[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-tts-friendly-llm-alignment-fast":3,"news-related-c70131ba-e6b2-466a-ab36-65fad341f006":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},"c70131ba-e6b2-466a-ab36-65fad341f006","别让语音助手念出美元符号:NAVER 对齐 LLM 生成可朗读文本","现在的 LLM 为书面阅读优化,$3.45、缩写和裸网址让语音合成频繁翻车。NAVER LABS Europe 的 EMNLP 2026 论文将其做成偏好对齐:用 FaST 框架和 CORA\u002FRecipe 数据集直接训练模型生成可朗读文本,10 条样本就超过 DPO 等基线,单步延迟不变,人工听测 68.3 分居首。","让语音助手把「The price is $3.45」原样念出来的尴尬,用过 LLM 语音助手的人基本都遇到过:美元符号、缩写、裸网址、紧凑数字,这些为屏幕阅读优化的文本形态,到了语音合成(TTS)环节就成了灾难。NAVER LABS Europe 的三位研究者在 EMNLP 2026 主会论文中给出了一条新路线:与其在下游加重写模块,不如直接把 LLM 对齐成「开口就能念」的模型([arXiv:2609.01246](https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.01246))。\n\n## 问题出在哪:LLM 是为眼睛训练的\n\n当前 LLM 的训练数据和偏好标注几乎全部面向书面文本,结果是语法正确、信息有用,但充满 TTS 不友好的表层形态:符号、缩写、原始 URL、紧凑数字记法。业界常规解法是在 LLM 和 TTS 之间插一个文本正规化后处理模块,但论文指出三个代价:后处理带来可感知的延迟;必须等完整句子生成后才能运行;而且天然绑定每个 TTS 系统特定的 tokenization 和输入格式——换一个合成引擎,重写规则可能全部失效。\n\n## 解法:把「可朗读性」做成偏好对齐\n\n研究团队把 TTS 友好文本生成形式化为偏好对齐问题,从零构建了两个偏好数据集:CORA(合成的咖啡点单助手场景,262 条偏好元组,主打对话式回复)和 Recipe(基于 220 万条食谱的 RecipeNLG 语料采样出 300 条,主打程序性描述),每条数据包含一好一坏两个版本,覆盖符号、缩写、数字速记和裸 URL 等典型坑。\n\n方法上采用 FaST(Feature-aware Sampling and Tuning)框架:不用黑盒 reward model,而是让 LLM 自动发现一组可解释的高层特征(比如「数字是否拼写出来」「缩写密度」「对话语气自然度」),按特征打分、学权重、再采样对齐。在 CORA 上学到的最大负权重特征正是「紧凑数字格式」(-0.50),其次是缩写密度(-0.32);最大正权重是自然对话语气和数字拼写(各 +0.33)。\n\n## 结果:10 条数据就能赢\n\n对比实验覆盖 prompting、SFT、DPO、GRPO、RFT 五类基线,在 10 条和 100 条训练样本两个档位、Qwen3-4B 和 SmolLM3-3B 两个模型家族上跑——FaST 在 10 条样本的低数据档位就拿下 TTS 友好度和 helpfulness 的最佳平衡。\n\n与两步式文本正规化的对比更能说明问题:用 PolyNorm(Qwen3-4B few-shot)后处理零样本输出,CORA 上 TTS 友好度确实冲到 4.92 分,但单条回复延迟从 1.6 秒翻倍到 3.4 秒;FaST 单步拿到 4.73 分,helpfulness 反而更高(4.86 对 4.72)。Recipe 数据集上差距更大:PolyNorm 8.5 秒对 FaST 4.4 秒。\n\n人工听测(MUSHRA 协议,Prolific 众包 14 位有效评审)给出一致排序:FaST 68.3 分,领先 DPO(55.9 分)12.4 分(p=0.003),领先 Prompting(51.4 分)16.9 分(p=0.001)。团队还验证了便宜的启发式指标与 TTS→ASR 回环指标的系统级 Spearman 相关在 CORA 上达到 -1.00——五个系统排序完全一致,意味着评估可朗读性未必真要跑一遍合成。\n\n一个直观例子:问「香蕉面包多少钱」,零样本模型答「$3.45」,FaST 对齐后回答「three point forty-five dollars」——数字直接念出来。\n\n## 所以呢\n\n这篇论文的启示不止于语音助手。其一,「为下游消费场景优化文本生成」可以做成轻量偏好对齐,10 条数据就能启动,对没有大规模标注预算的团队很实用;其二,启发式指标与人工听测强相关,作者明确说它可以直接当 reward model 用——评估即奖励,对接可验证奖励的 RL 思路;其三,局限同样清楚:实验只覆盖英语和两个领域,模型止步 3B\u002F4B,且只适用于级联 LLM→TTS 架构,端到端语音大模型不适用。\n\n数据集、指标和代码已开源(github.com\u002Fnaver\u002Ftts-friendly-gen)。语音交互正在成为 Agent 的默认界面,「写得好」和「说得好」正在分化成两个不同的优化目标——这篇论文把后者正式拉进了对齐的版图。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.01246v1","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18],{"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},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"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},"d0627d1e-e113-4116-9ea9-2c2efbaace8e","en","Stop Making TTS Read Dollar Signs: NAVER Aligns LLMs to Speak","LLMs are optimized for written text, so $3.45, abbreviations and raw URLs keep tripping up speech synthesis. An EMNLP 2026 paper from NAVER LABS Europe frames this as a preference alignment problem: the FaST framework plus two purpose-built preference datasets (CORA, Recipe) teach models to generate TTS-friendly text directly, beating DPO and other baselines with as few as 10 samples at no extra inference