[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-antalia-2-mini-turkish-tts-synthetic-data":3,"topics-all":38,"news-related-503fc4f0-755d-43f1-89cc-8d86b4cdaed9":67},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"503fc4f0-755d-43f1-89cc-8d86b4cdaed9","7.6M参数土耳其语TTS:WER 1.09%却没人做过盲听","PatientDesk AI 刚公开的土耳其语语音合成技术报告，把 7.6M 参数的模型做到字错率 1.09%、CPU 上实时 41 倍，但训练数据全部来自大约大 300 倍的教师模型，片段开头那个 ö 的发音缺陷被学生原样继承，最终分数还是在基准上挑出来的，而且至今从来没有做过任何真人盲听。","土耳其语有八千多万人使用，但在语音合成里一直算中等资源语言。PatientDesk AI 刚公开的技术报告把一个 7,620,931 参数的 TTS 模型推到台前：48 kHz 输出，在 Apple M5 笔记本 CPU 上只开两个线程就跑出实时 41 倍（RTF 0.024），权重 30.5 MB。训练数据没有一句真人录音，全部来自 VoxCPM2 这个大约大 300 倍的教师模型合成的语音，质检后留下约 1,790 小时。\n\n## 决定胜负的是数据，不是规模\n\n模型卡把参数拆成两块：3.69M 的声学模型从文本画出梅尔谱，3.93M 的声码器把它还原成 48 kHz 音频。总参数上它比对照组 EMA Lightning（8,556,172 参数）还少 10.93%，后者同样只用 VoxCPM2 的合成语音训练，同一教师不同音色。\n\nFreya-TR-Eval 是 495 句的公开基准。模型卡自带的 harness 记的是：Mini 1.0.1 的 WER 1.02%，EMA Lightning 1.00%，教师 VoxCPM2 0.92%。换到公开基准页那套 harness（换种子、单张 NVIDIA L4），Mini 是 1.09% 对 EMA 的 1.04%，CER 0.220% 对 0.225%。论文把 CER 0.220% 写成\"11 个开放权重系统里最低\"，但在跑过的 15 个里只排第二，前面压着自家闭源的 Alania-2 声音 Öncel（0.217%）。它自己补的那句更狠：这些差距在 80,922 个字符里只有 2 到 7 个错误，\"没有一个能与零区分开\"。\n\n姊妹报告 Alania-2 更直接：488M 版本与 2.3B 模型的字符错误率相差不到半分，7.6M 单音版本\"追平了我们测过的最可懂系统\"。\n\n## 声码器要用生成出来的梅尔谱训练\n\n最初的 2.16M 声码器只在真实梅尔谱上训练，结果越训越差：40 条语料上的复现合成 UTMOS 从第 20,000 步的 2.54 掉到第 148,000 步的 2.17，教师是 3.29。梅尔谱损失一路下降，UTMOS 一路往下掉，两者脱钩。诊断是对的：推理时声码器拿到的是声学模型生成的梅尔谱，和训练时见过的不一样。\n\n改成在声学模型自己的输出上微调后，词错误率从相邻预发布检查点的平均 3.89% 降到 1.75%；同期不动的声码器停在 3.63% 到 4.14%。最终出货的 3.93M 声码器为此付出约 0.24 UTMOS。\n\n## 教师的缺陷，被学生原样继承\n\n土耳其语有个特殊字母 ö。教师模型 VoxCPM2 在片段开头遇到 ö 时会把开头弄坏，且不同词坏的程度不一样：长集里 12,554 条以 ö（大小写均计）开头的保留片段，只有 26% 第一个词是对的（95% 置信区间 25.4 到 26.9），其他大写字母开头的有 94% 是对的，单看大写 Ö 只有 23.66%。\n\n学生把毛病学走了：质检最终保留了 9,271 条\"首个 ö 词被 Whisper 判错\"的行，过滤式微调没有任何改善。论文的临时解法很说明问题——在以 ö 或 ı 开头的片段前塞一句\"ee,\"，合成完再把这段音频剪掉。这是遮，不是修。模型卡的说法是\"这个技巧在每个片段开头把它藏住了，但并没有消除它\"，真正的修法得回数据源头。\n\n## 论文自己承认的三处不牢\n\n第一，最终分数是被挑过的。论文写得很直白：\"最终设置下的数字因为挑选而偏乐观。\"倍音规则来自基准的报错，温度、步数、语速、后缀和均衡器都在同一批 495 句上筛；公开 harness 只换种子，句子还是那批。\n\n第二，没有人做过真人盲听。论文写着\"No human listening test exists\"，错误分析那节又补了\"没有人听过这些 take\"。质量数字都来自自动指标，论文承认 WER 可能相对人类可懂度美化模型，而幅度\"我们没有测量过\"。\n\n第三，它的错误比对手更\"稳定\"。公开 harness 上，模型 70.3% 的词错误集中在三次全败的句子上，EMA Lightning 是 54.9%。有 16 条句子 Mini 三次全错、EMA Lightning 却不是，其中 12 条\"看起来是真的发音错误\"。UTMOS 方向也不一致：全频段 EMA Lightning 领先（3.30 对 3.18），8 kHz 频段 Mini 反而领先（3.05 对 2.88），论文说\"没有测量出符号为什么改变\"。\n\n## 六千步没有梯度的训练\n\n报告里还留了一段像事故复盘的记录。bf16 自动混合精度会缓存转换过的权重，而作者在同一个 autocast 区域里、在 no_grad 下算了快捷采样目标，于是训练拿到的是分离过的副本，生成器一个梯度都没拿到。签名很安静：流匹配损失平在 2.03 附近，滑动平均权重上的验证损失完全不动。主训练前 6,000 步、兄弟运行 15,000 步，生成器全都没在学；到第 13,000 步，输出权重范数仍是零。\n\n教训只有一句：第一步之后，断言每个可训练模块都有梯度。他们第一次启动就崩在\"参数未使用\"的报错上，用 find_unused_parameters=True 压下警告后过了很久才找到原因。\n\n## 公开了什么，又关上了什么\n\nApache-2.0 覆盖权重、推理代码、文本正则器、ONNX 图、页面代码，以及整套基准装置——适配器、results.json，还有逐条转写、得分和计时数据。论文说 results.json 里 20 行结果\"每一行都能从逐条文件零差异重算\"，不必重新渲染音频。\n\n没公开的是训练代码、数据生成脚本、训练数据、ONNX 导出工具和 Android 源码。你能跑它、能重新打分，但不能复现它。出货血脉花了 114.5 个 GPU 小时（声学 36.4，声码器 78.0），全部训练运行是 159.0。\n\n## 所以呢\n\n这份报告最值钱的不是 7.62M 这个数字，而是它愿意把\"教师的声音在开头会坏\"\"这个数字是挑出来的\"\"没人做过盲听\"写进正文。自动指标能告诉你模型在某个基准上错得少多少，却不会告诉你教师的声音里藏着什么缺陷。\n\n数据血缘是这个时代最容易被忽略的资产：模型可以小到在笔记本上跑、快到在浏览器里跑，但若训练数据不公开、教师的缺陷没被追到源头，\"小\"就只是把不可复现性也一起缩小了。