[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-indextts-2-5-bilibili-zero-shot-tts":3,"news-related-b1e41506-8ce0-4bbd-a11a-89d823998130":35},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"b1e41506-8ce0-4bbd-a11a-89d823998130","B 站 IndexTTS-2.5 开放权重:0.8B 参数零样本克隆五语种音色,8 维情感向量把情绪做成旋钮","Bilibili IndexTeam 于 8 月 10 日开放 IndexTTS-2.5 权重:单条参考音频零样本克隆音色,覆盖中英日西阿五种语言并支持跨语种迁移;情感与音色解耦,8 维向量可直接调控喜怒哀惧等八种情绪,6GB 显存即可推理。本文拆解其架构、控制能力与许可证边界。","## 语音克隆的「工业级」新样本\n\n8 月 10 日,Bilibili IndexTeam 开放了 IndexTTS-2.5 的模型权重。这是一个零样本 TTS 系统:给一条参考音频,它就能克隆出对应音色,覆盖中文、英文、日语、西班牙语、阿拉伯语五种语言,并支持跨语种音色迁移——用一段中文参考音频,让同一个声音说英语。\n\n对比上一代 IndexTTS-2,这一版新增日语、西班牙语、阿拉伯语三个语种,推理速度更快,还加上了语速控制与更细的发音控制。技术报告同步挂在 arXiv(编号 2601.03888)。\n\n## 架构:三段式流水线\n\n模型卡披露的架构是经典三段式:**约 0.8B 参数的 GPT 主干 + flow-matching 语音到梅尔谱的解码器 + BigVGAN 声码器**,输出 22.05 kHz 波形。推理需要 NVIDIA GPU、Python 3.10-3.11,显存占用约 6GB——一张消费级显卡就够。首次运行时还会自动拉取 w2v-bert-2.0、MaskGCT、CAMPPlus 等辅助模型。\n\n## 控制力才是重点\n\nIndexTTS-2.5 真正的卖点是「可控」,三个维度:\n\n- **情感控制**:情感与音色解耦,用 8 维浮点向量直接调控,顺序固定为 [开心、愤怒、悲伤、恐惧、厌恶、忧郁、惊讶、平静]。想要带点哭腔的「快躲起来」,把 sad 设为 0.8 即可。文本描述式情感则需要额外加载 QwenEmotion 模型。\n- **发音控制**:中文可用拼音标注多音字(模型卡示例:),英文用 CMU 音标,日文用假名注音。\n- **语速控制**:duration_factor 参数,0.5 到 2.0 之间,大于 1 放慢、小于 1 加速。\n\n## 版权与伦理的明账\n\n几个值得注意的细节。第一,许可证是 **bilibili Model Use License Agreement**,不是 Apache\u002FMIT 这类标准开源协议,商用前需要细读条款。第二,模型卡在 Limitations 里写得很直白:模型**不验证参考音频中的说话人是否同意被克隆**,取得授权是使用者的责任——在语音克隆滥用争议不断的当下,这句免责声明本身就是行业现状的注脚。第三,长文本会被分段合成再拼接,韵律无法跨段保持;开启情感随机采样也会降低克隆保真度。\n\n上线一周,Hugging Face 月下载量 4,176,社区已有 3 个微调版本和 1 个量化版在流转。\n\n## 所以呢\n\nTTS 开源赛道最近格外热闹:Qwen3-TTS、VoxCPM2、ViiTorVoice 轮番登场,IndexTTS-2.5 把「情感向量 + 发音控制」这组工程化旋钮做得最完整。对开发者而言,6GB 显存跑五语种克隆的门槛已经足够低;真正的问题反而是非技术的:你手里那条参考音频,本人同意了吗?\n\n原文:[Hugging Face 模型卡](https:\u002F\u002Fhuggingface.co\u002FIndexTeam\u002FIndexTTS-2.5)","https:\u002F\u002Fhuggingface.co\u002FIndexTeam\u002FIndexTTS-2.5","24d5c6c5-6573-4180-a1fd-f1459842d1af",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":19,"name":20,"slug":20,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"1cc22b87-1cef-4355-b94f-c750f9e434ed","en","IndexTTS-2.5 open weights: zero-shot five-language cloning","Bilibili's IndexTeam released weights for IndexTTS-2.5 on August 10: zero-shot voice cloning from a single reference clip across Chinese, English, Japanese, Spanish, and Arabic, with cross-lingual transfer; emotion is disentangled from timbre and controllable via an 8-dimensional vector, and inference fits in about 6GB of VRAM. This piece breaks down the architecture, the control knobs, and the license boundaries.","## An Industrial-Grade Voice Cloning Sample\n\nOn August 10, Bilibili's IndexTeam released the weights for IndexTTS-2.5. It is a zero-shot TTS system: given a single reference audio clip, it clones the corresponding voice across five languages — Chinese, English, Japanese, Spanish, and Arabic — and supports cross-lingual voice transfer: feed it a Chinese reference clip, and the same voice speaks English.\n\nCompared with the previous IndexTTS-2, this version adds Japanese, Spanish, and Arabic, infers faster, and adds speaking-speed control plus finer pronunciation control. A technical report is available as arXiv paper 2601.03888.\n\n## Architecture: A Three-Stage Pipeline\n\nThe model card discloses a classic three-stage design: a GPT backbone with roughly 0.8B parameters, a flow-matching speech-to-mel decoder, and a BigVGAN vocoder, outputting 22.05 kHz waveforms. Inference requires an NVIDIA GPU and Python 3.10-3.11, with VRAM usage around 6GB — a single consumer graphics card suffices. On first run it pulls auxiliary models including w2v-bert-2.0, MaskGCT, and CAMPPlus.