Natural Language Processing news and analysis on Web Pulse: 5422 curated articles tracking the latest Natural Language Processing developments, tools, and research, updated continuously from vetted sources.
A developer has published a practical guide to choosing audio formats and quality settings for voice AI pipelines, comparing MP3, AAC, WAV, FLAC, OGG Vorbis and Opus across bitrate, sample rate and use case. The writeup …
A voice agent named Alex negotiates freight rates over the phone using a pipeline of six small models built on LiveKit Agents, an open source framework, rather than a speech-to-speech model, because every price it speaks…
Aleph Alpha released Kolibri on October 3, a German-English mixture-of-experts model with 78.1 billion parameters, 3.46 billion active per token, and about 78 GB of FP8 weight memory under an Apache 2.0 license, accordin…
A developer built an internal meeting-notes utility called Meeting Minutes that uses WebAssembly and browser APIs to run transcription and summarization entirely on-device, keeping audio from ever leaving the browser tab…
A developer built a Digital Udhar Record System for a shopkeeper friend, digitizing handwritten credit notes into a searchable, secure record-keeping app. The project uses a Flask and SQLite backend with an HTML/CSS fron…
TypeSafe AI launched Jev on September 15, 2026, introducing "System One models" that take a state and a set of options and return a probability for each option in a single forward pass, with the same model able to make d…
A developer outlines System 2 Attention (S2A) prompting, a two-stage technique that first distills a raw user prompt by stripping emotional language, filler, and personal bias before sending the cleaned task to a large l…
A developer built Investment Memory, a personal investment decision journal for his father that records the reasoning behind each investment rather than generating buy or sell recommendations. The app accepts text or voi…
A developer walkthrough highlights eight underused features of the ElevenLabs voice API, including low-latency voice cloning, granular prosody controls, multilingual code-switching, and client-side voice activity detecti…
Cactus Compute released Whistle on October 2nd, a 16.9MB open speech-to-text model that runs locally on a CPU inside the company's existing Needle runtime. Co-founder and CTO Henry Ndubuaku led the work, which supports E…
A new arXiv paper (2610.00083v1) reports that large language models assign numerical uncertainty levels to verbal markers such as "possible" and "likely" that differ substantially from human judgments, based on a benchma…
A controlled ablation on character-level language modelling with a 6-layer GPT on TinyShakespeare found that the optimal single-scale bandwidth for FourierQK bandpass-filtered attention is sigma ~= 2 bins centred at para…
An autonomous AI agent running on a Seoul-based solo founder's PC built Reply Buddy, a single-file, standard-library-only Python tool that uses a local Ollama model (gemma3:4b) to summarize English emails in Korean, draf…
Microsoft launched MAI-Transcribe-2-Streaming on October 1, a WebSocket-based streaming speech-to-text model that ranked first for both Final Transcript and First Partial Transcript accuracy across 38 models on Artificia…
A developer compared OpenAI's tts-1 text-to-speech API against ElevenLabs for voice-enabled applications, citing ElevenLabs' sub-second streaming latency, studio-grade voice cloning, and custom voice ownership as reasons…
A developer published a tutorial showing how to build an AI voice assistant using the ElevenLabs text-to-speech API and Node.js, with a minimal Express server that accepts text and returns base64-encoded MP3 audio. The w…
Microsoft unveiled MAI-Transcribe-2-Streaming on Thursday, its first streaming transcription model, which delivers real-time transcripts in 60 languages with text appearing as soon as 100 milliseconds after audio is rece…
A review of recent research finds that AI models can detect AI-generated news text with high accuracy in narrow tests — a BERT-based classifier trained on 3,600 Turkish articles reached a 97.08% F1 score — but perform fa…
Cactus Compute released Whistle, a 16.9 MB on-device speech recognition model that runs on CPU with no dependencies and loads into the same C++ engine as Needle. Whistle handles transcription in seven languages, word tim…
A technology and human rights researcher ran three AI agents — Meta's Muse, Anthropic's Claude Cowork Opus 5.5 Medium, and OpenAI's GPT 6.1 Sol Medium — through the same World Bank Global Public Procurement Database task…