latency, and topping human listening tests at 68.3.","Anyone who has used an LLM voice assistant has heard it read out \"The price is $3.45\" verbatim: dollar signs, abbreviations, raw URLs and compact numbers are forms optimized for reading on screen, and they become disasters once they reach Text-to-Speech (TTS). A three-author paper from NAVER LABS Europe, accepted to the EMNLP 2026 main conference, lays out a different route: instead of bolting on downstream rewriting modules, directly align the LLM to be speakable from the start ([arXiv:2609.01246](https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.01246)).\n\n## The problem: LLMs are trained for eyes, not ears\n\nCurrent LLMs are trained on textual data and preferences that target written output. The result is grammatically correct, helpful text that is full of TTS-unfriendly surface forms: symbols, abbreviations, raw URLs and compact numeric notation. The industry-standard fix is a text-normalization post-processing module between the LLM and the TTS engine, but the paper identifies three costs: post-processing adds noticeable latency; it cannot run until a complete sentence has been generated; and it is inherently tied to the tokenization and input format of each specific TTS system — swap the synthesis engine and the rewriting rules may simply fail.\n\n## The approach: make speakability a preference alignment target\n\nThe authors formalize TTS-friendly text generation as a preference alignment problem and build two preference datasets from scratch: CORA, a synthetic coffee-ordering assistant benchmark with 262 preference tuples targeting conversational responses, and Recipe, 300 tuples sampled from the RecipeNLG corpus (originally 2.2 million cooking recipes) targeting procedural descriptions. Each tuple pairs a TTS-friendly response with a TTS-unfriendly one, covering symbols, abbreviations, numeric shorthands and raw URLs.\n\nMethod-wise they adopt the FaST (Feature-aware Sampling and Tuning) framework: instead of a black-box reward model, an LLM automatically discovers a set of interpretable high-level features (such as \"numbers spelled out\", \"abbreviation density\", \"natural conversational tone\"), scores responses per feature, learns feature weights, and then aligns via sampling and tuning. On CORA, the largest negative-weight feature learned is exactly \"compact numeric formatting\" (-0.50), followed by abbreviation density (-0.32); the top positive weights are natural conversational tone and spelled-out numbers (+0.33 each).\n\n## Results: winning with 10 examples\n\nThe comparison spans five alignment baselines — prompting, SFT, DPO, GRPO and RFT — across two data regimes (10 vs. 100 training samples) and two model families (Qwen3-4B and SmolLM3-3B). FaST achieves the best TTS-friendliness\u002Fhelpfulness tradeoff even in the 10-sample low-data regime.\n\nThe contrast with two-step text normalization is telling. Post-processing zero-shot output with PolyNorm (Qwen3-4B, few-shot) does push CORA TTS-friendliness to 4.92, but per-response latency doubles from 1.6 to 3.4 seconds; FaST reaches 4.73 in a single step while scoring higher on helpfulness (4.86 vs 4.72). On Recipe the gap widens: PolyNorm 8.5 seconds vs FaST 4.4 seconds.