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.13811","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"c4455801-0273-477d-9e92-00b4addfa439","asr",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"037a1be9-e7bd-4421-922f-d699b1736176","en","Antalia-2 Mini: A 7.6M Turkish TTS That Nobody Listened To","A 7.6M Turkish TTS reaches 1.09% WER at 41x real time on a laptop CPU, trained on synthetic speech from a 300x larger teacher, with no listening test.","Turkish is spoken by more than eighty million people, but it has remained a mid-resource language for speech synthesis. A technical report just published by PatientDesk AI puts a 7,620,931-parameter text-to-speech model in front of us: 48 kHz output, running at 41 times real time (RTF 0.024) on an Apple M5 laptop CPU with only two threads, and 30.5 MB of weights. Not one real human recording is in the training data. All of it comes from VoxCPM2, a teacher model roughly 300 times larger that generates synthetic speech, and about 1,790 hours survived automatic quality control.\n\nThe interesting part is not the parameters but the almost constant self-questioning in this report.\n\n## What decides the outcome is data, not scale\n\nThe model card splits the parameters in two: a 3.69M acoustic model that draws mel spectrograms from text, and a 3.93M vocoder that turns those mels back into 48 kHz audio. It has 10.93% fewer parameters than the comparison system EMA Lightning (8,556,172), and EMA Lightning is also trained purely on VoxCPM2 synthetic speech from the same teacher, with a different voice.\n\nFreya-TR-Eval is a public benchmark of 495 sentences. On the model card's own harness: Mini 1.0.1 scores WER 1.02%, EMA Lightning 1.00%, and the teacher VoxCPM2 0.92%. On the public benchmark page's harness instead (other seeds, one NVIDIA L4), Mini posts 1.09% against EMA's 1.04%, and CER 0.220% against 0.225%. The paper calls 0.220% \"the lowest of the 11 open-weights systems\", yet across all 15 systems it ran it comes second, behind the authors' own closed Alania-2 voice Öncel at 0.217%. The authors' own follow-up sentence is blunter still: these gaps amount to 2 to 7 character errors in 80,922 characters, and \"none of them is distinguishable from zero\".\n\nThe sibling report Alania-2 argues the same point: the 488M version lands within half a point of CER of a 2.3B model, and the 7.6M single-voice model \"matches the most intelligible systems we measured\".\n\n## A vocoder has to be trained on generated mels\n\nThe first 2.16M vocoder was trained only on real mel spectrograms, and got worse the longer it trained: copy-synthesis UTMOS on 40 corpus clips fell from 2.54 at step 20,000 to 2.17 at step 148,000, against 3.29 for the teacher. Mel loss fell the whole way down while UTMOS fell the whole way down too, and the two decoupled. The diagnosis is right: at inference the vocoder receives mels generated by the acoustic model, which are not the real mels it trained on.\n\nAfter fine-tuning on the acoustic model's own outputs instead, word error rate dropped from a mean of 3.89% at adjacent pre-release checkpoints to 1.75%, while a vocoder frozen over the same stretch stayed between 3.63% and 4.14%. The shipped 3.93M vocoder pays about 0.24 UTMOS for that.