English | Русский

BuzzMini

Hold a hotkey, speak, release — the text is pasted where the cursor is.

Offline push-to-talk for Windows, Linux, and macOS. No cloud, no subscription. Speech is recognized on this PC with faster-whisper.

Download for Windows (Setup.exe)

Why BuzzMini?

Vs Windows dictation	Stronger on Russian, no Microsoft cloud services, works without an online account.
Vs original Buzz	One job: dictate into any window. Lives in the tray, no extra workspace of windows.
Privacy	Audio never leaves the machine — not Yandex, Google, or OpenAI.

Default chord: Left Ctrl + Space. Change it in Settings. NVIDIA GPU (CUDA) is used when available; otherwise CPU.

This is speech-to-text, not a chat LLM. One Whisper model is loaded and turns audio into text.

Install (Windows)

1. Open Releases and download BuzzMini-Setup-*.exe.
2. Run the installer. The app starts in the system tray.
3. Tray → Models… → pick a model → Download, then Apply selection and reload model. For Russian, Small or Medium is a solid start; Large-V3-Turbo if you have the RAM/VRAM.
4. Click into any text field, hold the chord, speak, release. The transcript is pasted immediately.

Whisper weights are downloaded on first use (Hugging Face, with a GitHub release mirror as fallback) — not bundled in Setup.exe.

AMD / Intel GPUs are not used; transcription stays on CPU — prefer Tiny / Base / Small.

Features

• Tray icon; record while holding a two-key chord (left/right Ctrl + Space or Win/Cmd).
• Model picker and download in Models….
• Microphone and chord in Settings.
• Donate in the tray — CloudTips.

Models, disk size, and VRAM

Same sizes as typical OpenAI Whisper (.en = English-only):

Model (id)	Approx. disk	Notes
Tiny / Tiny.En	~75 MB	Fast on CPU; lower quality.
Base / Base.En	~150 MB	Everyday minimum.
Small / Small.En	~500 MB	Good CPU balance.
Medium / Medium.En	~1.5 GB	Better with ≥8 GB RAM on CPU.
Large (large-v1) / V2 / V3	~3 GB	Heavier RAM/VRAM.
Large-V3-Turbo	~1.5–2.5 GB	Faster than full large-v3.

Windows / Linux + NVIDIA: CUDA 12+ and a CUDA build of torch → device cuda. Rough VRAM: tiny/base ~1 GB; small ~2 GB; medium ~4–6 GB; large often 8 GB+. Tight VRAM: BUZZMINI_REDUCE_VRAM / BUZZ_REDUCE_GPU_MEMORY.

macOS: faster-whisper runs on CPU (no Metal path here). Tiny / base / small are the practical choices. Cmd as the second PTT key is supported.

Linux, macOS, and run from source (Python, uv, pip)

Python 3.12–3.14. Use a venv in the repo root.

uv (recommended)

cd BuzzMini
uv sync
uv run buzz-mini

On Windows, uv pulls the cu126 PyTorch wheel via [tool.uv.sources] in pyproject.toml.

pip

A plain pip install -e . on Windows usually installs CPU-only torch (torch.cuda=False). Install CUDA torch first:

cd BuzzMini
python -m venv .venv
.venv\Scripts\python.exe -m pip install -U pip
.venv\Scripts\python.exe -m pip install torch --index-url https://download.pytorch.org/whl/cu126
.venv\Scripts\python.exe -m pip install -e .

Already installed CPU torch:

.venv\Scripts\python.exe -m pip install --force-reinstall torch --index-url https://download.pytorch.org/whl/cu126

Launch

Windows: double-click scripts\run-buzz-mini.bat — creates .venv if needed, installs PyTorch cu126, pip install -e ., starts the tray app.

If the env is ready:

cd BuzzMini
.\.venv\Scripts\python.exe -m buzz_mini.app

Or buzz-mini with the venv activated.

Where models are stored

From source: models/ next to pyproject.toml. Installed app: user cache (compatible with Buzz’s directory). Override: BUZZMINI_MODEL_ROOT.

Environment variables

Variable	Purpose
BUZZMINI_MODEL	Default model if unset in the UI.
BUZZMINI_LANGUAGE	Recognition language, e.g. ru.
BUZZMINI_PTT_CHORD	Two-key chord, e.g. ctrl_l+space, ctrl_r+win; ctrl+space = left Ctrl + Space.
BUZZMINI_FORCE_CPU	Anything other than false forces CPU.
BUZZMINI_DEVICE	cuda, cpu, or auto.
BUZZMINI_PASTE_DELAY_MS	Delay before paste after clipboard write (ms).
BUZZMINI_LOG_LEVEL	e.g. DEBUG.

Windows build and installer (maintainers)

PyInstaller (dev): BuzzMini.spec + tools/build_windows.ps1. Same .venv with CUDA PyTorch; optional extra win-build. Output: dist/BuzzMini/ (BuzzMini.exe, several GB).

NSIS web installer — two GitHub Release assets:

File	Build with	Purpose
BuzzMini-<version>-win64.7z	.\tools\build_release_payload.ps1	PyInstaller onedir, downloaded at install time
BuzzMini-Setup-<version>.exe	.\tools\build_installer.ps1	Small installer (fetches the .7z)

Release checklist:

1. .\tools\build_windows.ps1 → dist\BuzzMini\
2. 7-Zip and NSIS (winget install NSIS.NSIS)
3. .\tools\build_release_payload.ps1 → dist\BuzzMini-1.1.0-win64.7z (name from installer\release.json)
4. GitHub Release tag 1.1.0 (no v prefix), attach the .7z
5. .\tools\build_installer.ps1 → dist\BuzzMini-Setup-1.1.0.exe, attach to the same release

Installing shows the GitHub URL, download, and extract under Show details. Whisper weights still come from Models, not the installer.

URL config: installer\release.json. Overrides on build_installer.ps1: -PayloadUrl, -GithubRepo, -ProductVersion.

Headless smoke test

Imports, models UI without showing, tiny engine, PTT chord parsing:

cd BuzzMini
.\.venv\Scripts\python.exe tools\smoke_test.py

If the package is not installed editable, set PYTHONPATH to the repo root.

Support the project

Spare-time work. If BuzzMini helps, support Timur K. via CloudTips (Russian cards, SBP, and other methods on the page).

In the app: tray → Donate, or the Donate tab.