WisprFlow Local

Free, offline, open-source alternative to WisprFlow for Linux. Speak anywhere, get clean text - no cloud, no subscription, no data leaves your machine.

Hold Ctrl+Shift+Space, speak, release - text appears in your clipboard and is saved to Obsidian.

Quick Start (One Command)

sudo dnf install ydotool libayatana-appindicator-gtk3 gnome-shell-extension-appindicator portaudio-devel && \
pip install faster-whisper pyaudio numpy evdev pystray Pillow && \
sudo usermod -aG input $USER && \
echo "Log out and back in, then run: python wisprflow.py --daemon"

How It Works

Hold Ctrl+Shift+Space → mic records → release → faster-whisper transcribes → clipboard + Obsidian

Two modes:

Mode	How to run	How it works
Daemon	python wisprflow.py --daemon	System tray icon + global hotkey. No terminal needed.
Terminal	python wisprflow.py	Press Enter to start/stop. Good for testing.

Features

• 100% offline - all processing happens locally, no internet needed
• Global hotkey - Ctrl+Shift+Space works from any app (browser, editor, chat, etc.)
• System tray icon - green = ready, red = recording, orange = transcribing
• Obsidian integration - daily notes with timestamped bullet points
• Clipboard copy - transcription auto-copied, ready to paste
• Wayland + X11 - works on both via evdev + parecord
• 100+ languages - auto-detection or specify with --language
• Multiple models - trade speed for accuracy based on your hardware

RAM & Performance

Tested on Intel i5-10210U (4 cores), 16GB RAM, Fedora 43.

Model	RAM Usage	Transcription Speed	Quality	Best For
tiny	~300 MB	~0.7s for 10s audio	Basic	Quick notes, fast hardware
base	~500 MB	~1.2s for 10s audio	OK	Everyday use on low-end hardware
small	~625 MB	~4s for 25s audio	Good	Recommended for most users
medium	~2.5 GB	~8s for 25s audio	Great	When accuracy matters
large-v3-turbo	~2.5 GB (INT8)	~6s for 25s audio	Near-best	16GB+ RAM systems

Models download automatically on first run. The small model (~500MB download) is the default.

Installation

Prerequisites

• Linux (tested on Fedora 43, should work on Ubuntu/Arch/etc.)
• Python 3.10+
• PipeWire or PulseAudio (default on modern distros)
• A microphone

Step 1: System packages

Fedora:

sudo dnf install ydotool libayatana-appindicator-gtk3 gnome-shell-extension-appindicator portaudio-devel

Ubuntu/Debian:

sudo apt install ydotool libayatana-appindicator3-1 gnome-shell-extension-appindicator portaudio19-dev pulseaudio-utils

Arch:

sudo pacman -S ydotool libappindicator-gtk3 portaudio

Step 2: Python packages

pip install faster-whisper pyaudio numpy evdev pystray Pillow

Step 3: Permissions

# Required for global hotkey detection (reads keyboard via /dev/input)
sudo usermod -aG input $USER

# IMPORTANT: Log out and log back in for this to take effect

Step 4: Run

git clone https://github.com/amitdevv/wisperflow-linux.git
cd wisperflow-linux
python wisprflow.py --daemon

Usage

Daemon Mode (recommended)

python wisprflow.py --daemon                    # default (small model)
python wisprflow.py --daemon --model tiny       # faster, less accurate
python wisprflow.py --daemon --model medium     # slower, more accurate
python wisprflow.py --daemon --language auto    # auto-detect language
python wisprflow.py --daemon --language hi      # Hindi
python wisprflow.py --daemon --no-save          # don't save to Obsidian
python wisprflow.py --daemon --no-clipboard     # don't copy to clipboard

Then from any app: hold Ctrl+Shift+Space, speak, release.

Terminal Mode

python wisprflow.py                             # interactive mode
python wisprflow.py --model tiny                # use tiny model

Press Enter to start recording, Enter again to stop.