\n\n## The teacher's defect, inherited by the student\n\nTurkish has a distinctive letter, ö. The teacher model VoxCPM2 corrupts the start of a clip when the first word begins with ö, and different words break by different amounts: of the 12,554 kept long-set lines starting with ö (counting both cases), only 26% have that first word right (95% CI 25.4 to 26.9), against 94% for lines starting with other capital letters and 23.66% for capital Ö alone. The same ö word scores about 98% later in a clip, which ties the defect to the clip start.\n\nThe student learned it. Quality control kept 9,271 lines whose first ö word Whisper judged wrong, and filtered fine-tunes showed no gain. The paper's stopgap is telling: prepend a spoken \"ee,\" to any chunk starting with ö or ı, then cut that audio back out. That is a cover-up, not a fix. The model card says the trick \"hides it at the start of each chunk but does not remove it\", and the real fix requires regenerating the data at the source.\n\n## Three places the paper admits it is weak\n\nFirst, the final numbers were selected. The paper says so plainly: \"Final-settings numbers are optimistic by selection.\" The doubling rule came from benchmark errors, and temperature, steps, speed, suffix and EQ were all screened on the same 495 sentences; the public harness only changes the seeds, not the sentences.\n\nSecond, no human listening test exists. The paper states \"No human listening test exists\", and the error analysis adds that nobody has listened to these takes. Every quality number comes from an automatic metric, and the authors concede WER may flatter the model relative to human intelligibility by an amount they have not measured.\n\nThird, its errors repeat more than the competitor's. On the public harness, 70.3% of the model's word errors sit in sentences that fail on all three takes, against 54.9% for EMA Lightning. Sixteen sentences fail all three takes for Mini but not for EMA Lightning, twelve of which \"look like real sound errors\" from the transcripts alone. UTMOS even points in opposite directions: EMA Lightning leads at full band (3.30 against 3.18) while Mini leads in the 8 kHz band (3.05 against 2.88), and the paper writes that it has no measurement of why the sign changes.\n\n## Six thousand steps with no gradients\n\nThe report also keeps what reads like an incident post-mortem. bf16 autocast caches the weights it casts, and the authors computed shortcut targets under no_grad inside the same autocast region, so the training pass received detached copies and the generator got no gradients. The signature is quiet: flow-matching loss sits flat near 2.03, validation loss on the moving-average weights never moves, and the shortcut loss is 0. For the first 6,000 steps of the main run and 15,000 of the sibling run, the generator was not learning at all; at step 13,000 of the sibling run, the generator's output weights still had zero norm.\n\nThe lesson is one line: after the first step, assert that every trainable module has a gradient. Their first launch crashed on an unused-parameter error, and they silenced it with find_unused_parameters=True long before finding the real cause.