Diarist

Meeting transcripts with speaker labels, on your own PC. Free forever.

A meeting transcriber for Windows that runs entirely on your own PC. It records your microphone and the sound your computer plays (Teams, Zoom, Meet, a browser), or imports an audio or video file, and turns it into a transcript where every line says who spoke. Transcripts live in a local, searchable database. Nothing is uploaded: there is no account, no cloud, no telemetry, no paid tier and no licence key.

The speech and speaker models are NVIDIA's open Nemotron models, run by NVIDIA's NeMo-Speech.cpp runtime on your GPU. No Python, PyTorch or CUDA toolkit is needed; the GPU driver is the only prerequisite.

Library and search	A meeting with its speakers
Recording, with live text	Dark theme

What it does

• Imports recordings (m4a, mp3, wav, flac, ogg, mp4, mkv and more): about 30 times faster than real time on an RTX 3080 (measured over the whole import), so an hour takes about two minutes.
• Records live: your microphone is "me", the system audio is split into the other speakers (up to 8). Text appears within about a second; when you stop, the recording is processed again for the final transcript, with punctuation.
• Speakers: rename "Speaker 2" to "Alice" once and the name sticks, merge labels, recolour, fix a line that went to the wrong person.
• Search every transcript at once (accent-insensitive, any language).
• Export Markdown, plain text, JSON, SRT or VTT.
• Keyboard first: R records, Ctrl+K searches, Ctrl+Alt+R records from any app, ? lists every shortcut.

Quick start

1. Download Diarist_0.1.0-beta.5_x64-setup.exe from the releases page and run it. It installs for your user only (no administrator rights) into %LOCALAPPDATA%\Diarist.
2. The first start shows a setup screen: it checks your GPU, driver and free space, then Install engine (98 MB) and Install recommended models (1.0 GB): the diarizer, the EN model and the punctuation model. Every file is checked against its published SHA-256.
3. Press Ctrl+O to import a recording, or R to record a meeting.

Using a coding agent (Claude Code, Codex or similar)? Point it at docs/SETUP-WITH-AN-AGENT.md: it checks your PC, picks the package and models, installs and verifies.

You can skip the setup ("Set up later"): Import and Record tell you exactly what is missing when you use them, with a one-click fix. If a download does not work on your network, every item shows its address, hash and folder, so you can fetch it with a browser and drop it in. Details, firewalls, offline machines and uninstalling: docs/INSTALL.md.

The installer is not code-signed yet, so Windows SmartScreen asks before it runs ("More info", then "Run anyway"). The NVIDIA runtime it downloads is not signed either; Windows Smart App Control, when it is on, may block it.

Choosing your package and models

From docs/04-model-guide.md. Numbers marked measured were taken on an RTX 3080 10 GB; the rest are estimates and say so.

What gets downloaded

Component	What it does	Size	Licence
NeMo-Speech.cpp 0.2.0 (NVIDIA)	The inference runtime	98 MB zip, 134 MB extracted (CUDA build)	Apache-2.0
Nemotron 3 Diarization, q8_0	Who spoke when, up to 8 speakers	102 MB	OpenMDW-1.1
Nemotron Speech Streaming EN 0.6B, q8_0	Speech to text, English (default)	668 MB	NVIDIA Open Model License
Nemotron 3.5 ASR Streaming 0.6B, q8_0	Speech to text, 40+ languages (optional)	742 MB	OpenMDW-1.1
Parakeet TDT 0.6B v3, q8_0	Experimental only (see below)	681 MB	CC-BY-4.0
Horizon-Labs punctuation-restoration-small, int8 ONNX, with ONNX Runtime 1.31.0 (Microsoft)	Punctuation and sentence capitals after Stop and on imports, on the CPU; never changes a word	306 MB	Apache-2.0 as declared in the model card (base model MIT); ONNX Runtime MIT

The default download (engine, diarizer, EN model, punctuation) is about 1.2 GB; with every optional model, 2.5 GB.

Which runtime package for which PC

The setup screen detects your GPU and preselects the package; you can change it.

Your GPU	Package	Status
NVIDIA with a driver new enough for CUDA 13 (R580 or newer is the estimate; 610.88 measured working)	CUDA	Measured on an RTX 3080 10 GB; kernels for Turing, Ampere, Ada and Blackwell
AMD or Intel, or an older NVIDIA driver	Vulkan (19 MB)	Measured on an RTX 3080 only (AMD and Intel untested): 25 to 28 times real time for files; live works
No usable GPU	CPU (6 MB)	Measured 5 times real time on a 12-core Ryzen 9 5900X (an hour in about 12 minutes); live not recommended

How much GPU memory

Measured on the RTX 3080, not counting the desktop's own use (about 1.5 GB):

