Natural Language Processing news and analysis on Web Pulse: 5409 curated articles tracking the latest Natural Language Processing developments, tools, and research, updated continuously from vetted sources.
Johan, a developer running the multilingual Magento 2 plant store Introgreen, built two custom AI tools: Plantcoach, a context-aware RAG assistant that retrieves advice from the store's own plant knowledge library and ad…
GLM-4.7, a language model formerly offered by Kagi, intermittently inserted Chinese ideograms into otherwise French-language output, according to user reports in a thread where one user said they could not reproduce the …
A civil servant at Seoul's Gwangjin-gu District Office built kordoc, an MIT-licensed TypeScript library, CLI and MCP server that converts Korean HWP/HWPX government documents into Markdown and structured data for LLM pip…
A developer-authored writeup describes how AI menu analysis, using natural language processing and computer vision, can parse restaurant menu text to estimate calories, flag allergens, and rank dishes against diner prefe…
Developer samkhawase released voxlocal, a fully local, low-latency voice agent for macOS written in Rust, on 9 Oct 2026. The CLI tool targets a narrow automotive-service use case in English and combines off-the-shelf Whi…
A developer benchmarked 13 common OCR preprocessing tricks against 10 screenshots with a known answer key, finding that none improved character error rate over the untouched original on clean images and several made resu…
A developer built Pocket Field Trip, a local-first app that uses the quantized ONNX conversion of the paraphrase-MiniLM-L3-v2 sentence embedding model to match a user's own wording to one of eight short outdoor activitie…
A developer walkthrough shows how to assemble a full voice agent — speech-to-text, LLM, and text-to-speech — using three OpenAI-compatible API calls on EcoHash with a single key and base URL. The pipeline pairs whisper-l…
Hikari Webworks detailed a streaming architecture that achieves sub-500ms end-to-end voice latency for AI phone agents by running the entire pipeline over full-duplex WebSockets rather than sequential HTTP request-respon…
A developer found that the bge-reranker-base cross-encoder silently truncates passages at 512 tokens, cutting the tail of 31% of their RAG chunks and causing the reranker to demote correct answers that retrieval had alre…
A developer argues that AI script-writing tools lose track of characters because they treat screenplays as isolated prompts rather than evolving sources of story facts, and proposes structuring continuity into five memor…
A developer built VirtualMentor, an open-source, audio-first AI interview coach designed for hands-free practice during outdoor walks. The project, hosted on GitHub, uses a local/self-hosted open-weight pipeline to keep …
A developer released jarvis-assistant-vocal, an open-source French-language voice assistant that runs fully offline on local LLMs via Ollama or connects to Claude for hybrid setups. The modular Python project combines na…
A developer published mms-forced-align, an npm package that runs Meta's MMS forced aligner in Node.js via ONNX Runtime with a from-scratch TypeScript CTC Viterbi decoder of roughly 170 lines. The package reports being ab…
A solo iOS developer built WhisperDirect, a one-time-purchase transcription app that has no backend and instead has users supply their own OpenAI API key, paying roughly $0.006 per minute of Whisper API transcription. Ov…
A developer walkthrough explains how neural text-to-speech pipelines work, from text normalization and acoustic modeling to vocoding, and demonstrates calling the ElevenLabs REST API to synthesize speech from Python and …
A developer published a step-by-step guide for integrating ElevenLabs' text-to-speech API into a Next.js application, using a server-side API route to keep the ElevenLabs API key hidden from the browser. The walkthrough …
Tiny Audio, a speech-to-text system from developer mazesmazes, connects a frozen pretrained speech encoder to a pretrained LLM via a small trainable projector and trains only about 80M parameters for roughly $25, achievi…
A developer published a set of ten practical tips for producing more natural-sounding text-to-speech output, covering voice selection, prosody controls such as speed and volume, SSML markup for emphasis and pauses, pitch…
Cactus Compute released Whistle on October 2, a 16.9 MB on-device speech recognition model that runs entirely on CPU with zero dependencies and, paired with Cactus's Needle runtime, converts audio directly into structure…