Natural Language Processing news and analysis on Web Pulse: 5439 curated articles tracking the latest Natural Language Processing developments, tools, and research, updated continuously from vetted sources.
A study of 12 post-training chains using supervised fine-tuning (SFT) and reinforcement learning (RL) found that removing the diagonal component of weight updates — the part that reshapes singular values — usually preser…
A training-free candidate-compression framework called Toollery improves LLM agent skill and tool selection across libraries scaling to roughly 79,000 capabilities, according to a new arXiv paper (2609.22218v1). Toollery…
OpenBMB released VoxCPM2, a 2-billion-parameter, Apache-2.0 licensed text-to-speech model that clones voices from a few seconds of audio, generates new voices from text descriptions, and outputs 48 kHz speech in 30 langu…
A maintainer of Oathra, an open-source runtime for AI phone-call agents, documented five ways voice agents falsely report a successful restaurant booking, including a case where a clerk's later "fully booked" remark fail…
The US government has authorized its first dedicated voice AI vendor for agency use under the new FedRAMP 20x automation framework, opening federal phone lines and citizen support channels to voice AI platforms. The auth…
A developer built a fully local, browser-based sleep monitoring system that uses Whisper.cpp compiled to WebAssembly to filter out speech for privacy and a TensorFlow.js CNN to classify Mel Spectrograms of breathing soun…
A developer set up self-hosted, frontier-level dictation on an iPhone keyboard by running Alibaba's Qwen3-ASR 1.7B model through the open-source Diction gateway and Agent CLI on a home RTX 3090, after the free tier of cl…
A developer compared five approaches to a banking-support classification task using the BANKING77 dataset: a TF-IDF lexical model, a small sentence transformer (MiniLM), a general-purpose language model run through DSPy,…
SageOx is building a real-time AI coworker on Media over QUIC (MoQ) that delivers context to teams within 1–3 seconds during live conversations, with Zoom Scribe handling transcription, according to the company's account…
Smallest.ai operates a real-time voice AI platform offering text-to-speech, speech-to-text, speech-to-speech, voice cloning, and a voice-agent platform built on small, specialized models for low latency, and the company …
Telli is an AI voice-agent platform for outbound and inbound phone calls that sell, support, and handle customer operations for B2C brands, according to the company's own site at telli.com. telli states it raised a seed …
A developer reports that many apparent AI agent reasoning failures are actually retrieval bugs, citing Anthropic's Contextual Retrieval research showing a 49% reduction in retrieval misses. The engineer argues that sessi…
Spirit AI co-founder and chief scientist Gao Yang told Reuters that humanoid robots could reach a "ChatGPT moment" by mid-2027, when machines understand natural-language instructions and attempt sequences of physical act…
Generative AI can bridge the gap between information access and comprehension by turning dense data — sales numbers, marketing results, expenses, trends — into plain-language narratives, according to Arria, which six yea…
Google's Gemini Notebook now offers a configurable Output Language setting, letting users pick a preferred language for chat responses, study guides, documentation, Audio Overviews and Video Overviews. The feature suppor…
The Transformer neural network architecture, introduced in the 2017 paper "Attention is All You Need," has become the go-to deep learning architecture for text-generative models including OpenAI's GPT, Meta's Llama, and …
Google is rolling out real-time voice conversations in Gemini Notebook for mobile users, supporting nearly 100 languages and grounding responses in a user's uploaded sources. The feature begins with Google AI Ultra subsc…
Rijul, a developer building the AI code review tool LiveReview, explains why retrieval-augmented generation systems fail to surface information that is present in a source document. The core issue, he writes, is chunking…
Modulate's ML team released the Entity Transcription Benchmark on Hugging Face, a free public dataset measuring per-entity transcription accuracy rather than word error rate. The benchmark contains 2,151 clips, six hours…
EXL launched a Medical intelligent document processing (IDP) solution built on AWS that combines its Xtrakto.AI document-extraction application with the EXL Insurance LLM to cut the more than 100 minutes per case insuran…