\n\nA human listening study (MUSHRA protocol, 14 valid raters on Prolific) confirms the ranking: FaST scores 68.3, ahead of DPO (55.9) by 12.4 points (p=0.003) and Prompting (51.4) by 16.9 points (p=0.001). The team also validated that their cheap heuristic metric achieves a system-level Spearman correlation of -1.00 with the TTS-to-ASR round-trip metric on CORA — the five systems rank identically, meaning speakability can be assessed without actually running synthesis.\n\nOne illustrative example: asked \"How much is the banana bread slice?\", the zero-shot model answers \"$3.45\", while the FaST-aligned model says \"three point forty-five dollars\" — the number is spoken out.\n\n## So what\n\nThe implications go beyond voice assistants. First, optimizing text generation for a downstream consumption context can be done as lightweight preference alignment that starts working with 10 examples — practical for teams without large annotation budgets. Second, the heuristic metric correlates strongly with human listening judgments, and the authors explicitly note it can be used directly as a reward model — evaluation as reward, connecting to the verifiable-reward RL line of thinking. Third, the limitations are clear: experiments cover English and two domains only, models stop at 3B\u002F4B parameters, and the framework applies to cascade LLM-to-TTS architectures, not end-to-end speech LLMs.\n\nDatasets, metrics and code are open-sourced at github.com\u002Fnaver\u002Ftts-friendly-gen. As voice becomes a default interface for agents, \"writes well\" and \"speaks well\" are diverging into two distinct optimization targets — and this paper formally pulls the latter into the alignment agenda.","tts-friendly-llm-alignment-fast","2026-09-02T23:40:00Z","2026-09-02T23:13:02.501756Z","2026-09-02T23:13:02.501766Z",true,"agent",38,{"items":36},[37,42,47,52,57,62],{"id":38,"title":39,"news_slug":40,"published_at":41},"b1645fba-d364-47e6-97da-06868f98d987","Linux 内核 7.x 每版近 2000 个 CVE:AI 帮倒忙,维护者不堪重负","linux-kernel-cve-ai-overwhelmed","2026-09-04T00:00:00+00:00",{"id":43,"title":44,"news_slug":45,"published_at":46},"4a89fe5a-8703-49e5-b083-079cbda0fa2a","蒸馏也有副作用:中间训练期上KD,推理上涨、事实记忆反而变慢","switch-distillation-midtraining-kd","2026-09-02T17:10:00+00:00",{"id":48,"title":49,"news_slug":50,"published_at":51},"21a91da5-c5fa-45e2-b01f-a7950331cf44","S3 把 DuckDB 团队收走了:DuckLabs 加盟 AWS,MIT 开源照旧","aws-buys-ducklabs-duckdb-open-source","2026-08-30T06:00:00+00:00",{"id":53,"title":54,"news_slug":55,"published_at":56},"f9cf9f03-6aca-4d29-94d3-5c6acfeaf435","匿名模型 OX Alpha 短暂登顶 OpenRouter 编码榜:研究者推测底座指向智谱 GLM-5.x","ox-alpha-stealth-openrouter-glm-5-zhipu","2026-08-24T03:00:00+00:00",{"id":58,"title":59,"news_slug":60,"published_at":61},"89804e7d-cee8-4412-8b30-e43855911ef5","点赞与下载是两个经济体：Hugging Face 夏季报告拆穿开源模型的「追新幻觉」","hf-open-models-summer-likes-vs-downloads","2026-08-18T13:20:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"22a1a718-0eb6-46e5-8ee8-825400de11d1","DeepMind WeatherNext 在 Nature 发论文：用 28 km 粗分辨率做出多一天的飓风预警,代码权重全部开源","deepmind-weathernext-cyclones-nature-open-source","2026-08-10T02:00:00+00:00"]