speech

Jack Clark delivers the Per Jacobsson lecture at the IMF meeting in Bangkok, arguing that convincing evidence shows the AI singularity has begun and urging economists and policymakers to prepare for rapid change while ensuring humanity shares in AI's benefits.

论文提出"主动性向量"（initiative vector），从全双工语音语言模型自身的残差流中无需训练即可提取的单一方向，用于读取和调控模型的抢话/让话决策，并通过参考录音实现说话人轮替风格的迁移。

本研究首次探索从韵律模式反向预测文本：通过仅使用最低 12 个 Mel 频带（约 450Hz 截止的低通滤波）微调 Whisper 模型，实现了 36% 的词错误率，表明低频语音特征与词汇内容的关联远强于此前认知。

The paper presents DuplexAgent-RSI, a full-duplex voice agent collaboration system whose coordination harness is modular and improved via a closed-loop recursive revision process driven by simulated interaction traces, achieving stronger spoken-knowledge and tool-use benchmark scores than comparable delegated systems.

本文研究流式说话人分离模型（Streaming Sortformer v2）在小数据适配时出现的'看似提升实则遗忘'现象，揭示聚合指标（DER）掩盖的身份一致性退化，并探讨rehearsal在恢复与跨域迁移之间的权衡。

Veris AI introduces VAmoS Energy, a harder voice-agent benchmark combining multi-request utility billing calls, stateful Stripe/Fineract twins, caller verification gates, and background speech — finding completion rates of 17.3-44.7% across 14 voice stacks, with Grok Voice leading and background TV slashing pooled completion from 38.7% to 8.6%.

A tweet recommends a speech for software engineers and all people, emphasizing keeping happy memories, facing reality, and being excited about the future.

A speech by the Jev founder, a former OpenAI researcher, presents Jev as the next phase for LLMs, claiming it is 200 times faster, 400 times cheaper, and hallucination-free compared to current methods like RLHF.

Ex-Omni-2D is an omni-modal dialogue framework that generates coordinated text, speech, and reference-conditioned video responses via a visual thought plan and a distilled streaming video generator, achieving a practical quality-efficiency trade-off.

LISA is a compact, interpretable MEG decoder that retrieves perceived speech from non-invasive brain recordings with 39.75% Top-1 accuracy among 1,005 candidates using ~20x fewer parameters, while revealing cortical sources and speech features.

ByteDance announced SeedRealtime, a native audio-visual model that can process continuous video, audio, and text while speaking in real time.

NVIDIA released NemotronLabs VoiceChat 11B, an open end-to-end full-duplex speech model enabling real-time conversational AI with ~450ms turn-taking latency, barge-in, and live tool calling, the first open full-duplex model to support tool calling.

This paper presents a cross-lingual analysis of entrainment behavior in Mandarin-English, Hindi-English, and Spanish-English code-switched dialogues, finding that lexical entrainment generalizes but acoustic-prosodic and style entrainment vary. It also evaluates how classification models capture these behaviors, showing they prioritize different features than humans.

GigaChat Audio 10B is an audio-native LLM built on GigaChat 3.1 Lightning, integrating a Conformer speech encoder and modality adapter for audio question answering, temporal grounding, and tool use.

Heard is a product that gives Claude Code and Codex a voice, enabling voice interaction with these AI coding assistants.

NVIDIA released Nemotron-Labs-Audex-30B-A3B, a unified audio-text LLM built on a 30B MoE backbone with 3B activated parameters, offering strong performance on audio understanding, speech recognition/translation, and generation while preserving text reasoning and alignment capabilities.

This paper proposes a speaker-disentangled syllabic tokenizer that uses regression of perturbed student representations toward clean teacher targets, achieving state-of-the-art syllable boundary detection and a 7% relative improvement in speech language model understanding over SpiRit-LM.

This paper introduces CLeaD, a supervised contrastive alignment framework for cross-lingual depression detection from speech using WavLM embeddings. It reveals that previous results were inflated due to speaker identity leakage and achieves modest improvements on Mandarin speakers.

FlexiSLM introduces dynamic frame rate capabilities for speech input and output in spoken language models, outperforming fixed-frame-rate models and enabling controllable inference speed.

Open-sourced Humanize PPT, an HTML PPT skill designed for presentations. It reorganizes outlines using AST (Audience, State, Transfer) logic, combined with downstream rendering skills to produce PPTs truly usable for presentations, addressing issues like uneven information distribution and content emptiness.