Jotter is a local recording and transcription tool built with Rust, designed for on-device processing without requiring external services.
Two PhD candidates at the University of Pennsylvania explored handwritten text recognition (HTR) technology to transcribe premodern Italian and Devanāgarī manuscripts from Penn's collections during a 2025-2026 fellowship. Using the open-source platform eScriptorium and the Kraken ATR system, they created ground truth data through manual transcription to train models, investigating how HTR might streamline manuscript work and enable new research approaches across different scripts.
The Socioscope project conducts large-scale qualitative research on food systems using AI to analyze hundreds of multimedia case studies across 31 countries. The paper documents the complete data pipeline infrastructure, including sampling, interviewing, transcription, quality control, and ethical compliance, reporting 686 cases with 1,430 hours of recordings and 12.6 million words of transcript from 2023-2026.
A collection of free browser-based tools that operate privately without requiring signup or file uploads, including background removal, PDF editing, audio/video transcription, code generation, and image conversion.
Notula for Google Meet is a browser extension that provides real-time transcription of Google Meet calls with speaker identification, saving transcripts as Markdown files to local folders. The tool leverages Google Meet's built-in captions, stores data locally without sending audio to external servers, and integrates with note-taking workflows by saving meetings to version-controlled repositories.
Next Notes is a macOS dictation and meeting notes app that uses push-to-talk for on-device transcription, automatically records calendar meetings and calls with speaker attribution, and generates meeting summaries and actions using local LLMs. The app integrates with Apple Calendar and Google Calendar, runs entirely on-device except for model downloads, and supports both Apple and Parakeet transcription backends.
Nari launches Qwen3-ASR, a speech-to-text model ranked in the top 5 for accuracy on Coval Benchmarks with sub-50ms latency and the lowest pricing at $0.06 per hour. The company entered public beta today with free limited-time access to its Fast and Standard endpoints for real-time transcription.
Epilude introduces EP-1, an on-device AI model that enables meeting transcription and summarization entirely on Mac without cloud processing. While EP-1 achieves 62% accuracy compared to cloud mode's 92%, it offers privacy-conscious users local alternatives for dictation and meeting notes.
SpeechText.AI offers a free, browser-based Bates numbering tool for PDFs that processes documents entirely locally without uploading to servers, addressing privacy concerns for legal professionals handling confidential materials under HIPAA and GDPR. Users can customize prefixes, sequential numbering, fonts, and alignment for court-ready discovery files and evidence exhibits.
YuE2 is a frontier music generation model achieving state-of-the-art results on SongBench with a score of 6.9632, outperforming Suno v5. The system includes MERT2 for music understanding, SheetSage2 for transcription across six tasks, and supports agentic music editing through conversational interaction.
Apple is introducing Live Rewind on Apple Watch Series 12 and Ultra 4, a feature that transcribes the last 15 seconds of audio when activated via Digital Crown press. The company has implemented safeguards including an audible chime, visual indicators, and on-device processing in a Secure Exclave to prevent secret activation and address privacy concerns.