multilingual

Show HN: Wonderhop is a free daily word game with four modes per day where players guess the most common, rarest, oldest, or most recent answer to shared questions within 30 seconds.

LifeSearch introduces a benchmark of 203 Chinese and English tasks for evaluating deep research agents in real-world life and productivity settings, with over 4,000 expert-written rubrics and experiments on 12 advanced agents, finding that even the best frontier agent (GPT-5.6 Sol) scores only 63.4%.

Fish Audio 推出了 Drama 3 语音模型，支持音频修复（audio inpainting），可以只重新生成音频中的单个词或短语，其余部分保持不变，并支持 83+ 种语言。

论文提出MixLID，一个支持111种语言-文字组合的token级混合语言识别模型，采用两阶段训练（先单语再混合语言数据），在单语和混合语言识别基准上均超越现有基线，模型已在Hugging Face开源。

论文提出 HexaBee-4B，一个基于 Gemma3 4BIT 通过三阶段训练（维基百科持续预训练、多语言问答与答案翻译联合训练、双向翻译专业化）改编的六语种翻译模型，在阿拉伯语、中文、英语、希伯来语、波斯语和西班牙语全部 30 个翻译方向上相比基线均有 chrF++ 提升。

Cactus Compute releases Whistle: a 16.9 MB speech recognition model that runs on pure CPU, supporting seven European languages. Measured first-token latency of 5-8 ms and decoding speed of ~1300 tokens/sec. However, for unsupported languages (such as Chinese and Japanese), it does not throw errors — instead it silently outputs fake words, posing a risk of silent data corruption in production environments.

本文研究了端到端语音解析架构（Wav2tree）中中间神经网络单元的作用，提出了一种移除这些单元的简化架构，在减少12%参数的同时在ASR和解析任务上达到相当或更优的性能，并对法语、斯洛文尼亚语和Naija语进行了综合评估。

This paper introduces Local Prototype Reconstruction (LPR), a lightweight training regularizer that improves the transferability of bridge modules connecting frozen speech encoders to frozen LLMs, showing gains in multilingual ASR and speech translation by enforcing local lexical manifold compatibility.

Invited 是一款 AI 提醒扫描工具，可将账单、邀请函、支票、截图和 PDF 转换为智能提醒，自动提取日期、时间、地点和金额，支持 80 多种语言和 9 种日历。

该论文提出三阶段流程识别多语言大模型中的"Bridge Routing Heads"（BRH），发现五种语言对应的头集合几乎完全互斥，证明多语言多跳推理由语言特异电路实现；消融实验显示因果效应，并通过激活级干预无需训练即可挽救高达51.7%的跨语言推理失败。

该论文提出 Language-Aware Skeleton Exploration Framework (LASEF)，系统研究多层数学推理中骨架提示语言的选择问题，发现英语骨架平均略有优势但并非普遍最优，骨架语言是一个依赖上下文的设计变量。

提出一种通过LLM生成方言文本并用标准语TTS合成伪方言语音的零真实方言数据增强方法，并引入中间标准文本预测以缩小语义差距，提升语音语言模型在日语、德语和中文方言上的翻译表现。

An empirical study of adapting adult ASR models to child speech across Arabic and English while retaining adult performance, comparing fine-tuning, LoRA, and weight-space merging (LERP/TIES), with findings that bilingual adaptation and merging improve the adaptation-retention trade-off.

The paper introduces Tokka-Bench, an open-source framework for evaluating tokenizers across 100 natural languages and 20 programming languages using five metrics, finding that vocabulary allocation strategy matters more than raw vocabulary size among seven BPE tokenizers including GPT-4, Llama 3.1, Gemma 3, Qwen3, and Kimi K2.

Technology Innovation Institute (TII) introduces Falcon-ASR, a 1.6B parameter speech recognition model focused on Arabic, including the Emirati dialect, with support for English, French, Spanish and Portuguese. It reports a 20.92% average Arabic WER, outperforming the best published leaderboard result, and offers word-level timestamps.

An essay arguing that AI, particularly LLMs, has quietly achieved near-perfect machine translation, effectively cracking the language barrier after decades of mediocre progress.

This paper extends prior-fitted networks (TabPFN) to natural language by training a 300M-parameter byte-level transformer entirely on synthetic sequences drawn from randomly sampled recurrent causal models — a 'prior over languages'. With frozen weights, the model learns to predict real languages in context across six languages, and can also count, compare numbers, add approximately, and predict deterministic sequences like primes and Kolakoski, showing in-context language learning can emerge from a synthetic non-linguistic prior.

A practitioner shares how their ASR pipeline failed on code-switching calls because language detection ran once at the start, mangling the Hindi portion of English-Hindi utterances; they now run continuous detection and route unstable chunks to a review queue, exposing how benchmark numbers diverged from real-world caller behavior.

This paper investigates whether forcing LLMs to reason in the same language as the query and retrieved documents improves monolingual non-English RAG performance. Using a German RAG testbed built over the fictional world of The Dark Eye, the authors find that aligning the reasoning language with the evidence helps (German reasoning beats French despite the model's stronger French proficiency), but native unconstrained English reasoning still matches the aligned setting, highlighting the need for truly multilingual reasoning models.

Introduces the Multilingual Distractor Interference (MDI) protocol showing that foreign-language distractor sentences cause LLMs to switch output scripts (e.g., Devanagari) or trigger abstention, inflating exact-match hallucination metrics — Llama-3.1-8B's apparent 0.710 hallucination rate drops to 0.470 once script-switched but semantically correct answers are accounted for.