音频均为 AI 合成，发布时须明确标注“AI生成”（《人工智能生成合成内容标识办法》，2025-09-01 起施行）。在 bilibili 投稿时，请在上传页勾选 AI 标识。
• 禁止商用。禁止冒充声优或官方，禁止用于损害角色及声优形象的内容。
• BanG Dream! 官方二创指南（https://bang-dream.com/bdp-guideline/）第 10 条禁止使用作品音声制作、公开二次创作。本模型是在此前提下的非官方粉丝作品。权利方或声优提出要求时，模型与演示将立即下架。

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	中文名	Band	Japanese data	Japanese identity	Mandarin identity	Mandarin voice
ppp_kasumi	戸山香澄	户山香澄	Poppin'Party	35 min	0.84	0.25	generic
ppp_tae	花園たえ	花园多惠	Poppin'Party	31 min	1.00	0.94	own voice
ppp_rimi	牛込りみ	牛込里美	Poppin'Party	24 min	1.00	0.25	generic
ppp_saaya	山吹沙綾	山吹沙绫	Poppin'Party	47 min	1.00	1.00	own voice
ppp_arisa	市ヶ谷有咲	市谷有咲	Poppin'Party	61 min	0.98	0.94	own voice
ag_ran	美竹蘭	美竹兰	Afterglow	26 min	1.00	0.91	own voice
ag_moca	青葉モカ	青叶摩卡	Afterglow	59 min	0.98	1.00	own voice
ag_himari	上原ひまり	上原绯玛丽	Afterglow	45 min	1.00	0.88	own voice
ag_tomoe	宇田川巴	宇田川巴	Afterglow	31 min	1.00	0.97	own voice
ag_tsugumi	羽沢つぐみ	羽泽鸫	Afterglow	30 min	1.00	0.88	own voice
pp_aya	丸山彩	丸山彩	Pastel*Palettes	52 min	0.98	0.84	partly generic
pp_hina	氷川日菜	冰川日菜	Pastel*Palettes	34 min	1.00	0.97	own voice
pp_chisato	白鷺千聖	白鹭千圣	Pastel*Palettes	26 min	1.00	1.00	own voice
pp_maya	大和麻弥	大和麻弥	Pastel*Palettes	45 min	1.00	0.94	own voice
pp_eve	若宮イヴ	若宫伊芙	Pastel*Palettes	46 min	1.00	0.62	partly generic
ro_yukina	湊友希那	凑友希那	Roselia	19 min	1.00	1.00	own voice
ro_sayo	氷川紗夜	冰川纱夜	Roselia	31 min	1.00	1.00	own voice
ro_lisa	今井リサ	今井莉莎	Roselia	39 min	1.00	1.00	own voice
ro_ako	宇田川あこ	宇田川亚子	Roselia	8 min	1.00	0.88	own voice
ro_ako_maou	宇田川あこ	宇田川亚子（大魔姬）	Roselia	5 min	0.88	0.31	generic
ro_rinko	白金燐子	白金燐子	Roselia	42 min	1.00	0.53	generic
hhw_kokoro	弦巻こころ	弦卷心	Hello, Happy World!	29 min	1.00	0.94	own voice
hhw_kaoru	瀬田薫	濑田薰	Hello, Happy World!	30 min	1.00	0.84	partly generic
hhw_hagumi	北沢はぐみ	北泽育美	Hello, Happy World!	23 min	1.00	0.81	partly generic
hhw_kanon	松原花音	松原花音	Hello, Happy World!	36 min	1.00	0.66	partly generic
hhw_misaki	奥沢美咲	奥泽美咲	Hello, Happy World!	37 min	1.00	0.97	own voice
mo_mashiro	倉田ましろ	仓田真白	Morfonica	23 min	1.00	0.97	own voice
mo_touko	桐ヶ谷透子	桐谷透子	Morfonica	22 min	1.00	0.91	own voice
mo_nanami	広町七深	广町七深	Morfonica	34 min	1.00	0.94	own voice
mo_tsukushi	二葉つくし	二叶筑紫	Morfonica	25 min	1.00	1.00	own voice
mo_rui	八潮瑠唯	八潮瑠唯	Morfonica	26 min	1.00	1.00	own voice
ras_layer	和奏レイ	和奏瑞依	RAISE A SUILEN	33 min	1.00	1.00	own voice
ras_lock	朝日六花	朝日六花	RAISE A SUILEN	34 min	1.00	0.50	generic
ras_masking	佐藤ますき	佐藤益木	RAISE A SUILEN	23 min	1.00	0.81	partly generic
ras_pareo	鳰原令王那	鳰原令王那	RAISE A SUILEN	22 min	1.00	0.97	own voice
ras_chuchu	珠手ちゆ	珠手知由	RAISE A SUILEN	42 min	0.93	0.75	partly generic
mygo_tomori	高松燈	高松灯	MyGO!!!!!	30 min	1.00	0.44	generic
mygo_anon	千早愛音	千早爱音	MyGO!!!!!	11 min	1.00	0.94	own voice
mygo_raana	要楽奈	要乐奈	MyGO!!!!!	13 min	1.00	0.91	own voice
mygo_soyo	長崎そよ	长崎爽世〔素世〕	MyGO!!!!!	10 min	0.88	1.00	own voice
mygo_soyo_tsuki	長崎そよ	长崎爽世〔素世〕（月之森）	MyGO!!!!!	16 min	1.00	0.44	generic
mygo_taki	椎名立希	椎名立希	MyGO!!!!!	27 min	1.00	0.97	own voice
mujica_uika	三角初華	三角初华	Ave Mujica	6 min	1.00	1.00	own voice
mujica_mutsumi	若葉睦	若叶睦	Ave Mujica	3 min	1.00	0.91	own voice
mujica_mortis	若葉睦	Mortis〔墨缇丝〕	Ave Mujica	5 min	1.00	0.94	own voice
mujica_umiri	八幡海鈴	八幡海铃	Ave Mujica	3 min	0.97	1.00	own voice
mujica_nyamu	祐天寺にゃむ	祐天寺若麦〔喵梦〕	Ave Mujica	6 min	1.00	1.00	own voice
mujica_sakiko	豊川祥子	丰川祥子〔小祥〕	Ave Mujica	6 min	0.83	0.94	own voice
mujica_sakiko_tsuki	豊川祥子	丰川祥子〔小祥〕（月之森）	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.

