FlowLocal

A local, private, free Wispr Flow clone for Windows. Hold a hotkey, speak, release — clean text appears wherever your cursor is. All processing happens on your PC; no audio ever leaves the machine, no subscription.

⬇ Download the latest release — unzip, run FlowLocal.exe, hold Ctrl+Win and talk. No Python needed.

Quick start

1. Install dependencies (one time):

   pip install -r requirements.txt
   

2. Double-click run.bat (or python -m flowlocal.main). First launch downloads the Whisper model (~460 MB for small), then it's fully offline.
3. Click into any text field, hold Ctrl+Space, speak, release.

While dictating, a floating pill appears at the bottom of the screen — red pulsing dot + live mic level while listening, orange while transcribing — then hides. A mic icon also lives in the system tray (grey = idle, red = recording, orange = transcribing); right-click it to edit the config, view the log, or quit. Troubleshooting: check flowlocal.log next to run.bat.

Keyboard hooks and keystroke injection are implemented directly against the Windows API in flowlocal/winkb.py (the popular keyboard PyPI package is silently broken on Python 3.14).

Configuration (config.json)

Created automatically on first run. Restart the app after editing.

Key	Default	Notes
hotkey	ctrl+space	Any keyboard-lib combo, e.g. f9, right ctrl
hotkey_mode	hold	hold (push-to-talk) or toggle
model	distil-small.en	English-only, fast; small/medium for other languages or accuracy
device	cuda	GPU (needs pip install nvidia-cublas-cu12 nvidia-cudnn-cu12); cpu also works
compute_type	int8_float16	for cuda; use int8 on cpu
beam_size	1	raise to 5 for slightly more careful decoding
mute_apps_while_recording	true	Mutes other apps' audio while you dictate, restores after
language	en	Set null to auto-detect any language
inject_mode	paste	paste (fast) or type (for apps that block Ctrl+V)
remove_fillers	true	Strips um/uh/erm
replacements	{}	Voice snippets: {"my email": "you@example.com"}
vocabulary	[]	Names/jargon to bias recognition toward (e.g. "Helius", "Xbot")
llm_cleanup	false	Extra grammar pass via local Ollama, if you run it

How it maps to Wispr Flow

Wispr Flow	FlowLocal
Push-to-talk hotkey	keyboard global hook, hold or toggle
Works in every app	Clipboard paste / SendInput typing into focused window
Cloud AI transcription	faster-whisper running locally
Filler removal & formatting	Regex cleanup layer (+ optional Ollama pass)
Snippets	replacements in config
100+ languages	Whisper supports ~100 languages (language: null)

Notes

• Pre-roll: the mic keeps a rolling 0.5 s buffer while idle, so words you start saying a moment before pressing the hotkey aren't clipped.
• Speed: on the GTX 1660 Super (cuda/int8_float16) a 20 s dictation transcribes in ~0.5 s with distil-small.en. CPU fallback (device: cpu, compute_type: int8) does the same in ~3 s.
• Elevated apps: to dictate into admin-level windows, run run.bat as administrator (Windows blocks keystroke injection across privilege levels).
• Autostart: put a shortcut to run.bat in shell:startup.

Licence

FlowLocal's own source is MIT — use it, fork it, ship it.

Dependencies keep their own licences; most are MIT or BSD. One exception worth knowing: pystray is LGPL-3.0, so the packaged release carries the LGPL's relinking obligation. The full source here plus requirements.txt reproduces the build, so anyone can substitute their own pystray. Speech models are downloaded at first run, not redistributed. Full details in THIRD-PARTY-NOTICES.md.

Related

FlowClip — the same idea for video. A local clip editor for TikTok and YouTube Shorts: trim a clip, then compose a vertical layout from multiple regions of the frame (facecam, gameplay, minimap) instead of cropping most of it away. Runs entirely on your machine, no subscription.