Tag
CineSubBench is a benchmark for evaluating LLMs on long-form narrative and cultural understanding from multilingual movie subtitles, with tasks covering narrative reconstruction, genre prediction, and safety across six languages.
Lattice is a writing assistant tool that reshapes text across multiple languages to create fresh expressions while preserving the original meaning.
An open-source voice AI agent with a live avatar can see, talk, think, and draw in real-time, and supports multilingual interactions, such as switching to Hindi mid-call.
Open-source System 1 AI models are rapidly emerging just 7 days after Jev's release, gaining over 15,000 GitHub stars and topping Hugging Face trending, showcasing the fast pace of open-source AI development.
BRIDGE ASR 2.0 is a benchmark that evaluates speech recognition models on real two-person conversations in 21 languages, focusing on code-switching and multilingual performance to enhance voice interaction in AI and robotics.
YODAS v3 is a weakly-labeled speech corpus containing over 1.1 million hours of high-bandwidth, stereophonic, multilingual audio in 147 languages, designed to advance speech foundation models.
This paper examines script knowledge distribution in LLM layers, revealing that script recognition happens early while commitment to output script emerges in late layers, highlighting the importance of model depth for multilingual architectures.
Google introduces Gemini 3.8 Live with Live Avatar, an update that integrates real-time visual avatars into conversational AI, enhancing interactions with features like lip-syncing, multilingual support, and asynchronous tool execution for enterprise applications.
This paper introduces a generation benchmark for evaluating large language models on culturally specific kinship terms in Hindi, Tamil, and Korean, revealing significant gaps between recognition and generation capabilities.
Speechka is a real-time voice translation tool that mimics the user's voice across 44 languages, available on macOS, Windows, and browsers.
The paper proposes MMSafe, a multi-layer framework for identifying and filtering safety-degrading data in multilingual large language models, reducing harmful responses by 60% compared to random filtering.
Onepin 是一款 AI 语音合成辅助工具，接入 ElevenLabs、OpenAI、Google 等 30 多种现有语音引擎，通过 400 万词发音词典校正人名、价格和日期的读音，并支持逐句评分与无需重渲染的修正，面向多语言视频配音和课程制作团队，已在 Product Hunt 上线。
The number one trending model on Hugging Face is an open-source multilingual system 1 decision model, underscoring the vibrant open-source AI community.
GLiNER2.5 is a multilingual boundary checkpoint for unified schema-based information extraction, supporting entities, classification, structured records, relations, and span attributes in one model.
This paper introduces μ²-Bench, a benchmark for evaluating multilingual machine unlearning in large language models, aiming to ensure that undesired information is effectively removed across diverse languages.
Grok Voice Transcribe 2.0 is a new speech-to-text model from x.ai that doubles the accuracy of its predecessor and excels in multilingual and real-world audio transcription, ranking first on public leaderboards.
Qwen3.8-LiveTranslate is an AI model that reduces translation lag and improves quality using an interleaved audio-text architecture, featuring real-time speaker separation and voice cloning in 60 languages.
gg-friggin-ez is an open-source Node.js tool for fast, multilingual profanity and toxicity screening, supporting languages like Hindi and Bengali with evasion-aware features.
Laya is an open-source, fast multilingual decision engine that offers non-autoregressive, calibrated probabilities for structured schemas, claiming to be 6-8 times faster than Jev with full openness.
Laya Multilingual is a multilingual AI decision model that provides typed answers with probabilities across 100+ languages in a single forward pass, improving accuracy over English-only models.