Changelog

All notable changes to Diarist. Versions follow semantic versioning.

[Unreleased]

The first public release, as a beta pre-release.

The project was called meet-assist until 15 Oct 2026 and is now Diarist: app diarist.exe (installer Diarist_<version>_x64-setup.exe), data folder %LOCALAPPDATA%\app.diarist.desktop, database diarist.db, registry key HKCU\Software\Diarist\DataRoot, CLI diarist-cli, crates diarist-core and diarist-cli, environment variables DIARIST_* (MEET_ASSIST_CIMC_WAV is now DIARIST_LONGFILE_WAV), repository ebt55/diarist.
The first start under the new name offers once to move existing meet-assist data (%LOCALAPPDATA%\app.meetassist.desktop, or the folder named by HKCU\Software\meet-assist\DataRoot) into Diarist's data folder, with the same copy, check and cleanup as Move data folder, or to start fresh.

Import audio and video files (AAC, MP3, FLAC, WAV, Ogg/Vorbis directly; other formats through Windows Media Foundation or ffmpeg) and transcribe them with speaker labels on the GPU, about 27 times faster than real time on an RTX 3080.
Live recording of the microphone ("me") and the system audio (diarized into up to 8 speakers), with text in about a second, level meters, automatic gain for quiet system audio, and a post-pass over the recorded channels when you stop. Names given during the meeting carry over.
Crash recovery: an interrupted recording or post-pass is finished at the next start; an interrupted import is marked failed; a notice says what was recovered.

Full-text search across all transcripts, speaker rename, merge, recolour, reassign and split, playback with a waveform, export to Markdown, text, JSON, SRT and VTT.
Recordings kept as lossless FLAC; audio retention (30 days by default, pinned meetings exempt, transcripts never deleted), at launch and daily, with the reclaimable space in Settings; diarist-cli retention.

First-run screen: GPU (name, VRAM, driver, CUDA version), disk, the runtime package for this GPU (CUDA, Vulkan or CPU) and the recommended models, each with progress, Retry (resumes), Copy link, SHA-256, size and folder. "Set up later" is always possible; Import and Record say what they miss.
Files put into models\ or engine\ by hand are recognised on Refresh (hash-checked; runtime zips extracted). Add from URL and Add from file for models the app does not list; models.user.json for extra recommended entries; unrecognised files can be used with a warning.
Check for newer runtime lists NeMo-Speech.cpp releases from GitHub and installs a package next to the others.

Per-user NSIS installer (no administrator rights); the uninstaller keeps transcripts unless "Delete the application data" is ticked, and otherwise offers to remove the engine and models.
Data folder relocatable from Settings (copy, verify, restart, clean up); portable mode with portable.txt.
Tray icon with a recording indicator; Ctrl+Alt+R records from any app and asks before stopping; one instance at a time; start with Windows; window size and position remembered.
The live engine loads on the first recording and unloads after 10 idle minutes; low-VRAM mode for imports.
Diagnostics: "Why isn't it working?" checklist with fixes, Copy diagnostics, the app log and the engine's last errors.
Check for updates (manual), signed with the project's updater key.

Not code-signed yet (SmartScreen warns once); the NVIDIA runtime is unsigned too.
Vulkan and CPU runtime packages are untested; at most 8 speakers per recording; the Nemotron models punctuate little.