All Options

Flag	Description
--daemon	Run as background daemon with tray icon + hotkey
--model MODEL	Whisper model: tiny, base, small (default), medium, large-v3-turbo
--language LANG	Language code (en, hi, es, etc.) or auto for detection
--save-dir PATH	Where to save transcripts (default: Obsidian vault)
--no-save	Don't save transcripts to disk
--no-clipboard	Don't copy to clipboard
--type	Auto-type text into focused app via ydotool
--device N	Use specific audio input device (see --devices)
--devices	List available audio input devices

Auto-Start on Login (systemd)

mkdir -p ~/.config/systemd/user

cat > ~/.config/systemd/user/wisprflow.service << 'EOF'
[Unit]
Description=WisprFlow Local - Voice Dictation
After=graphical-session.target pipewire.service

[Service]
Type=simple
ExecStart=/usr/bin/sg input -c "/usr/bin/python3 /path/to/wisprflow.py --daemon --model small"
WorkingDirectory=/path/to/wisperflow-linux
Restart=on-failure
RestartSec=3
Environment=DISPLAY=:0
Environment=XDG_SESSION_TYPE=wayland
Environment=PYTHONUNBUFFERED=1

[Install]
WantedBy=graphical-session.target
EOF

# Edit the paths above, then:
systemctl --user daemon-reload
systemctl --user enable --now wisprflow.service

Service Commands

Command	What it does
systemctl --user start wisprflow	Start
systemctl --user stop wisprflow	Stop
systemctl --user restart wisprflow	Restart
systemctl --user status wisprflow	Check status
journalctl --user -u wisprflow -f	View live logs

Obsidian Integration

Transcriptions are saved as daily notes (configurable with --save-dir).

Each day gets one file (2026-03-20.md) with bullet points:

---
date: 2026-03-20
type: voice-transcripts
---

# Transcripts - March 20, 2026

- **15:45:46** - Hi, hello hello
- **15:45:57** - What are you guys doing currently today?
- **16:10:30** - I need to finish the API integration by tomorrow

Works whether Obsidian is open or not - it's just markdown files.

Architecture

┌─────────────┐     ┌──────────┐     ┌───────────────┐     ┌──────────┐
│ evdev        │────>│ parecord │────>│ faster-whisper │────>│ wl-copy  │
│ (hotkey)     │     │ (mic)    │     │ (STT)          │     │ (clipboard)│
└─────────────┘     └──────────┘     └───────────────┘     └──────────┘
                                            │
                                            ▼
                                     ┌──────────────┐
                                     │ Obsidian .md  │
                                     └──────────────┘

Component	Tool	Why
Global hotkey	evdev	Works on Wayland + X11 (kernel-level)
Audio capture	parecord	Native PipeWire/PulseAudio, separate process
Speech-to-text	faster-whisper	4x faster than OpenAI Whisper, INT8 quantization
Clipboard	wl-copy / xclip	Wayland-first with X11 fallback
System tray	pystray	Cross-desktop (GNOME, KDE, etc.)
Auto-typing	ydotool	Works on Wayland via /dev/uinput

vs WisprFlow

	WisprFlow	WisprFlow Local
Price	$15/month	Free
Privacy	Cloud (audio sent to servers)	100% local
Internet	Required	Not needed
RAM	~800 MB (idle)	~300-625 MB (active)
Platforms	Mac, Windows, iOS, Android	Linux
AI cleanup	Yes (Flow mode)	Coming soon (Ollama)
Languages	100+	100+

Troubleshooting

"No keyboard found" - You're not in the input group. Run sudo usermod -aG input $USER and log out/in. Quick workaround: newgrp input before running.

No tray icon on GNOME - Install and enable the AppIndicator extension:

sudo dnf install gnome-shell-extension-appindicator
# Then enable "AppIndicator and KStatusNotifierItem Support" in GNOME Extensions app

ALSA/GTK warnings in logs - Cosmetic and harmless. They don't affect functionality.

Wrong language detected - Use --language en (or your language code) instead of auto-detect.

High latency - Switch to a smaller model: --model tiny or --model base.

Future Plans

• LLM text cleanup via Ollama (remove filler words, fix grammar - "Flow mode")
• Auto-type into focused app (ydotool integration)
• Custom hotkey configuration
• Per-app tone adjustment

License

MIT