Scenario	VRAM
Live recording, two channels, Nemotron 3.5 and the diarizer	1.05 GB idle, 1.26 GB with two sessions, spikes of 0.5 to 0.6 GB per session; budget 2 GB (the EN model sits at 1.18 GB)
File transcription, Nemotron 3.5	1.82 GB peak (77-minute file)
File transcription, EN model	2.6 GB peak
Experimental Parakeet	4.9 GB peak
Live recording plus a file job	never at the same time: file jobs wait until recording stops

Estimated tiers (the app reads free VRAM before it starts, and the live engine unloads after 10 idle minutes):

Card	What works
4 GB	File mode with Nemotron 3.5 (turn on low-VRAM mode); live is borderline next to the desktop's own use
6 GB	Live and file mode comfortably; Parakeet only with nothing else on the GPU
8 GB and up	Everything

Which speech model

• English meetings: the EN model (default). It keeps more of what was said: cpWER 4.7 to 9.5 on four English Zoom meetings against 5.1 to 10.3 for Nemotron 3.5, and 18.0 against 20.7 on the AMI meeting (measured). Speaker attribution is about the same. It drops quiet speech, though: for a soft-spoken participant or a file recorded at a low level, use Nemotron 3.5.
• Other or mixed languages, or a 4 GB card: Nemotron 3.5. 40+ languages with automatic detection, 30 percent less VRAM. It drops about 40 percent of filler words ("um", "uh").
• Parakeet TDT v3 is experimental and off by default. It punctuates well and is about 0.6 WER points better, but it silently skipped a 46-second span of a 77-minute recording (measured), and no chunk size avoids it. Use it only for short clips you check yourself.

Neither Nemotron model punctuates much, so since 0.1.0-beta.5 a small punctuation model adds . , ? and sentence capitals after Stop and on imports, on the CPU (about 5 to 6 seconds per hour of audio), without changing a word; lines end at sentence ends where the pause allows. About two in three of its periods and question marks are where a person would put them (docs/04-model-guide.md, section 8). On clean meeting audio expect about one word in 15 to 20 to be wrong or missing; on far-field room recordings about one in 9.

Speakers: limits worth knowing

• Speaker diarization is Nemotron 3 Diarization, the best system on the only independent cross-system benchmark (DER 14.7 against 19 to 31 for the alternatives). On our own Zoom meetings 0.2 to 0.3 percent of words went to the wrong speaker, 5.5 to 7.8 percent within a second of a speaker change with Nemotron 3.5 and 5.6 to 12.7 with the EN model (measured).
• At most 8 speakers per recording. A ninth person is merged into an existing label; split very long multi-meeting recordings.
• Speaker labels in the live view are provisional; the processing after Stop produces the stored transcript, and names you gave during the meeting carry over.
• Heavy overlap, far-field microphones and noisy rooms degrade the result.

Other requirements

• Windows 10 (1803 or newer) or Windows 11, 64-bit. WebView2 is preinstalled on these; the installer fetches it otherwise.
• About 1.3 GB of disk for the engine and the default models; recordings use about 90 MB per hour (lossless FLAC), kept 30 days by default (transcripts are kept forever).
• RAM: the engine uses about 300 MB (measured); 8 GB in total is comfortable.

You own the folders

Like ComfyUI, models and runtimes are plain files in plain folders under the data folder (%LOCALAPPDATA%\app.diarist.desktop unless you move it in Settings):

• Drop any GGUF the runtime supports anywhere under models\ and press Refresh. The app reads its header to know what it is. Settings, Models also has Add from URL and Add from file.
• Drop a newer runtime (the zip or the extracted folder) into engine\, or use Check for newer runtime in Settings, which lists NVIDIA's releases.
• Diagnostics has a Why isn't it working? checklist with a fix per row, and Copy diagnostics for a bug report.

Privacy

Everything stays on your computer: audio, transcripts, names, settings. The app makes network requests only when you ask for a download (the runtime from GitHub, the models from Hugging Face), when you press Check for newer runtime (the GitHub API), or when you press Check for updates. It never checks anything by itself and sends nothing about you. A crash leaves a report in the data folder's logs folder that is never sent anywhere and is deleted with the data folder. See PRIVACY.md.

Status

Version 0.1.0-beta.5, the fifth beta pre-release: file import and live recording work end to end on the development machine (Windows 11, RTX 3080). The Vulkan and CPU runtime packages were tested on that machine's NVIDIA card and CPU only, and machines other than the author's are untested. Next on the roadmap: better speaker-change handling, and recognising the same person across meetings (docs/plans/00-roadmap.md). Changes: CHANGELOG.md.

Building from source

Rust, bun and the VS Build Tools: see docs/DEV.md and CONTRIBUTING.md. Design: docs/01-architecture.md.

Licence

MIT for the code in this repository (LICENSE). The runtime (Apache-2.0), ONNX Runtime (MIT) and the models (OpenMDW-1.1, NVIDIA Open Model License, CC-BY-4.0, Apache-2.0) are not part of the installer; you download them at first run under their own licences. The Diarist name, logo and icons are not MIT: use them to refer to Diarist or to package the official builds, not on forks or other products (branding/README.md). Third-party notices: LICENSES.md.