\n\n## What is open, and what is shut\n\nApache-2.0 covers the weights, inference code, text normalizer, ONNX graphs, page code, and the whole benchmark apparatus: adapters, results.json, plus per-clip transcripts, scores and timing. The paper says every one of the 20 rows in results.json \"can be recomputed with zero difference from the per-clip files\" without rendering any audio.\n\nNot published: the training code, data generation scripts, training data, ONNX export tools and Android source. You can run it and rescore it, but you cannot reproduce it. The shipped lineage cost 114.5 GPU-hours (36.4 for the acoustic model, 78.0 for the vocoder), and all training runs came to 159.0.\n\n## So what\n\nThe most valuable thing in this report is not the 7.62M number but its willingness to write \"the teacher's voice breaks at the start\", \"this number was picked\" and \"nobody did a listening test\" into the body. Automatic metrics tell you how many mistakes a model makes on a benchmark; they do not tell you what is hiding inside the teacher's voice, nor that a training run sat idle for six thousand steps.\n\nData lineage is the most overlooked asset of this era. A model can be small enough to run on a laptop, fast enough to run in a browser, and downloadable in full — but if the training data is closed and the teacher's defect was never traced to its source, \"small\" merely means the irreproducibility was scaled down too.","antalia-2-mini-turkish-tts-synthetic-data","2026-11-12T01:40:00Z","2026-11-12T01:31:25.887633Z","2026-11-12T01:31:25.887643Z",true,"agent",200,[39,49,58],{"slug":40,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":35,"created_at":47,"modified_at":48},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":50,"tag_slug":50,"title_zh":51,"title_en":52,"intro_zh":53,"intro_en":54,"id":55,"is_active":35,"created_at":56,"modified_at":57},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":59,"tag_slug":59,"title_zh":60,"title_en":61,"intro_zh":62,"intro_en":63,"id":64,"is_active":35,"created_at":65,"modified_at":66},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":68},[69,74,79,84,89,94],{"id":70,"title":71,"news_slug":72,"published_at":73},"1208ec8f-77b0-476f-8309-720517d22d32","8B语音基座追平4倍大Omni:索尼把力气全花在数据上","sony-walrus-8b-speech-audio-foundation","2026-11-11T21:06:18+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"34bd9f83-04c2-4ffa-8e68-f65a74dfb490","开源语音智能体破50分:Apriel Voice 27B","servicenow-apriel-voice-27b-preview","2026-11-02T13:19:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"039d2494-af49-4d5d-a89f-e3ca91f60df5","转写没有标准答案:英伟达开源43亿激活语音大模型","nvidia-nemotron-3-5-transcribe","2026-10-29T15:15:27+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"35e9f682-7952-4f6a-a8e9-7f8bdd47e791","IBM 语音三件套更新:2B 旗舰 ASR 精度开源第二","granite-speech-5-collection-update","2026-10-22T21:08:48+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"d83c4910-6dce-4171-8a44-9f68c9f841af","Qwen-Audio-3.1-ASR开源:转写自带说话人","qwen-audio-3-1-asr-speaker-diarization","2026-10-22T17:13:29+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"7fd694d4-7eec-4dd4-8fb7-b6e547da98f4","Falcon ASR 开源:1.6B 刷新阿语 WER","falcon-asr-arabic-speech-recognition","2026-10-11T23:08:05+00:00"]