FAQ

Does it need reference audio? No. Every voice is a built-in speaker row, called by its id (speaker="mygo_anon").

Can the characters speak Chinese? Yes. Pass language="Chinese". Every voice speaks Mandarin with a native accent. Most voices also keep their own timbre in Mandarin; those that drift toward a generic timbre are marked in the table and listed in the Chinese section.

Which revision should I load?

• main for the 49 Girls Band Party / MyGO!!!!! / Ave Mujica voices.
• revision="ournotes" for the 16 Our Notes characters and the *_ournotes voices.
• The demo picks the right one automatically.

How is it different from GPT-SoVITS, RVC or Bert-VITS2 character models?

• It is one checkpoint with every voice built in, so it needs no reference clip (GPT-SoVITS) and no source recording (RVC).
• Bert-VITS2 character models also call built-in speakers; against those, the difference is Mandarin and evaluation.
• It was trained so that its Mandarin is not Japanese-accented.
• Its checkpoints were chosen by rules fixed before training, and the later ones by blind listening.

What hardware does it need? It is a fine-tune of Qwen3-TTS-12Hz-1.7B-Base; the bf16 weights are 3.8 GB, and inference runs on a CUDA GPU (pip install qwen-tts). Without a GPU, use the demo.

Can I download it from mainland China? Set HF_ENDPOINT=https://hf-mirror.com before from_pretrained("starrydark/BanG_Dream_Qwen3_TTS", ...).

Can I post what I make? Only as a labelled, non-commercial fan work. The BanG Dream! fan-work guideline prohibits derivative works that use the franchise's audio (item 10), and China requires an explicit AI label on generated audio. Never impersonate the cast or present output as official. See Rights and responsible use.

Is this official, or related to 夢ノ結唱 (梦的结唱)? No. This is an unofficial fan project. 夢ノ結唱 is Bushiroad Music's official AI singing product and is unrelated.

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.
• Six 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.
  • E6 and E7 (2026-11-08) retrained on BanG Dream! Our Notes story audio: the MyGO!!!!! / Ave Mujica voices were replaced, and 16 new characters were added (E7 with own-voice Mandarin rows from the start). Neither passed its pre-registered guards. The main blockers were one new voice's Japanese identity (yomogi), Japanese CER on the new text (which real recordings themselves only just clear) and Mandarin accent on three voices. So v2 was not released. Details: e6/E6_RESULTS.md and e7/E7_RESULTS.md in the code repository.
  • E8 and E9 (2026-11-09) were judged by blind listening against this model.
    • E9 trained both eras of the MyGO!!!!! / Ave Mujica characters, with one id per era.
    • Its 25 Our Notes voices were all judged in character, but its versions of this model's MyGO!!!!! / Ave Mujica voices were judged worse than these (mean −0.58 on −3…+3, p = 0.046).
    • So this model stays the default, and E9 is published only as the opt-in ournotes branch for its 25 Our Notes voices.
    • Details: e9/E9_RESULTS.md in the code repository and eval/ on that branch.

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–E9) 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/) prohibits derivative works that use the franchise's audio. Item 10 reads 「本コンテンツのうち、音声を利用した二次的著作物の制作、展示、頒布、公開。」. This model is an unofficial fan work made despite that. Read the guideline before using outputs anywhere public.
• AI labels: in mainland China, AI-generated audio must carry an explicit AI label (《人工智能生成合成内容标识办法》, in force since 2025-09-01), and bilibili asks uploaders to set its AI标识 when uploading.

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}
}