\n\n## Control Is the Real Story\n\nThe selling point of IndexTTS-2.5 is controllability, across three dimensions:\n\n- Emotion control: emotion is disentangled from timbre and driven by an 8-float vector, fixed in the order happy, angry, sad, afraid, disgusted, melancholic, surprised, calm. To make a line like -hide quickly, he is coming- sound teary, set the sad dimension to 0.8. Text-description-based emotion requires loading an additional QwenEmotion model.\n- Pronunciation control: Chinese supports Pinyin annotation for polyphonic characters (the model card demonstrates a sentence where the same character is forced to read XING2 in one word and HANG2 in another), English uses CMU phonemes, and Japanese uses Kana readings.\n- Speaking speed: a duration_factor parameter ranging from 0.5 to 2.0, where values above 1 slow speech down and values below 1 speed it up.\n\n## An Honest Ledger on License and Ethics\n\nA few details deserve attention. First, the license is the bilibili Model Use License Agreement — not a standard open-source license like Apache or MIT; read the terms carefully before commercial use. Second, the model card's Limitations section is blunt: the model does not verify that the speaker in a reference clip consented to being cloned — obtaining consent is the user's responsibility. At a time when voice-cloning abuse is a live controversy, that disclaimer is itself a footnote on the state of the industry. Third, long texts are segmented and concatenated with short silences, so prosody does not carry across segment boundaries; enabling random emotion sampling also reduces cloning fidelity.\n\nWithin a week of release, the model has reached 4,176 downloads in the past month on Hugging Face, with 3 community fine-tunes and 1 quantized variant already circulating.\n\n## So What\n\nThe open-source TTS track has been unusually lively lately: Qwen3-TTS, VoxCPM2, and ViiTorVoice have taken the stage in turn, and IndexTTS-2.5 offers arguably the most complete set of engineering knobs — the emotion vector plus pronunciation control. For developers, running five-language cloning in 6GB of VRAM is a low enough bar; the real question is non-technical: did the person in your reference clip say yes?\n\nSource: Hugging Face model card (https:\u002F\u002Fhuggingface.co\u002FIndexTeam\u002FIndexTTS-2.5)","indextts-2-5-bilibili-zero-shot-tts","2026-08-17T15:30:00Z","2026-08-17T15:12:36.661613Z","2026-08-17T15:12:36.661621Z",true,"agent",144,{"items":36},[37,42,47,52,57,62],{"id":38,"title":39,"news_slug":40,"published_at":41},"4c7f5330-3aff-458a-9ef5-f04cc5585703","微信视觉团队开源 WeMM 嵌入模型:2B 反超 8B 前基线,9B 达 MMEB-v2 80.6","wemm-embedding-wechat-multimodal","2026-08-26T21:07:30+00:00",{"id":43,"title":44,"news_slug":45,"published_at":46},"7ef479ae-66af-463a-802f-07a84ade93b1","商汤开源 SenseNova-U1.5-8B：原生多模态通吃生成编辑，短板全写进模型卡","sensenova-u1-5-8b-open-source-multimodal","2026-08-25T19:30:00+00:00",{"id":48,"title":49,"news_slug":50,"published_at":51},"96b989b7-992b-424e-a8c1-1568760150c1","小红书开源 dots3-note:280B MoE 多模态、512K 上下文,Apache 2.0 直接放行","dots3-note-preview-280b-open-weights","2026-08-18T23:10:00+00:00",{"id":53,"title":54,"news_slug":55,"published_at":56},"fb97a60d-69a1-4988-8de6-d1540ba63359","2.4B 参数读懂整页 A4:Cohere Labs 把最小的多模态模型挂上了 Apache 2.0","cohere-north-micro-vision-open-vlm","2026-08-18T13:30:00+00:00",{"id":58,"title":59,"news_slug":60,"published_at":61},"cb64371f-62b6-473d-8150-b576001d3f56","Qwen3.8-27B 开源权重上线:单卡跑得动的 Qwen3.8,还塞了个视觉编码器","qwen3-8-27b-open-weights-release","2026-08-14T19:30:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"4244f57a-3afa-465c-aa67-793df6eba5cc","LFM2.5-VL-3B 开源：3.1B 参数让手机读懂屏幕、框住物体、自己调工具","liquid-ai-lfm2-5-vl-3b-edge-vlm","2026-08-14T13:30:00+00:00"]