Diarist 0.1.0-beta.1

Diarist 0.1.0-beta.1 is the first public release, as a beta.

Fully local, speaker-separated meeting transcripts for Windows 10/11 with an NVIDIA GPU. Record a meeting live (your microphone as "me", the system audio diarized into up to 8 speakers) or import an audio or video file. Everything is stored in a searchable SQLite database on your PC. Nothing leaves it.

Why beta. Measured on one machine (RTX 3080, 10 GB) and on the author's own meetings. Other GPUs, the Vulkan and CPU runtime packages and first-run on other PCs have not been tried by anyone else yet. Expect rough edges. If something fights you, open an issue and paste the output of Diagnostics, Copy diagnostics.

Install

Download Diarist_0.1.0-beta.1_x64-setup.exe below and run it. It installs per user, needs no administrator rights, and Windows SmartScreen warns once because the installer is not code-signed yet.

First run downloads the NeMo-Speech.cpp runtime (about 98 MB) and two models, Nemotron Speech Streaming EN 0.6B for speech recognition and Nemotron 3 Diarization for speaker separation (about 770 MB together), into your data folder. Every download shows its hash, folder, Retry and Copy link, so a browser plus a file copy always works. If you run a per-app firewall, allow diarist.exe when it asks.

docs/INSTALL.md covers runtime packages, VRAM tiers and models; docs/04-model-guide.md explains how to pick or add models; PRIVACY.md lists the only network calls the app ever makes.

SHA-256 of Diarist_0.1.0-beta.1_x64-setup.exe:

c42e34913eb52de0c6a51226128bab196323225b882284c22f3af8a104a6737c

What is in the box

• Import audio and video files (AAC, MP3, FLAC, WAV, Ogg directly; other formats through 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 and the system audio with text in about a second, level meters, automatic gain for quiet system audio, and a full post-pass when you stop so the stored transcript has full-context speaker labels. Names given during the meeting carry over.
• Crash recovery: an interrupted recording or post-pass is finished at the next start.
• Library: 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 are kept as lossless FLAC with a 30-day retention (pinned meetings exempt, transcripts never deleted).
• Setup and models, ComfyUI style: first-run screen with your GPU, disk and the right runtime package (CUDA, Vulkan or CPU); files dropped into models\ or engine\ are recognised on Refresh; Add from URL and Add from file for models the app does not list; Check for newer runtime lists NeMo-Speech.cpp releases. "Set up later" is always possible.
• Desktop: per-user installer whose uninstaller keeps your transcripts unless you tick otherwise, relocatable data folder, portable mode, tray icon with a recording indicator, Ctrl+Alt+R to record from any app, start with Windows, "Why isn't it working?" checklist with fixes, Copy diagnostics, and a manual Check for updates signed with the project's updater key.

Known limits

• English by default. Nemotron 3.5 ASR (40+ languages) can be downloaded and selected in Settings.
• At most 8 speakers per recording. The Nemotron models punctuate little.
• In-room and far-field audio are harder than remote calls; measured numbers are in docs/eval.
• Headphones are recommended for live recording until echo handling lands.
• The Vulkan and CPU runtime packages ship with the NVIDIA runtime but are untested by us.
• Not code-signed yet, and the NVIDIA runtime is unsigned too.

Updating

Settings, Check for updates reads updater/latest.json from this repository's main branch and installs signed builds only. Beta installs are offered the next beta.

Full details: CHANGELOG.md.